A Novel Noise Clustering Based on Local Outlier Factor

Abstract

Reducing the impact of outliers is an essential issue in machine learning, including clustering. There are two main approaches to reducing the impact of outliers: one is to build robust models, and the other is to remove outliers through preprocessing. In this paper, we propose a new noise clustering method that combines noise clustering, which builds a model robust to outliers, and local outlier factor, which removes outliers as a preprocessing step. The proposed method is an optimization problem of noise clustering with a weighting of dissimilarities by LOF. Numerical experiments were conducted using four artificial datasets to verify the effectiveness of the proposed method. In the experiments, the proposed method was compared with k-medoids clustering, DBSCAN, and noise clustering. The results show that the proposed method yields good results regarding both clustering performances and detecting outliers. The guideline for determining k and $\varepsilon$ among the three parameters D, k, and $\varepsilon$ required by the proposed method was also suggested.