On achieving semi-supervised pattern recognition by utilizing tree-based SOMs

Years of research in the field of Pattern Recognition (PR) has led to scores of algorithms which can achieve supervised pattern classification. Such algorithms assume the knowledge of well-defined training sets with a clear specification of the identity of all the training samples. However, more recently, a new stream has emerged, namely, the so-called "semi-supervised" paradigm, i.e., one that uses a combination of labeled and unlabeled samples to perform classification [41]. Classifiers based on the latter, do not demand the specification of the class labels of every sample. Rather, a clustering-like mechanism processes the manifold, and attempts to distinguish the training samples into the separate classes, subsequent to which a supervised classifier is derived using a small subset of the training samples whose class identities are known. In this paper we will venture to utilize the Tree-based Topology Oriented SOM (TTOSOM) [3] for semi-supervised pattern classification. We first train a TTOSOM in which the neurons collectively obey the stochastic, topological and structural distribution of all the classes. Subsequently, we make use of the information provided in the labeled dataset. By using this information, we assign a class label to every single node in the Neural Network (NN), which, in turn, partitions the space into its Voronoi regions. On receiving the testing data, the task at hand is rather straightforward. One nearly determines the closest neuron to the testing sample and assigns the sample to the corresponding class. The complexity of the testing is linear, not in cardinality of the training set, but rather in the size of the TTOSOM tree! Our experimental results show that on average, the classification capabilities of our proposed strategy, even with a small number of neurons, are reasonably comparable to those obtained by some of the state-of-the-art classification schemes that only use labeled instances during the training phase. The experiments also show that improved levels of accuracy can be obtained by imposing trees with a larger number of nodes. (C) 2012 Elsevier Ltd. All rights reserved.