Abstract: Today, the ability to track users' sequence of online activities, makes identifying their evolving preferences for recommendation practicable. However, despite the myriad of available online activity information, most existing time-based recommender systems either focus on predicting some user rating, or rely on information from similar users. These systems, therefore, disregard the temporal and contextual aspects of users preferences, revealed in the rich and useful historical sequential information, which can potentially increase recommendation accuracy. In this work, we consider such rich, user online activity sequence, as a complex dependency of each user's consumption sequence, and combine the concept of collaborative filtering with long short-term memory recurrent neural network (LSTM-RNN), to make personalized recommendations. Specifically, we use encoder-decoder LSTM-RNN, to make sequence-to-sequence recommendations. Our proposed model builds on the strength of collaborative filtering while preserving individual user preferences for personalized recommendation. We conduct experiments using Movielens dataset to evaluate our proposed model and empirically demonstrate that it improves recommendation accuracy when compared to state-of-the-art recommender systems.
Authors: Benjamin A Kwapong, Richard Anarfi and Kenneth K Fletcher (University of Massachusetts Boston, USA)
Email: benjamin.kwapong001@umb.edu, richard.anarfi001@umb.edu, kenneth.fletcher@umb.edu