Partial correlation metric based classifier for food product characterization

Abstract

Data classification algorithms applied to multidimensional and multiclass food characterization problems mainly assume feature independency to quantify intra-class similarity or inter-class dissimilarities. As an alternative, possible class specific inter-relations among the feature vectors can be exploited for distinguishing samples into specific classes. Based on this idea, a new partial correlation coefficient metric (PCCM) based classification method is proposed. Existence of such inter-variable correlations as signatures of unique classes is established with illustrative problems. Categorized variable dependency structures are hypothesized as the basis for class discrimination. Two food quality analysis datasets with chemometrics importance are utilized as benchmark problems to compare the performance of new method with classification algorithms like LDA (linear discriminant analysis), CART, Treenet and SVM (support vector machines). The PCCM method is observed to perform well for different tests over large sets of classification experiments. Discriminating PCCM classifier also provides a quick visualization tool to diagnose complex classification problems.