TL;DR
This paper introduces ScoreNet, a novel framework for session-based recommendation that tackles the challenge of semantic heterogeneity between item IDs and side information. By explicitly modeling user re-engagement patterns and ensuring consistency in the decoding space, ScoreNet significantly improves prediction accuracy, especially for items with limited interaction history.
What Problem Does This Paper Address?
Session-based recommender systems (SBRs) aim to predict a user's next action based on their recent, anonymous browsing history. While incorporating "side information" like item categories or brands can enrich recommendations, existing methods often struggle with the inherent semantic differences between item IDs and this auxiliary data. This mismatch can lead to noisy attention calculations, inconsistent user interest modeling, and ultimately, inaccurate predictions. Furthermore, many approaches overlook the valuable patterns of users re-engaging with items that share similar attributes, limiting their ability to capture persistent user preferences.
Methods
ScoreNet proposes a unified framework to address these issues by integrating multi-side information directly into the candidate item scoring process, ensuring decoding consistency. Its core components are:
- Multi-Path Re-Engagement Network (MPRE): This network independently processes various side information sequences to capture re-interaction behavior patterns. It models users' persistent interests in items with similar attributes in a semantic-agnostic way, preventing interference from semantic variations. It also incorporates a score redistribution mechanism to adjust re-engagement scores based on overall user behavior.
- Position-Enhanced Consistent Scoring Network (PECS): This component specifically targets items with limited user interactions. It creates a consistent semantic space between session and item ID representations by using a position-enhanced encoder. This helps to align session representations and item ID embeddings, improving prediction accuracy. PECS also redistributes attention scores within sessions to further refine predictions.
The final score for a candidate item is a combination of scores from the generalization mode (handling items with limited interactions) and the re-engagement mode (capturing persistent interests).
Datasets
The authors evaluated ScoreNet on three real-world e-commerce datasets:
- Diginetica: From the 2016 CIKM Cup, containing transaction data and item categories.
- Tmall: From IJCAI-15, including anonymous shopping logs with category and brand as side information.
- Retailrocket: Released by a Kaggle competition, focusing on browsing activity with limited side information beyond item categories.
The paper follows standard data preprocessing steps from previous studies for fair comparison.
Key Insights and Ideas
- Semantic Consistency is Key: The paper strongly emphasizes that aligning the semantic space of item IDs and side information is crucial for accurate SBRs. Inconsistent representations lead to noisy attention and poor predictions.
- Re-engagement Patterns Matter: Users often revisit items with similar attributes. ScoreNet's MPRE module effectively captures these persistent interests, which is a significant improvement over methods that only focus on immediate transitions.
- Handling Cold-Start Items: PECS provides a novel way to improve recommendations for items with few interactions by leveraging positional information and ensuring consistent semantic spaces.
- Balancing Side Information: While side information is beneficial, over-reliance can degrade performance. The paper shows an inverted U-shaped relationship with a hyperparameter controlling side information influence, highlighting the need for careful balancing.
Key Results
ScoreNet significantly outperforms state-of-the-art baseline models across all three datasets on P@K and MRR@K metrics. For example, on the Diginetica dataset, ScoreNet achieved 42.76% P@10 and 19.38% MRR@10, outperforming the best baseline by a notable margin. Ablation studies confirmed the effectiveness of both the MPRE and PECS modules, with removing MPRE leading to a significant performance drop. The visualization of attention distribution also demonstrated ScoreNet's interpretability, showing how it prioritizes recent re-interactions.
Limitations and Future Work
The authors acknowledge that the hyperparameter $\alpha_2$, which balances the contribution of re-engagement behavior and semantic preference modeling, requires careful tuning. They suggest that future work could explore more advanced methods for automatically learning these hyperparameters or developing more adaptive fusion strategies. Additionally, while ScoreNet shows strong performance, further research into more sophisticated ways to handle diverse and complex side information, as well as exploring its application in other recommendation scenarios, would be valuable.
Summary updated