Haina Wang

Haina Wang

he/him

Haina Wang is an M.S. student in Computer Science at UC San Diego, having earned his B.S. from Zhejiang University. His research focuses on Embodied AI and Vision-Language Models (VLM), aiming to translate cutting-edge AI theory into robust, reliable real-world systems. He is looking for a Ph.D. position in 2026.

📊 Autodebias GCN: Dual-Model Correction for Exposure Bias in Implicit Feedback

Project Summary

  • Goal:To address Exposure Bias in recommender systems—the issue where implicit user feedback is a joint product of Preference ($R$) and Exposure ($O$), leading to biased preference estimates.

  • Core Innovation: Designed and implemented a novel dual-Graph Convolutional Network (GCN) model, based on LightGCN, that jointly trains separate models for the Preference Matrix ($R$) and the Exposure Matrix ($O$) using Expectation-Maximization (EM) principles.

  • Key Contributions: Theoretically demonstrated that the proposed method significantly reduces the variance of the preference estimation compared to state-of-the-art methods like Item Propensity Weighted (IPW) models.

Validated model uniqueness by leveraging the hypothesized difference in properties (e.g., Eigenvalue distribution) between $R$ and $O$ via customized regularization terms.

Authors: Haina Wang (Undergraduate Thesis)

Institution: Zhejiang University

Problem & Model Motivation

Modern recommender systems rely on implicit feedback (clicks, views), which suffer from exposure bias because we cannot distinguish if a non-click means ‘dislike but exposed’ or ‘like but unexposed’. Existing debiasing methods, such as Weighted Matrix Factorization (WMF) and Propensity-Based models (IPW), often lead to biased estimates or suffer from high variance.

Our approach explicitly models the observation (Y_{ui}) as the product of two latent variables: (Y_{ui} = R_{ui} \cdot O_{ui}).

Technical Approach: Dual-LightGCN Architecture

1. Dual-Model Training

We utilize two parallel LightGCN models—a powerful and simplified GCN architecture—to learn the embeddings for (R) and (O).

  • Preference Model ((R)): Trained to predict the user’s true interest (\xi_{ui}=P(r_{ui}=1)).
  • Exposure Model ((O)): Trained to predict the probability of exposure (\theta_{ui}=P(o_{ui}=1)).

The combined output, (P(Y_{ui}=1) = \text{sigmoid}(R) \cdot \text{sigmoid}(O)), is used for the overall loss calculation against the training data. The Preference Model output is used for final testing.

2. Variance Reduction (Theoretical Proof)

A crucial theoretical contribution of this work is proving that our method effectively bounds the variance of the preference estimation loss (L_{\text{prefer}}).

  • IPW Issue: The IPW estimator contains the inverse of the propensity score (1 / P(O_{ui}=1)), which can be very small, leading to unbounded variance in extreme cases.
  • Our Solution: We proved that the variance of our proposed loss function is upper-bounded, demonstrating a theoretical advantage in estimation stability.

3. Hypothesis-Driven Regularization

To prevent the two parallel models ((R) and (O)) from becoming redundant, we enforced structural differences based on hypotheses about user behavior:

  • Hypothesis: The Exposure Matrix ((O)) has a more decentralized and polarized structure compared to the Preference Matrix ((R)), which exhibits stronger long-tail effects (fewer popular items dominate preference).
  • Customization: We introduced Eigenvalue Regularization into the loss function, specifically leveraging the difference between the top (k) eigenvalues and the subsequent eigenvalues to guide the (R) and (O) models toward their hypothesized structural properties.

Experimental Results & Discussion

  • Performance: Comprehensive experiments on unbiased benchmark datasets (Yahoo!R3 and Coat) showed that the dual-model approach significantly outperformed single-model baselines (MF, ExpoMF) and IPW variants in terms of Recall@K and NDCG@K.

  • Ablation Studies: We performed extensive ablation tests on LightGCN layer count and eigenvalue coefficients, confirming that introducing the Exposure Model consistently boosts the performance of the Preference Model.

  • Reflection: While the Exposure Model itself proved challenging to train to its full theoretical potential, its inclusion served as a powerful regularizer, effectively separating the Preference signal from the Exposure noise. The analysis also revealed complex relationships between LightGCN layer counts and performance on different datasets (e.g., over-smoothing sensitivity), guiding future research directions.

Citation: Haina Wang. “Autodebias GCN: Dual-Model Correction for Exposure Bias in Implicit Feedback.” Undergraduate Thesis, Zhejiang University (2024).