Abstract:Recently, recommender systems have witnessed the fast evolution of Internet services. However, it suffers hugely from inherent bias and sparsity issues in interactions. The conventional uniform embedding learning policies fail to utilize the imbalanced interaction clue and produce suboptimal representations to users and items for recommendation. Towards the issue, this work is dedicated to bias-aware embedding learning in a decomposed manner and proposes a counterfactual graph convolutional learning (CGCL) model for personalized recommendation. Instead of debiasing with uniform interaction sampling, we follow the natural interaction bias to model users’ interests with a counterfactual hypothesis. CGCL introduces bias-aware counterfactual masking on interactions to distinguish the effects between majority and minority causes on the counterfactual gap. It forms multiple counterfactual worlds to extract users’ interests in minority causes compared to the factual world. Concretely, users and items are represented with a causal decomposed embedding of majority and minority interests for recommendation. Experiments show that the proposed CGCL is superior to the state-of-the-art baselines. The performance illustrates the rationality of the counterfactual hypothesis in bias-aware embedding learning for personalized recommendation.

What problem does this paper attempt to address?

This paper attempts to solve the problem of latent interaction bias in recommendation systems. Specifically, traditional embedding learning methods often produce biases when dealing with interactions between users and items due to the unbalanced distribution of interaction data (such as long - tail distribution), being biased towards majority interests and ignoring minority interests. This bias will lead to a decline in the performance of the recommendation system and fail to fully explore the diverse interests of users. To solve this problem, the paper proposes a Counterfactual Graph Convolutional Learning (CGCL) model based on counterfactual hypothesis. The following are the core objectives and solutions of the paper: --- ### **Core Problems in the Paper** 1. **Interaction Bias Problem**: In the recommendation system, the interaction data between users and items usually has an unbalanced distribution (for example, some attributes or interest points appear more frequently, while others are less frequent). This imbalance will cause the learned user and item embeddings to be biased towards high - frequency attributes and ignore low - frequency attributes. 2. **Challenges**: - How to extract users' diverse interests from unbalanced interaction data? - How to model majority and minority interests in a balanced way without violating the existing bias environment? --- ### **Solutions** The paper solves the above problems through the following steps: 1. **Counterfactual Hypothesis**: - Based on counterfactual reasoning, CGCL constructs multiple counterfactual worlds by masking part of the interaction data. - In these counterfactual worlds, the influence of majority and minority interests can be distinguished. For example, by comparing the embeddings of the factual world and the counterfactual world, the effect of minority interests can be extracted. 2. **Causal Decomposition**: - Decompose the user and item embeddings into two parts: - **Majority Embeddings**: Extracted from multiple counterfactual worlds by average pooling. - **Minority Embeddings**: Extracted by comparing the differences between the real world and the counterfactual world. - The formulas are as follows: - Majority Embeddings: $$ \dot{\mathbf{E}}_u^{(l)}=\frac{1}{T} \sum_{t = 1}^T \mathbf{E}_u^{(l)-t}, \quad \dot{\mathbf{E}}_v^{(l)}=\frac{1}{T} \sum_{t = 1}^T \mathbf{E}_v^{(l)-t} $$ - Minority Embeddings: $$ \ddot{\mathbf{E}}_u^{(l)}=\frac{1}{T} \sum_{t = 1}^T (\mathbf{E}_u^{(l)}-\mathbf{E}_u^{(l)-t}), \quad \ddot{\mathbf{E}}_v^{(l)}=\frac{1}{T} \sum_{t = 1}^T (\mathbf{E}_v^{(l)}-\mathbf{E}_v^{(l)-t}) $$ 3. **Aggregation and Interaction Modeling**

Counterfactual Graph Convolutional Learning for Personalized Recommendation

Personalized Counterfactual Fairness in Recommendation

CounterCLR: Counterfactual Contrastive Learning with Non-random Missing Data in Recommendation

A Personalized Collaborative Filtering Recommendation System Based on Bi-Graph Embedding and Causal Reasoning

Enhanced Graph Learning for Recommendation via Causal Inference

Mitigating Confounding and Selection Biases in Personalized Recommendation: A Causal Approach.

FairGap: Fairness-aware Recommendation via Generating Counterfactual Graph

Data-Augmented Counterfactual Learning for Bundle Recommendation

IntentGC: a Scalable Graph Convolution Framework Fusing Heterogeneous Information for Recommendation

Heterogeneous Graph Contrastive Learning for Recommendation

Multi-behavior-based graph contrastive learning recommendation

Counterfactual Adversarial Learning for Recommendation

Intent-guided Heterogeneous Graph Contrastive Learning for Recommendation

Neural Causal Graph Collaborative Filtering

Attentive Knowledge-aware Graph Convolutional Networks with Collaborative Guidance for Personalized Recommendation

Towards Robust Recommendation via Decision Boundary-aware Graph Contrastive Learning

Bilateral Unsymmetrical Graph Contrastive Learning for Recommendation

Graph Heterogeneous Multi-Relational Recommendation

Causal Structure Representation Learning of Confounders in Latent Space for Recommendation

Counterfactual Learning-Driven Representation Disentanglement for Search-Enhanced Recommendation

Counterfactual Graph Augmentation for Consumer Unfairness Mitigation in Recommender Systems