Isotonic Layer: A Universal Framework for Generic Recommendation Debiasing

TL;DR

The Isotonic Layer is introduced, a novel, differentiable framework that integrates piecewise linear fitting directly into neural architectures that effectively mitigates systematic bias and enhances calibration fidelity, significantly outperforming production baselines in both predictive accuracy and ranking consistency.

Abstract

Model calibration and debiasing are fundamental to the reliability and fairness of large scale recommendation systems. We introduce the Isotonic Layer, a novel, differentiable framework that integrates piecewise linear fitting directly into neural architectures. By partitioning the feature space into discrete segments and optimizing non negative slopes via a constrained dot product mechanism, we enforce a global monotonic inductive bias. This ensures model outputs remain logically consistent with critical features such as latent relevance, recency, or quality scores. We further generalize this architecture by parameterizing segment wise slopes as learnable embeddings. This enables the model to adaptively capture context specific distortions, such as position based CTR bias through specialized isotonic profiles. Our approach utilizes a dual task formulation that decouples the recommendation objective into latent relevance estimation and bias aware calibration. A major contribution of this work is the ability to perform highly granular, customized calibration for arbitrary combinations of context features, a level of control difficult to achieve with traditional non parametric methods. We also extend this to Multi Task Learning environments with dedicated embeddings for distinct objectives. Extensive empirical evaluations on real world datasets and production AB tests demonstrate that the Isotonic Layer effectively mitigates systematic bias and enhances calibration fidelity, significantly outperforming production baselines in both predictive accuracy and ranking consistency.

Figures (5)

• Figure 1: Example use cases of the Isotonic Layer in ranking and calibration. (a) Standard prediction model without monotonic calibration. (b) Global isotonic calibration applied to model outputs with learnable bucket weights $\mathbf{w}$ and bias $\mathbf{b}$. (c) Context-conditioned isotonic calibration via learned embeddings, where a context feature $c$ is mapped to a context-conditioned isotonic embedding $(\mathbf{w}_c, b_c)$. (d) Dual-tower architecture with joint optimization of relevance and context-aware isotonic calibration for debiasing.
• Figure 2: Expressive capacity of the Isotonic Layer on a monotonic synthetic task ($y = x^2$).
• Figure 3: Robustness to non-monotonic noise. The Isotonic Layer enforces global monotonic structure while smoothing local distortions in the target signal.
• Figure 4: Comparison of Online Prediction Scores. It monitors model average output scores in a daily basis, where X axis represents date and Y axis represents the probabilities.
• Figure 5: Dual-head architecture for isotonic position debiasing. Each task group consists of a position-neutral inference head and a context-conditioned isotonic head. The isotonic head utilizes position and platform features during training to decouple exposure bias from intrinsic relevance. Loss weights for each head are optimized as hyperparameters to balance ranking performance and model stability.