SiD2Re Adaptive Learning on Non-stationary Regression Tasks

This dataset is a collection of synthetic non‑stationary regression streams designed for evaluating concept and data drift detection methods. All streams are generated with the SiD2ReGenerator / DataGeneratorGraph (10.1109/INDIN51400.2023.10218255 ; https://github.com/FraunhoferIOSB/SiD2Re ) and simulate industrially motivated scenarios with controlled, annotated drifts. Each individual dataset consists of a univariate regression target and 10 input features, ordered in time: Length: 10,000 samples per stream Inputs: 10 features (0_0 – 2_9) Target: 1 continuous output (O_0) Feature types: Mixture of Uniform, Gaussian, Constant, and Periodical distributions (assigned per seed) Drift is injected at known timestamps, with full ground truth: Concept drift types: S1 – Sudden single drift G1 – Gradual single drift I1 – Incremental single drift S2 – Two successive sudden drifts in one stream DC1 – Sudden concept + data drift (one feature affected) Drift timing: Single-drift streams (S1, G1, I1, DC1) drift around a centre at sample 5000 with a seed‑dependent radius (transition zone). Two-drift streams (S2) have centres at 2500 and 7499, each with its own radius. Ground truth files: concept_drifts.csv provides drift centre, radius, parameter shift, and drift class. data_drifts.csv (DC1 only) additionally annotates the affected feature and data drift parameters. For every drifted stream there is a corresponding non‑drift baseline (data_nonDrift.csv) with identical feature and target definitions but no drift. In total, the benchmark comprises: 5 drift types × 5 severities × 5 noise levels × 10 seeds = 1,250 datasets Each dataset directory additionally contains an automatically generated overview plot (plot_overview.png) that visualizes the target, drift windows, and (for DC1) the affected feature, along with a corresponding plot for the non‑drift baseline.