Institution: University of Technology Sydney
Supervisor: Longbing Cao

Research Question: 

• How to improve the accuracy and workability of SVM-based classification algorithms on uncertain data?

Research Motivation:

• The standard SVM were originally designed for binary classifications. However, many real-world problems fall into multi-class classifications, which are usually converted into binary classifications.
• Outlier detection has attracted increasing attention in machine learning and data mining areas due to its wide-ranging applications from machine fault detection, credit card fraud detection, network intrusion to medical diagnosis.  

High-level Research Design:

The first stage of this research will be focusing on the design of a more efficient and accurate on-line SVM-based multi-class classification algorithm. The algorithm will improve the overall accuracy of the multi-class problems by adopting a novel framework, called multi-space-mapping (MSM) with SVM, which maps the data set with all classes into different feature spaces in terms of hierarchical tree architecture.

The second stage of this research will be focusing on the design of a novel and robust on-line support vector data description (SVDD) approach for outlier detection by introducing a fuzzy membership to each input data into the SVDD training phase. The so-called Robust SVDD will explicitly handle the uncertainty of the input data and enhance the ability of SVDD in reducing the effect of noise, therefore providing a better tradeoff between the detection rate and false alarm rate for outlier detection.

Finally, the proposed multi-class classification algorithm and robust SVDD method will be tested in real-life data, in particular, the health insurance claim data to improve outlier detection.