Alleviating Behavior Data Imbalance for Multi-Behavior Graph Collaborative Filtering

Yijie Zhang, Yuanchen Bei, Shiqi Yang, Hao Chen, Zhiqing Li, Lijia Chen, Feiran Huang

Introduction: The Data Funnel Problem

In modern e-commerce, user behavior is multi-dimensional, encompassing actions like clicking, adding to carts, and purchasing. While Graph Collaborative Filtering (GCF) is effective for modeling these interactions, most existing models focus on a single behavior type.

The primary challenge is behavior data imbalance. In a typical dataset, clicks significantly outnumber purchases. If a model focuses only on sparse purchase data, it misses rich context; if it mixes all data, the high volume of clicks can overwhelm the specific signal needed to predict purchases. This "data funnel" effect creates a structural bottleneck for recommendation systems.

The multi-behavior data funnel and the resulting data scale imbalance.

The IMGCF Framework

To address this, the authors propose IMGCF (Imbalance Multi-behavior Graph Collaborative Filtering). The framework explicitly transfers knowledge from high-volume behavior graphs to sparse ones through three main stages:

  1. Multi-Task Graph Encoding: Using LightGCN as a backbone, the model constructs separate bipartite graphs for each behavior (e.g., click, cart, purchase). Each graph learns behavior-specific user and item embeddings.
  2. Sparse Behavior Enhancement (SBE): This component "helps" the sparse purchase task by aggregating embeddings from the "rich" behavior domain (clicks) into the "sparse" domain (purchases) using functions like mean pooling or concatenation.
  3. Joint Optimization with Gradient Isolation: The model is trained via a Multi-Task Learning (MTL) framework. Crucially, it implements a "one-way valve" for information: while the purchase task receives information from the click task, a gradient isolation mechanism prevents the sparse, potentially noisy purchase data from backpropagating into and "polluting" the high-quality click embeddings.

Experimental Evaluation

The model was evaluated on two real-world datasets: Beibei and Taobao. Both exhibit the classic funnel imbalance, with Taobao being significantly sparser.

Key Results

IMGCF consistently outperformed state-of-the-art baselines like LightGCN and MMOE:

  • Significant Gains: Across both datasets, IMGCF improved Recall@20 by 29.34% and NDCG@20 by 37.85% on average.
  • Resilience to Sparsity: On the extremely sparse Taobao dataset, the improvement in NDCG@20 reached 49.09%, demonstrating the effectiveness of the SBE component.
  • Ablation Study: Removing the SBE component led to a significant performance drop, confirming that cross-domain aggregation is the primary driver of success.

Insights and Future Directions

The success of IMGCF highlights that simplicity can be a feature; parameter-free aggregation (like mean pooling) performed exceptionally well, suggesting the bottleneck is structural rather than algorithmic complexity. The gradient isolation ensures a stable "foundation" of user interest based on high-volume data.

Future work will explore more nuanced, non-linear relationships between behaviors (e.g., recognizing that a "cart" action is a stronger purchase signal than a "click") and testing the framework on non-e-commerce data like social media interactions.

Summary updated

BibTeX
@inproceedings{zhang2023alleviating,
  author = {Yijie Zhang and Yuanchen Bei and Shiqi Yang and Hao Chen and Zhiqing Li and Lijia Chen and Feiran Huang},
  title = {{Alleviating Behavior Data Imbalance for Multi-Behavior Graph Collaborative Filtering}},
  booktitle = {Workshops @ ICDM},
  year = {2023},
  doi = {10.1109/icdmw60847.2023.00075},
}
Figure