Introduction
Sequential Recommender Systems (SRSs) have become pivotal in e-commerce, aiming to predict a user's next interaction based on their historical sequence of purchases. However, traditional models often suffer from two critical oversights. First, they focus primarily on dynamic item sequences, neglecting static interests—stable preferences for brands or price points—that define a user's long-term behavior. Second, they operate under a "positive bias," assuming every past interaction only increases the likelihood of future ones.
In reality, purchasing an item can create negative excitation. For instance, buying a smartphone significantly decreases the immediate probability of purchasing another smartphone (a substitute), even while it increases the likelihood of buying a protective case (a complement). Failing to model this inhibitory effect leads to redundant or illogical recommendations.
The Static-Dynamic Interest Learning (SDIL) Framework
To address these gaps, the Static-Dynamic Interest Learning (SDIL) framework was developed. SDIL integrates long-term user traits with short-term behavioral shifts while explicitly modeling both positive and negative temporal excitations.
1. Static Interest Modeling (SIM)
The SIM module captures the user’s "vibe" by analyzing item attributes such as category, brand, and price. Using a multi-head self-attention mechanism, it extracts stable signals from the user's history, ensuring that recommendations align with their general preferences regardless of recent trends.
2. Temporal Positive and Negative Excitation (TPNE)
The core innovation of SDIL is the TPNE module, which extends the classical Hawkes Process. While standard Hawkes models only account for how events trigger future occurrences (positive excitation), TPNE introduces a negative term to model inhibitory effects.
- Relation-Based Logic: Items are categorized as complements (triggering positive excitation) or substitutes (triggering negative excitation).
- Time-Decay Kernels: The model employs Gaussian-style kernels to determine the temporal window of these signals. For example, the negative excitation for a substitute item is strongest immediately after a purchase and decays over time.
3. Gate Fusion and Prediction
A gating mechanism intelligently weighs the outputs of the static and dynamic modules. This ensures the final prediction is a balanced reflection of who the user is (static) and what they are currently seeking (dynamic).
Experimental Evaluation
The SDIL framework was evaluated on three Amazon datasets: Beauty, Cell Phones and Accessories, and Toys and Games. These datasets provide the rich metadata and relationship information necessary to test the TPNE module.
Performance Results
SDIL consistently outperformed state-of-the-art baselines, including SASRec and TiSASRec. Notably, on the Cellphones dataset, the model achieved a 10.45% improvement in Mean Reciprocal Rank (MRR) over the strongest baseline. Ablation studies confirmed that both the static interest module and the negative excitation modeling are vital; removing either component resulted in a significant performance decline.
Conclusion
By acknowledging that recommendations are a balancing act between static identity and dynamic behavior—and by recognizing that some purchases inhibit others—the SDIL framework provides a more logical and accurate approach to sequential recommendation. Future work will focus on automatically learning negative relationships from behavior to reduce dependency on explicit metadata.
Summary updated