TL;DR
This paper tackles the challenge of data sparsity in search-enhanced recommendation systems, where users with limited search history are often underserved. They propose GSERec, a novel method that leverages message passing on user-code graphs to transfer knowledge from "rich" search users to "sparse" ones, significantly improving recommendations for the latter.
What Problem Does This Paper Address?
Modern online platforms often integrate search and recommendation (S&R) functionalities to cater to diverse user needs. While incorporating search signals into recommendation models has shown promise, existing methods heavily rely on users with abundant search interactions. This leads to a significant performance gap: users with rich search histories benefit greatly, while those with sparse search activity see minimal improvement. This is problematic because sparse search users are the majority. The core challenge is twofold: (1) how to extract meaningful search features from limited user interactions, and (2) how to design effective training objectives that don't neglect these sparse users.
Methods
GSERec addresses this by constructing User-Code Graphs and employing Message Passing to transfer knowledge. The process involves:
- User Preference Summarization: A Large Language Model (LLM) is used to summarize each user's search and recommendation preferences based on their history.
- User Preference Quantization: These summarized preferences are encoded and then quantized into discrete "codes" using a Residual Quantized Variational Autoencoder (RQ-VAE). This step also incorporates contrastive learning to align user preferences before quantization.
- Graph Construction: Users are connected to their corresponding codes, forming bipartite graphs (User-Code Graphs). Users who share similar codes are implicitly linked.
- Message Passing over User-Code Graphs: A graph neural network (specifically LightGCN) is used to propagate embeddings. This allows users with rich search interactions to "teach" users with sparse interactions by sharing their learned representations through the shared codes.
- History Modeling and Prediction: The enhanced user embeddings are then integrated with the user's S&R history embeddings. This combined representation is used to predict the next item. Contrastive losses are also used to align user embeddings and history representations.
Datasets
The authors evaluated GSERec on three datasets:
- Amazon: A widely used semi-synthetic dataset, with "CDs and Vinyl" and "Electronics" subsets. Synthetic search behaviors were generated.
- Qilin: A dataset collected from Xiaohongshu, a Chinese lifestyle search engine, containing user behavior data from both S&R scenarios and multimodal item information.
- Electronics: (Implied to be another subset of the Amazon dataset, based on Table 1).
Key Insights and Ideas
- The "Rich-to-Sparse" Knowledge Transfer: The core idea of using users with abundant search data to bolster the representations of users with sparse data is a powerful way to combat sparsity. The user-code graph acts as the conduit for this transfer.
- LLMs for Preference Summarization: Leveraging LLMs to distill complex user preferences from raw interaction logs is a promising direction for feature engineering in recommendation systems.
- Quantization for Graph Construction: Discretizing user preferences into codes provides a structured way to link users and build the graph, enabling efficient message passing.
- User-Code Graphs vs. User-Item Graphs: The paper highlights that user-code graphs offer a more efficient and effective way to model user-user relationships for this specific problem compared to traditional user-item graphs used in GNNs.
Key Results
- State-of-the-Art Performance: GSERec consistently outperforms traditional recommendation models and existing search-enhanced recommendation methods across three datasets.
- Significant Gains for Sparse Users: Crucially, GSERec shows substantially larger improvements for users with sparse search interactions, directly addressing the paper's core problem. For example, Figure 2 shows GSERec achieving much larger gains than UniSAR for users with fewer search interactions.
- Ablation Studies Validate Components: Ablation studies (Table 3, Figure 5) confirm the importance of each module: removing the user-code graph or alignment losses significantly degrades performance, especially for sparse users.
- LLM-based methods excel: The paper notes that methods incorporating LLMs (like GSERec) generally achieve better results than traditional approaches, underscoring the value of LLM-driven features.
Limitations and Future Work
- LLM Inference Cost: While LLM summarization can be done offline, it still represents a computational cost. The paper mentions this is done offline, mitigating online serving impact.
- Hyperparameter Tuning: The performance is sensitive to hyperparameter tuning, particularly for the contrastive loss weights (e.g., AU-CL, His-CL), as shown in Figure 7.
- Generalizability of LLM Prompts: The effectiveness of the LLM summarization relies on well-crafted prompts, which might need adaptation for different domains or LLMs.
- Future Work: The paper doesn't explicitly detail future work, but the success of user-code graphs and LLM integration suggests further exploration in these areas for more complex S&R scenarios.
Summary updated