ConsisRec: Enhancing GNN for Social Recommendation via Consistent Neighbor Aggregation

Liangwei Yang, Zhiwei Liu, Yingtong Dou, Jing Ma, Philip S. Yu

TL;DR

ConsisRec is a novel Graph Neural Network (GNN) framework that improves social recommendation by addressing "social inconsistency"—the fact that your social media friends don't always share your specific consumer tastes. By dynamically sampling only "consistent" neighbors based on the specific item being recommended and applying relation-aware attention, ConsisRec significantly outperforms existing state-of-the-art models on real-world datasets.

What Problem Does This Paper Address?

Social recommendation systems try to solve the "cold-start" problem (where we have little data on new users) by looking at a user's social circle. The underlying assumption is homophily: friends likely have similar tastes.

However, the authors argue that this assumption is often too simple, leading to social inconsistency. They categorize this into two levels:

  1. Context-level inconsistency: You might be friends with someone on a professional network, but while you buy sports equipment, they only buy rare books. Aggregating their "book-heavy" data into your profile hurts the system's ability to recommend sports gear to you.
  2. Relation-level inconsistency: Even if you and a friend buy the same item, you might have opposite reactions. If you gave a movie 5 stars and your friend gave it 1 star, their preference is actually a negative indicator for yours, despite the social link.

Existing GNN-based models often treat all social neighbors equally, effectively "spoiling" user embeddings with irrelevant or contradictory information.

A toy example of social inconsistency and the ConsisRec architecture

Methods

ConsisRec introduces three main components to a standard GNN architecture to filter out this social noise:

  • Query Layer: Instead of a static user representation, the model generates a query embedding ($q_{u,t}$) that combines the user's features with the specific target item's features. This reflects the intuition that when you're looking for a new laptop, you only care about the opinions of your "techie" friends, not your "foodie" friends.
  • Consistent Neighbor Sampling: Rather than aggregating all neighbors, ConsisRec calculates a consistency score between the query embedding and each neighbor's embedding. It then uses these scores as probabilities to sample a subset of neighbors. This explicitly discards inconsistent neighbors during the learning process.
  • Relation Attention: Not all links are created equal. The model uses a self-attention mechanism that considers the type of relation (e.g., a social "trust" link vs. a specific "4-star rating" link). This allows the model to assign higher importance to neighbors who share consistent relations with the user.

Datasets

The authors evaluated ConsisRec on two widely used real-world social recommendation benchmarks:

  • Ciao: 7,317 users and 104,975 items.
  • Epinions: 18,069 users and 261,246 items.

In both cases, the researchers pre-processed the data to focus specifically on the social recommendation task by removing users without social links and linking items that share a significant number of common neighbors.

Key Insights and Ideas

  • Dynamic Relevance: The most important "aha" moment is that neighbor relevance is query-dependent. A friend's influence on your profile shouldn't be static; it should change depending on what item the system is currently trying to predict your interest in.
  • Sampling as a Denoising Tool: While many GNNs use attention to weight neighbors, ConsisRec uses sampling. This is a stronger form of filtering that can more effectively ignore highly inconsistent neighbors that would otherwise introduce noise into the node embeddings.
  • Multi-Relational Heterogeneity: By treating different rating values (1-5 stars) as distinct relation types in a heterogeneous graph, the model can distinguish between "friends who like what I like" and "friends who dislike what I like."

Key Results

ConsisRec achieved the best results across all metrics (RMSE and MAE) on both datasets, outperforming strong baselines like GraphRec and CUNE.

  • Performance Gain: The model showed an average relative improvement of 1.7% to 2.8% over the second-best methods. In the world of recommendation systems, even a 1% improvement in RMSE is considered significant for top-N recommendation quality.
  • Ablation Success: The authors tested versions of the model with various parts removed. They found that Variant B (removing neighbor sampling) caused the most dramatic drop in performance, proving that selecting consistent neighbors is the most critical part of their framework.
Ablation study showing the impact of different components

Limitations and Future Work

While ConsisRec is effective, the authors note a few areas for growth:

  • Complexity: The query-based sampling adds computational overhead, as neighbor selection must be re-evaluated for different user-item pairs.
  • Hyper-parameter Sensitivity: The model's performance is sensitive to the "neighbor percent" (how many neighbors are sampled). Aggregating too few loses information, while aggregating too many (approaching 100%) re-introduces the social inconsistency problem.
  • Future Directions: The authors suggest exploring more sophisticated ways to filter informative neighbors and applying these "consistency" principles to other graph-related research areas beyond recommendation.
Sensitivity analysis of key hyper-parameters

Summary updated

Abstract

Social recommendation aims to fuse social links with user-item interactions to alleviate the cold-start problem for rating prediction. Recent developments of Graph Neural Networks (GNNs) motivate endeavors to design GNN-based social recommendation frameworks to aggregate both social and user-item interaction information simultaneously. However, most existing methods neglect the social inconsistency problem, which intuitively suggests that social links are not necessarily consistent with the rating prediction process. Social inconsistency can be observed from both context-level and relation-level. Therefore, we intend to empower the GNN model with the ability to tackle the social inconsistency problem. We propose to sample consistent neighbors by relating sampling probability with consistency scores between neighbors. Besides, we employ the relation attention mechanism to assign consistent relations with high importance factors for aggregation. Experiments on two real-world datasets verify the model effectiveness.
BibTeX
@inproceedings{yang2021consisrec,
  author = {Liangwei Yang and Zhiwei Liu and Yingtong Dou and Jing Ma and Philip S. Yu},
  title = {{ConsisRec: Enhancing GNN for Social Recommendation via Consistent Neighbor Aggregation}},
  booktitle = {International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year = {2021},
  url = {https://openreview.net/forum?id=BZAwmO4806I},
}

Loading related content...

Figure