Locality Sensitive Clustering in High Dimensional Space

In high dimension space, many conventional clustering algorithms do not work well in effectiveness and efficiency, especially for image data set. For example, k-means is widely used in image clustering especially visual clustering. But its drawback such as long clustering time and high memory cost seriously deteriorates feasibility in incremental large image set. To improve the feasibility, we proposed a Locality Sensitive Clustering method. Firstly, multiple hashing functions are generated. Secondly, data points are projected to get bucket indices. Thirdly, proper quantification interval is selected to merge the bucket indices, and the cluster labels are assigned for each point. Experimental results show that on synthetic data set this method performs almost as well as k-means, and on image data set it performs slightly worse than k-means algorithm about accuracy. But its advantage is in low memory cost, fast running speed and incremental clustering. So Locality Sensitive Clustering can be used to clustering data, especially in high dimensional space.