Learning with fuzzy labels : a random set approach towards intelligent data mining systems

Traditionally machine learning and data mining research has focused on learning algorithms with high classification or prediction accuracy. This is not always sufficient for some applications that require good algorithm transparency. In this thesis, we apply a random set based high-level knowledge representation language to some well-known data mining algorithms in order to build transparent models. This new framework is referred to as Label Semantics, which captures the idea of computation on linguistic expressions rather than numerical quantities. Based on this framework, we proposed linguistic decision tree, fuzzy Bayesian estimation tree and linguistic FOIL algorithms. Empirical experiments on real-world data sets suggest that our algorithms have better or equivalent classification (or prediction) accuracy and much better transparency compared to other well known machine learning algorithms such as C4.5, naive Bayes, neural networks and support vector machines.