Browsing by Author "Mehmet Emin, Akşehirli"

Filter results by typing the first few letters
Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    New cluster ensemble algorithm with automatic cluster number and new pruning technique for fast detection of neighbors on binary data
    (Bahçeşehir Üniversitesi Fen Bilimleri Enstitüsü, 2011-06) Mehmet Emin, Akşehirli; Mimaroğlu, Selim Necdet
    Cluster analysis is to group similar, real or abstract data objects together in an unsupervised way. Cluster analysis, or clustering is a very important tool for data analysis and widely-used in almost every scientific field including data mining, machine learning, bioinformatics, and social network analysis. Unsupervised nature of clustering comes with unique opportunities and challenges. Applying the optimum clustering algorithm with correct parameters is not straight forward. Moreover, unlike classification algorithms which use the provided labels, clustering algorithms extract the information from the data itself, therefore most of the algorithms suffer from long execution times. Combining multiple clusterings methods emerge as a promising solution that not only ease the algorithm and parameter selection for cluster analysis but also solve some unique clustering problems. In this theses we discuss the methods that combine multiple clusterings to obtain a better overall clustering of the data, including a recent method: Di- CLENS. DiCLENS does not take any input arguments and finds the number of clusters automatically using objective measures. Although finding the co-associations between objects is a computationally expensive task, it is one of the strongest similarities in the field. DiCLENS utilizes a recent method to compute the similarities in an efficient way. Our experiments show that DiCLENS produces a better final clustering at almost all of the scenarios. Moreover execution time of the DiCLENS is very good compared to other methods. We also discuss DBSCAN BV, a novel method that improves the execution time performance of DBSCAN clustering algorithm by utilizing a pruning method on binary data and Hamming distance. DBSCAN is a well-known density-based algorithm. Even though space indexing techniques are widely used with DBSCAN, they do not perform well on categorical and binary data sets. Extensive tests show that DBSCAN BV works up to 40 times faster than DBSCAN while keeping the same clustering accuracy. Tests also show that the new pruning method allows the application of DBSCAN to resource limited environments.