Invariant debiasing learning for recommendation via biased imputation

TL;DR

The paper demonstrates that invariant-only debiasing can discard informative variant cues and harm generalization in recommender systems. It introduces KD-Debias, a two-part framework combining disentangled invariant/variant preferences with distance-aware knowledge distillation to transfer biased information into a lightweight unbiased predictor. A distance-based fusion and a teacher–student distillation setup produce unbiased recommendations without requiring unbiased data, achieving state-of-the-art results on Yahoo!R3, Coat, and Mind with substantially fewer parameters. The work highlights the value of leveraging biased variant information to enhance debiasing and offers a practical unsupervised solution for robust recommendations.

Abstract

Previous debiasing studies utilize unbiased data to make supervision of model training. They suffer from the high trial risks and experimental costs to obtain unbiased data. Recent research attempts to use invariant learning to detach the invariant preference of users for unbiased recommendations in an unsupervised way. However, it faces the drawbacks of low model accuracy and unstable prediction performance due to the losing cooperation with variant preference. In this paper, we experimentally demonstrate that invariant learning causes information loss by directly discarding the variant information, which reduces the generalization ability and results in the degradation of model performance in unbiased recommendations. Based on this consideration, we propose a novel lightweight knowledge distillation framework (KDDebias) to automatically learn the unbiased preference of users from both invariant and variant information. Specifically, the variant information is imputed to the invariant user preference in the distance-aware knowledge distillation process. Extensive experiments on three public datasets, i.e., Yahoo!R3, Coat, and MIND, show that with the biased imputation from the variant preference of users, our proposed method achieves significant improvements with less than 50% learning parameters compared to the SOTA unsupervised debiasing model in recommender systems. Our code is publicly available at https://github.com/BAI-LAB/KD-Debias.

Figures (5)

• Figure 1: The drawbacks (i.e., low model accuracy and unstable prediction ability) of the invariant learning method InvPref which solely relying on invariant information to make unbiased recommendations. (a) The model performances of our model KD-Debias and InvPref on the Yahoo!R3 dataset with random seeds in the 20 repeated experiments,the horizontal axis represents different experimental groups; (b) The distributions of invariant and variant preference in InvPref; (c) The information entropies of invariant and variant preference; (d) The distance between the unbiased true label and the predicted label from invariant and variant preference: the smaller the distance is, the more accurate the prediction is. The cases in which the prediction from variant preference is more accurate to the unbiased true label are circled with the dotted line, implying their potential utility to enhance the model performance, the horizontal axis refers to different data samples.
• Figure 2: An Overview of KD-Debias framework. It contains two components: the Disentangled Preference Modeling component and the Distance-aware Knowledge Distillation component. The disentangled preference modeling is conducted on the left causal graph with an environment classifier to identify the biases from different environments. Then the knowledge distillation are conducted on the disentangled invariant preference and variant preference with a distance-aware fusion strategy.
• Figure 3: Stability comparison on Yahoo!R3 dataset. The standard deviation is computed based on 20 times repeated experiments with different random seeds.
• Figure 4: The performances comparison of NDCG@5 with different hyper-parameter $\gamma$.
• Figure 5: The distance between the prediction results of invariant preferences (marked by the red circles), variant preferences (marked by the blue triangles), and the fused preference (marked by the green stars). The dots under the red dashed line (y = 0.5) can be regarded as correct predictions.