Label Ranking Methods Based On Gaussian Mixture Model

Classification problem has been widely studied in machine learning community, while the significance and prevalence of multi-label data is gradually realized until the recent years. There are many practical applications where the objective is to learn an exact label preference of instance in form of a total order. For example, goods or services recommendation, search engine and gene selection. In contrast to conventional classification problem, label ranking is a more complicated learning problem. It extends the conventional classification in sense that it does not only predict a most likely candidate label but instead gives a total order of all class labels. More specifically, label ranking has a larger search space and more complicated evaluation of prediction than that in the classification setting.The main works in this thesis include three main aspects:Firstly, we review and analyse the currently existing label ranking algorithms, and provide a basic taxonomy of these approaches. Specifically, we divide these approaches into four categories, including reduction methods, proba-bilistic methods, similarity methods and other methods. Secondly, by using Pearson and Spearman correlation techniques, we analyse the relationships among seven widely used performance metrics for classification algorithms. In this paper, we divide these seven performance metrics into three groups by analyzing their definitions. Experimental results confirm the correctness of the classifi-cation, i.e., metrics inside the same group have high correlation, and metrics from different groups have low correlation. Additionally, we also investigate and analyse the difference of two widely-used performance metrics for evaluating the label ranking algorithms. Finally, we propose a novel label ranking method based on Gaussian mixture model, which can effectively relieve the cost of storage memory and prediction time of instance-based label ranking algorithms. Also, we compare our method with two state-of-the-art label ranking algorithms. The experimental results are quite promising and suggest that our approach is particularly strong in terms of predictive accuracy.