Free Essay

Marketing

In:

Submitted By FHERZALLAH
Words 10413
Pages 42
Community Specific Temporal Topic Discovery from Social Media
Zhiting Hu

Chong Wang

Junjie Yao

Department of Computer
Science,Peking University

Machine Learning Department
Carnegie Mellon University

huzhiting@pku.edu.cn

chongw@cs.cmu.edu

Department of Computer
Science,University of
California, Santa Barbara

arXiv:1312.0860v1 [cs.SI] 3 Dec 2013

Eric Xing

Hongzhi Yin

jjyao@cs.ucsb.edu
Bin Cui

Machine Learning Department
Carnegie Mellon University

Department of Computer
Science,Peking University

Department of Computer
Science,Peking University

epxing@cs.cmu.edu

bestzhi@pku.edu.cn

bin.cui@pku.edu.cn

ABSTRACT

quick-updated texts, exhibiting rich temporal dynamics. Understanding these dynamics provides important insights into people’s changing online behaviors. Extensive research is devoted to uncover the temporal dynamics of online content [11, 26, 23].
However, most of these existing work only explores global temporal variation, or the overall trends of topics. This ignores an important aspect of social media—the communities. A community is a collection of users with more or/and better interactions amongst its members than the rest of the global network [10]. Communities play a crucial role in social media, and provide the basis for user participation and engagement. Members in the same community typically bear similar content preferences and often communicate on shared topics [21, 20]. Given that the content of social media is so dynamic, it is expected that different communities tend to have different temporal dynamics of topics. One example from our experiments on the microblog data is illustrated in Figure 1, where we show the temporal distributions of topic “food security” in a community interested in “food”, and another community mainly focusing on “law”, respectively. We can clearly observe different patterns. The huge burst in community “law” coincides with the scandal of New Zealand substandard milk powder erupted on Jan
25th, 2013. By distinguishing the patterns of temporal variations across different communities, we gain deeper insights on how topics change over time, and how different pieces of content attract attentions from different communities. This can be potentially used for various applications. For instance, in online marketing, advertisers are allowed to achieve exact targeting for their advertisements with different subjects, or design community-specific content to effectively catch the eyes of community members.
Extracting community-specific temporal dynamics of topics can be very challenging. First, user community membership is often unknown, and the topics are also hidden. With the boom of both links and text in social media, it is necessary to extract community structures and latent topics simultaneously to attain a complete view of the social media. However, even though there are a bulk of studies on community detection and topic modeling respectively, only a limited number of work [12, 8] aims at jointly modeling these two important aspects. Furthermore, the interactions between community and topic are not straightforward to model. For example, communities in [12] do not have direct relations with topics, and in [8] one community corresponds to only one topic, violating the fact that a community in social media typically has varying degrees of interests in different topics. Second, human activities

Studying temporal dynamics of topics in social media is very useful to understand online user behaviors. Most of the existing work on this subject usually monitors the global trends, ignoring variation among communities. Since users from different communities tend to have varying tastes and interests, capturing communitylevel temporal change can improve the understanding and management of social content. Additionally, it can further facilitate the applications such as community discovery, temporal prediction and online marketing. However, this kind of extraction becomes challenging due to the intricate interactions between community and topic, and intractable computational complexity.
In this paper, we take a unified solution towards the communitylevel topic dynamic extraction. A probabilistic model, CosTot (Community Specific Topics-over-Time) is proposed to uncover the hidden topics and communities, as well as capture community-specific temporal dynamics. Specifically, CosTot considers text, time, and network information simultaneously, and well discovers the interactions between community and topic over time. We then discuss the approximate inference implementation to enable scalable computation of model parameters, especially for large social data.
Based on this, the application layer support for multi-scale temporal analysis and community exploration is also investigated.
We conduct extensive experimental studies on a large real microblog dataset, and demonstrate the superiority of proposed model on tasks of time stamp prediction, link prediction and topic perplexity.

1.

INTRODUCTION

With the prevalence of online social networks, such as Twitter and Facebook, social media has become a ubiquitous part of people’s daily lives. It provides a platform for users to post short and

1

terms of time prediction, link prediction and text perplexity, which indicates the advantage of CosTot in modeling temporal dynamics, network structure and text, respectively.
To summarize, we make the following contributions in our work.
1. We identify the problem of community-specific topic dynamic extraction from social media data with text, time and network. The rich features and interaction can be used in the framework of new setting. To the best of our knowledge, such a problem has not been investigated before.
2. We propose a unified probabilistic model, CosTot, which uncovers the topics and communities as well as captures the community-specific temporal dynamics of topics. We well study the features required for this model, and design an efficient inference algorithm to guarantee the scalability of our method. Figure 1: Temporal distributions of topic “food security” in community “food” (top) and community “law” (bottom), over a period of 90 days. We see different communities exhibit very different patterns on the same topic.

3. We deploy our approach to facilitate a set of applications on real social media data. We also present a comprehensive study of community-specific topic temporal variations, and show its usefulness to social media analysis.
4. We conduct extensive experiments to evaluate the performance of our approach on large-scale real dataset. The results show the superiority of our model in terms of modeling text, network structure and temporal dynamics.

are highly volatile, especially on the online social networks where diverse content grows and fades rapidly over time. Although the aggregate temporal dynamics of individual behaviors may exhibit certain patterns, different communities with various interests tend to have their own temporal patterns. Third, the unprecedented data scale, including enormous volume of text and large sparse network, poses new computational challenges.
In this paper, we propose a probabilistic framework to address the above challenges. A probabilistic longitudinal model, CosTot (Community Specific Topics-over-Time) is developed to uncover the hidden topics and communities, as well as capture the community-specific temporal topic variations, from large-scale social media data. Our model defines a generative process for text, time and network to accurately characterize social media. Specifically, our model assumes that 1) each user can belong to multiple communities with different degrees of affiliation strength; 2) each community can have varying levels of interests in multiple topics;
3) each topic exhibits different temporal variations within different communities, and these variations are determined by the content of user posts along the time line. Different from existing temporal models such as Topics over Time (TOT) [23], CosTot can provide a finer-grained exploration of community-specific temporal dynamics of topics. CosTot is also more accurate than [8] in terms of modeling the correlation between community and topic by allowing communities to have mixture of topics rather than one-to-one correspondence. To enable our proposed CosTot scalable to largescale social media datasets, we design a Gibbs Sampler: by implicitly modeling negative links in Bayesian prior, it takes, in each iteration, linear time in terms of the size of total words and number of positive links in the dataset. It can usually converge after a small number of iterations in our experiments.
We then deploy our approach to facilitate a set of applications on real microblog data. The patterns and knowledge learnt by CosTot enables the application layer to explore fine-grained topic temporal dynamics, analyze community temporal characteristics, as well as detect bursty events.
We further conduct extensive experiments on both synthetic and large-scale real datasets to evaluate the performance of our proposed method against state-of-the-art approaches. The real dataset consists of 11M posts generated by 53K users between 12/01/2012 to 02/28/2013. The results show the superiority of our method in

The rest of the paper is organized as follows: Section 2 reviews related literature;Section 3 formulates the problem, and introduces the proposed model and the inference algorithm; Section 4 reveals the analysis and applications enabled by the proposed new model;
Section 5 presents our experimental results; and finally we conclude this paper and outline future work.

2.

RELATED WORK

In this section, we describe the related work in two areas: temporal dynamics of topics and community detection.
Temporal Dynamics of Topics: Topic models, such as latent
Dirichlet allocation (LDA) [3] are usually utilized to find latent topics from text collections. In topic models, documents are modeled as a distribution over a shared set of topics, while topics themselves are distributions over words. Modeling temporal dynamics using topic models has attracted huge interest. A number of temporal topic models were proposed. Topics Over Time (TOT) [23] models the text and time stamp of a document jointly, assuming that latent topics generate time stamp according to a Beta distribution.
One main shortcoming of Beta distribution is that it is unimodal, and thus limits the available patterns of topic temporal variation.
TOT can also be seen as a special form of a more flexible model, supervised latent Dirichlet allocation (sLDA) [15]. Another set of approaches [2, 30, 19] makes Markovian assumption on topic variation. They divide time into epoches, and assume that topics evolve based on their states in the previous epoch,
In addition to topic models, there are also a bulk of other approaches modeling temporal variation of information diffusion. [26] develops the K-Spectral Centroid (K-SC) clustering algorithm which finds six classes of patterns of temporal variation. [14] proposes
SPICKM, a flexible analytical model that generalizes and explains earlier theoretical models for the rise and fall patterns of influence propagation. A closely related line of work to temporal dynamics of topics is bursty event detection [9, 27, 5, 29]. More specifically, [5] applies a state machine-based method to detect bursty topics discovered by
2

the social relationship between user i and user i . While we can adapt our model to describe either directed links or undirected links, we focus on directed links in this paper, since this is more common in microblog platforms such as twitter and weibo where eii ∈ Ei means user i follows user i . We assume the links between users are constant within the time period we focus on. (We would like to relax this assumption in the future.)

