Multi-Attribute Decision Fusion for Pattern Classification

Classification is an important problem that is found in several domains. It has a wide range of application in medical diagnosis, decision making, and target classification. Several methods have been used for classification with varying degrees of success. Although multi-criteria decision analysis (MCDA) has been applied to classification problem, its application to pattern classification is relatively new. In this work, we propose a multi attribute decision fusion based on the classical multi-criteria decision analysis (MADF-MCDA) for pattern classification problems. Multi-criteria decision making is a technique for selecting an alternative from a group of known alternatives based on certain criteria. The MCDA concept has been applied to many problems in sciences, business and engineering disciplines. In the proposed method, every attribute is considered as a criterion upon which a final decision is taken. For every attribute, an attribute model is built using a set of Gaussian membership functions for different classes. We assign a weight to each attribute model. This is because each attribute used in the modelling of an attribute model may have a different contribution to the final decision. It is expected that a model used for classification should have high distance among classes. The assigned weights are determined by the areas of intersection among the classes. This study also compares the performance of the proposed method with standard classification techniques such as K-nearest neighbor (K-NN), and support vector machine (SVM).