Page 427 - Contributed Paper Session (CPS) - Volume 2

CPS1915 Han G. et al.
ranges from around 65% to 74% under all conditions for Model I and Model
II (see Figure 4).

Figure 4. Prediction accuracy for Model I and Model II when n_h = 10, 20, 30,
40, 50 and learning rate = (a) 2.0; (b) 0.5; (c) 0.1; (d) 0.05; I 0.01; (f) 0.005

4. Discussion and Conclusion
The neural network is a highly complex system with various parameters as
well as hyperparameter. In this research, neural network models with different
activation functions in hidden layer were discussed. The results showed that
the prediction accuracies for Model I and Model II with different hidden layer
neurons and learning rates mainly ranges from 60% to 70%. Model II had a
relatively stable prediction performance compared to Model I. Overall, the
prediction power between Model I and Model II is not quite obvious. In other
words, the comparison result is prone to be negative. However, the research
result can still provide some intuitions on how the neural network works. Due
to the limited sample number and input features, the performance of the
models we proposed was relatively poorer than the NN model proposed by
More et al. (2014). Although the models we proposed have quite promising
prediction power, there is still some future work for improvement such as
follows:
(1) The raw data was only split into training set and test set. The validation
dataset can provide an unbiased evaluation for the model to fit on the
training set while tuning the model’s hyperparameters. The split ratio
can be 60-20-20 or 70-15-15.
(2) Feature selection was not conducted in this research. Since the dataset
available from UCI Machine Learning Repository includes 20 input
features, while the raw dataset includes more than 100 features.

416 | I S I W S C 2 0 1 9