We propose two distinct analyses based on items re-consumed in the test set. The first one refers to the \textit{recall} of these items. Note that among the training items of each user, there are \textit{(1)} unexpected items; \textit{(2)} expected ones that users re-consumed in the test set; and \textit{(3)} expected items that users did not re-consume. Thus, given the whole training set as input, a good recommender should assign high scores to the re-consumed items. This analysis shows the capability of our heuristics to identify items effectively re-consumed instead of unexpected items potentially relevant. Hence, as the second analysis, we measure how relevant the recommended items might be, considering all users. A good evidence about relevance of an item is the percentage of users who consumed it and re-consumed again after $x$ time units. We assume that the higher this percentage, which we name re-consumption rate (\textit{rerate}), the higher the probability of this item to be a potentially relevant item after $x$ time units. Thus, a good RS of known items should present simultaneously high \textit{recall} and \textit{rerate} levels.\looseness=-1

We evaluate the \textit{recall} of re-consumed items through the following experiment. First, we use the whole training set of each user as input for each proposed heuristic. Then, we retrieve as output recommendations of …show more content…
We observe that TDH-2 presents the best results w.r.t. both \textit{recall} and \textit{rerate} in both datasets. Besides recovering significant percentages of re-consumed items, this strategy identifies items potentially more relevant in each collection. Again, we observe that ACT-R could not improve the individual information of recency, pointing out that the approach used to weight and combine recency, relevance and context diverges between learning and consumption