Effective techniques for handling incomplete data using decision trees

Abstract

Decision Trees (DTs) have been recognized as one of the most successful formalisms for knowledge representation and reasoning and are currently applied to a variety of data mining or knowledge discovery applications, particularly for classification problems. There are several efficient methods to learn a DT from data. However, these methods are often limited to the assumption that data are complete. In this thesis, some contributions to the field of machine learning and statistics that solve the problem of extracting DTs for learning and classification tasks from incomplete databases are presented. The methodology underlying the thesis blends together well-established statistical theories with the most advanced techniques for machine learning and automated reasoning with uncertainty. The first contribution is the extensive simulations which study the impact of missing data on predictive accuracy of existing DTs which can cope with missing values, when missing values are in both the training and test sets or when they are in either of the two sets. All simulations are performed under missing completely at random, missing at random and informatively missing mechanisms and for different missing data patterns and proportions. The proposal of a simple, novel, yet effective proposed procedure for training and testing using decision trees in the presence of missing data is the next contribution. Original and simple splitting criteria for attribute selection in tree building are put forward. The proposed technique is evaluated and validated in empirical tests over many real world application domains. In this work, the proposed algorithm maintains (sometimes exceeds) the outstanding accuracy of multiple imputation, especially on datasets containing mixed attributes and purely nominal attributes. Also, the proposed algorithm greatly improves in accuracy for IM data. Another major advantage of this method over multiple imputation is the important saving in computational resources due to it simplicity. The next contribution is the proposal of three versions of simple probabilistic techniques that could be used for classifying incomplete vectors using decision trees based on complete data. The proposed procedure is superficially similar to that of fractional cases but more effective. The experimental results demonstrate that these approaches can achieve comparative quality to sophisticated algorithms like multiple imputation and therefore are applicable to all kinds of datasets. Finally, novel uses of two proposed ensemble procedures for handling incomplete training and test data are proposed and discussed. The algorithms combine the two best approaches either with resampling (REMIMIA) or without resampling (EMIMIA) of the training data before growing the decision trees. Experiments are used to evaluate and validate the success of the proposed ensemble methods with respect to individual missing data techniques in the form of empirical tests. EMIMIA attains the highest overall level of prediction accuracy.