Page 422 - Contributed Paper Session (CPS) - Volume 2
P. 422
CPS1915 Han G. et al.
semi-automatic extraction of key information from raw data (Written & Frank,
2005; Moro et al.,2014). Kukar et al. (2010) proposed an integrative approach
of data mining and decision support: data mining based decision support
system, which facilitate the users to make more beneficial decisions.
In Moro et al.’s (2014) research, they compared four DM models, namely,
logistic regression (LR), decision tree (DT), artificial neural network (ANN) as
well as support vector machine (SVM), and the results showed that ANN
outperformed other models. However, they failed to specify the activation
function used in their ANN model. In this research, we applied ANN models
with two different activation functions in the hidden layer, namely, sigmoid
function and hyperbolic tangent (tanh) function, to predict and compare the
bank client’s behaviour, i.e. whether the client will subscribe a term deposit or
not.
The rest of the paper is organized as follows: Section 2 presents the
methodology (ANN) and dataset (bank telemarketing data) used in this
research, and Section 2.1 and 2.2 discuss the sigmoid and tanh functions,
respectively; Section 3 describes the results from two ANN models with two
different activation functions; a brief discussion and conclusion are drawn in
Section 4.
2. Methodology
ANN is a highly popular technique in machine learning area, such as image
recognition (Sun et al., 2005) and natural language processing (Petitto et al.,
2000). The idea of ANN was inspired by the biological neural networks (Jain et
al. 1996). Artificial and biological NN share some similarities, especially the
transmission mechanism (see Figure 1).
Figure 1. Biological (left) and artificial (right) neural networks
Building a NN model contains a training and testing procedure. During the
training process, it is vital to select the initial hyperparameter values and keep
adjusting them in order to obtain a reliable and robust model. In the testing
procedure, the model performance should be evaluated using some indices,
such as confusion matrix. The entire model-building process can be divided
into five major steps:
(1) Defining the neural network structure;
411 | I S I W S C 2 0 1 9