Performance Evaluation of Decision Tree Graph-Based Induction

Abstract

A machine learning technique called Decision tree Graph-Based Induction (DT-GBI) constructs a classifier (decision tree) for graph-structured data, which are usually not explicitly expressed with attribute-value pairs. Substructures (patterns) are extracted at each node of a decision tree by stepwise pair expansion (pairwise chunking) in GBI and they are used as attributes for testing. DT-GBI is efficient since GBI is used to extract patterns by greedy search and the obtained result (decision tree) is easy to understand. However, experiments against a DNA dataset from UCI repository revealed that the predictive accuracy of the classifier constructed by DT-GBI was not high enough compared with other approaches. Improvement is made on its predictive accuracy and the performance evaluation of the improved DT-GBI is reported against the DNA dataset. The predictive accuracy of a decision tree is affected by which attributes (patterns) are used and how it is constructed. To extract good enough discriminative patterns, search capability is enhanced by incorporating a beam search into the pairwise chunking within the greedy search framework. Pessimistic pruning is incorporated to avoid overfitting to the training data. Experiments using a DNA dataset were conducted to see the effect of the beam width, the number of chunking at each node of a decision tree, and the pruning. The results indicate that DT-GBI that does not use any prior domain knowledge can construct a decision tree that is comparable to other classifiers constructed using the domain knowledge.