[Paper Review] WSFE: Wasserstein Sub-graph Feature Encoder for Effective User Segmentation in Collaborative Filtering

Yankai Chen, Yifei Zhang+4 more|arXiv (Cornell University)|May 8, 2023

Recommender Systems and Techniques4 citations

TL;DR

WSFE is a model-agnostic, training-free representation encoder that leverages optimal transport theory via Wasserstein distance to measure user-wise similarity in collaborative filtering by encoding high-order sub-graph patterns from user-item interaction graphs. It significantly improves user segmentation performance across six state-of-the-art models on six real-world datasets without retraining, with gains up to +13.60% in recall and NDCG for user clustering tasks.

ABSTRACT

Maximizing the user-item engagement based on vectorized embeddings is a standard procedure of recent recommender models. Despite the superior performance for item recommendations, these methods however implicitly deprioritize the modeling of user-wise similarity in the embedding space; consequently, identifying similar users is underperforming, and additional processing schemes are usually required otherwise. To avoid thorough model re-training, we propose WSFE, a model-agnostic and training-free representation encoder, to be flexibly employed on the fly for effective user segmentation. Underpinned by the optimal transport theory, the encoded representations from WSFE present a matched user-wise similarity/distance measurement between the realistic and embedding space. We incorporate WSFE into six state-of-the-art recommender models and conduct extensive experiments on six real-world datasets. The empirical analyses well demonstrate the superiority and generality of WSFE to fuel multiple downstream tasks with diverse underlying targets in recommendation.

Motivation & Objective

• To address the lack of explicit user-wise similarity modeling in modern collaborative filtering models, which hinders effective user segmentation.
• To develop a representation encoder that improves user similarity measurement without requiring model retraining or architectural changes.
• To enable flexible, on-the-fly deployment of user segmentation across diverse graph-based recommender models.
• To leverage high-order sub-graph patterns and optimal transport theory for robust, distribution-aware user embedding encoding.
• To validate the method’s generality and effectiveness across multiple models and real-world datasets.

Proposed method

• WSFE models user preferences as an unknown high-dimensional probability distribution derived from learned item interaction patterns.
• It encodes user representations by computing Wasserstein distances between empirical and embedded feature distributions using optimal transport theory.
• The method extracts high-order sub-graph features from user-item interaction graphs to capture complex behavioral similarities.
• It applies a slicing-based approximation (S) to compute cumulative Wasserstein distances efficiently, balancing accuracy and computational cost.
• Layer-wise embeddings are aggregated via sum or concatenation to reduce dimensionality and mitigate memory issues.
• The encoder is fully compatible with any pre-trained graph-based recommender model, enabling on-the-fly user segmentation.

Experimental results

Research questions

• RQ1 Can a training-free, model-agnostic method improve user segmentation performance in collaborative filtering without retraining?
• RQ2 How effective is WSFE in capturing user-wise similarity through sub-graph pattern encoding and optimal transport?
• RQ3 Does WSFE generalize across diverse graph-based recommender models and real-world datasets?
• RQ4 What is the impact of the slicing number S on performance and computational efficiency?
• RQ5 How does feature aggregation (sum vs. concatenation) affect dimensionality and segmentation accuracy?

Key findings

• WSFE improves user segmentation performance across all six real-world datasets and six state-of-the-art models, with recall gains ranging from +0.83% to +13.60%.
• The largest improvements are observed on MovieLens, where users have higher average interactions (165.60), indicating richer preference distributions for WSFE to exploit.
• Contrastive learning-based models (e.g., SGL, SimGCL, NCL) benefit most from WSFE, with gains up to +12.60% in recall, due to enhanced embedding diversity from data augmentation.
• Setting the slicing number S to 64 provides the best trade-off between performance and computational cost, as further increases yield diminishing returns.
• The sum aggregation method achieves competitive performance compared to concatenation, making it a viable option for resource-constrained environments.
• WSFE enables effective user segmentation without retraining, demonstrating strong generality and flexibility across diverse model architectures.