Approximation of kernel k-means for streaming data

Streaming data are any data that are sequentially presented to a system such that future data cannot be accessed. By their nature, streaming data are often large data sets and can quickly outgrow the working memory for a typical computer. Clustering is one of the primary tasks used in the pattern recognition and data mining communities and kernel k-means is a well-studied and popular algorithm. However, kernel k-means is ill-suited to streaming data. In this paper, I develop an approximation to the kernel k-means that only needs access to data from time-step t and (t-1); hence, the memory requirement of the proposed algorithm is only O(n²), where n is the size of the data chunk at time t. Empirical results show that streaming kernel k-means (skKM) produces partitions similar to those produced by kernel k-means run on the entire data set.