Density-Based Clustering by Double-Bit Quantization Hashing

Abstract

Grouping data into the different parts, while the objects in the same part have the most similarity with each other and cannot belong to the other parts, called data clustering. Clustering used for data analysis in data mining, so far, many different algorithms for clustering have been offered. Density-based algorithms are one of the useful clustering approaches, which used for databases with different shapes. This algorithms have a short response time for small databases and also be able to extract clusters with arbitrary shapes. DBSCAN is a density-based clustering algorithm that can detect and extend clusters based on a restricted neighbor radius and the number of near objects in neighbor radius. The time complexity of this algorithm belongs to \(O(n^2)\) for large datasets. We used data indexing technique Local Sensitive Hashing (LSH) to reduce the implementation time of the algorithm, this data structure can be used to found neighbor points in the DBSCAN algorithm, so, the response time of the algorithm, will be reduced, because LSH be able to approximate the K-nearest neighbors algorithm in linear time complexity. We used this data structure to detect neighbor points quickly by mapped data to a binary space. We used the influence space idea to detect clusters, to improve the response time of the algorithm, this concept can reduce the search space to expand the clusters. We evaluated our algorithm by two density-based clustering algorithms DBSCAN and BLSH-DBSCAN. We can improve both mentioned algorithms in terms of response time for large datasets.