Sequential Recommendation via Joint Modeling of Temporal Dynamics With LSTM and Hawkes Process

The recommendation system serves as a pivotal component of numerous online systems, playing a crucial role in enhancing user experience. Nevertheless, the majority of these models make the assumption that the user interaction history sequence is continuous, without taking into account the event intervals between user behaviors during modeling. The combination of Long Short Term Memory (LSTM) and Hawkes process incorporating the power law decay characteristics, enables the modeling of complex temporal dynamics in sequential data. To achieve this, we leverage the capabilities of a LSTM to automatically acquire a representation of the influences derived from the behavior history. Through this hybrid model, we are able to predict the next item a user is likely to interact with, taking into consideration the temporal dimension of user-item interactions. We evaluate our model on three real-world datasets, demonstrating significant improvements in recommendation performance compared to state-of-the-art baseline methods. Additionally, the inclusion of the power-law feature within the Hawkes process provides deeper insights into the underlying temporal patterns of user behavior, which can be leveraged for more refined and timely recommendations.