2.3.2 Back Propagation (BP) Neural
Network
Ma et al. developed a lending behavior evaluation
system for online lending platforms, taking into
account the peculiarities of online lending and past
research (Ma et al., 2021). They then designed a BP
neural network using the BP algorithm (Ma et al.,
2021). A BP neural network, also known as a
multilayer neural network, consists of three or more
layers, which include many hidden layers in addition
to the input and output layers. The neural network
employs an error backpropagation method to
facilitate learning, enabling it to extract additional
information from input samples and effectively
perform intricate data processing tasks (Ma et al.,
2021). During the training process of a neural
network, the weights of the neurons are adjusted (Ma
et al., 2021). This adjustment occurs by propagating
the signals in the opposite direction of error reduction
(Ma et al., 2021). The weights of the connections
between neurons in each layer are modified starting
from the output layer and moving through the hidden
layer (Ma et al., 2021). The output values, which have
been altered through the process of backward
propagation of mistakes, are once again linked to the
input neurons as inputs for the subsequent
computation (Ma et al., 2021). During successive
iterations, the neural network's output value
progressively diminishes until it reaches a state of
stability (Ma et al., 2021). Upon comparing the
performance curves of neural networks utilizing the
Levenberg-Marquardt (LM) algorithm, Scaled
Conjugate Gradient, and Bayesian Regularization as
training functions, it is evident that the network model
based on the LM algorithm exhibits superior
performance (Ma et al., 2021). This is due to the fact
that both the number of iterations and the best test
error are smaller in comparison to the models based
on scaled conjugate gradient and Bayesian
regularization. Hence, the LM algorithm is selected
as the training technique for the neural network
model.
3 DISCUSSIONS
Traditional Machine Learning emphasizes the
selection of suitable algorithms to train models based
on the features of a given dataset, extending its
applicability to a broad spectrum of data predictions.
This domain encompasses a variety of algorithms,
including Linear Regression, Logistic Regression,
Decision Trees, Random Forests, Support Vector
Machines, among others. Each algorithm targets
specific tasks. For instance, Decision Trees are
particularly effective for classification purposes.
Machine Learning facilitates predictions and
informed decision-making on new, unseen data using
existing datasets, playing a crucial role in bolstering
the economic sector. However, the rapid
advancement of digitalization has highlighted the
limitations of traditional Machine Learning, raising
concerns among experts. One significant drawback is
its limited flexibility in processing high-dimensional
data, its generalization capabilities are constrained,
making it susceptible to issues like overfitting and
underfitting, which can result in inaccurate
predictions (Grieve, 2023). In complex environments,
such as large-scale lending scenarios, traditional
Machine Learning falls short in effectively analyzing
and predicting outcomes. In response to these
challenges, Deep Learning has emerged as an
advanced subset of Machine Learning, gaining
widespread adoption among experts. Deep Learning
utilizes sophisticated architectures, such as
Convolutional Neural Networks (CNNs) and
Recurrent Neural Networks (RNNs), to handle high
dimensional data sets more effectively (Grieve,
2023). By incorporating multi-layered neural
networks and complex nonlinear activation functions,
Deep Learning enhances the model's generalization
ability and prediction accuracy, thereby addressing
the limitations of traditional Machine Learning
techniques.
Various techniques have been investigated to
improve the usefulness, comprehensibility, and
confidentiality of machine learning models in order to
maximize the effectiveness of artificial intelligence.
An exemplary instance is the advancement of expert
systems, which are intricate information systems
engineered to emulate the decision-making
procedures of human experts in well-defined domains
(Waterman, 1985). These systems incorporate expert
knowledge and experiential learning to replicate the
decision-making processes of professionals. Machine
learning poses a considerable obstacle because to its
demanding computing requirements, necessitating
extensive time and top-of-the-line technology. Expert
systems, in contrast, provide a cost-effective solution
by diminishing the time and resources required for
machine learning activities. Furthermore, the
incorporation of expert systems into machine learning
can enhance the comprehensibility of models, thereby
aligning artificial intelligence more closely with
human knowledge systems (Malekipirbazari &
Aksakalli, 2015). In addition to expert systems, the
SHapley Additive exPlanations (SHAP) method has