Generating Fixed-Size Training Sets for Large and Streaming Datasets

The k Nearest Neighbor is a popular and versatile classifier but requires a relatively small training set in order to perform adequately, a prerequisite not satisfiable with the large volumes of training data that are nowadays available from streaming environments. Conventional Data Reduction Techniques that select or generate training prototypes are also inappropriate in such environments. Dynamic RHC (dRHC) is a prototype generation algorithm that can update its condensing set when new training data arrives. However, after repetitive updates, the size of the condensing set may become unpredictably large. This paper proposes dRHC2, a new variation of dRHC, which remedies the aforementioned drawback. dRHC2 keeps the size of the condensing set in a convenient, manageable by the classifier, level by ranking the prototypes and removing the least important ones. dRHC2 is tested on several datasets and the experimental results reveal that it is more efficient and noise tolerant than dRHC and is comparable to dRHC in terms of accuracy.