Accurate modeling of influence relationships between events is critical to achieving effective sequential recommendation. Representing event sequences as a Hawkes process is a widely adopted approach; however, it is suboptimal in practical applications because it overlooks the contextual information of events and assumes that all historical events exert a positive influence on current events. To address these limitations, we propose a category-aware next event prediction method. This method computes both the event-level and category-level influences of all historical events associated with the target user on candidate events and recommends the Top-k ones with the highest predicted event intensities. Specifically, it captures the temporal influence between events from the perspectives of event features and the features of the categories to which the events belong. Using embedding learning, the method derives feature vectors for both events and their corresponding categories, as well as their mutual influence relationships, thereby allowing more accurate prediction of the next event. Extensive experiments on real-world datasets demonstrate that the proposed method achieves superior performance across multiple evaluation metrics compared to conventional methods.

As one of the most important topics in recommender systems, sequential recommendation aims to predict users' next behavior based on their historical event sequences and provide a personalized recommendation list. At present, sequential recommendation has been widely used in many fields. However, two main challenges still remain in sequence data mining and modeling. Firstly, the users' behavioral preferences usually change dynamically over time. Secondly, there may be complex negative effects between the users' events. Traditional sequential recommendation methods have difficulty in dealing with those problems. In this paper, we propose a Hawkes process based sequential recommendation model to capture users' dynamic preferences as well as the complex correlation from historical behavior sequences. Specifically, the proposed model is able to mine the intrinsic relationship between events, including both positive and negative effects, which enables it to further improve the performance of recommendation. The experimental results on real-world dataset show that the proposed model achieves better performance than the baselines.