TimeUserLDA [5]. [29] simultaneously detects stable topics (e.g. topics on user interest) and bursty topics (e.g. topics on emergencies) in a unified PLSA-based model.
Our work here is distinct from all the above methods as we not only find the global trends of topics, but more importantly distinguish patterns of temporal variation across different communities, and thus provide a more thorough and fine-grained view for temporal characteristics of social media. [11] follows a similar line by tracking opinion shift of members from two different groups, whereas a group is defined by some pre-defined features and is far away from the network community as our work.
Community Detection: This has been a hot topic, especially in recent social community analysis [10]. Numerous techniques have been developed to detect disjoint communities, i.e., each user in the network is assigned to a single community. However, in real life users are usually characterized by multiple community memberships, leading to overlapping communities. A review and comparative study of overlapping community detection is presented by
[25]. Among all these work, blockmodeling [6] is based on statistical inference, and [1] introduces a mixed membership stochastic blockmodel in which each user has a probability distribution over communities drawn from a Dirichlet distribution. Different from our approach, these work uses network structures alone to extract communities. A growing number of recent papers incorporate both the network structure and content to improve community detection performance. For example, Topic-Link LDA [12] jointly models underlying topics of documents and author communities, and achieves good performance in link prediction task. However, it does not uncover the relations between communities and topics. There is also a line of work aiming at modeling documents and links between them (e.g. citations), such as Pairwise-Link-LDA and Link-PLSALDA [17], RTM [4] and PMTLM [28].
In these models, words and links are both generated by the same latent factor, which can be thought of as communities when generating links, and as topics when generating words. Hence there is a one-to-one correspondence between communities and topics. In contrast, our approach allows each community to have a mixture of topics, which is more reasonable due to the fact that communities in social media tend to have multiple interests.

3.

A community c ∈ {1, . . . , C} has two components: a topic probability vector θc where each component θck represents the probability that a post from the community is related with corresponding topic k, and a link formation probability vector ηc where each component ηcc is the probability that a user in community c follows a user in community c . Here C is the total number of communities.
Each user i can belong to different communities (mixed membership). That is to say, each user i is associated with a community probability vector πi .
We define a topic k ∈ {1, . . . , K} as a multinomial distribution over the vocabulary, denoted as φk . Here K is the number of topics. For each topic k, the community-specific temporal dynamics is represented by a set of C multinomial vectors ϕk =
{ϕk1 , ϕk2 , . . . , ϕkC }, each of which represents the time variation of topic k within the corresponding community, i.e. a probability distribution over discrete time slices. This kind of vector represents how the popularity or attention to topic k in community c changes over time.
Given the text data with time stamps and network data in social media, our goal is to uncover hidden communities and topics, and infer the community-specific temporal dynamics of topics. This extraction can improve the understanding of information changes and community characteristics, which further benefit several important tasks in social media, e.g., link prediction, time stamp prediction and multi-scale dynamic analysis.

3.2

A PROBABILISTIC MODEL FOR COMMUNITY SPECIFIC TOPIC DYNAMICS

1. Feature Extraction: We extract the text and network features of users from raw social media records. Here we discretize time line by dividing the time span of all users’ posts into time slices. We then select the records from each consecutive time slots, and associate each post with corresponding time stamp.

In this section, we introduce our framework for community-specific temporal dynamic discovery in social media, which can effectively support applications of temporal understanding and community analysis. We first formally define the problem we are interested in, then we propose a probabilistic model to uncover hidden topics and communities, as well as capture community-level topic temporal variations. Based on the model, we further design an efficient approximate inference algorithm.

3.1

Framework Overview

In this paper, we propose a probabilistic framework to achieve these goals and support the upper layer applications. The topics and communities are both hidden factors to be extracted, and the correlation between them are always omitted due to the modeling and inference complexity. Here we use a multiple stage approach to tackle this challenge. Specifically, our work can be listed as the following stages, and the framework is shown in Figure 2.

2. Model Training: We introduce a probabilistic model to uncover the correlations between communities and topics over time. In this model, we combine the community extraction, topic identification, and community-specific topic dynamic discovery in a unified way. Though seeming complex, we tackle it with a well defined multiple component strategy and design an effective inference algorithm.

Problem Formulation

We consider a social network G consisting of U users. Each user i ∈ {1, . . . , U } is associated with two types of features.
1. Text Data with Time Stamps: a set of posts Di = {di1 , di2 ,
. . . , di|Di | } generated by user i. Each post dij ∈ Di contains a bag of words from a given vocabulary, along with a time stamp tij , meaning that user i generated post dij at time tij .

3. Model Output: We can get the temporal, topic and community information from the previous designed probabilistic model. We organize these intermediate output for later process.

2. Network Data: network links Ei = {eii |1 ≤ i ≤ U } between i and other users. Each link eii ∈ Ei represents

4. Dynamic Analysis: The probabilistic model enables a set of novel applications based on the new data extraction. For
3

Figure 2: Framework of CosTot

User membership component. Our proposed model uses a mixed membership approach [1], to capture the fact that people bear multiple roles in social media and their behaviors are influenced by different community context [25]. Note that in our model user membership integrates two aspects of user behaviors, i.e., posting (which generates text) and friending/following (which generates links). It is coherent with the observation that members in a community not only have denser links among each other than those from different communities, but also tend to be interested in similar topics. For each user i, we would like to infer the probability that i belongs to each of the C communities. Hence i is associated with a community probability vector πi . Each post dij ∈ Di is assigned to a single community cij , denoting the community membership of user i when she writes the post. In addition, each link eii ∈ Ei is associated with two communities sii and sii , one for each of the two users i and i respectively, denoting their community memberships when user i builds relationship with user i .
Text-time component. Content of community c is generated by a mixture θc of topics. The distribution θc models community’s varying levels of interests in multiple topics, and hence plays the critical role of connecting the two aspects of community and topic in social media.
Each post dij ∈ Di contains a bag of words {wij1 , . . . , wij|dij | } where |dij | denotes the length of the post. In traditional topic models such as latent Dirichlet allocation (LDA) [3], a document is associated with a mixture of topics and each word has a topic label.
This is reasonable for long documents such as academic papers.
However, on social media like twitter or weibo, a post is usually very short, and thus is most likely to be about a single topic [5, 31,
8]. We therefore assume a single latent topic variable zij with dij to indicate its topic. In addition, posts are typically noisy, containing words irrelevant to the main topics [31, 8]. Hence, we assume a background word distribution φB to capture such common words, and associate each word wijl with a background boolean indicator fijl to indicate the word wijl is drawn from the background topic or not.
To model the discretized time stamps of posts, we use a multinomial distribution ϕkc over time stamps to model the time variation specific to each topic k and each community c. Thus, a post dij is

example, we are allowed to gain a multi-scale view of topic temporal dynamics. In addition, by utilizing the fine-grained data representation, our approach further improves the extensively studied tasks of time and link prediction. We demonstrate how our approach supports these various applications in Section 4.
For the model training part, we can enumerate the requirements and necessary steps towards this goal.
1. For each user i, we need to infer the community probabilities πi .
2. For each community c, we are required to infer the topic probabilities θc , and link probabilities ηc ;
3. For each topic k, infer the word probabilities φk . This is necessary for topic representation.
4. Infer the community-specific temporal dynamics ϕkc of each topic k within each community c. Based on the above three steps, we can unify them into the final stage.

