Explaining Concept Shift with Interpretable Feature Attribution

TL;DR

This work addresses concept shift in tabular data by proposing SGShift, a sparse GAM-based framework that attributes degraded model performance to a small set of shifted features via an interpretable, additive update to the source distribution $P(y|X)$. It extends the basic SGShift with an absorption term to handle misspecification (SGShift-A) and with knockoffs to control false discoveries (SGShift-K) and their combination (SGShift-KA). Through semi-synthetic and real healthcare data, SGShift achieves high detection power (AUC $>0.9$) and recall (often $>90\%$), frequently matching or exceeding baselines by substantial margins, and demonstrates sparsity in the true concept shift. The proposed framework provides interpretable diagnostics for distribution shift and robust feature-level attribution, enabling targeted modeling fixes with practical impact on deployment in dynamic domains.

Abstract

Regardless the amount of data a machine learning (ML) model is trained on, there will inevitably be data that differs from their training set, lowering model performance. Concept shift occurs when the distribution of labels conditioned on the features changes, making even a well-tuned ML model to have learned a fundamentally incorrect representation. Identifying these shifted features provides unique insight into how one dataset differs from another, considering the difference may be across a scientifically relevant dimension, such as time, disease status, population, etc. In this paper, we propose SGShift, a model for detecting concept shift in tabular data and attributing reduced model performance to a sparse set of shifted features. SGShift models concept shift with a Generalized Additive Model (GAM) and performs subsequent feature selection to identify shifted features. We propose further extensions of SGShift by incorporating knockoffs to control false discoveries and an absorption term to account for models with poor fit to the data. We conduct extensive experiments in synthetic and real data across various ML models and find SGShift can identify shifted features with AUC $>0.9$ and recall $>90\%$, often 2 or 3 times as high as baseline methods.

Figures (7)

• Figure 1: Evaluation of recall at FPR 10% in semi-synthetic simulations for matched and mismatched model settings. Results are aggregated across all model configurations. Standard 95% confidence intervals are shown, calculated as the standard error across model configurations and 100 replicates.
• Figure 2: Performance increase as we update more features in the source model identified by SGShift. Each dot represents a penalty between 0 and 1. 2 penalties may have the same number of features selected. Gradient boosting (A, B, C) and logistic/linear regression (D, E, F) are shown as examples in our 3 datasets.
• Figure 3: Real data results showing the ordering of selected features for each model as the penalty term increases for COVID-19 severity. Positive (+) and negative (-) coefficients are treated as 2 distinct features. Only features selected in the top 5 for any model are shown.
• Figure 4: Individual model configuration results for diabetes readmission semi-synthetic simulations, measured in recall. 2 standard error intervals are shown over 100 replicates.
• Figure 5: Individual model configuration results for COVID-19 semi-synthetic simulations, measured in recall. 2 standard error intervals are shown over 100 replicates.

Theorems & Definitions (6)

• Theorem 4.1: Convergence Guarantee for Estimation Error Under RSC
• Lemma 4.2: Stability Selection Control
• Lemma 1: RSC for $\ell_1$-Penalized Loss
• proof
• proof
• proof