Benefit from Rich: Tackling Search Interaction Sparsity in Search Enhanced Recommendation

Teng Shi, Weijie Yu, Xiao Zhang, Ming He, Jianping Fan, Jun Xu

Abstract

In modern online platforms, search and recommendation (S&R) often coexist, offering opportunities for performance improvement through search-enhanced approaches. Existing studies show that incorporating search signals boosts recommendation performance. However, the effectiveness of these methods relies heavily on rich search interactions. They primarily benefit a small subset of users with abundant search behavior, while offering limited improvements for the majority of users who exhibit only sparse search activity. To address the problem of sparse search data in search-enhanced recommendation, we face two key challenges: (1) how to learn useful search features for users with sparse search interactions, and (2) how to design effective training objectives under sparse conditions. Our idea is to leverage the features of users with rich search interactions to enhance those of users with sparse search interactions. Based on this idea, we propose GSERec, a method that utilizes message passing on the User-Code Graphs to alleviate data sparsity in Search-Enhanced Recommendation.

Introduction

In modern applications like TikTok or Amazon, search and recommendation (S&R) coexist. Ideally, a user's search queries should help the recommendation engine understand their immediate intent. However, current search-enhanced models suffer from a "rich-get-richer" problem: they show significant gains only for the small subset of power users who search frequently. For the majority of users, search data is incredibly sparse. When a user has only searched once or twice, traditional models struggle to learn meaningful features or optimize training objectives. GSERec addresses this by allowing the system to "borrow" insights from data-rich users to improve recommendations for those with sparse search histories.

Methodology: The GSERec Framework

The proposed GSERec (Graph-based Search-Enhanced Recommendation) operates in four main stages to bridge the gap between data-rich and data-sparse users.

1. User Preference Summarization and Quantization

Instead of relying on raw IDs or keywords, the model uses a Large Language Model (LLM) to summarize a user's S&R history into natural language descriptions of their interests. These summaries are then encoded into dense vectors and "quantized" into discrete codes using a Residual Quantized Variational Autoencoder (RQ-VAE). This maps users into a shared semantic "code space."

2. User-Code Graph and Message Passing

A bipartite graph is created connecting users to these discrete codes. If two users share a code (e.g., they both like "minimalist tech aesthetics"), they are connected. Using LightGCN, the model performs message passing across this graph. This allows the well-defined embeddings of "rich" users to flow through the shared codes and enhance the representations of "sparse" users.

The overall architecture of GSERec, illustrating the four-stage pipeline from preference summarization to model prediction.

3. Contrastive Alignment and Prediction

To ensure the message passing captures meaningful information, contrastive learning losses (InfoNCE) are introduced to align search preferences with recommendation preferences. Finally, these enhanced embeddings are combined with sequential modeling (Transformers) to predict the user's next interaction.

Experimental Results

The model was evaluated on three datasets, including Qilin, a real-world dataset from Xiaohongshu containing actual paired search and recommendation logs. GSERec consistently outperformed state-of-the-art baselines like UniSAR and SASRec across all metrics.

Sparsity Resilience

The most striking result is found in the "sparse" user groups (users with <10 search interactions). As shown in the performance analysis, GSERec achieved an 11.64% improvement over the traditional SASRec model in the most sparse category, significantly outperforming other search-enhanced models that struggled with the same users.

Performance comparison across different levels of search interaction sparsity, highlighting GSERec's significant gains for users with sparse data.

Key Insights and Conclusion

The core innovation of GSERec lies in using discrete codes as bridges. By forcing users into a finite set of interest codes, the model creates a topology where data-rich users act as anchors for data-sparse users. Furthermore, using LLMs as semantic translators ensures the model starts with clean, semantically dense representations rather than noisy raw interaction data. While the model currently relies on static graph structures, it demonstrates a powerful path forward for making search-enhanced recommendation inclusive of the "silent majority" of users.

Summary updated

Abstract

In modern online platforms, search and recommendation (S&R) often coexist, offering opportunities for performance improvement through search-enhanced approaches. Existing studies show that incorporating search signals boosts recommendation performance. However, the effectiveness of these methods relies heavily on rich search interactions. They primarily benefit a small subset of users with abundant search behavior, while offering limited improvements for the majority of users who exhibit only sparse search activity. To address the problem of sparse search data in search-enhanced recommendation, we face two key challenges: (1) how to learn useful search features for users with sparse search interactions, and (2) how to design effective training objectives under sparse conditions. Our idea is to leverage the features of users with rich search interactions to enhance those of users with sparse search interactions. Based on this idea, we propose GSERec, a method that utilizes message passing on the User-Code Graphs to alleviate data sparsity in Search-Enhanced Recommendation. Specifically, we utilize Large Language Models (LLMs) with vector quantization to generate discrete codes, which connect similar users and thereby construct the graph. Through message passing on this graph, embeddings of users with rich search data are propagated to enhance the embeddings of users with sparse interactions. To further ensure that the message passing captures meaningful information from truly similar users, we introduce a contrastive loss to better model user similarities. The enhanced user representations are then integrated into downstream search-enhanced recommendation models. Experiments on three real-world datasets show that GSERec consistently outperforms baselines, especially for users with sparse search behaviors.
BibTeX
@inproceedings{shi2025benefit,
  author = {Teng Shi and Weijie Yu and Xiao Zhang and Ming He and Jianping Fan and Jun Xu},
  title = {{Benefit from Rich: Tackling Search Interaction Sparsity in Search Enhanced Recommendation}},
  booktitle = {ACM International Conference on Information and Knowledge Management},
  year = {2025},
  doi = {10.1145/3746252.3761314},
}
Figure