When Product Search Meets Collaborative Filtering: A Hierarchical Heterogeneous Graph Neural Network Approach

Xiangkun Yin, Yangyang Guo, Liqiang Nie, Zhiyong Cheng

TL;DR

This paper introduces HHGNN, a novel Hierarchical Heterogeneous Graph Neural Network that effectively integrates collaborative filtering signals with semantic matching for personalized product search. By modeling user-product interactions and textual information within a multi-layered graph structure, HHGNN significantly outperforms existing methods, demonstrating the crucial role of collaborative filtering in enhancing product search relevance.

What Problem Does This Paper Address?

The core challenge this paper tackles is improving the personalization of product search. Traditional product search systems primarily rely on matching the semantic content of user queries with product descriptions. While effective to some extent, this approach often overlooks a rich source of information: user collaborative behaviors. The authors argue that understanding how users interact with products collectively (e.g., what they buy together) can provide valuable signals that complement semantic matching, leading to more accurate and personalized search results. The gap lies in effectively bridging the gap between these two distinct but complementary information sources.

Methods

The paper proposes a Hierarchical Heterogeneous Graph Neural Network (HHGNN). The novelty lies in its graph construction and hierarchical processing:

  • Heterogeneous Graph: The system models product search as a graph where nodes represent different entities like words, reviews, queries, users, and products. Edges represent their relationships.
  • Hierarchical Structure: The graph is organized into three layers based on edge types:
  • Sequence Edge (Bottom Layer): Connects words to form sentences/queries/reviews, capturing textual syntax. This layer uses a Multi-Head Attention (MHA) mechanism to process word embeddings and generate query/review representations.
  • Composition Edge (Middle Layer): Aggregates review semantic features to form user and product node representations.
  • Interaction Edge (Top Layer): Models collaborative filtering signals by connecting user and product nodes. This layer employs Graph Convolutional Networks (GCNs) to propagate and aggregate information, capturing higher-order user-product interactions.
  • Integration: The model fuses semantic features (from MHA) and collaborative filtering features (from GCNs) for a comprehensive representation.

Datasets

The experiments were conducted on six subsets of the publicly available Amazon review dataset. These subsets include:

  • Musical Instruments
  • Automotive
  • Office Products
  • Toys & Games
  • Cell Phones & Accessories
  • Clothing, Shoes & Jewelry

The authors preprocessed these datasets by applying a 5-core filter (ensuring each user and product has at least 5 records) and removing stop words and low-frequency words. Synthetic queries were generated based on product categories.

Key Insights and Ideas

  • Collaborative Filtering is Key: The most significant insight is that collaborative filtering signals are not just useful but complementary and mutually beneficial to semantic matching in product search. Ignoring these signals leads to suboptimal performance.
  • Hierarchical Graph for Richness: The hierarchical graph structure is crucial. It allows for distinct processing of textual information (sequence layer), semantic aggregation (composition layer), and interaction modeling (interaction layer), creating a more nuanced understanding of user preferences and product relevance.
  • Higher-Order Interactions Matter: The GCN layers effectively capture higher-order collaborative filtering signals, which previous methods might miss, leading to improved personalization.
  • Semantic Enhancement: The model can uncover semantic relationships that might not be explicitly present in a user's direct review history but are implied through their collaborative behavior or the behavior of similar users.

Key Results

  • State-of-the-Art Performance: HHGNN consistently outperformed all baselines across the six Amazon datasets in terms of Hit Ratio (HR@10), Mean Reciprocal Rank (MRR), and Normalized Discounted Cumulative Gain (NDCG@10). For example, it achieved significant improvements in NDCG@10, with gains ranging from 6% to 22% over the best baseline.
  • Ablation Studies Confirm Value:
    • Removing the graph convolutional layers (HHGNNno_conv) resulted in the worst performance, strongly supporting the benefit of collaborative filtering.
  • Variants that combined semantic and ID-based collaborative filtering (HHGNNid_conv, HHGNNtext_conv) also showed significant improvements over the no-convolution baseline, highlighting the synergistic effect of combining different signal types.
  • Layer Depth Trade-off: While increasing GCN layers generally improved performance, there was a point of diminishing returns and potential over-smoothing, suggesting a need to balance depth with model complexity.

Limitations and Future Work

  • Over-smoothing: The paper acknowledges that deeper GCN architectures can lead to over-smoothing, potentially introducing noise. This suggests that careful tuning of the number of convolutional layers is necessary.
  • Synthetic Queries: The use of synthetically generated queries, rather than real user queries, might limit the direct applicability to real-world scenarios where query generation is more complex.
  • Future Work: The authors suggest exploring more advanced tools and techniques for future research, implying that the field is still evolving and there's room for further innovation in integrating collaborative filtering with product search. They also mention releasing their code, which will facilitate further research.

Summary updated

Abstract

Personalization lies at the core of boosting the product search system performance. Prior studies mainly resorted to the semantic matching between textual queries and user/product related documents, leaving the user collaborative behaviors untapped. In fact, the collaborative filtering signals between users intuitively offer a complementary information for the semantic matching. To close the gap between collaborative filtering and product search, we propose a Hierarchical Heterogeneous Graph Neural Network (HHGNN) approach in this paper. Specifically, we organize HHGNN with a hierarchical graph structure according to the three edge types. The sequence edge accounts for the syntax formulation from word nodes to sentence nodes; the composition edge aggregates the semantic features to the user and product nodes; and the interaction edge on the top performs graph convolutional operation between user and product nodes. At last, we integrate the higher-order neighboring collaborative features and the semantic features for better representation learning. We conduct extensive experiments on six Amazon review datasets. The results show that our proposed method can outperform the state-of-the-art baselines with a large margin. In addition, we empirically prove that collaborative filtering and semantic matching are complementary to each other in product search performance enhancement.
BibTeX
@misc{yin2021when,
  author = {Xiangkun Yin and Yangyang Guo and Liqiang Nie and Zhiyong Cheng},
  title = {{When Product Search Meets Collaborative Filtering: A Hierarchical Heterogeneous Graph Neural Network Approach}},
  year = {2021},
  eprint = {2108.07574},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2108.07574},
}
Figure