Outlier detection via optimized density peaks clustering and K-means-derived objective function

This paper proposes an outlier detection method that integrates optimized Density Peak Clustering (DPC) with K-means-derived objective function analysis, aiming to address the limitations of existing techniques in detecting both global and local outliers as well as sparse clusters. Our approach combines the cluster center selection strategy of DPC with the initialization of K-means, further enhanced by kernel density estimation and k-nearest neighbor techniques to improve both computational efficiency and accuracy in identifying cluster centers. Following K-means clustering, a novel anomaly scoring mechanism is developed through three key steps: 1) within-cluster ascending sorting of objective values, 2) least squares-based function fitting and derivative analysis to estimate rate of change, and 3) comprehensive anomaly scoring through a weighted summation of objective values and their corresponding slopes. The effectiveness of the proposed method is validated through extensive experiments on six complex synthetic datasets and fourteen publicly available real-world datasets, with performance compared against ten state-of-the-art outlier detection algorithms.