Filters (Mathematics)

Machine learning techniques allow useful insight to be distilled from the increasingly massive repositories of data being stored. As these data mining techniques can only learn patterns actually present in the data, it is important that the desired knowledge be faithfully and discernibly contained therein. Two common data quality issues that often affect important real life classification applications are class noise and class imbalance. Class noise, where dependent attribute values are recorded erroneously, misleads a classifier and reduces predictive performance. Class imbalance occurs when one class represents only a small portion of the examples in a dataset, and, in such cases, classifiers often display poor accuracy on the minority class. The reduction in classification performance becomes even worse when the two issues occur simultaneously. To address the magnified difficulty caused by this interaction, this dissertation performs thorough empirical investigations of several techniques for dealing with class noise and imbalanced data. Comprehensive experiments are performed to assess the effects of the classification techniques on classifier performance, as well as how the level of class imbalance, level of class noise, and distribution of class noise among the classes affects results. An empirical analysis of classifier based noise detection efficiency appears first. Subsequently, an intelligent data sampling technique, based on noise detection, is proposed and tested. Several hybrid classifier ensemble techniques for addressing class noise and imbalance are introduced. Finally, a detailed empirical investigation of classification filtering is performed to determine best practices.

Possibly the largest problem when working in bioinformatics is the large amount of data to sift through to find useful information. This thesis shows that the use of feature selection (a method of removing irrelevant and redundant information from the dataset) is a useful and even necessary technique to use in these large datasets. This thesis also presents a new method in comparing classes to each other through the use of their features. It also provides a thorough analysis of the use of various feature selection techniques and classifier in different scenarios from bioinformatics. Overall, this thesis shows the importance of the use of feature selection in bioinformatics.

One of the greatest challenges to data mining is erroneous or noisy data. Several studies have noted the weak performance of classification models trained from low quality data. This dissertation shows that low quality data can also impact the effectiveness of feature selection, and considers the effect of class noise on various feature ranking techniques. It presents a novel approach to feature ranking based on ensemble learning and assesses these ensemble feature selection techniques in terms of their robustness to class noise. It presents a noise-based stability analysis that measures the degree of agreement between a feature ranking techniques output on a clean dataset versus its outputs on the same dataset but corrupted with different combinations of noise level and noise distribution. It then considers classification performances from models built with a subset of the original features obtained after applying feature ranking techniques on noisy data. It proposes the focused ensemble feature ranking as a noise-tolerant approach to feature selection and compares focused ensembles with general ensembles in terms of the ability of the selected features to withstand the impact of class noise when used to build classification models. Finally, it explores three approaches for addressing the combined problem of high dimensionality and class imbalance. Collectively, this research shows the importance of considering class noise when performing feature selection.

Collaborative filtering (CF), a very successful recommender system, is one of the applications of data mining for incomplete data. The main objective of CF is to make accurate recommendations from highly sparse user rating data. My contributions to this research topic include proposing the frameworks of imputation-boosted collaborative filtering (IBCF) and imputed neighborhood based collaborative filtering (INCF). We also proposed a model-based CF technique, TAN-ELR CF, and two hybrid CF algorithms, sequential mixture CF and joint mixture CF. Empirical results show that our proposed CF algorithms have very good predictive performances. In the investigation of applying imputation techniques in mining incomplete data, we proposed imputation-helped classifiers, and VCI predictors (voting on classifications from imputed learning sets), both of which resulted in significant improvement in classification performance for incomplete data over conventional machine learned classifiers, including kNN, neural network, one rule, decision table, SVM, logistic regression, decision tree (C4.5), random forest, and decision list (PART), and the well known Bagging predictors. The main imputation techniques involved in these algorithms include EM (expectation maximization) and BMI (Bayesian multiple imputation).