Local Outlier Detection Based on K-distance Variation to Enhancing Imaging-aided Diagnosis

Local outlier detection is a critical task in data mining, which identifies outliers by estimating the anomaly score for each data point. The traditional density-based methods assign large k-distances to boundary data, which leads to the "boundary bias issue". Due to this issue, the anomaly scores of the boundary data are overestimated while the scores of their neighboring data points are underestimated, limiting the performance of local outlier detection. In this paper, we propose a novel local outlier detection method based on k-distance variation. This technique replaces the typical k-distance by measuring the k-nearest neighbor distance variations, eliminating the boundary bias by using adjacent nearest neighbors to transfer distances. To evaluate the performance of our proposed approach, we conduct plenty of experiments based on synthetic dataset, and also transform local anomaly scores into attention maps in the breast cancer detection. Experiments demonstrate that the proposed method not only outperforms other state-of-the-art methods in terms of the area under the receiver operating characteristic curve and precision on various synthetic datasets, but also generates more accurate attention maps to focus on multiple lesion areas for medical imaging-based diagnosis.