The k-means forest classifier for high dimensional data

The priority search k-means tree algorithm is the most effective k-nearest neighbor algorithm for high dimensional data as far as we know. However, this algorithm is sensitive to attribute noise which is common in high dimensional spaces. Therefore, this paper presents a new method named k-means forest that combines the priority search k-means tree algorithm with random forest. The main idea is to create multiple priority search k-means trees by randomly selecting a fixed number of attributes to make decisions and get the final result by voting. We also design a parallel version for the algorithm. The experimental results on artificial and public benchmark data sets demonstrate the effectiveness of the proposed method.