Engineering HDR Seminar Series 5, 2012: Development of High-Dimensional ADTree Algorithm

Mr Sok Hong Kuan, PSCT postgraduate student

Date: 2012-04-13
Time: 10:30 to 11:30
Venue: Classroom 9-3-04


The alternating decision tree or ADTree, is a powerful machine learning algorithm that efficiently combines the decision tree algorithm with an accuracy enhancement technique known as boosting. Unfortunately, the original ADTree algorithm has several issues that limit its implementation. First, it only considers univariate tests. Thus, there is currently no multidimensional ADTree algorithm. Second, there is no existing framework on the selection of the optimal ADTree size. This makes it difficult to define and generate the optimal ADTree for a given classification problem. Third, the existing research on ADTree has not adequately addressed memory constraints and limitations faced by large scale domain problems. Finally, the current ADTree induction methodology uses a greedy approach due to the underlying boosting mechanism. This approach tends to grow suboptimal trees. This research aims to extend the existing ADTree model and develop a high-dimensional ADTree algorithm by addressing these limitations. The proposed research will generalize the existing univariate ADTree by developing multivariate decision tests and designing an omnivariate ADTree for optimal selection of these decision tests. Next, the overall ADTree complexity will be controlled through careful design of the stopping criteria and pruning mechanism. These developments are important because the ADTree complexity will influence the process of knowledge discovery and model validation. The research will go on to develop a scalable ADTree that accounts for memory limitations faced by large scale and high dimensional domain problems. This can be achieved by optimizing the computational complexity and implementing an efficient information extraction algorithm. The final stage of the research will address the problem with using the existing greedy approach by proposing an alternative tree growing approach using evolutionary algorithm. The success of this research will lead to a new class of ADTree that is high-dimensional, efficient, optimal and scalable. This will lead to wider usage of the ADTree algorithm in classification and knowledge discovery applications.

About the Speaker

Sok Hong Kuan received his BEng (Hons) degree with first class honors in Electrical and Computer Systems engineering from Monash University, Malaysia in 2010. He is currently pursuing his MEngSc (Research) with Monash University under the supervision of Dr Kuang Ye Chow and Dr Melanie Ooi Po-Leen. His research focus is on developing high-dimensional ADTree algorithm, a machine learning algorithm that combines decision tree and boosting methodology. His main research areas include machine learning, classification and statistical approximation.