1. We formalize the problem of top-k group recommendation and use a model for group consensus similar to the social value functions developed in [18] to incorporate various predicted rating and disagreement models.
2. We propose the use of pairwise disagreement lists, and design and implement efficient group recommendation algorithms based on the merging and effective pruning of individual predicted rating lists and pairwise disagreement lists.
3. Given the potentially large number of disagreement lists, we exploit shared user behavior to reduce the space
requirement of those lists. As a result, we extend the group recommendation algorithms to process factored lists.We showthat factoring common entries in disagreement lists can drastically reduce storage space without
incurring I/O overhead.
4. The factoring strategy does not always guarantee reaching a fixed space budget. To achieve a certain space budget, we develop a partial materialization strategy which exploits the size of each disagreement list and their impact on query processing: it skips disagreement lists in order to minimize space, while incurring small processing time overhead.We formalize this question as an adaptation of the Knapsack problem and develop an algorithm to solve it.
5. We run an extensive set of experiments with different group sizes on MovieLens data sets. We perform extensive user study in Amazon’s Mechanical Turk to demonstrate the effectiveness of our group recommendation semantics and howsatisfied users are with recommended group ratings compared to individual ones.Our elaborate performance experiments exhibit the efficiency of group recommendation computation. We also demonstrate the benefit of behavior factoring and partial materialization on space.