Mm-websom: a Variant of Websom Based on Order Statistics

A variant of the WEBSOM architecture for information retrieval is proposed in this paper. WEBSOM is based on the self-organizing map that employs a linear LMS adaptation rule for updating the weight vector of each neuron. Accordingly, the weight vector converges asymptotically to the conditional cluster mean of the feature vectors assigned to the class represented by the weight vector of the neuron. We propose to replace the updating rule by employing the marginal median. The objective is to overcome the drawbacks of the standard technique in the presence of outliers in the training set and to use robust estimators of the reference vectors for each class. Experimental results demonstrate a superior performance of the proposed variant against the standard algorithm, in terms of the number of training iterations needed so that the mean square error (i.e., the average distortion) drops to the 1 e of its initial value. We provide precision-recall curves as a measure of the quality of the clustering procedur as well. Both techniques are tested using a corpus that comprises web pages selected over the Internet.