Abstract
Session-based recommender systems (SBRSs) often struggle to capture the complex, evolving social dynamics of user preferences. We present LSUG (Long- and Short-term latent User Group modeling), a framework that moves beyond independent user analysis by dynamically mapping individuals to latent "like-minded" groups. LSUG incorporates an adaptive mechanism to evolve the number of groups automatically, addressing the limitations of static clustering. Experimental results on Tmall and Gowalla datasets demonstrate that LSUG significantly outperforms existing state-of-the-art models in next-item prediction.
Introduction
Modern recommendation engines, such as those powering "what to watch next" features, typically analyze a user's session history in isolation. This approach overlooks the fact that human behavior is inherently social; our preferences often mirror those of similar individuals. Existing systems face four primary hurdles: the over-independence of user models, the multi-faceted nature of user identity (e.g., being both a tech enthusiast and a cartoon fan), temporal shifts in group affiliation, and the "K problem"—the difficulty of pre-determining the optimal number of user clusters in a dynamic environment.
The LSUG Framework
The proposed LSUG model utilizes a hierarchical neural network to fuse individual history with group-level insights. The architecture is built upon several key innovations:
- Long- and Short-term Splitting: The model bifurcates user history into long-term preferences (historical sessions) and short-term demands (the current session).
- Group Influence Component (GIC): Inspired by key-value memory networks, the GIC uses "anchor" embeddings to determine a user's current latent group affiliations. It calculates a probability distribution to identify which "crowd" most influences the user at any given moment.
- Hybrid Representation: The final user profile is a weighted combination of four distinct factors: personal long-term, personal short-term, group long-term, and group short-term influences. These are fused using either a Multi-Layer Perceptron (MLP) or an Attention mechanism, depending on the dataset's characteristics.
Adaptive Grouping Mechanism
A significant contribution of this work is the A-LSUG (Adaptive LSUG) variant. Traditional models often require researchers to fix the number of groups ($L$) beforehand. A-LSUG removes this constraint by assessing whether a user's current behavior is sufficiently represented by existing groups. If a similarity threshold is not met, the system automatically spawns a new latent group. This allows the model to grow organically with the data, mimicking the emergence of new subcultures or trends in the real world.
Experimental Evaluation
The model was evaluated on two large-scale real-world datasets: Tmall (e-commerce) and Gowalla (location-based check-ins). LSUG was compared against several state-of-the-art baselines, including GRU4Rec, SHAN, and CMN.
Key Findings
- Superior Accuracy: LSUG consistently outperformed all baselines across Recall, Mean Average Precision (mAP), and Area Under Curve (AUC). On the Tmall dataset, LSUG achieved a 16.9% improvement in Recall@20 compared to the SHAN model.
- Adaptive Gains: The A-LSUG variant showed even higher predictive accuracy, with Recall@20 on Tmall jumping from 0.159 to 0.184, proving that automated group discovery is superior to manual tuning.
- Contextual Fusion: The researchers noted that the optimal fusion mechanism varies by context. The Attention mechanism performed best on Gowalla, where users frequently revisit locations, while the MLP was superior for the more varied shopping behaviors found in Tmall.
Conclusion
LSUG demonstrates that integrating latent group dynamics into session-based recommendations significantly enhances predictive power. By allowing group affiliations to be both dynamic and adaptive, the model captures the fluid nature of human preferences more effectively than static or independent models. Future work will focus on implementing "pruning" mechanisms to remove obsolete groups and utilizing contrastive learning to further refine group representations.
Summary updated