Data complexity measures for classification of a multi-concept dataset

Classification algorithms design predictive models that classify data under one of the predefined categories. The data can be text, image, audio, video, or animation. The Data Complexity Metrics (DCM) gives insight into the different aspects of the data characteristics like data distribution, noise, overlap, and separability. Existing data complexity metrics consider linearly separable datasets to be less complex. However, for a separable multi-concept dataset, these metrics fail to generate low complexity values despite the high accuracy yielded by most of the state-of-the-art classification models. The erroneous complexity estimates generated by the existing metrics make it difficult to choose the efficient classifier model or fine-tune its parameter settings. This work addresses the complexity-accuracy discrepancy issue by formulating novel data complexity metrics for multi-concept and simple datasets. A density-based clustering algorithm (OPTICS) is initially utilized to identify the concepts in each class, which is later utilized in the formulation of the DCM in multi-concept datasets. This work also explores the relationship between the metrics and classifiers in multi-concept datasets. The classifiers’ accuracy for different data complexity levels is studied to elect a highly accurate, robust classifier for any dataset. Comprehensive coverage of the proposed technique in different data distributions with varied degrees of overlapping scenarios is examined in this paper using synthetic datasets.