Improving css-KNN Classification Performance by Shifts in Training Data

This paper presents a new approach to improve the performance of a css-k-NN classifier for categorization of text documents. The css-k-NN classifier (i.e., a threshold-based variation of a standard k-NN classifier we proposed in [1]) is a lazy-learning instance-based classifier. It does not have parameters associated with features and/or classes of objects, that would be optimized during off-line learning. In this paper we propose a training data preprocessing phase that tries to alleviate the lack of learning. The idea is to compute training data modifications, such that class representative instances are optimized before the actual k-NN algorithm is employed. The empirical text classification experiments using mid-size Wikipedia data sets show that carefully cross-validated settings of such preprocessing yields significant improvements in k-NN performance compared to classification without this step. The proposed approach can be useful for improving the effectivenes of other classifiers as well as it can find applications in domain of recommendation systems and keyword-based search.