Novel approaches for hierarchical classification with case studies in protein function prediction

A very large amount of research in the data mining, machine learning, statistical pattern recognition and related research communities has focused on flat classification problems. However, many problems in the real world such as hierarchical protein function prediction have their classes naturally organised into hierarchies. The task of hierarchical classification, however, needs to be better defined as researchers into one application domain are often unaware of similar efforts developed in other research areas.

The first contribution of this thesis is to survey the task of hierarchical classification across different application domains and present an unifying framework for the task. After clearly defining the problem, we explore novel approaches to the task.

Based on the understanding gained by surveying the task of hierarchical classification, there are three major approaches to deal with hierarchical classification problems. The first approach is to use one of the many existing flat classification algorithms to predict only the leaf classes in the hierarchy. Note that, in the training phase, this approach completely ignores the hierarchical class relationships, i.e. the parent-child and sibling class relationships, but in the testing phase the ancestral classes of an instance can be inferred from its predicted leaf classes. The second approach is to build a set of local models, by training one flat classification algorithm for each local view of the hierarchy. The two main variations of this approach are: (a) training a local flat multi-class classifier at each non-leaf class node, where each classifier discriminates among the child classes of its associated class; or (b) training a local fiat binary classifier at each node of the class hierarchy, where each classifier predicts whether or not a new instance has the classifier’s associated class. In both these variations, in the testing phase a procedure is used to combine the predictions of the set of local classifiers in a coherent way, avoiding inconsistent predictions. The third approach is to use a global-model hierarchical classification algorithm, which builds one single classification model by taking into account all the hierarchical class relationships in the training phase. In the context of this categorization of hierarchical classification approaches, the other contributions of this thesis are as follows.

The second contribution of this thesis is a novel algorithm which is based on the local classifier per parent node approach. The novel algorithm is the selective representation approach that automatically selects the best protein representation to use at each non-leaf class node.

The third contribution is a global-model hierarchical classification extension of the well known naive Bayes algorithm. Given the good predictive performance of the global-model hierarchical-classification naive Bayes algorithm, we relax the Naive Bayes’ assumption that attributes are independent from each other given the class by using the concept of k dependencies. Hence, we extend the flat classification /¿-Dependence Bayesian network classifier to the task of hierarchical classification, which is the fourth contribution of this thesis.

Both the proposed global-model hierarchical classification Naive Bayes and the proposed global-model hierarchical /¿-Dependence Bayesian network classifier have achieved predictive accuracies that were, overall, significantly higher than the predictive accuracies obtained by their corresponding local hierarchical classification versions, across a number of datasets for the task of hierarchical protein function prediction.