Data for Evaluation of Stream Data Analysis Algorithms

Published: 1 November 2021| Version 1 | DOI: 10.17632/c43kr4t7h8.1
Contributor:
Félix Iglesias Vázquez

Description

This collection of datasets have been generated with MDCStream for evaluating clustering and outlier detection algorithms in stream data analysis. The temporal behavior is assumed in the order in which data points appear in the file, this means: simultaneity is not considered and the time-difference between consecutive data points (i.e., consecutive rows) is the unit. Datasets are arranged in 9 folders according to the data challenge: base (baseline), nonstationary (clusters coexist, appear and disappear randomly), sequential (clusters happen sequentially), moving (cluster centroids move as time passes), medium-outliers (outliers account for 5% to 15% of the data), many-outliers (outliers account for 15% to 40% of the data), close (the space is reduced and clusters are very close each other), density-differences (distributions underlying point generation are highly varied), overlap (the data generation favors cluster overlap).

Files

Steps to reproduce

Datasets are in ARFF format. Features are numerical. The ground truth is provided in the "class" feature, where natural numbers above '0' identify clusters and '0' remains for outliers

Institutions

Technische Universitat Wien Institute of Telecommunications

Categories

Unsupervised Learning, Clustering, Multivariate Analysis, Stream, Concept Drift, Outlier

License