3.3

Community Specific Topics-Over-Time
Model

Here we describe the proposed model, CosTot (Community Specific Topics-over-Time), later we will show how to perform inference with this model using a Gibbs sampling algorithm which scales linearly with respect to the size of data.
CosTot is a latent space model jointly over text, time and network, and infers the patterns mentioned above. Some of its building blocks are inspired by earlier successful attempts, including the
Mixed Membership Stochastic Blockmodel (MMSB) [1] over networks, and Topics over Time (TOT) [23] over text and time. Specifically, CosTot contains three closely linked components. The user membership component models user membership to communities; the network component explains the link structure; the text-time component uncovers the semantic contexts, and captures the temporal variations in different communities. We will describe each component in detail later in this section.
Our model is summarized as the generative process shown in
Algorithm 1. Figure 3 displays its graphical model representation.
4

Algorithm 1 Generative Process for CosTot
1. Sample foreground-background distribution, χ|δ0 , δ1 ∼ Beta(δ0 , δ1 ).
2. Sample the background word topic, φB |β ∼ Dirichlet(β).
3. For each topic k = 1, 2, . . . , K,

ρ

πi α (a) Sample the distribution over words, φk |β ∼
Dirichlet(β).
(b) For each community c = 1, 2, . . . , C,
i. Sample the distribution over time stamps, ψkc | ∼
Dirichlet( ).

cij

zij

θc

sii

C

χ

sii

λ0, λ1

fijl ηcc eii

tij

4. For each community c = 1, 2, . . . , C,

C2

Ei

(a) Sample the distribution over topics, θc |α

Dirichlet(α).
(b) For each community c = 1, 2, . . . , C,
i. Sample community-community link probability, ηcc |λ0 , λ1 ∼ Beta(λ0 , λ1 ).

wijl

δ0, δ1

Wij
Di
U

φB

5. For each user i = 1, 2, . . . , U

φk

ψkc
C

(a) Sample the distribution over communities, πi |ρ ∼ Dirichlet(ρ).
(b) For each post j = 1, 2, . . . ,
i. Sample community indicator, cij |πi ∼ Multi(πi ). ii. Sample topic indicator, zij |θcij ∼ Multi(θcij ). iii. For each word l = 1, 2, . . . ,
A. Sample foreground indicator, fikl

Bernoulli(χ).
B. Sample word, wijl |φzij ∼ Multi(φzij ) if fikl = 1, or wijl |φB ∼ Multi(φB ) if fikl =
0.
iv. Sample time stamp, tij |ψzij cij ∼ Multi(ψzij cij ).
(c) For each user i = 1, 2, . . . , U
i. Sample community indicator, sii |πi

Multi(πi ). ii. Sample community indicator, sii |πi

Multi(πi ). iii. Sample link, eii |ηsii s ∼ Bernoulli(ηsii s ). ii K

β

Figure 3: Graphical Model Representation of CosTot. A double circle indicates a hyperparameter; a single hollow circle indicates a latent variable; and a filled circle indicates an observed variable.
The latent variable sii is represented as dashed circle since it is drawn from πi which is not shown in the graphical model.

nneg = U (U − 1) − i |Ei | is the number of negative links. In this way, we reduce large amount of computation and achieve linear complexity on network modeling, as explained in Section 3.4.1.

3.4

Approximate Inference Implementation

We then proceed to propose a collapsed Gibbs sampler for approximate inference of the CosTot model. At each iteration of our
Gibbs sampler, we need to sample, for each post dij by user i, both the corresponding community indicator cij and the topic indicator zij , and for each link eii the corresponding community indicators sii and sii . We are also required to sample the per-word foreground indicator fijl . We discuss these separately.
Sampling community indicator cij for post dij . We sample the community indicator cij for post dij according to ,

ii

generated at the time tij drawn from ϕzij cij . Compared to Topics over Time (TOT) [23] which uses a Beta distribution to model time variations and only allows a unimodal distribution over time for each topic, our use of multinomial distribution can capture multimodal variations. It is more flexible in capturing real-life topics which usually rise and fall for many times.
Network component. Different from text data collected from individual users, the network data is relational and thus violates the classical independence or exchangeability assumptions [1]. To address the problem, the network component uses pairwise community Bernoulli distributions to model the presence and absence of links between pairs of users. For link eii , the indicator sii and sii denotes the community membership of user i and user i respectively, when user i built relationship with user i . Then eii is drawn from ηsii s which represents the relationship strength ii between community sii and sii .
The network of a social media is typically sparse, thus we only model positive links: the variables sii , sii exist if and only if eii ∈ Ei . As in [8], the negative links eii ∈ Ei are implicitly modeled in a Bayesian fashion: we use a Beta(λ0 , λ1 ) prior on each ηcc , and set λ0 = ln(nneg /C 2 ) and λ1 = 0.1, where

P (cij = c|zij = k, tij = t, c−ij , s, z −ij , t−ij , .)
(c)



(c)

ni

(·) ni +ρ

+ Cρ

(k)

·

nc

(·) nc +α

+ Kα

(t)

·

nck +
(.)
nck

,

(1)

+T

where ni denotes the number of posts and links of user i gener(·) ated by community c; ni is the total number of posts and links of
(k)
user i; nc is the number of posts assigned to community c and
(·)
generated by topic k; nc is the total number of posts generated
(t)
by community c; nck denotes the number of times that timestamp
(·)
t is generated by community c and topic k, and nck denotes the total number of timestamps generated by community c and topic k.
All the counters mentioned above are calculated with the post dij excluded. Sampling community indicators sii and sii for link eii . Recall that we only model sii , sii and eii for positive links eii ∈
5

4.

Ei . The resulting conditional posterior distribution is:

In this section, we demonstrate the usefulness of our approach by various representative applications on real-world social media data. We show that CosTot can effectively uncover communities and topics, and simultaneously capture temporal dynamics of topics in different communities. Based upon the patterns inferred by our model, we are allowed to explore topic dynamics at multiple granularities, identify bursty events, and give an in-depth analysis of particular communities. Our method can also support time stamp prediction and link prediction.
Real Data Setting: We first introduce the data set used in the following study. Our data is crawled from Sina Weibo1 , one of the most popular microblog platforms in China. After removing stop words and low active users with fewer than 20 posts, we obtain our dataset consisting of about 53K users, 2.1M links, 11M posts and 91M words with a vocabulary of size 89K. The posts are distributed evenly in the time period from December 1st 2012 through
February 28th 2013. Each post is labeled by the date it was posted.
Therefore, time stamps of the dataset range from 1 to 90.

P (sii = c, sii = c |s−ii , c, e, .)
(c)

ni



(·) ni +ρ

(c )

·

+ Cρ

ni

(·) ni +ρ

(2)

ncc + λ1
,
ncc + λ0 + λ1

·

+ Cρ

where ncc is the number of positive links, with eii excluded, whose communities indicators are {c, c }.
Sampling topic indicator zij for post dij . This is done through the conditional posterior probability
P (zij = k|cij = c, tij = t, c−ij , z −ij , f , w, t, .)
(v)

(k)



nc

(·) nc +α

+ Kα

(t)

·

nck +
(.)
nck

+T

·

nij −1
(v)
(nk q=0 V v=1 + q + β)

(·)

nij −1
(·)
(nk q=0 ,

APPLICATIONS BASED ON COSTOT

(3)

+ q + V β)

(v)

where nij is the number of times word v occurs in the post dij
(·)

and is labeled as a foreground word; nij is the total number of
(v)
foreground words in the post dij ; nk denotes the number of times
(·)
word v is assigned to topic k, and nk is the total number of words
(v)
(·) assigned to topic k. Note that nk and nk are calculated with the post dij excluded.

4.1

Multi-scale Topic Temporal Dynamics

In addition to most of existing works that only captures global trends of topics, CosTot can detect community-specific trends of topics. This allows us to have a multi-scale view of temporal dynamics, as well as gain deeper insight on how topics attract attentions from different communities.

Sampling foreground indicator fijl for word wijl . The conditional posterior distributions for the foreground indicator fijl are,
P (fijl = 1|zij = k, wijl = v, z −ij , w−ijl , .)
(v)



n +β n(1) + δ1
· k
,
+ δ0 + δ1 n(·) + V β k (4)

n(·)

and
P (fijl = 0|wijl = v, z, w−ijl , .)
(v)



n +β n(0) + δ0
· B
,
n(·) + δ0 + δ1 n(·) + V β
B

(5)
Figure 4: Global temporal dynamic. The “Top Words” on the top left shows the top five most probable words in the topic, based on which we give the topic a concise name “New Year”. Popular peaks are labeled with lettered flags. By manually examining the data, we give a brief explanation for each peak to the right of the timeline.

where n(0) and n(1) are the number of background words and fore(v) ground words respectively; n(·) = n(0) + n(1) ; nB denotes the
(·)
number of times word v is generated by background topic, and nB is the total number of words generated by background topic. Again, all the above counters are calculated with word wijl excluded.

3.4.1

Figure 4 shows the global trends of topic “New Year” in Sina
Weibo, where the global popularity of topic k at time stamp t is obtained by summing over all communities (denoted as c) and users
(denoted as i),

Linear Time Complexity

After the inference description, here we analyze the time complexity of this inference algorithm. It is shown that the chosen algorithm scales linearly in terms of the size of data, i.e. the number of words and positive links, achieving satisfying performance.
Be ware that, in each iteration, the communities of each user’s
(c)
posts are first sampled. Since all the counters (e.g. ni ) involved in Eq.(1) can be cached and updated in constant time for each cij being sampled, Eq.(1) can be calculated in constant time. Thus, sampling all c takes linear time w.r.t the number of posts. Next, we sample community indicators s using Eq.(2). Since we have implicitly modeled negative links in Bayesian piror (i.e., the Beta prior for ηcc ), we only need to sample sii and sii for positive links eii ∈ Ei . Hence the complexity is reduced from quadratic
(w.r.t the number of users) to linear (w.r.t the number of links). It significantly saves computation cost due to the sparseness of networks. Finally, sampling all z and f by Eq.(3), (4) and (5) is linear in the number of words. Overall the inference algorithm takes linear time in the amount of data.

