Introduction
Recommender systems (RS) are central to modern digital experiences but are frequently plagued by selection bias. Because training data is collected from past user interactions—which are themselves influenced by previous recommendation policies—models often learn from a "biased" slice of reality. This leads to a "rich get richer" phenomenon, where popular items receive disproportionate exposure while niche items are ignored.
Traditional debiasing solutions, such as Inverse Propensity Scoring (IPS) or Doubly Robust (DR) methods, rely on the unconfoundedness assumption: the idea that every variable influencing whether a user sees an item has been recorded. In practice, this is rarely true. Hidden factors—unobserved confounders—such as social media trends or unmeasured financial status, influence behavior but remain absent from datasets. This paper bridges the gap between theoretical bounds and empirical success by reframing unbiased recommendation as a distribution shift problem.
Adversarial Self-Training (AST)
The authors introduce Adversarial Self-Training (AST), a unified framework that treats unbiased recommendation as a domain adaptation task. By deriving new generalization bounds that explicitly account for the shift between biased training distributions and unbiased test distributions, the authors provide a rigorous foundation for modern debiasing techniques.
The AST method consists of three primary components designed to minimize these theoretical bounds:
- Adversarial Feature Adaptation: A "Critic" network ($\theta$) plays a mini-max game with the feature encoder ($\phi$) to ensure that latent representations of biased and unbiased data are indistinguishable, thereby minimizing covariate shift.
- Self-Training & Entropy Minimization: To handle "concept shift" caused by unobserved confounders, the model generates pseudo-labels for unlabeled data. This encourages the model to find a labeling function ($\psi$) that is invariant across different exposure environments.
- Unified Objective: Unlike meta-learning approaches that require expensive second-order gradients, AST uses a standard optimization flow, making it computationally efficient for large-scale applications.
Experimental Evaluation
The researchers evaluated AST on three primary datasets: Yahoo! R3 (songs), Coat (clothing), and a semi-synthetic Goodreads environment designed to simulate unobserved confounders using Gaussian noise.
Key Results
AST consistently outperformed state-of-the-art baselines, including IPS, DRJL, and AutoDebias, across all benchmarks.
- Ranking Accuracy: On the Yahoo! dataset, AST achieved an NDCG@5 of 0.7147, significantly higher than the "Biased" baseline (0.6542) and the strong AutoDebias (0.6959).
- Robustness to Data Scarcity: As shown in the performance comparison, AST maintains its lead even when the amount of unbiased uniform data is drastically reduced. This proves that its adversarial and self-training components are effective even in purely observational settings where ground-truth unbiased data is scarce.
- Distribution Alignment: The framework successfully "bridges the gap" in feature space between biased and unbiased samples, ensuring the model generalizes well to the true underlying user preferences.
Conclusion and Future Work
By viewing the problem through the lens of Out-of-Distribution (OOD) generalization, AST provides a robust answer to the "unconfoundedness lie." While unobserved confounders remain a significant challenge in the field, AST effectively alleviates their impact without requiring unrealistic assumptions. Future work may involve exploring more complex adversarial strategies or integrating AST with dynamic reinforcement learning environments where logging policies evolve over time.
Summary updated