Direct Mining of Discriminative Patterns for Classifying Uncertain Data

摘要

Classification is one of the most essential tasks in data mining. Unlike other methods, associative classification tries to find all the frequent patterns existing in the input categorical data satisfying a user-specified minimum support and/or other discrimination measures like minimum confidence or information-gain. Those patterns are used later either as rules for rule-based classifier or training features for support vector machine (SVM) classifier, after a feature selection procedure which usually tries to cover as many as the input instances with the most discriminative patterns in various manners. Several algorithms have also been proposed to mine the most discriminative patterns directly without costly feature selection. Previous empirical results show that associative classification could provide better classification accuracy over many datasets.Recently, many studies have been conducted on uncertain data, where fields of uncertain attributes no longer have certain values. Instead probability distribution functions are adopted to represent the possible values and their corresponding probabilities. The uncertainty is usually caused by noise, measurement limits, or other possible factors. Several algorithms have been proposed to solve the classification problem on uncertain data recently, for example by extending traditional rule-based classifier and decision tree to work on uncertain data. In this paper, we propose a novel algorithm uHARMONY which mines discriminative patterns directly and effectively from uncertain data as classification features/rules, to help train either SVM or rule-based classifier. Since patterns are discovered directly from the input database, feature selection usually taking a great amount of time could be avoided completely. Effective method for computation of expected confidence of the mined patterns used as the measurement of discrimination is also proposed. Empirical results show that using SVM classifier our algorithm uHARMONY outperforms the state-of-the-art uncertain data classification algorithms significantly with 4% to 10% improvements on average in accuracy on 30 categorical datasets under varying uncertain degree and uncertain attribute number.