P (t|k) =

P (t|c, k)P (c|k) c ∝

P (t|ψkc )P (k|θc ) c P (c|πi ), i where we assume the prior distribution P (i) of each user i is constant. We label the spikes of timeline with lettered flags. To better understanding the semantic context of the spikes, we manually check the posts generated around corresponding time stamps and give a brief explanation for each spike. For example, the spike “B” on Feb 10th, 2013 is the Chinese New Year’s Day, while the lower spike “A” on Jan 1st, 2013 corresponds to the first day of 2013.
(Note our data was from Sina Weibo, a Chinese website.)
Figure 5 shows part of the communities in Sina Weibo, where each node represents a community. We choose a label for each
1

6

http://weibo.com

model captures mixed membership of users, we define the contribution of a user i to a community c by considering both the membership probability and number of posts generated by i, as contribution(i, c) = πic · log|Di |.
For community “E-commerce”, we calculate contributions of all users, and find that the contribution distribution approximately follows a classic power law. Due to the space constraint, we omit this distribution figure. We select top users with contributions larger than 1.0, which yields a subset of about 750 users (Figure 6). The size of each node is proportional to the user’s contribution. We further recognize the central actor [24] of the community as the user with highest contribution (the yellow node), and as in [18], we highlight the nodes that can reach the central actor within 2 steps along the directed edges (i.e. “following” relationship). From the figure we see that most of the nodes are highlighted, suggesting that the members are closely connected.
Figure 5: Community-specific temporal dynamics. The global and community-specific temporal dynamics of topic “New Year”. We focus on the time period starting on 12/25/2012 since the popularity is zero before that time.

community based on its distribution over topics. For example, community “Movie” puts high probabilities on movie-related topics 2 .
The darkness of the color of edges indicates the link probability between corresponding communities—the darker the color is, the higher the probability. For instance, there is a probability of
0.7 that the users in community “Family” follow the users in community “Life”, while the probability that the users in community
“Transport” follow the users in community “’Movie” only achieves
0.2. For comparison, we put the community-specific dynamics of topic “New Year” near the corresponding nodes (communities). We see that while most of the timelines peak around the Chinese New
Year’s Day (i.e. spike “B” in Figure 4), only those communities on the left (e.g. community “Movie”) have spikes around the first day of 2013 (i.e. spike “A” in Figure 5), and communities on the right
(e.g. community “Transport”) pay little attention to the topic at that time. One possible explanation for the phenomenon is that the communities on the left are more relevant to entertainment, while those on the right seem to be more concerned with professional stuff. Furthermore, strong links between communities tend to make the temporal dynamics in corresponding communities more similar, because they provide more effective channels for information diffusion. The community-specific temporal dynamics also suggest that the huge spike “B” in the global dynamic is contributed by most communities of the social media, and by contrast, the small spike
“A” is formed due to the attentions from part of the communities.

Figure 6: Nodes and Temporal Change in Community “Ecommerce”. Size of each node is proportional to the user’s contribution (see text for more details). The node in yellow represents the central actor of the community. Nodes that can reach the central actor within 2 steps along the directed edges are highlighted in dark green.
Topic dynamics within the same community.
To give a holistic view of the temporal dynamics of the community’s attention to different topics, we compute the community’s distribution over topics given time with the Bayes rule:
P (k|t, c) =

4.2

Characters of Extracted Communities

Here we demonstrate that, by focusing on community-specific results, CosTot model enables us to step into a finer granularity and get in-depth characteristics of particular communities.
Community connectivities and user contributions. Figure 6 provides a visualization for community “E-commerce”. Since our

P (t|ψkc )P (k|θc )
.
k P (t|ψkc )P (k|θc )

Figure 7 shows the resulting patterns. The height of a topic’s region indicates the relative popularity of the topic in the community “Ecommerce” at given time. We can observe that the attentions of the community members change over time, while topic “E-commerce” dominates the focus. (That is why we name the community as “Ecommerce”.) It is also notable that other minor topics are also competing for the attention from the community. For example, at time stamp A, topic “Economy” gains more concern among these minor topics, while afterwards (at time stamp B) topic “Movie” successfully catches the eyes of the members.

2
It is worth mentioning that, in the results inferred by our model, there is no community whose topic probability vector is dominated by topic “New Year”. This is consistent with the fact that New Year is a temporary event and unlikely to be some community’s major interest. 7

(a) Topic “Sports” in community “Sports”

Figure 7: The distribution over topics as a function of time . The legend shows the name for each topic and its corresponding color.
Top 10 topics w.r.t community-topic distribution are plotted, while
“Other” represents the remaining 90 topics.
It is also interesting to step further by focusing on individual topics and analyzing their temporal dynamics in the particular community. We select two topics and display their variation within community “E-commerce” in Figure 8. We find that it is Ang Lee’s win for best director in the Academy Award that boosts topic “Movie” at time stamp B in Figure 7. It can also be observed that timelines of topics with higher levels of interest tend to be smoother, while topics with low levels of interest receive spiky attentions from the members. The result indicates that temporal dynamics of topics in
“unrelated” communities can facilitate bursty event detection. We further verify this claim in the next section.

(b) Topic “Sports” in community “Movie”

Figure 9: Temporal dynamics of Topic “Sports” in Different Communities. Bursty events can be easily identified in (b).

An example is shown in Figure 9, where (a) demonstrates the temporal dynamic of topic “Sports” in community “Sports” (i.e. the community takes major interest in topic “Sports”), whereas (b) is the temporal dynamic of the same topic in community “Movie”.
The percentages shown on the top left corner represent the topic probability of corresponding communities. From Figure 9(b) we can more easily identify the bursts, such as the burst “C” which coincides with the Australian Open Final the famous Chinese professional tennis player Na Li attended. On the contrary, the timeline in Figure 9(a) is much smoother and without clear spikes. By examining the data, we found that community “Sports” did talk about the particular events detected by Figure 9(b) when they happened. However, these bursty behaviors are concealed since members keep talking about sports throughout the time period, and this results in no clear spikes. On the other hand, members in community “Movie” are not concerned with sports-related topic in the daily life, but their attentions would still be drawn when significant events happened. It is worth mentioning that although the probability that community “Movie” generates sports-related posts is low, there are still many members and large number of such posts involved due to the large size of the data. Thus, the spikes represent the attentions from this whole community, rather than from a small portion of its members.

(a) Popularity of topic “Life” peaked when the new year is drawing near.

4.4
(b) Movie-related events significantly drew attentions of the community.

Figure 8: Community “E-commerce”-specific temporal dynamics of two topics with various levels of interest from the community.

4.3

Time Stamp/Link Prediction

Time Stamp Prediction. Another application of modeling temporal dynamics is to predict the time stamp of a previously unseen document based on its content. It can be used to recover time stamps of documents with missing or incorrect meta-data [22].
Given the words in a post and its author, we predict its time stamp by choosing the one that gives maximum likelihood. Specifically, for a post d by user i, its predicted time stamp is:

Event Identification

Bursty event detection [9, 5] aims to capture the most popular events that have drawn the public’s attentions. When we want to find such events related to a certain topic, one may intuitively focus on the community whose major interests lie in the particular topic, and analyze the behaviors of its members to detect bursts. Our results, however, suggest that analyzing communities that are not regularly concerned with the topic may provide a easier way for bursty event detection.

ˆ td = arg max t P (c|πi ) c P (k|θc )P (t|ψkc ) k (χP (wdl |φk ) + (1 − χ)P (wdl |φB )). l Link Prediction. Our framework also supports to predict the probability of a link between two users. Link prediction in social
8

media not only helps in analyzing networks with missing data, but also can be used to recommend friends or followees by identifying very likely but not yet existent links [13].
Based on the results of CosTot, the link prediction algorithm works as follows: for a pair of users (i, i ), we compute the probability of a link from user i to i as

Eq.(6) allows easy comparison with the estimated link probabilities from our model.
Experimental Results. We set C = 5, K = 30, V = 100,
T = 30, U = 250, Di = 50, W = 20, P0 = 0.7, Pslope = 0.3 and Pmin = 0.1. All Gaussian variances are set to 1.0. We train
CosTot on this synthetic data, with the numbers of communities and topics set to the true values.

