Monitoring and model drift

Naming the shifts precisely#

Let the training distribution be $p_{\text{tr}}(x, y)$ and the deployment distribution $p_{\text{te}}(x, y)$. The joint factorises two ways, and each factorisation names a family:

• Covariate shift — $p_{\text{tr}}(x) \neq p_{\text{te}}(x)$, with $p(y \mid x)$ unchanged. The population changed; the underlying rule did not.
• Label shift (prior probability shift) — $p_{\text{tr}}(y) \neq p_{\text{te}}(y)$, with $p(x \mid y)$ unchanged. Natural where $y$ causes $x$, as in disease prevalence changing while symptoms do not.
• Concept drift — $p(y \mid x)$ itself changes. No reweighting of inputs fixes this; the target function moved and you need new labels.

Only covariate shift is correctable without labels. Under covariate shift with shared support, importance weighting by $w(x) = p_{\text{te}}(x) / p_{\text{tr}}(x)$ gives an unbiased risk estimate (Shimodaira, 2000). Estimating $w$ well is itself hard, and the variance of the weighted estimator scales with the Rényi divergence between the distributions — so the correction degrades exactly when it is most needed.

Two-sample testing#

Kolmogorov–Smirnov. $D_{n,m} = \sup_x |F_n(x) - G_m(x)|$, the maximum vertical gap between empirical CDFs. Univariate and continuous only, insensitive in the tails, and it detects a difference without telling you where.

Population Stability Index. With reference proportions $e_i$ and actual proportions $a_i$ over $B$ bins:

$$\mathrm{PSI} = \sum_{i=1}^{B} (a_i - e_i) \ln\frac{a_i}{e_i}$$

This is the symmetrised Kullback–Leibler divergence, $\mathrm{KL}(a | e) + \mathrm{KL}(e | a)$, also called Jeffreys divergence. It is a divergence, not a test — no p-value, no null distribution. The 0.1 and 0.25 thresholds are credit-industry convention with no distributional justification, and they depend on bin count: more bins mechanically increase PSI. Fix $B$ and compare like with like.

Maximum Mean Discrepancy. For a characteristic kernel $k$, $\mathrm{MMD}^2(P, Q) = |\mu_P - \mu_Q|^2_{\mathcal{H}}$ in the RKHS, with an unbiased estimator computable from the Gram matrices (Gretton et al., 2012). Multivariate, kernel-based, and a permutation test gives a valid p-value. Cost is $O(n^2)$ naively, with linear-time variants available. This is the principled choice when you need one test over the whole feature vector rather than $d$ marginal tests.

Classifier two-sample test. Train a discriminator to separate reference from live samples. Its held-out AUC is an interpretable effect size — 0.5 means indistinguishable — and the feature importances name the culprits. Lopez-Paz and Oquab (2017) formalise it.

Multiple testing#

Running a KS test per feature per day on $d$ features is $d$ tests per day. At $d = 200$ and $\alpha = 0.05$, ten false alarms daily by construction. Apply Benjamini–Hochberg FDR control across features, or move to a single multivariate test.

Better still, decouple the two questions. Statistical significance answers "did anything change". Effect size answers "does it matter". Only the second belongs on an alert.

Label shift estimation without labels#

Black Box Shift Estimation (Lipton, Wang and Smola, 2018) recovers the deployment label distribution from an existing classifier's confusion matrix on held-out training data plus its predicted-label distribution on unlabelled deployment data:

$$\hat{w} = \hat{C}^{-1} \hat{q}$$

where $\hat{C}$ is the confusion matrix (predicted against true, from held-out training data), $\hat{q}$ is the vector of predicted-label proportions in deployment, and $\hat{w}$ estimates the ratio $p_{\text{te}}(y) / p_{\text{tr}}(y)$. Requires $\hat{C}$ invertible and the label-shift assumption to hold. RLLS (Azizzadenesheli et al., 2019) adds regularisation for stability with few samples.

Streaming detectors#

For sequential settings with fast labels:

• DDM (Gama et al., 2004) monitors the online error rate, flagging warning and drift when it exceeds a running minimum by two and three standard deviations respectively.
• ADWIN (Bifet and Gavaldà, 2007) maintains an adaptive window, splitting it whenever two sub-windows have significantly different means, with a rigorous false-positive bound.
• Page–Hinkley — a cumulative-sum test for a shift in mean.

All three need labels promptly, which rules them out for the delayed-label problems that dominate industrial ML.

Performance estimation without labels#

An active area, and one to treat sceptically. Confidence-based methods (CBPE) assume the model stays calibrated under shift, which is exactly what shift breaks. Importance-weighted estimation (DLE, and reverse-classifier approaches) is better founded but degrades as the distributions separate.

The honest summary: these methods are useful as early-warning indicators and unsound as replacements for measured performance. Design the label collection path instead of hoping to avoid needing it.

Papers#

• Gretton et al., A Kernel Two-Sample Test, JMLR 2012 — jmlr.org/papers/v13/gretton12a.html
• Lipton, Wang and Smola, Detecting and Correcting for Label Shift with Black Box Predictors, ICML 2018 — arxiv.org/abs/1802.03916
• Rabanser, Günnemann and Lipton, Failing Loudly: An Empirical Study of Methods for Detecting Dataset Shift, NeurIPS 2019 — arxiv.org/abs/1810.11953
• Lopez-Paz and Oquab, Revisiting Classifier Two-Sample Tests, ICLR 2017 — arxiv.org/abs/1610.06545
• Gama et al., Learning with Drift Detection, SBIA 2004
• Bifet and Gavaldà, Learning from Time-Changing Data with Adaptive Windowing, SDM 2007
• Quiñonero-Candela et al., Dataset Shift in Machine Learning, MIT Press 2009 — the book-length treatment.

What to learn next#

• CI/CD for machine learning — turning "we should retrain" into something that runs and gets checked.
• Model evaluation — which numbers are worth tracking over time.
• Model serving — where the predictions you are monitoring come from.