TL;DR
The BST-CDSR framework significantly advances cross-domain sequential recommendation by explicitly modeling user behavior and item semantics over continuous time. By combining a dual-behavior Neural Ordinary Differential Equation (ODE) with a counterfactual-enhanced Large Language Model (LLM), the model successfully adapts to domain-specific interaction frequencies and filters out temporary, situational interests, setting a new state-of-the-art.
What Problem Does This Paper Address?
Cross-Domain Sequential Recommendation (CDSR) helps combat the notorious "data sparsity" problem in recommendation systems by combining a user's interaction history across multiple domains (e.g., using someone's movie viewing habits to recommend books). However, current state-of-the-art models ignore two critical temporal realities:
- Domain-Specific Time Dynamics: Different domains operate on different time scales. A user might watch three movies a week but read only one book a month. A one-week gap between actions represents a massive pause (and potential interest shift) in the movie domain, but complete continuity in the book domain.
- Time-Varying Semantics (Situational Interests): User preferences are often driven by temporary contexts, like watching holiday movies during Christmas. If a model treats these "situational semantics" as permanent, stable preferences, it risks transferring irrelevant data to another domain out of season (e.g., recommending Christmas books in July), causing what researchers call negative transfer.
Methods
To bridge behavior and semantics across time, the authors introduce BST-CDSR (Bridging Behavior and Semantics for Time-aware Cross-Domain Sequential Recommendation). The architecture consists of three core innovations:
- Behavioral Preference Evolution Module: Instead of updating user states only at discrete interaction points, this module leverages a Neural ODE to model how user preferences drift smoothly and continuously over time. It uniquely decouples user behavior into long-term interests and short-term intentions, using temporal regularization to keep the two distinct.
- Temporal Counterfactual-Enhanced Semantic Generator: LLMs are notoriously bad at processing continuous numerical timestamps. The authors overcome this by mapping exact time gaps into discrete tokens (e.g.,
[Gap]) within the text prompt. To force the LLM to actually pay attention to these time tokens rather than just the item text, they introduce counterfactual perturbations—randomly altering the time tokens at small and large scales—and optimize the model using a temporal contrastive ranking loss. - Time-Preference Guided Domain Transfer Module: Instead of using static cross-domain transfer weights, a gated network dynamically computes personalized transfer weights based on the user's specific cross-domain switching patterns and preference relevance.
Datasets
The model was evaluated on three real-world Amazon dataset pairs, all characterized by extreme interaction sparsity (over 99.9%):
- Food-Kitchen
- Movie-Book
- Art-Office
Key Insights and Ideas
- Preferences Evolve Between Actions: Traditional sequence encoders (like Transformers or GRUs) only update when a user interacts with an item. Neural ODEs treat user preference as a continuous dynamical system, capturing the gradual decay of interest during long periods of inactivity.
- Making Time "LLM-Friendly": Discretizing continuous time into structured tokens allows LLMs to effectively reason about temporal context alongside semantic text descriptions.
- Counterfactuals as a Sensitivity Dial: Without explicit contrastive constraints, powerful LLMs will ignore subtle time tokens and over-index on text semantics. Swapping time tokens during training forces the model to learn when a specific semantic preference is valid.
Key Results
- Consistent State-of-the-Art Performance: BST-CDSR outperformed all baseline models—including advanced single-domain models (SASRec, TiSASRec) and advanced LLM-based CDSR models (LLM-EDT)—across all datasets. Notably, it achieved up to a 30.54% relative improvement in MRR in the Food domain.
- Superiority in Irregular Sequences: When test data was bucketed by temporal irregularity, BST-CDSR showed its largest performance gains in the most irregular buckets, validating the strength of the continuous-time Neural ODE approach.
- Resilience to Noise: In robustness tests where 10% and 20% random noise interactions were injected into the training data, BST-CDSR maintained its performance edge, proving that its time-guided transfer module successfully mitigates negative transfer.
Limitations and Future Work
The authors do not explicitly detail limitations within the paper. However, from an engineering standpoint, deploying a framework that relies on both continuous-time Neural ODE solvers and inference from a Large Language Model (such as Llama2-7B) introduces substantial computational overhead. Future work will likely need to address how to distill or optimize these components for real-time, low-latency production environments, or extend the framework to handle multi-domain scenarios beyond simple pairs.
Summary updated