An anomaly detection method based on Lasso

OFCOD: On the Fly Clustering Based Outlier Detection Framework

“…We compute the TPR and FPR of the proposed algorithm on a test set which results in 99.58%. The accuracy of the proposed OFCOD framework is compared to K-SVM approach [39], KNN [40] and Lasso [40] on the same dataset to evaluate the performance of the proposed framework. In K-SVM approach [39], a k one-class support vector machine models are trained for normal data.…”

“…Thereafter, the SVM model of the closest hypersphere to the point, is queried to classify either the point is inlier or outlier. Similarly, the K-Nearest Neighbor (KNN [40]) approach creates clusters inside each class of points. Therefore, any new instance is compared to the created clusters to find the most similar one to which the new instance belongs.…”

“…Therefore, any new instance is compared to the created clusters to find the most similar one to which the new instance belongs. On the other hand, the Lasso approach [40] transforms the anomaly detection problem into a linear regression model. It defines the detection parameters as regression variables and constructs the model (finding the regression variables) that maximizes the detection accuracy.…”

“…This assumption cannot be justified and guaranteed since outliers can be grouped in small hyperspheres (clusters) as discussed in Sections 2.1.3 and 2.1.4. The KNN approach [40] has the similar problem in addition to its limited generalization ability to unseen test instances. These issues lead to the worst results among others.…”

“…These issues lead to the worst results among others. On the contrary, the Lasso approach [40] relaxed this assumption leading to better results compared to the K-SVM and KNN approaches. In conclusion, the proposed OFCOD framework has the advantage of finding outliers in both outlier clusters and normal clusters which is the general case.…”