Rule extraction from multi-layer perceptron neural network using decision tree for currency exchange rates forecasting
Neural network can be used in acquiring hidden knowledge in datasets. However, knowledge acquired by neural network was presented in its topology, the weights on the connections and by the activation functions of the hidden and output nodes. These representations are not easily understandable since...
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English English |
| Published: |
2015
|
| Subjects: | |
| Online Access: | http://umpir.ump.edu.my/id/eprint/12627/ http://umpir.ump.edu.my/id/eprint/12627/ http://umpir.ump.edu.my/id/eprint/12627/1/FSKKP%20-%20SOLEH%20ARDIANSYAH%20-%20CD%209677.pdf http://umpir.ump.edu.my/id/eprint/12627/2/FSKKP%20-%20SOLEH%20ARDIANSYAH%20-%20CD%209677%20-%20CHAP%201.pdf http://umpir.ump.edu.my/id/eprint/12627/3/FSKKP%20-%20SOLEH%20ARDIANSYAH%20-%20CD%209677%20-%20CHAP%203.pdf |
| Summary: | Neural network can be used in acquiring hidden knowledge in datasets. However, knowledge acquired by neural network was presented in its topology, the weights on the connections and by the activation functions of the hidden and output nodes. These representations are not easily understandable since neural networks act as a black box. The black box problem can be solved by extracting rule from trained neural network. Thus, the aim of this study was to extract valuable information (rule) from trained multi-layer perceptron (MLP) neural networks using decision tree. The main process in extracting rules from MLP using decision tree for currency exchange rate forecasting can be divided into two stages. In the first stage, the MLP network was built based on the parameter that was defined in the previous chapter. We also perform training and testing process experimentally and then the performance was evaluated in order to obtain the network with the best performance. The MLP network which provides the best prediction performance will be extracted by decision tree in the second stage by mapping input-output of the network directly. The forecasting result have shown that MLP network of EUR/USD produced a significant results compared to MLP network of GBP/USD and USD/JPY in term of MSE, RMSE, MAPE, and DS. It is quite evident that as the number of hidden neurons increases, MSE and MAPE decrease. In addition, the number of iterations for each model continues to increase along with the increasing number of hidden neurons. The results on decision tree induction show that C4.5 algorithm induction produced a significant result in term of accuracy 84.07% - 86.34%, precision and recall 93.17% and 81.97% respectively. This study has shown how rule can be extracted from MLP network by decision tree without making any assumptions about the networks activations function or having initial knowledge about the problem domain. The extracted rule can be used to explain the process of the neural network systems and also can be used in other systems like expert systems. |
|---|