Linear-PAL: A Lightweight Ranker for Mitigating Shortcut Learning in Personalized, High-Bias Tabular Ranking

TL;DR

This work tackles the problem of extreme position bias in e-commerce ranking, where standard deep learning models exhibit shortcut learning and overfit to rank signals. It proposes Linear-PAL, a lightweight linear framework that enforces de-biasing through explicit cross-feature interactions and strong regularization, leveraging a causal intervention viewpoint with a do-operation to simulate the optimal display position. Key innovations include a vectorized integer hashing kernel for fast, interpretable feature construction, COEC/UCOEC-based feature normalization, and an interaction kernel Phi(x) otimes Psi(k) that captures rank-dependent utility in a tractable linear model. Empirically, Linear-PAL achieves a substantial improvement in Relevance AUC (+~13%) over deep ensembles while delivering a 43x reduction in training time, enabling frequent retraining and robust personalization on large-scale tabular data.

Abstract

In e-commerce ranking, implicit user feedback is systematically confounded by Position Bias -- the strong propensity of users to interact with top-ranked items regardless of relevance. While Deep Learning architectures (e.g., Two-Tower Networks) are the standard solution for de-biasing, we demonstrate that in High-Bias Regimes, state-of-the-art Deep Ensembles suffer from Shortcut Learning: they minimize training loss by overfitting to the rank signal, leading to degraded ranking quality despite high prediction accuracy. We propose Linear Position-bias Aware Learning (Linear-PAL), a lightweight framework that enforces de-biasing through structural constraints: explicit feature conjunctions and aggressive regularization. We further introduce a Vectorized Integer Hashing technique for feature generation, replacing string-based operations with $O(N)$ vectorized arithmetic. Evaluating on a large-scale dataset (4.2M samples), Linear-PAL achieves Pareto Dominance: it outperforms Deep Ensembles in de-biased ranking quality (Relevance AUC: 0.7626 vs. 0.6736) while reducing training latency by 43x (40s vs 1762s). This computational efficiency enables high-frequency retraining, allowing the system to capture user-specific emerging market trends and deliver robust, personalized ranking in near real-time.

Figures (5)

• Figure 1: Linear-PAL Architecture. Explicit feature conjunctions (Cross Product) replace the hidden layers of Deep Neural Networks, allowing the model to learn position-dependent relevance weights linearly.
• Figure 2: Empirical Position Bias. The heatmap reveals a non-monotonic "Golden Triangle" attention pattern typical of grid layouts. Values are normalized to Rank 1 ($P=1.0$).
• Figure 3: Efficiency Frontier. Linear-PAL (Blue) dominates Deep Architectures (Red/Green) in Relevance Quality, which collapse due to overfitting the strong position bias.
• Figure 4: Ablation Study: Impact of Regularization on Quality. Stronger regularization ($C=10^{-5}$) penalizes sparse rank interactions, forcing the model to rely on dense content features and maximizing Relevance AUC.
• Figure 5: The Bias-Accuracy Trade-off. Standard AUC peaks at $C=10^{-2}$ (0.8168), where the model fits the historical position bias. As regularization tightens to $C=10^{-5}$, Standard AUC drops slightly, but this loss represents the successful suppression of bias, enabling the rise in Relevance AUC.