A Novel Fusion of Temporal Convolutional Networks and Neural Hawkes Processes for Event-Based Sequential Recommendation

The exponential growth of user interaction data on digital platforms has underscored the limitations of traditional recommendation systems, which often fail to capture the rich, continuous-time nature of user behaviour. Event-based temporal data, where each action is a time stamped event, requires models that can understand both the semantic context and the precise temporal dynamics of user sequences. This paper introduces the Temporal Convolutional Hawkes Network (TCHN), a novel architecture that integrates Temporal Convolutional Networks (TCNs) with a Neural Hawkes Process to address this challenge. The TCN component excels at capturing multi-scale, long-range dependencies within the user's interaction history, while the Neural Hawkes Process models the underlying continuous-time intensity, accounting for the excitation and decay effects of past events. We evaluate TCHN on three real-world datasets: Amazon Electronics, LastFM, and Reddit. Extensive experiments demonstrate that TCHN significantly outperforms a suite of state-of-the-art baselines, including GRU4Rec, SASRec, and RMTPP. Our model achieves an average improvement of 13.7% in HR@10 and 16.1% in NDCG@10 on the Amazon dataset, and 11.2% in HR@10 and 14.5% in NDCG@10 on the LastFM dataset. Ablation studies confirm the critical contribution of both the TCN's receptive field and the Hawkes-based intensity function. The results firmly establish that jointly modelling hierarchical patterns and continuous-time dynamics is paramount for the next generation of sequential recommenders.