Summary: | 碩士 === 國立臺灣大學 === 工業工程學研究所 === 90 === In the data mining field, constructing a predictive model from a large volume of data is important for the classification problem. Neural networks and decision trees are the two main tools to solve this problem. Although the predictive accuracy of the neural network is better than that of the traditional statistical methods, such as the Bayesian classifier and the discriminant analysis, and the neural network has better tolerance to noises than the decision tree. The knowledge learned by a trained neural network is encoded in both the neural network architecture and the weights and biases of it. Hardly could we understand the decision process of the neural network due to the complex nonlinear mapping of the data, which is the mechanism of operation of the neural network. This thesis presents two different rule extraction algorithms to extract comprehensible rules from supervised learning feedforward neural networks to solve the above black-box problem.
We first use decomposional approach and based on the concept of treating a neuron in the neural network as a “Boolean neuron” that has only two states to develop the BAB-BB rule extraction algorithm, which can extract Boolean functions (rules) from multilayer feedforward neural networks with binary or bipolar inputs. XOR problem shows that this algorithm is practicable. Next, according to the concept of discretizing continuous hidden neuron activation values, we develop the BAB-G rule extraction algorithm, which can extract if-then classification rules from three-layer feedforward neural networks with discrete, continuous, or mixed inputs. The antecedent parts of the if-then rules obtained from this algorithm are slanting hyperplanes. These hyperplanes are formed by the linear combinations of the inputs of the neural networks. During the rule extraction procedure, redundant hidden neurons can be removed without affecting the functionality of the neural networks. Some empirical results on the data sets from the UCI repository of machine learning database are given for comparing our rule extraction algorithm and C5.0 decision tree algorithm. For these datasets, statistical hypothesis tests show that the rules obtained from our algorithm achieve the same classification accuracy as the neural networks. Moreover, our rules are better than the C5.0 decision tree both on comprehensibility and on accuracy for these datasets.
|