P (s|πi )P (s |πi )ηss ,

Pi→i = s s

and predict that the link exists if Pi→i exceeds some threshold.
We will present the empirical improvements of time stamp and link prediction tasks in the later experimental study.

5.

EXPERIMENTS

(a.1)

5.1

Pij = max{P0 − Pslope · |i − j|, Pmin },

(b.2)

(c.1)

(c.2)

(d.1)

Experiments on Synthetic Data

Topic and community Generation. We generate K topics each associated with a multinomial distribution over V words obtained by discretizing Gaussian distributions with means sampled uniformly on [0, V ]. One example of generated topic-word distribution is shown in Figure 10(b.1). We then generate C synthetic communities each associated with a multinomial distribution over topics obtained by the similar method to topic-word distribution but with the mean sampled on [0, K]. Figure 10(a.1) shows the topic distribution of one of the communities. For each (topic, community) pair, we also use a discretized Gaussian distribution to mimic the temporal variation of the topic in the community, with the mean sampled uniformly on [0, T ]. Here T is the number of time stamps. One example of temporal variation is demonstrated in Figure 10(c.1), where we use smoothed curve to fit the discrete values.
We also generate U users, each of which is randomly assigned a community label denoting the major community that the user belongs to.
Text Generation. Each user has Di posts. To mimic the property of mixed membership in social media, the community that a user belongs to when she creates the post is sampled from a discretized Gaussian distribution over communities, with the mean equal to her community label. After that, topic, words and time stamp of the post are generated according to the generative process described above (with background topic omitted). Each post contains W words.
Link Generation. First we devise the following link probability between two communities i and j,

(a.2)

(b.1)

In this section, we conduct extensive experiments on both synthetic and real-world data to evaluate our proposed approach. For synthetic dataset, our experimental results show that CosTot is able to precisely uncover the hidden patterns. For real-world dataset, we demonstrate the superiority of our method by comparing it with state-of-the-art methods in various aspects.

(d.2)

Figure 10: Comparisons of ground-truth values and those inferred by CosTot. (a.1/2) The ground-truth/inferred distribution over topics for a particular community on the synthetic dataset. (b.1/2) The ground-truth/inferred distribution over words for a particular topic.
(c.1/2) The ground-truth/inferred distribution over time stamps for a particular (topic, community) pair. (d.1/2) The ground-truth adjacency matrix/inferred link probability matrix of the network.
Results are shown in the right column of Figure 10. We observe that the distributions inferred by CosTot well match the ground truth. Figure 10(d.2) shows the matrix with each cell representing the link formation probability between two users. CosTot also well recovers the network structure.

(6)

where i and j ∈ {1, . . . , C} are the natural number indexes of the communities; P0 denotes the link probability within a community
(i.e. Pij = P0 if i = j); Pslope is the slope; and Pmin is the threshold minimum link probability between two communities. Eq.(6) implies that links are more likely generated between communities indexed by consecutive numbers. Second, for each user pair, we find their major communities of each and use the link probability between those two communities to generate links between the user pair. Figure 10(d.1) shows the adjacency matrix of users indicated by the generated links. Here rows (i.e. users) are reordered according to their community labels. The resulting block structure due to

5.2

Experiments on Real-world Data

We now present empirical results of our approach on Sina Weibo dataset (Section 4). We first describe the settings of experiments, then quantitatively evaluate our model in terms of three aspects: time stamp prediction for evaluating the capacity of capturing temporal dynamics, link prediction for measuring the capacity of modeling network, and perplexity for evaluating the capacity of modeling text. We only report the optimal results of different methods after tuning the parameters, and leave the study of parameter impacts for the end of the section.
9

5.2.1

Experimental Settings

smoothing, performs best in the time stamp prediction task among a host of competitors.
CosTot without Link (CosTot-NoLink). CosTot-NoLink is a subpart of CosTot with the network component (Section 3.3) removed. Hence it provides a more fairly comparison with TOT. We can also take a look at the impact of considering link information by comparing to CosTot.
Simple Approach of Community-specific Temporal Variations
(Simple-CTV). One alternative method to analyze communityspecific topic dynamics is to exploit network and text data step by step: we begin by dividing users into communities using wellestablished network community detection techniques, then uncover topic variation in certain community by running TOT on the posts generated by its members. By comparing with the simple approach, we gain insight into the benefits of combining the two aspects of community and topic in the CosTot way. In the experiment, we capture user multiple memberships by running MMSB on user network and assigning each user to two communities [25] with highest probabilities. We randomly select 20% of the posts as the test set, while the remaining 80% posts and all links are used to train the models. We set the hyperparameters in above baselines closely resemble those in our model. Here we only report the best results of each with tuned parameters. The impact of model parameters is discussed in later section.
Figure 12 shows the prediction accuracy as a function of tolerance range for these models. From the figure, we see that our model performs better than all competitors. Moreover, CosTotNoLink outperforms TOT and EUTB, justifying the advantage of distinguishing temporal variations of topics in different communities, while the superiority of CosTot to CosTot-NoLink shows the benefit brought by incorporating link structures in social media.
We also observe that, Simple-CTV, despite taking into account community-specific topic dynamics, has poor performance as TOT.
The reason is that it exploits network and content information separately, while ignores the correlations between them. Another drawback of Simple-CTV worth mentioning is that, since it runs TOT separately on different corpus, topics in a certain community are not shared by others. Therefore, it only provides us with disjoint views of different communities, and fails to consider the social media as a whole.

We fix the hyperparameters to ρ = α = β = = δ0 = 0.01 and δ1 = 1, and λ0 and λ1 are determined as described in Section 3.3.
We calculate the complete log-likelihood of the data as the proxy to monitor the convergence of the Gibbs sampling algorithm,
L =log

P (cij |πi )P (zij |θcij )P (tij |ϕzij cij ) i j

P (wijl |φzij )fijl P (wijl |φB )1−fijl l P (sii |πi )P (sii |πi )P (eii |ηsii

+ log i s

ii

) ,

i eii ∈Ei

where the first part on the RHS is the log-likelihood of text and time, and the second part is that of links. Figure 11 shows that the log-likelihood as a function of the number of iterations. It converges after a small number of iterations. The convergence behavior is roughly the same under different configurations of the number of communities and the number of topics. In particular, we set the number of iterations to 500 in the following experiments. All our experiments are conducted on a Linux Server with eight 2.4GHz
CPU cores and 32G memory.

Figure 11: The complete log-likelihood of the model as a function of the number of Gibbs Sampling iterations. The convergence is roughly the same under different model configurations.

5.2.2

Time Stamp Prediction

The task of time stamp prediction provides a way to quantitatively evaluate the capacity of capturing temporal dynamics of topics [23]. We compare our proposed CosTot with the following four competitors, where the first two are the existing methods providing state-of-the-art performance on this task; the third one corresponds to a subpart of CosTot; and the last one is an alternative approach of uncovering community-specific temporal variations.
Topics over Time (TOT). Similar to CosTot, TOT [23] jointly models the text and time stamp of a document by treating both words and time stamps as variables generated by latent topics. It employs a Beta distribution to model the time distribution of each topic. TOT does not exploits link data.
Enhanced User-Temporal Model with Burst-weighted Smoothing (EUTB). EUTB [29] incorporates time information by assuming that a topic is generated either by a user or a time stamp.
Hence it models the topic distributions of users and time stamps.
Network data is exploited as a regularization based on the observation that neighbors in social network tend to have similar interests. Note that an array of regularization methods are proposed in
[29], while EUTB, with the link regularization and burst-weighted

Figure 12: The prediction accuracy as a function of the tolerance range in days. For example, if we see that the difference between predicted time stamp and ground truth time stamp is within 10 days as accurate, our model achieves accuracy of 26.7% while TOT gives 10.7%. The best results are obtained by setting K = 100 for all the four models, and C = 100 for CosTot and CosTot-NoLink.

