Mrs Zahra Mirzamomen is going to defense his Phd thesis on “To Learn Stable Decision Tree Based Classifiers for Data Streams” on Wednesday Sep. 07, 2016. The session will be held in Phd defense hall, Computer Engineering Department, Iran University of Science and Technology, Tehran, Iran.
In this thesis, we have investigated the instability issue in decision tree learning algorithms and the causes of it. Along with proposing a general abstract algorithm to induce more stable decision trees, we have also proposed detailed algorithms for both the static and the stream contexts. As there is no definition for the stability in the stream context, in this thesis, we have illustrated the working space by resolving the confusions in defining the stability in this context. Although several references have declared that there is strong instability in the decision tree learning algorithms in the static context, but this issue is not investigated for the incremental learning algorithms in the stream context. In this thesis, we have illustrated the presence of the instability issue in the incremental decision tree learning algorithms, theoretically and experimentally. To improve structural stability, i. e. to minimize the sensitivity of the decision tree structure to the training instances in both the static and the stream contexts, had been our focus in this thesis.
The key solution of this thesis for improving the structural stability of decision trees, is to use non-monolithic split tests based on multiple attributes, designed with the aim of eliminating the competition between the attributes with close merits, localizing the effect of the training instances on the split test and, making the split test trainable. In this thesis, we have proposed a high-level algorithm to induce decision trees by applying such split tests and based on it, we have proposed detailed algorithms for both the static and the stream contexts, in which fuzzy min-max neural networks are employed as the split tests, in a way that provides the desired attributes.
The proposed models, not only provide more structural stability in comparison with available decision trees, but also create smaller and shallower models, because of non-linearly splitting the feature space at the internal nodes. Theoretical analysis and experimental evidence show that the structural stability is improved in the proposed algorithms and in the meanwhile, they present comparable precision and efficiency.
Phd cadidate:Zahra Mirzamomen
Supervisor: : Dr. Morteza Kangavari
Jury Committee:Dr. Naser Mozayani, Dr. Behrooz Minaei, Dr.
Dr.Hamid Beigi, Dr. Mirmohsen Pedram
Time: 13:00 AM, Wednesday Sep. 07, 2016, Phd defense hall, Computer Engineering Department, Iran University of Science and Technology, Tehran, Iran