Understanding the Evaluation Abilities of External Cluster Validity Indices to Internal Ones

Xiaonan Gao ,Sen Wu ,Guiying Wei ,Falong Fan

doi:10.17559/tv-20200828055024

Abstract

Understanding the Evaluation Abilities of External Cluster Validity Indices to Internal Ones

Highlights

Clustering is to discover unknown clusters in mass data, which directly depends on some prior choices, such as clustering algorithm, similarity metric and parameter configuration, etc [1, 2]
We propose a novel approach, named external Cluster Validity Index (CVI)'s evaluation Ability MEasurement approach through Ranking consistency (CAMER), by formulating the external CVI's evaluation ability measurement problem as a ranking consistency task.This approach can assess the evaluation abilities of external CVIs to internal CVIs quantitatively and help choose reasonable external CVIs to evaluate in different scenarios
To illustrate the effect of the correct number of clusters on the measurement result, we conduct comparison based on two kinds of partitions: (1) Partitions including the correct number of clusters; (2) Partitions excluding the correct number of clusters. 4.2.1 Partitions Including the Correct Number of Clusters

Summary

Introduction

Clustering is to discover unknown clusters in mass data, which directly depends on some prior choices, such as clustering algorithm, similarity metric and parameter configuration, etc [1, 2]. Different algorithms, similarity metrics or parameter configurations would produce various clustering results. Cluster Validity Indices (CVIs) are demanded to estimate which one is the best. Existing studies on CVIs can be divided into external CVIs [3] and internal CVIs [4]. The difference between them is whether the data labels or other external information are required. External CVIs evaluate clustering results by assessing the consistency between the clustering label and the ground truth label. Internal CVIs do evaluation according to the compactness of intra-cluster and separation of inter-clusters

Objectives

Results

Conclusion