Text mining based on conditional probability output networks

Abstract

Multi-labeled classification presents a challenging problem in data mining. Furthermore, it has be-come a very important research field due to the need of handling large scale databases, with some ofthem having an incredible amount of information in text format. Thus, ecient methods and automatictools had recently gained relevance. This work focuses on the efforts to improve the performance ofthe automatic text multi-categorical multi-labeled classification using SVM classifiers. We introduce anew method of multi-labeled classification based on a class probability output network called 2 layerConditional Probability Output Networks. With the objective of refining the classification accuracy,the output of the support vector machine is considered in order to get a complete distribution inde-pendent algorithm, both kernel and probability distribution parameters are finely tuned to improve itsperformance, furthermore a new method for multi-labeled classification based on a complete distributionand an uncertainty measure is proposed. Experiments are done using 2 different data frameworks forclassification problems: multimedia data filtering and Reuters-21578 modapte as benchmark data-sets,the effectiveness of the method is compared in terms accuracy and micro and macro averaging F1- mea-sure.