Increasing the Quality and Performance of N-Dimensional Point Anomaly Detection in Traffic Using PCA and DBSCAN

Detection of point anomalies is a very important issue in a large scale of fields from Astronomy and Biology to network intrusions. Clustering has been employed by many researchers to solve such problems and DBSCAN seems like the most efficient technique. Due to its high computational complexity, this work focused on decreasing it by decreasing the dimensionality of the data points. For this reason, Principal Components Analysis used and then DBSCAN applied on the new data sets provided by PCA. The quality of the experimental results was very promising proving that such an approach can be adapted. In addition, the performance of the combined PCA and DBSCAN was examined. Our analysis shows that a speedup of 25% was achieved while the quality was 80% reducing the dimensionality of data set to half.