Most tools for learning on a changing data stream answer a yes-or-no question: has the data distribution changed? When a detector fires, the model is retrained on recent data and the old data is thrown away. Many real shifts are not like this. A sensor degrades slowly, user behaviour moves gradually, and the pace of the change can itself speed up or slow down. For that kind of slow, steady change, a single alarm is the wrong summary. What a learner needs to know is how fast things are moving, because that speed decides how much history is still useful. Epidemiology has a standard tool for exactly this question about a scalar series. Joinpoint regression fits a chain of connected straight lines to a time series, finds the times where the slope changes, and reports the slope of each segment. That tool is offline. It needs the whole series in hand, and its tests are only valid for a fixed sample size.
We propose to turn joinpoint regression into an online drift-rate estimator and to use its output directly. The method tracks a scalar that grows roughly linearly while the model's optimal parameters move at constant speed. It fits a continuous piecewise-linear trend to that scalar as data arrives and accepts new slope changes only through tests that stay valid under continuous monitoring. The current slope then sets the training window through a closed-form bias-variance trade-off. Put in two sentences: drift detectors tell you that the world changed, but under gradual change the useful quantity is the rate of change. We estimate that rate online with a statistically valid piecewise-linear trend model and convert it into the training window size that minimizes a known generalization bound.
Core Claim
Treat the drift rate, not the drift alarm, as the quantity to estimate. The theory of weighted empirical risk minimization under nonstationarity in Fast Rates for Nonstationary Weighted Risk Minimization splits excess risk into two parts. The first is a learning error that shrinks with the effective sample size $1/\|w\|_2^2$. The second is a drift error measured by the distance between past and future optimal models. Suppose the optimal parameters move at speed $v$ inside a segment and we train on a uniform window of $n$ recent samples with per-sample noise level $\sigma^2$. Then the learning term scales like $\sigma^2/n$. The squared drift term scales like $(vn)^2$, because the average lag of the window is proportional to $n$. Minimizing the sum gives
$$n^\star \propto \left(\sigma^2 / v^2\right)^{1/3},$$
so the right window shrinks as the cube root of the inverse squared speed. The prediction is specific. A learner that knows $v$ for the current segment, and notices when $v$ changes, should beat both a fixed window and an alarm-and-reset policy whenever drift is incremental with a speed that varies over time. Joinpoint regression provides $v$ as the segment slope and the change points as the joinpoints. We expect no gain under purely abrupt drift. In that regime the method reduces to an ordinary change detector, and we treat that case as a boundary condition, not a selling point.
Research Program
The first component is the tracked scalar. A plain two-sample statistic does not work. The Kolmogorov-Smirnov distance computed by Fast Unsupervised Online Drift Detection Using Incremental Kolmogorov-Smirnov Test, which makes the KS test streamable through an augmented search tree, saturates at one. Under steady drift its trend is therefore curved, not linear. This is the "model rigidity" problem listed for joinpoint methods. We instead fit a cheap model on each small block of the stream and track the displacement $d_b = \|\hat\theta_b - \hat\theta_{\text{ref}}\|$ of block $b$'s estimate from a reference estimate. Under constant-speed drift this displacement grows linearly plus a noise floor, so the hinge model's continuity assumption is appropriate. For the label-free setting we also track the mean-embedding displacement of the inputs. The cost is that this variant only sees covariate change, and we report it separately.
The second component is the online fit. The offline joinpoint procedure grid-searches knot locations and accepts extra knots by permutation test or by the Bayesian information criterion. Both are invalid when the test is repeated every time a new block arrives. Correcting Split Selection in Online Decision Trees via Anytime-Valid Inference shows that this optional-continuation failure is real for Hoeffding trees and repairs it with betting-based confidence sequences. We reuse that technique. At each block we maintain a confidence sequence for the slope of the current segment, fitted from the last accepted joinpoint onward. We open a new joinpoint when a slope test, run against the hypothesis "one line since the last knot" and evaluated over a bounded set of candidate knot positions, rejects at an anytime-valid level. Restricting the search to the current segment makes each update cost proportional to the segment length, not the full stream. That addresses the scalability limitation of grid search.
The third component is the controller. The estimated slope $\hat v$ and its confidence interval set the training window through the cube-root rule. We use the conservative end of the interval so that a noisy slope does not collapse the window. The constant in $n^\star$ is calibrated once on a held-out synthetic stream. When a jump is detected, meaning a discontinuity in $d_b$ that a hinge cannot express, the controller falls back to a reset. The analysis asks whether the joinpoint estimates of $v$ are consistent under the dependence conditions assumed in the weighted-risk bound, and how large the excess risk becomes when $\hat v$ is wrong by a constant factor. The cube-root form suggests the excess risk is fairly forgiving of such errors, and we will check that.
