Waveform-5000 Classification Benchmark

## Overview Waveform-5000 is a synthetic 3-class classification benchmark generated by Breiman et al. (1984) and widely used in classical ML literature. The dataset contains 5,000 rows and 41 columns (40 numeric features + 1 class label). Each class corresponds to a combination of two of three base 'waveforms' with added Gaussian noise. Of the 40 attributes, 21 carry signal and 19 are pure noise (mean 0, variance 1), making it useful for studying feature selection and noise robustness. Format: ARFF (OpenML id 60). ## Schema - attribute_0 … attribute_39 — float — numeric waveform samples; first 21 are informative, last 19 are noise - class — categorical {0, 1, 2} — target waveform class (roughly balanced, ~33% each) ## Sources - OpenML dataset 60 — https://www.openml.org/d/60 — Public Domain (CC0) - Originally from UCI Machine Learning Repository (Waveform Database Generator v2) — https://archive.ics.uci.edu/dataset/108/waveform+database+generator+version+2 - Reference: Breiman, L., Friedman, J., Olshen, R., Stone, C. (1984). Classification and Regression Trees. Wadsworth. ## Methodology The data is fully synthetic. Each instance is constructed as a random convex combination of two of three predefined triangular base waveforms, sampled at 21 equally spaced points, with independent N(0,1) noise added at every point. An additional 19 pure-noise attributes are appended. Class labels indicate which pair of base waveforms was combined. OpenML stores the data unchanged from the UCI generator output in ARFF form. ## Known gaps & limitations Because the data is synthetic with a known generative process, results do not transfer directly to real-world signal classification. Class boundaries are well-studied and the Bayes-optimal error rate is approximately 14%, so this dataset is mainly useful as a controlled benchmark rather than a difficult modern challenge. Source does not document additional gaps; buyers should validate empirically. ## Intended use & out-of-scope - IS for: tabular ML benchmarking, feature-selection studies, noise-robustness experiments, teaching examples, AutoML pipeline validation. - NOT for: real-world signal/waveform analysis, production model training, or any application requiring realistic, non-synthetic data. Original supplier listing: Waveform-5000 (UCI / OpenML) Classic 3-class synthetic waveform classification benchmark with 5,000 instances and 40 numeric attributes (21 informative + 19 noise). Originally from Breiman et al. 1984, distributed via UCI and OpenML.