Outlier detection by network flow

Outliers are those points which are different from or inconsistent with the rest of the data. Novel, new, abnormal, unusual or noisy information can all be called outliers. Sometimes the outliers are more interesting than the majority of the data, such as the applications of intrusion detection and unusual usage of credit cards. With the increase of the complexity and variety of datasets, the challenges of outlier detection are how to catch similar outliers as a group, and how to evaluate the outliers. The research described in this dissertation has investigated the above challenges and has addressed the solutions. A novel method for outlier and outlier group identification which is based on network flow has been proposed in this dissertation. It uses the Maximum Flow Minimum Cut theorem from graph theory to find the outliers and strong outlier groups, and evaluate the outliers by outlier degrees. This algorithm uses k-nearest neighbors to find density connected points of the dataset. Data relationships are magnified to expose loosely connected outliers by the network settings. It can repair poor quality clusters generated by any clustering algorithm, in particular, to solve the problem that points supposed to be separated are in one cluster. As part of the experimental evaluation, our algorithm has been compared with density-based, distribution-based, and conventional statistic outlier detection algorithms, and it is more effective at finding outlier groups. For high dimensional data, this algorithm is also effective at detecting outliers and outlier groups. The effectiveness of this algorithm is demonstrated by improving the query accuracy of content-based image retrieval.