Research On Clustering And Anomaly Detection Based On Soft Computing

Inaccuracy of data and non-unique results will lead to uncertainty,making the problem more complex and changeable in real-life.To solve the uncertain problem,people have introduced soft computing methods,which is an emerging computer technology and often get more adaptable and robust result.Soft computing has now become a research focus in the fields of engineering,science,education,and management.Moreover,there are two situations widely exist in real life,data without label or data with insufficient labels.Cluster analysis is a common method for modeling unlabeled data,while transfer learning can solve the problem of insufficient labeling of target domain data by making full use of source domain data.In this paper,we propose an improved maximum entropy clustering algorithm and an anomaly detection method based on soft instance transfer learning.The contributions of this article mainly include the following two aspects.1)In this paper,we propose an minimal dependence-based maximum entropy clustering alogrithm(MDMEC).The original maximum entropy clustering methods only considering the within-cluster similarity but ignoring the between-cluster dissimilarity.This paper introduces the Hilbert-Schmidt independence criterion to solve this problem.In the process of algorithm iteration,the within-cluster distance and the between-cluster distance are optimized at the same time to ensure the maximum dependence within the same cluster and the minimum dependence between different clusters.In the experiment,several commonly used experimental data sets are selected to test the performance of the improved algorithm,and the experimental results are compared with the results of the traditional clustering algorithms.Finally,this paper analyzes the convergence and parameter sensitivity of the improved maximum entropy clustering algorithm,which proves that the algorithm can converge quickly.2)We also propose an soft instance transfer based anomaly detection method to find the abnormal points in the data set.In the field of anomaly detection,due to the lack of reliable labels in the target domain,we get to learn a transfer learning model to transfer useful knowledge from the source domain,In the experiment,we also verify that the performance of the anomaly detection algorithm combined with soft instance transfer learning is better than the traditional anomaly detection algorithm on a credit card fraud data set.