Long- and short-term collaborative attention networks for sequential recommendation

Sequential recommendation models the users’ historical interaction sequence and predicts which item the user will click next. To better capture the users’ hobbies, most models utilize the users’ interaction sequence to capture the users’ long-term hobbies, while ignoring the users’ short-term intentions. Recently, some work has focused on capturing users’ long-term hobbies and short-term intentions to predict the next item recommendation. However, they only consider the information about the user’s interaction sequence and ignore the information about the collaboration between different user interaction sequences. This paper proposes a Long- and Short-Term Collaborative attention network for Sequential Recommendation (LSTCSR) to better capture and integrate users’ long-term and short-term hobbies. Specifically, we construct an item–item graph with the interaction sequences of different users to obtain information on the collaboration between items in different sequences. It then uses a self-attention network to capture the users’ long-term hobbies and utilizes convolutional filters of different sizes to capture the users’ multiple short-term intentions. Finally, the users’ long-term hobbies and short-term intentions are integrated through the collaborative information of the item–item graph to predict the next item recommendation. Experiments on 3 public benchmark datasets show that LSTCSR model outperforms several state-of-the-art methods, further demonstrating the effectiveness of the LSTCSR model.