A note on scope. The research file did not show me prior work that already fits joinpoint models online to drift statistics. Its intersection section was not available to me, so I cannot rule out overlap. The closest work I could see is the Bayesian run-length method of A Bayesian Approach to Concept Drift, which keeps a posterior over how long the current concept has lasted. That posterior is still about segment boundaries, not within-segment velocity.
Baselines
The naive reference point is a fixed sliding window, together with its strongest version: an oracle that picks the single best window length in hindsight on each test stream. If our method cannot beat the oracle fixed window on streams with changing drift speed, the rate estimate adds nothing. We give the oracle its full grid at no extra cost, and the comparison counts as a win only if our method's cumulative prequential loss (each sample is predicted before it is used for training) is lower with a block-bootstrap interval that excludes zero.
The strongest alarm-based competitor is Dynamic Threshold Determination from Autonomous Concept Drift Threshold Determination. It runs an early-adaptation model, a reactive model and a no-change model in parallel after each alarm, and retunes its threshold based on which wins. It is the strongest competitor because it already adapts how aggressively it responds, which is the closest thing in the file to rate awareness. We run it with the same base learner, the same block size, and the same number of hyperparameter trials as our method. Its triple-model comparison is charged against its compute.
The strongest probabilistic competitor is the pruned Bayesian run-length method from A Bayesian Approach to Concept Drift. The research file reports no numbers for any of these methods: no reported result found on Lacuna for DTD, the Bayesian run-length method, IKS, or anytime-valid trees on the streams we plan to use. We will therefore reproduce all baselines ourselves under a shared tuning budget.
Two ablations isolate the key ingredient. The first, "joinpoint-alarm-only", uses our detector to trigger resets but ignores the slope. It tests whether the gain comes from rate-informed windows or merely from a better detector. The second, "naive-BIC", replaces anytime-valid knot tests with BIC re-evaluated at every block. It measures how much the false-knot rate rises without valid sequential inference.
Why not just…?
Why not just tune exponential forgetting or a fixed window? When the drift speed is constant across the whole stream, a well-tuned forgetting factor implements the cube-root rule implicitly, and it is a real threat. The proposal's claim is restricted to streams whose speed changes. The oracle fixed-window baseline settles the question. If our method only matches the oracle when speed varies by a factor of four or more between segments, the idea fails.
Why not just use an adaptive-window or alarm-based detector? Detectors such as IKS or DTD shrink the window after an alarm but carry no estimate of speed. Under slow drift they either never fire or fire repeatedly, so they reset again and again. Most detectors, the file notes, emit point alarms and do not parameterize velocity. The joinpoint-alarm-only ablation shows directly whether the slope matters beyond detection.
Why not just use Bayesian online change-point detection with linear segments? This is the most serious alternative, because a run-length posterior with linear-trend segments could also produce a slope. We include such a variant built on the Bayesian run-length baseline. Our method has two possible advantages. It enforces continuity at knots, which suits incremental drift, and it gives frequentist anytime-valid error control without a hazard-rate prior. If the Bayesian variant matches us at equal compute, we will report that and call the method a cheaper, prior-free alternative, not a better one.
Proposed Experiments
The main testbed is a synthetic regression and classification stream whose optimal parameters move along piecewise-constant-speed paths, mixed with abrupt jumps and recurring concepts. Here the true $v$ is known, so we can measure slope-estimation error and knot-location error alongside prequential loss. We will also use the tabular streams evaluated in Bounded Context Management for Tabular Foundation Models on Stream Learning and in the anytime-valid tree paper. On those streams we additionally test whether the rate-derived window can set the context budget of a frozen tabular foundation model, as a replacement for CURE's fixed bank sizes.
Four checks decide the outcome. On varying-speed synthetic streams, our method must beat the oracle fixed window. It must beat its own alarm-only ablation, which shows that the slope, not the detector, carries the gain. Under no drift, its false-knot rate must stay at the nominal level while naive BIC exceeds it. And on purely abrupt streams it must not lose to DTD by more than the detection delay of a single block. If the second check fails, the core claim is wrong and the work reduces to yet another change detector.
References
- Fast Rates for Nonstationary Weighted Risk Minimization
- Fast Unsupervised Online Drift Detection Using Incremental Kolmogorov-Smirnov Test
- Correcting Split Selection in Online Decision Trees via Anytime-Valid Inference
- A Bayesian Approach to Concept Drift
- Autonomous Concept Drift Threshold Determination
- Bounded Context Management for Tabular Foundation Models on Stream Learning