Word Cloud Model for Text Categorization
Tam Thanh Nguyen, Kuiyu Chang, Siu Cheung Hui · 2011
In centroid-based classification, each class is represented by a prototype or centroid document vector that is formed by averaging all member vectors during the training phase. In the prediction phase, the label of a test document vector is assigned to that of its nearest class prototype. Recently there has been revived interest in reformulating the prototype/centroid to improve classification performance. In this paper, we study the theoretical properties of the recently proposed Class Feature Centroid (CFC) classifier by considering the rate of change of each prototype vector with respect to individual dimensions (terms). The implication of our theoretical finding is that CFC is inherently biased towards large (dominant majority) classes, which means it is destined to perform poorly for highly class-imbalanced data. Another practical concern about CFC lies in its overly-aggressive design in weeding out terms that appear in all classes. To overcome these CFC limitations while retaining its intrinsic and worthy design goals, we propose an improved and robust centroid-based classifier that uses precise term-class distribution properties instead of simple presence or absence of terms in classes. Specifically, terms are weighted based on the Kullback-Leibler divergence measure between pairs of class-conditional term probabilities, we call this the CFC-KL centroid classifier. We then generalized CFC-KL to handle multi-class data by summing pair wise class-conditioned word probability ratios. Our proposed approach has been evaluated on 5 datasets, on which it consistently outperforms CFC and the baseline Support Vector Machine classifier. We also devise a word cloud visualization approach to highlight the important class-specific words picked out by our CFC-KL, and visually compare it with other popular term weigthing approaches. Our encouraging results show that the centroid based generalized CFC-KL classifier is both robust and efficient to deal with real-world text classification.