Comparing Clustering Algorithms.ppt

KD Trees K Dimensional Trees Space Partitioning Data Structure Splitting planes perpendicular to Coordinate Axes Useful in Nearest Neighbor Search Reduces the Overall Time Complexity to O(log n)? Has been used in many clustering algorithms and other domains DBSCAN (Using KD Trees)? Density based Clustering (Maximal Set of Density Connected Points)? O(m) Space Complexity Using KD Trees the overall Time Complexity reduces to O(m * logm) from O(m^2)? Pros Fast for low dimensional data Can discover clusters of arbitrary shapes Robust towards Outlier Detection (Noise)? DBSCAN - Issues DBSCAN is very sensitive to clustering parameters MinPoints (Min Neighborhood Points) and EPS (Images Next)? The Algorithm is not partitionable for multi-processor systems. DBSCAN fails to identify clusters if density varies and if the data set is too sparse. (Images Next)? Sampling Affects Density Measures DBSCAN (Contd.)? Performance Measurements Compiler Used - Java 1.6 Hardware Used Intel Pentium IV 1.8 Ghz (Duo Core)? 1 GB RAM No. of Points 1572 3568 7502 10256 Clustering Time (sec) 3.5 10.9 39.5 78.4 CURE – Hierarchical Clustering Involves Two Pass clustering Uses Efficient Sampling Algorithms Scalable for Large Datasets First pass of Algorithm is partitionable so that it can run concurrently on multiple processors (Higher number of partitions help keeping execution time linear as size of dataset increase)? CURE (Contd.)? Outperforms Basic Hierarchical Clustering by reducing the Time Complexity to O(n^2) from O(n^2*logn)? Two Steps of Outlier Elimination After Pre-clustering Assigning label to data which was not part of Sample Captures the shape of clusters by selecting the notion of representative points (well scattered points which determine the boundary of cluster)? CURE - Benefits against Popular Algorithms K-Means ( Centroid based Algorithms) : Unsuitable for non-spherical and size differing clusters. CLARANS :