Context-Aware and Adaptive Multi-Scale Interest Modeling for Sequential Recommendation

Abstract

Sequential recommendation aims to predict a user's next interaction by modeling ordered user-item behavior sequences and plays a critical role in modern recommender systems. In real-world scenarios, user behavior is influenced by multiple contextual factors, among which temporal dynamics and item popularity are particularly important. Time intervals between interactions reflect the evolution and decay of user interests, while item popularity introduces frequency bias that may cause recommender systems to overemphasize popular items and underrepresent long-tail preferences. Moreover, user interests naturally exist at different temporal scales: long-term behaviors capture stable and persistent preferences, while recent interactions often reflect short-term, context-dependent intents. Effectively modeling these signals and integrating long-term and short-term interests remains a central challenge in sequential recommendation.

To address these challenges, this thesis presents two sequential recommendation models with increasing modeling capability. The first model, Dual-Gated Time Frequency Co-Modeling for Sequential Recommendation (TiIfSRec), is designed to explicitly incorporate temporal intervals and item-frequency information into long-term and short-term interest modeling. TiIfSRec employs a dual-gated recurrent architecture in which time-interval signals control the decay of historical preferences, while item-frequency signals regulate the direction of state updates to mitigate popularity bias and preserve long-tail information. An attention mechanism is further introduced to highlight informative historical interactions and improve long-range dependency modeling. Experiments on multiple Amazon benchmark datasets demonstrate that TiIfSRec consistently outperforms representative time-aware and popularity-aware baselines, validating the effectiveness of jointly modeling temporal and frequency signals for sequential recommendation.

While TiIfSRec improves long-term and short-term interest modeling, its fusion strategy relies on deterministic mechanisms, which limits flexibility when balancing stable preferences and rapidly changing intents. Motivated by this limitation, the second model, Diffusion-based Long-Short Interest Fusion for Sequential Recommendation (DiffLSRec), formulates long-short interest integration as a generative process. DiffLSRec introduces a diffusion-based framework in which the long-term interest representation serves as a generative prior, and the short-term interest representation is incorporated as conditional guidance during a multi-step denoising process. This progressive fusion strategy enables the model to adaptively adjust the contribution of long-term and short-term interests at different stages, rather than relying on a single static fusion operation. To further enhance contextual modeling and stability, DiffLSRec incorporates token-level contextual enhancement to capture fine-grained recent behavioral patterns, as well as a monotonic signal-to-noise ratio-adaptive guidance mechanism to regulate the influence of short-term signals throughout the diffusion trajectory. Extensive experiments show that DiffLSRec consistently outperforms strong sequential recommendation baselines, including diffusion-based models, across multiple evaluation metrics.

Overall, this thesis demonstrates that explicitly modeling contextual behavioral signals and progressively integrating user interests across different temporal scales can substantially improve the accuracy, robustness, and adaptability of sequential recommender systems. The proposed models provide complementary contributions, with TiIfSRec offering effective context-aware interest encoding and DiffLSRec introducing a flexible generative paradigm for long-short interest fusion.