Abstract
The problem of learning from noisy data sets has been the focus of much attention for many years. Three different types of noise could be defined that generate difficulties in data classification. The first type is related to the noisy features and labels where data entry and data acquisition are inherently prone to errors. The second type is from the redundant features, which may confuse the classification algorithm and degrade the classification performance. The last type could be generated by insufficient features where some features may become quite ambiguous in the absence of related hidden complementary features. In order to address these problems, robust methods for data classification have been studied in many areas, such as bio-informatics, genetics, medicine, education and electronic engineering. This thesis aims to study classification methods that are robust for noisy data sets. Different problems caused by the three types of noise listed above are investigated. New robust methods for data classification are proposed.