Aykırı değer tespitinde yoğunluk tabanlı kümeleme yöntemleri

Abstract

Fraud that causes high amounts of finance loss, has became one of the serious problems. Either proactive efforts that focuses on prevention of fraud or working on fraud detection always use data mining approaches.Outlier detection, which is one of the data mining studies, detects objects that has different behavior in similar elements. These elements are usually nominated to be fraudulent elements. Clustering methods are mostly used for outlier detection. Clustering algorithms that are sensitive to noise or the inconsistent elements, are playing an active role in the detection of fraudulent behavior.Clustering is one of the data mining methods that is used for the unsupervised analysis of the data. Especially, if the data has not enough information(foreknowledge), similar data is grouped by the help of the clustering methods. DBSCAN, which is the one of the density-based clustering methods, does the process of clustering, according to density of data.Although DBSCAN method seems effective in the small data sets, its efficiency decreases with the growing of data volumes. Because of this reason, DBSCAN as a clustering method is not considered a suitable clustering method for large data sets.In the scope of this thesis, R-P-DBSCAN (Recursive-Partitioned DBSCAN) algorithm is proposed. The new algorithm is based on partitioning & combining and DBSCAN algorithm is used for data clustering. Large-volume data sets are divided into smaller pieces and clustered by DBSCAN. Then, combining each clustered piece, until whole set of data is clustered. Each cluster obtained by R-P-DBSCAN, is the same as the clusters obtained with the classical DBSCAN.The results obtained with R-P-DBSCAN have shown that, the proposed algorithm has better clustering performance (until 85%) according to classical DBSCAN algorithm

Fraud that causes high amounts of finance loss, has became one of the serious problems. Either proactive efforts that focuses on prevention of fraud or working on fraud detection always use data mining approaches.Outlier detection, which is one of the data mining studies, detects objects that has different behavior in similar elements. These elements are usually nominated to be fraudulent elements. Clustering methods are mostly used for outlier detection. Clustering algorithms that are sensitive to noise or the inconsistent elements, are playing an active role in the detection of fraudulent behavior.Clustering is one of the data mining methods that is used for the unsupervised analysis of the data. Especially, if the data has not enough information(foreknowledge), similar data is grouped by the help of the clustering methods. DBSCAN, which is the one of the density-based clustering methods, does the process of clustering, according to density of data.Although DBSCAN method seems effective in the small data sets, its efficiency decreases with the growing of data volumes. Because of this reason, DBSCAN as a clustering method is not considered a suitable clustering method for large data sets.In the scope of this thesis, R-P-DBSCAN (Recursive-Partitioned DBSCAN) algorithm is proposed. The new algorithm is based on partitioning & combining and DBSCAN algorithm is used for data clustering. Large-volume data sets are divided into smaller pieces and clustered by DBSCAN. Then, combining each clustered piece, until whole set of data is clustered. Each cluster obtained by R-P-DBSCAN, is the same as the clusters obtained with the classical DBSCAN.The results obtained with R-P-DBSCAN have shown that, the proposed algorithm has better clustering performance (until 85%) according to classical DBSCAN algorithm