DIP: Density-increasing Neighbor Path for Unsupervised Anomaly Detection

This repository contains the code for paper "Searching Density-increasing Path to Density Peaks for Unsupervised Anomaly Detection".

Quick start

Run the main.py in the Evaulate dictionary.

We offer a demo to evaluate single DIP on musk dataset and evaluate ensemble DIP on arrhythmia dataset.

Main idea of DIP

Idea of DIP. (A) plots the data points; (B) shows the anomaly scores by color; (C) and (D) show the paths of two normal point 12 and 29; (E) and (F) show the paths of two abnormal point 43 and 38; (G)-(I) show graphs learned by DIP and its variants. (Fig. 1 in the paper)

Experimental results

DIP achieved better average results than a lot of existing traditional and deep methods. The evaluation metrics is F1 score. This repository only show the F1 score performance, while the other results can be found in the original paper. The dataset information and other datasets used in the paper can be download from here.

F1 score performance

ROC	arrhythmia	ionosphere	lympho	mnist	musk	pima	satellite	satimage-2	thyroid	vowels	wbc	Average
ABOD	0.762	0.922	0.946	0.733	0.167	0.678	0.554	0.793	0.909	0.965	0.908	0.757
CBLOF	0.786	0.890	0.979	0.802	1.000	0.646	0.719	0.999	0.931	0.915	0.939	0.873
FB	0.782	0.868	0.980	0.698	0.357	0.635	0.552	0.457	0.950	0.946	0.945	0.742
HBOS	0.815	0.653	0.998	0.613	1.000	0.686	0.750	0.983	0.959	0.686	0.953	0.827
ISF	0.808	0.846	0.995	0.807	0.997	0.664	0.730	0.992	0.979	0.752	0.919	0.862
LODA	0.757	0.801	0.667	0.724	0.893	0.623	0.611	0.986	0.828	0.713	0.926	0.775
LSCP	0.797	0.887	0.986	0.751	0.346	0.642	0.567	0.671	0.947	0.944	0.94	0.770
PCA	0.778	0.793	0.986	0.834	1.000	0.625	0.628	0.977	0.956	0.617	0.906	0.827
SOD	0.731	0.889	0.919	0.589	0.654	0.589	0.644	0.840	0.921	0.887	0.919	0.780
DAE	0.801	0.947	0.886	0.794	0.758	0.665	0.638	0.799	0.943	0.555	0.875	0.787
DAGMM	0.279	0.369	0.52	0.332	0.314	0.503	0.305	0.862	0.536	0.57	0.761	0.486
DSVDD	0.684	0.730	0.796	0.688	0.767	0.481	0.670	0.733	0.693	0.500	0.911	0.695
DIP-euc	0.811	0.928	0.913	0.858	1.000	0.728	0.652	0.999	0.946	0.958	0.943	0.885
EDIP-cos	0.816	0.941	0.927	0.945	1.000	0.616	0.800	0.991	0.962	0.909	0.583	0.863
EDIP-euc	0.812	0.916	0.959	0.861	1.000	0.73	0.779	0.999	0.942	0.933	0.943	0.900
EDIP-man	0.829	0.875	0.979	0.824	1.000	0.738	0.789	0.998	0.958	0.921	0.948	0.900