10

5.2.3

Link Prediction

posts, the perplexity is:

Link prediction is a natural generalization task in networks, and a way to measure the quality of our model in modeling link structures. As discussed in Section 4.4, we can predict links between users where the probability is above some threshold. However, since we are unaware of this threshold, we turn to area under the receiver operating characteristic curve (AUC) [7] as the metric of the accuracy of prediction algorithm. Given the rank of all nonobserved links, the AUC value can be interpreted as the probability that a randomly chosen true positive link is ranked above a randomly chosen true nonexistent link.
We compare CosTot with MMSB [1] and Link-PLSA-LDA [16].
MMSB exploits only the network data, while Link-PLSA-LDA incorporates both network and text information.
Mixed Membership Stochastic Blockmodel (MMSB). Similar to CosTot, MMSB infers a probability distribution over communities for each user, and a link formation probability for each community pair. MMSB does not exploits text data.
Link-PLSA-LDA. Link-PLSA-LDA defines a generative process for both text and citations between documents, where text generation is following the LDA approach, and citations are models as multinomial sampling of the target document from a topic-specific distribution over documents. We can also interpret Link-PLSALDA in another way by regarding documents as users in social media, words as user-generated text, and citations as directed links between users. In this perspective, links and text are generated by the same latent factor, which means one community is bound to one topic (as we see the latent factor as community when generating links, and as topic when generating text).
We randomly select 20% of the positive links and 1% of the negative links to evaluate the AUC; models are trained on the remaining links and all posts. Figure 13 gives the AUC values for the three models. We see that by incorporating content information of users, Link-PLSA-LDA and CosTot outperform MMSB significantly. CosTot outperforms Link-PLSA-LDA, since it is coherent with the fact that a community in the real world have varying levels of interests in multiple topics.

perplexity(Dtest ) = exp

M d=1 log p(wd )
M
d=1 Nd

,

where p(wd ) is the probability of the words in the test post; for
CosTot, it is computed as: p(wd ) =

P (c|πi ) c P (k|θc ) k (χP (wdl |φk ) + (1 − χ)P (wdl |φB )), l where i is the author of the post.
We compare CosTot with three competitors: TOT, EUTB and
Link-PLSA-LDA, and the results are shown in Figure 14. We see that CosTot has the lowest perplexity (i.e. best text prediction performance) among all the competitors. In contrast, Link-PLSALDA shows a poor performance, since its topics are tangled with communities in the same latent factor, and thus their fitness in modeling text is weakened by links.

Figure 14: The perplexity values. The best results are obtained by setting K = 100 for all the four models, and C = 100 for CosTot.

5.2.5

Parameter Study

The two parameters, i.e., the number of communities C and the number of topics K, are critical to the performance of CosTot. We therefore study the impacts of these parameters in different tasks.
Here we show the experimental results for link prediction and perplexity, which demonstrate the different roles of C and K in determining the model performance.
Figure 15 shows the AUC values of link prediction under different settings, C ∈ {20, 50, 100, 150} and K ∈ {20, 50, 100, 150}.
We see that given a fixed K, the AUC value at first increases as
C increases, and there is an intermediate value of C (i.e. 100) at which CosTot has the best performance. After that the AUC value decreases as C continues to increase. On the other hand, given any fixed C, the result fluctuates slightly without a clear pattern as K varies, indicating that the number of topics is less important for link prediction than the number of communities. The underlying reason is that, in CosTot links are generated by mixture of communities, hence the number of communities directly impacts the capacity of modeling network. In contrast, although there exists correlations between text and network, the influence of topics on network modeling is indirect.
Figure 16 shows the impacts of C and K with regard to the quality of CosTot in text modeling. We see that perplexity decreases with the increasing number of topics, while remain stable as the number of communities varies. The result is reasonable since topics account for generating text. It is also worth mentioning that the

Figure 13: The AUC values of the link prediction task. The best results are obtained by setting K = 100 for Link-PLSA-LDA and
CosTot, and C = 100 for MMSB and CosTot.

5.2.4



Perplexity

We evaluate the quality of our proposed CosTos in modeling text by computing the perplexity [3] of a held-out test set. As a widely used metric in language modeling, perplexity monotonically decreases in the likelihood of the test data. A lower perplexity value indicates better generalization performance. For a test set of M
11

[6] P. Doreian, V. Batagelj, and A. Ferligoj. Generalized blockmodeling, volume 25. Cambridge University Press, 2005.
[7] J. A. Hanely and B. J. McNeil. The meaning and use of the area under a receiver operating characteristic (roc) curve. Radiology,
143(1).
[8] Q. Ho, R. Yan, R. Raina, and E. P. Xing. Understanding the interaction between interests, conversations and friendships in facebook. CoRR, abs/1211.0028, 2012.
[9] J. Kleinberg. Bursty and hierarchical structure in streams. Data
Mining and Knowledge Discovery, 7(4):373–397, 2003.
[10] J. Leskovec, K. J. Lang, and M. Mahoney. Empirical comparison of algorithms for network community detection. In Proc. of WWW, pages 631–640, 2010.
[11] Y.-R. Lin, D. Margolin, B. Keegan, and D. Lazer. Voices of victory: a computational focus group framework for tracking opinion shift in real time. In Proc. of WWW, pages 737–748, 2013.
[12] Y. Liu, A. Niculescu-Mizil, and W. Gryc. Topic-link lda: joint models of topic and author community. In Proc. of ICML, pages
665–672, 2009.
[13] L. L¨ and T. Zhou. Link prediction in complex networks: A survey. u Physica A: Statistical Mechanics and its Applications,
390(6):1150–1170, 2011.
[14] Y. Matsubara, Y. Sakurai, B. A. Prakash, L. Li, and C. Faloutsos.
Rise and fall patterns of information diffusion: model and implications. In Proc. of KDD, pages 6–14, 2012.
[15] J. D. Mcauliffe and D. M. Blei. Supervised topic models. In Proc. of
NIPS, pages 121–128, 2007.
[16] R. Nallapati and W. W. Cohen. Link-plsa-lda: A new unsupervised model for topics and influence of blogs. In ICWSM, 2008.
[17] R. M. Nallapati, A. Ahmed, E. P. Xing, and W. W. Cohen. Joint latent topic models for text and citations. In Proc. of KDD, pages 542–550,
2008.
[18] N. Pathak, C. DeLong, A. Banerjee, and K. Erickson. Social topic models for community extraction. In The 2nd SNA-KDD Workshop, volume 8, 2008.
[19] L. Ren, D. B. Dunson, and L. Carin. The dynamic hierarchical dirichlet process. In Proc. of ICML, pages 824–831, 2008.
[20] Y. Ruan, D. Fuhry, and S. Parthasarathy. Efficient community detection in large networks using content and links. In Proc. of
WWW, pages 1089–1098, 2013.
[21] M. Sachan, D. Contractor, T. A. Faruquie, and L. V. Subramaniam.
Using content and interactions for discovering communities in social networks. In Proc. of WWW, pages 331–340, 2012.
[22] D. D. Walker, K. Seppi, and E. K. Ringger. Topics over nonparametric time: A supervised topic model using bayesian nonparametric density estimation. In UAI Applications Workshop,
2012.
[23] X. Wang and A. McCallum. Topics over time: a non-markov continuous-time model of topical trends. In Proc. of KDD, pages
424–433, 2006.
[24] S. Wasserman. Social network analysis: Methods and applications, volume 8. Cambridge university press, 1994.
[25] J. Xie, S. Kelley, and B. K. Szymanski. Overlapping community detection in networks: The state-of-the-art and comparative study.
ACM Comput. Surv., 45(4):43:1–43:35, Aug 2013.
[26] J. Yang and J. Leskovec. Patterns of temporal variation in online media. In Proc. of WSDM, pages 177–186, 2011.
[27] J. Yao, B. Cui, Y. Huang, and X. Jin. Temporal and social context based burst detection from folksonomies. In Proc. of AAAI, pages
1474–1479, 2010.
[28] Z. Yaojia, Y. Xiaoran, G. Lise, and M. Cristopher. Scalable text and link analysis with mixed-topic link models. In Proc. of KDD, 2013.
[29] H. Yin, B. Cui, H. Lu, Y. Huang, and J. Yao. A unified model for stable and temporal topic detection from social media data. In Proc. of ICDE, 2013.
[30] J. Zhang, Y. Song, C. Zhang, and S. Liu. Evolutionary hierarchical dirichlet processes for multiple correlated time-varying corpora. In
Proc. of KDD, pages 1079–1088, 2010.
[31] W. X. Zhao, J. Jiang, J. Weng, J. He, E.-P. Lim, H. Yan, and X. Li.
Comparing twitter and traditional media using topic models. In Proc. of ECIR, pages 338–349, 2011.

Figure 15: The impact of model parameters C and K in the task of link prediction.

Figure 16: The impact of model parameters C and K in perplexity.

performance does not change significantly when K is larger than
100.

6.

CONCLUSION

In this paper, we have addressed the problem of temporal topic dynamics within different communities in social media. We presented CosTot (Community Specific Topics-over-Time), a probabilistic longitudinal model jointly over network, text and time, to simultaneously uncover the hidden topics and communities, and capture the community-specific temporal variation of topics. We provided efficient inference implementation and abundant applications to demonstrate the feasibility and usefulness of this model.
In the empirical study, our model achieved best performance in the tasks of time stamp prediction, link prediction and text perplexity among several competitors. We also provided several novel visualization examples of topic temporal patterns at different granularities, which clearly show how topics attract attentions from different communities. For future work, we are interested in extending the model to further capture network dynamics. We also would like to incorporate information diffusion among different communities that leads to the observed temporal patterns.

