02232nas a2200241   4500000000100000000000100001008004100002260001200043653001800055653003100073653003000104653002700134653001900161653003000180100001500210700001700225245008800242856007900330300001000409490000600419520155100425022001401976       2023                            d  c12/202310aDeep Learning10aGated Graph Neural Network10aKnowledge Graph Embedding10aRecommendation Systems10aSelf-Attention10aSequential Recommendation1 aMin-Ji Seo1 aMyung-Ho Kim00aAttentive Flexible Translation Embedding in Top-N Sparse Sequential Recommendations  uhttps://www.ijimai.org/journal/sites/default/files/2023-11/ijimai8_4_5.pdf  a56-660 v83 aSequential recommendation aims to predict the user’s next action based on personal action sequences. The major challenge in this task is how to achieve high performance recommendation under the data sparsity problem. Translation-based recommendations, which learn distance metrics to capture interactions between users and items in sequential recommendations, are a promising method to overcome this issue. However, a disadvantage of translation-based recommendations is that they capture long-term preferences of the user and complex item transitions. In this paper, we propose attentive flexible translation for recommendations (AFTRec) to tackle data sparsity problem by capturing a user’s dynamic preferences and complex interactions between items in user’s purchasing behaviors. In particular, we first encode semantic information of an item related to user’s purchasing behaviors as the user-specific item translation vectors. We also design a transition graph and encode complex item transitions as correlation-specific item translation vectors. Finally, we adopt a flexible distance metric that considers directions with respect to the translation vectors in the same space for predicting the next item. To evaluate the performance of our method, we conducted experiments on four sparse datasets and one dense dataset with different domains. The experimental results demonstrate that our proposed AFTRec outperforms the state-of-the-art baselines in terms of normalized discounted cumulative
gain and hit rate on sparse datasets.  a1989-1660