Text Categorization Based on Granular Partition

Two factors strongly influence the quality of text categorization: (1) the class ambiguity of texts, i.e., some texts in one category may have greater similarities with some other texts in another category, (2) the diversity of discriminability of different type of feature. A classification approach that exploits the same type of feature at all steps of classification, or performs a single level classification, would suffer from the problems related to these factors. To deal with these problems, this paper proposes a text categorization model based on granular partition. This approach transforms text categorization to an optimization problem: given n feature types, to search an optimal partition solution, in which the collection is partitioned into many sub-parts, when every sub-part is represented by features with the suitable feature type that ensures the sub-part has the highest categorization performance, the global categorization performance is the best one in all impossible partition solutions. To get an approximate solution of the proposed model, a multi-level segmentation algorithm is developed, which employs dimidiate strategy, i.e., it uses a classifier with a given feature type to classify the test collection, then divide the test collection into two parts according to the output of the classifier, the part that the classification result is reliable is assigned to the given feature type as a match sub-part, the other part is considered as a new test collection at the next level. The n sub-parts generated for n feature types are considered as an approximate optimal partition solution. The experiments show that the proposed method can consider effectively the two factors and achieve a better performance.