7.

REFERENCES

[1] E. M. Airoldi, D. M. Blei, S. E. Fienberg, and E. P. Xing. Mixed membership stochastic blockmodels. JMLR, 9:1981–2014, 2008.
[2] D. M. Blei and J. D. Lafferty. Dynamic topic models. In Proc. of
ICML, pages 113–120, 2006.
[3] D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent dirichlet allocation.
JMLR, 3:993–1022, 2003.
[4] J. Chang and D. M. Blei. Relational topic models for document networks. In Proc. of Artificial Intelligence and Statistics, pages
81–88, 2009.
[5] Q. Diao, J. Jiang, F. Zhu, and E.-P. Lim. Finding bursty topics from microblogs. In Proc. of ACL, pages 536–544, 2012.

12

Similar Documents

Premium Essay

Marketing

...ASB-1104 Introduction to Marketing Assignment 1 In view of the dynamic nature of the marketing environment, to what extent do you consider consumers to be, in practice, central to marketing activities? Name: ZHUOMING AN Student No: 500356688 Tutor: David James Introduction What is marketing? The answer is not changeless. There are some different definitions about marketing. The Chartered Institute of Marketing define that "Marketing is the management process responsible for identifying, anticipating and satisfying consumers' requirements profitably." (CIM). Taking a concern into this definition, it indicates that marketing begins before a product or service is developed. In additional, it also explain that marketing involves identifying an unsatisfied consumer need or want and determining if a profitable opportunity exists. Another definition is that “A social and managerial process by which individuals and groups obtain what they need and want through creating and exchanging products and value with others." (Kotler et al., 2005). The basic idea of this definition is that core to all marketing activities is customer satisfaction, which means marketing is an ongoing process as consumer demands and the environment is constantly changing. Products need to adapt as demands change. At the same time, marketing does not involve misleading, tricking or manipulation the customer. The Jobber also define the marketing is "The achievement of corporate goals through meeting...

Words: 1933 - Pages: 8

Premium Essay

Marketing

...*/introduction Critically evaluates the marketing planning process Discusses impediments to effective implementation of marketing plan Introduction The leading exponents of the marketing planning have been warned of the communications factors, operational, cultural and managerial in which frequently impede the effective implementation of the marketing planning programmers in the past two decades. (Cravens, 1998; Doyle, 1998; Greenley, 1982; Leeflang and de Mortanges, 1996; McDonald, 1992a, b; 1995; Piercy and Morgan, 1994; Jain, 1993; Simkin, 1996a, b; Verhage and Waarts, 1988). There have some specific guidance are offered in the recent years to assist marketing managers overcoming those internal organisational and in pre-empting forces (cf. Cravens, 1998; Dibb et al., 1996; Lings, 1999; Piercy, 1997; 1998; Simkin, 2000). Yet, the recent research has shows barriers to the implementation of programmes and marketing strategies. (Dibb and Simkin, 2001; Simkin, 2000). Another key barrier is indicating impeding the deployment of effective marketing practices used to be the lack in most marketing function or either in organisations. (cf. McDonald, 1992a, b; Piercy and Morgan, 1994). The research are shows this is a no longer to the case with the bulkiness businesses professing to have a marketing department undertaking not only promotion and customer research,but are relate to the Kotleresque textbook approach to marketing management (Dibb and Simkin, 1997; Piercy...

Words: 1580 - Pages: 7

Premium Essay

Marketing

...A high-level data model in business or for any functional area is an abstract model that documents and organizes the business data for communication between functional and technical people. It is used to show the data needed and created by business processes. A data model in software engineering is an abstract model that documents and organizes the business data for communication between team members and is used as a plan for developing applications, specifically how data are stored and accessed. An entity-relationship model (ERM) is an abstract conceptual data model (or semantic data model) used in software engineering to represent structured data. There are several notations used for ERMs. Methodology: 1. Use E-R model to get a high-level graphical view of essential components of enterprise and how they are related 2. Then convert E-R diagram to SQL DDL, or whatever database model you are using E-R Model is not SQL based. It's not limited to any particular DBMS. It is a conceptual and semantic model – captures meanings rather than an actual implementation The E-R Model: The enterprise is viewed as set of * Entities * Relationships among entities Symbols used in E-R Diagram * Entity – rectangle * Attribute – oval * Relationship – diamond * Link - line Ellipsis (plural ellipses; from the Ancient Greek: ἔλλειψις, élleipsis, "omission" or "falling short") is a series of dots that usually indicate an intentional omission of a word, sentence...

Words: 1759 - Pages: 8

Premium Essay

Marketing

...Defining Marketing Paper ShuMikki Stinson MKT/351 Thomas Collins August 26, 2013 Marketing is a tool that can help business owners to prosper with their product. Marketing can be defined as an instrument that is helpful to individuals to promote their goods in their unique way that other people items would not be an eye-catcher. It is vital to know that the key thing for every marketing tool that the individual uses will be used for the proper reasons. By the individual using marketing tools in the proper way, the individual will reach their marketing level in which they wish to be at. There are side effects when it comes to the marketing tools and the techniques which they do not work properly as it was intended too. There are numerous of definitions that relate to the marketing field. According to William D. Perreault, the author of Basic Marketing, the best definition of marketing is the performance of activities that seek to accomplish an organization’s objectives by anticipating customer or client needs and directing a flow of need satisfying goods and services from producer to customer or client. (William D. Perreault, 2011). Marketing is important because it can be a benefit for a business or a hardship. Respectable marketing methods can make a transformation between a concrete upsurges in sales to an impasse circumstances on a quality merchandise. Marketing can be as simple as having a general conversation, which is usually the case in the long run. Marketing is vital...

Words: 767 - Pages: 4

Premium Essay

Marketing

...Omar Rochell Marketing MKT/421 April 7, 2011 Nikki Jackson Introduction Marketing is exposed to someone every day, even when they do not seem realize it. Driving down the roads you see billboards everywhere and that is part of marketing. Logos people were on their shirts and signs in the middle or on the sign of football fields are all part of marketing. Even when a child is marketing themselves to their parents to borrow the car or go to a party they are marketing themselves to their parents in exchange for the car or the party. A set of activities that will benefit both parties’ objectives is my own personal definition of marketing. This paper will be defining marketing in different perspectives. Discussing the importance of marketing in a organizational success will also be discussed with examples included from different organizations. As an organization it is important to know what marketing is and how to establish success. What is Marketing “Marketing is defined as the activity, set of institutions, and processes for creating, communicating, delivering, and exchanging offerings that will have value for customers, clients, partners, and society at large.”(American Marketing Association, 2011) Marketing is a process that helps links the consumer, customer, and public to information that will help identify and market opportunities. Marketing research will generate, and evaluate different types of market actions, monitor marketing performance, and help improve...

Words: 1088 - Pages: 5

Premium Essay

Marketing

...focus inward on the organization’s needs instead of outward (the customer’s needs). • Product is aimed at everyone. • Firms want to profit through maximizing sales volume. • Promotion to achieve goals. 2. Describe some of the characteristics of a firm that would follow a marketing orientation. Marketing orientation is “a philosophy that assumes that a sale does not depend on an aggressive sales force but rather than on a customer’s decision to purchase a product; it is synonymous with the marketing concept.” • Unlike sales orientation, a firm would focus outward on the customers wants and needs. • The goal of a firm is to satisfy customers wants and needs and delivering superior value. • The target is specific groups of people. • Where sales orientation profits by sales volume, marketing orientation firms profit with good feedback from customers or customer satisfaction. • It’s more about marketing and less about selling (less persuasion). • Firms identify what customers want and have businesses give them what they want efficiently. 3. In what ways does McDonald's embody both a marketing and a societal marketing orientation? Do some internet research if necessary. McDonald’s embodies a marketing orientation...

Words: 1110 - Pages: 5

Premium Essay

Marketing

...customer-focused and heavily committed to marketing. These companies share a passion for understanding and satisfying customer needs in well-defined target markets. They motivate everyone in the organization to help build lasting customer relationships based on creating value. Marketing is just as important for non-profit-making organizations as it is for profit-making ones. It is very important to realize that at the heart of marketing is the customer. It is the management process responsible for identifying, anticipating and satisfying consumer requirements profitability. Background The term ‘‘marketing’’ is derived from the word ‘‘market’’, which refers to a group of sellers and buyers that cooperate to exchange goods and services. The modern concept of marketing evolved during and after the revolution in the 19th and 20th centuries. During that period, the proliferation of goods and services, increased worker specialization and technological advances in transportation, refrigeration and other factors that facilitate the transfer of goods over long distances resulted in the need for more advance market mechanisms and selling techniques. But it was not until the 1930s that companies began to place a greater emphasis on advertising and promoting their products and began striving to tailor their goods to specific consumer needs. By the 1950s, many larger companies were sporting entire marketing departments charged with devising and implementing marketing strategies that would complement...

Words: 2190 - Pages: 9

Premium Essay

Marketing

...MARKETING PLAN RESEARCH DEFINITION: A marketing plan is a business document written for the purpose of describing the current market position of a business and its marketing strategy for the period covered by the marketing plan. Marketing plans usually have a life of from one to five years. PURPOSE: The purpose of creating a marketing plan is to clearly show what steps will be undertaken to achieve the business' marketing objectives. CONTENT OF MARKETING: A marketing plan for a small business typically includes Small Business Administration Description of competitors, including the level of demand for the product or service and the strengths and weaknesses of competitors. 1. Description of the product or service, including special features 2. Marketing budget, including the advertising and promotional plan 3. Description of the business location, including advantages and disadvantages for marketing 4. Pricing strategy 5. Market Segmentation The main contents in marketing plan are: * Executive Summary Brief statement of goals and recommendations based on hard data. * Environmental Analysis Presents data on the market, product, competition, distribution, macro-environment. (Product fact book) S.P.I.N.S. Situation “Where am I”, Problem identification/Implications “What is happening”, Needs Assessment “Why is it happening”, Solutions “What can I do about it” Market Situation: Data on target market, size and growth for past years...

Words: 579 - Pages: 3

Premium Essay

Marketing

...Marketing MKT 421 Marketing According to “American Marketing Association” (2013), “Marketing is the activity, set of institutions, and processes for creating, communicating, delivering, and exchanging offerings that have value for customer, clients, partners, and society at large.” The American Marketing Society has grown to be the largest marketing associations in the world. The members work, teach, and study in the field of marketing across the globe. Another definition of marketing is according to “About.com Investors” (2013), “Marketing is an activity. Marketing activities and strategies result in making products available that satisfy customers while making profits for the companies that offer those products.” Organizations success lies in marketing and it is the heart of the success. The marketing introduces a product or service to potential customers. An organization can offer the best service or product in the industry but the potential customers would not know about it without marketing. Sales could crash and organizations may close without marketing. For a business to succeed the product or service that is provided needs to be known to the potential buyers. Getting the word out is important part of marketing in any organizational success. Product or service awareness is created by marketing strategies. If marketing is not used the potential customers will never be aware of the organizational offerings and the organization will not have the opportunity to succeed...

Words: 776 - Pages: 4

Premium Essay

Marketing

...chapter 1 Marketing’s Role in the Global Economy When You Finish This Chapter, You Should 1. Know what marketing is and why you should learn about it. 2. Understand the difference between micro-marketing and macro-marketing. 3. Know why and how macromarketing systems develop. 4. Understand why marketing is crucial to economic development and our global economy. 5. Know why marketing special— ists—including middlemen and — facilitators—develop. 6. Know the marketing functions and who performs them. 7. Understand the important new terms (shown in red). www.mhhe. When it’s time to roll out of bed in the morning, does your General Electric alarm wake you with a buzzer—or by playing your favorite radio station? Is the station playing rock, classical, or country music—or perhaps a Red Cross ad asking you to contribute blood? Will you slip into your Levi’s jeans, your shirt from L. L. Bean, and your Reeboks, or does the day call for your Brooks Brothers interviewing suit? Will breakfast be Lender’s Bagels with cream cheese or Kellogg’s Frosted Flakes—made with grain from America’s heartland—or some extra large eggs and Oscar Mayer bacon cooked in a Panasonic microwave oven imported from Japan? Will you drink decaffeinated Maxwell House coffee—grown in Colombia—or some Tang instant juice? Will you eat at home or is this a day to meet a friend at the Marriott-run cafeteria—where you’ll pay someone else to serve your breakfast? After breakfast, will you head off to school...

Words: 14069 - Pages: 57

Premium Essay

Marketing

...Abstract In the world of today with rude competition everywhere, customers’ expectations have become higher than ever. It is not the customers who come towards the products but it is the products which should make their way to the customers. And for this, only competitive businesses that are able to stimulate customers’ interests survive in the market. Therefore firms need to increase customers’ awareness about their products or services to be able to pull and encourage them to engage in purchase of their products. And as such, the promotional mix used by a company is really important for this task. The promotional mix in itself is very broad, consisting of various tools, like advertising, personal selling, direct marketing, public relation and sales promotion. To make the optimum use of these tools, marketers usually select them, depending on their budget and objectives, as well as the sector in which they operate (Kotler & Armstrong 1997). As such, research has been conducted on the use of promotional mix and research questions and objectives have been set. The methodology which will be used has been devised. We shall be doing a descriptive study through a survey questionnaire, in which there will be open as well as close ended questions and the questionnaire will be administered through personal interview that is direct, face-to-face. The sample size will be 100 persons and will all be customers of J Kalachand & Co Ltd. After the research, we will be...

Words: 4233 - Pages: 17

Premium Essay

Marketing

...Marketing is the process of communicating the value of a product or service to customers, for the purpose of selling that product or service. Marketing can be looked at as an organizational function and a set of processes for creating, delivering and communicating value to customers, and customer relationship management that also benefits the organization. Marketing is the science of choosing target markets through market analysis and market segmentation, as well as understanding consumer behavior and providing superior customer value. From a societal point of view, marketing is the link between a society’s material requirements and its economic patterns of response. Marketing satisfies these needs and wants through exchange processes and building long term relationships. Organizations may choose to operate a business under five competing concepts: the production concept, the product concept, the selling concept, the marketing concept, and the holistic marketing concept.[1] The four components of holistic marketing are relationship marketing, internal marketing, integrated marketing, and socially responsive marketing. The set of engagements necessary for successful marketing management includes capturing marketing insights, connecting with customers, building strong brands, shaping the market offerings, delivering and communicating value, creating long-term growth, and developing marketing strategies and plans.[2] Marketing may be defined in several ways, depending on...

Words: 270 - Pages: 2

Premium Essay

Marketing

...oriented philosophy is so important. The phrase market-oriented is used in marketing conversations as an adjective describing a company with a marketing orientation. Market orientation more describes the company's approach to doing business. Market-oriented defines the company itself. If a company is market-oriented, its board and executive leadership believe that the best way to succeed is to prioritize the marketplace above products. This usually goes over well with customers, but the company also must have adequate research and development to provide what the market wants. Hence, a market-oriented organization is one whose actions are consistent with the marketing concept. Difference Between Marketing Orientation & Market Oriented by Neil Kokemuller, Demand Media http://smallbusiness.chron.com/difference-between-marketing-orientation-market-oriented-14387.html Marketing is a management process and management support for marketing concept is very important element in success. If a company wants to be successful then it is market oriented. Marketing involves identifying the customer requirements and estimate the customer requirements in future. It requires planning which is very important process of marketing. To satisfy the needs the business should provide benefits – offering right marketing at right time at right place. Generally market based companies adopt strategic level marketing that defines the mission and long term objectives of the company. Market oriented...

Words: 716 - Pages: 3

Premium Essay

Marketing

...qwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmrtyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmrtyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmrtyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmqwertyuiopasdfghjklzxcvbnmrtyuiopasdfghjklzxcvbnmqwer...

Words: 789 - Pages: 4

Premium Essay

Marketing

...Assessment: MKC1 Market Environmental Variables Reading: Contemporary Marketing: Chapter 3 Questions: 1. How would you categorize Generation X using the five segments of the marketing environment? A: Competitive Environment B: Political-legal environment C: Economic environment D: Technological environment E: Social-cultural environment 2. Joe and Ryan both have storefronts in the local mall. Joe sells candies and Ryan sells pretzels. Are Joe and Ryan in direct competition with each other? A: Yes B: No Consumer Behavior and Marketing Reading: Contemporary Marketing: Chapter 5 Questions: 1. Rachel and Sarah’s parents always purchased groceries from the local Aldi marketplace. What is this type of behavior an example of? A: Cultural influences B: Social Influences C: Personal factors 2. Maryanne purchases Maxwell House coffee every two weeks from the grocery. What is this type of behavior an example of? A: Routinized Problem Solving B: Limited problem solving C: Extended problem solving 3. Aaron does research on several local colleges before applying to his first three choices. This is an example of: A: High – involvement purchase decision B: Low – involvement purchase decision Marketing Plans Reading: Contemporary Marketing: Chapter 2 + Ch. 2 Appendix Web sites: http://www.jpec.org/handouts/jpec33.pdf http://www.netmba.com/marketing/process/ Questions: 1. Strategies are designed to meet objectives...

Words: 8933 - Pages: 36