Application of Neural Network for the Enhancement of Digital Marketing in Banking Sector

Application of Neural Network for the Enhancement of Digital Marketing in Banking Sector

^1*EZUGWU Lilian Martina, ¹OZIOKO Frank Ekene, ²MBA Chioma Juliet

¹Department of Computer Sciences, Enugu State College of Education, Enugu, Nigeria

¹Department of Computer Sciences; Enugu State University of Science and Technology (ESUT), Enugu, Nigeria

²Department of Computer Sciences, Enugu State Polytechnique, Iwolo, Enugu, Nigeria

DOI: https://doi.org/10.51584/IJRIAS.2025.10060073

Received: 27 May 2025; Accepted: 09 June 2025; Published: 11 July 2025

ABSTRACT

This study presents the design, implementation and evaluation of a Feedforward Neural Network (FFNN) aimed at enhancing digital marketing strategies in the banking sector by accurately predicting customer subscription behaviour. The artificial neuron model was defined using weighted inputs, bias parameters and a sigmoid activation function, forming the foundation for a multi-layer FFNN architecture optimized for learning efficiency. In order to address the issues of overfitting and improve generalization, a regularization technique which includes the traditional dropout algorithm were incorporated into the proposed model. The network was trained using the gradient descent back-propagation method for effective weight adjustment and big data analytics were employed to uncover customer financial patterns across quarterly periods, revealing the fourth quarter as the most favourable for product marketing due to higher account balances. A correlation matrix analysis showed weak linear relationships between input features and the target variable, validating the use of a non-linear model like FFNN. Experimental results demonstrated the model’s high predictive accuracy, with low error metrics (MAE: 0.003, MSE: 0.0013, RMSE: 0.000492), confirming the robustness and generalization ability of the trained network.

Keywords: Digital Marketing; Customer Behaviour; Feed forward Neural Network; Banking Sector

INTRODUCTION

This 21^st century presents an era where digital marketing has evolved into a highly sophisticated and data-driven field where business organizations leverage on a variety of online platforms and channels to engage targeted customers (Turkmen, 2021; Yufeng, 2022). The increased volume of information generated by customers across different business domains presents both challenges and opportunities for marketers. For instance, the banking sector has increasingly embraced digital marketing as a tool to enhance the engagement of customers and marketing of goods and services (Sergio et al., 2013).All over the world, commercial banks generate huge volumes of data from customer information, banking behaviour, and customer interactions, thus making big data a valuable asset for identifying patterns in specific customer behaviours to enhance customer personalization and improve customer services (Vafeiadis et al., 2015; Salman, 2020a).

Big data is defined as an extremely huge volume of datasets that are difficult to process using the traditional data processing approaches (Muneeb, 2018). In recent times, big data has continued to play a crucial role in helping the banking sector optimize digital marketing. This is achieved using data analytical tools to collect, process, and analyze information from multiple sources. This exact information is mined to gain actionable insights about the behavior and preferences of a targeted audience (Namulia, 2011).

In the context of the banking sector, due to the dynamic behavior of bank customers, big data can be applied to help banks identify the right customer for a particular product, thus allowing for more precise and personalized marketing (Salman, 2020b). However, the complexity in customer characterization, the changing business landscape, and customer preference has made the traditional big data analytical platforms struggle to correctly identify the right customers for a particular product, hence remaining a major issue in the scientific community.

Machine learning algorithms have continued to dominate approaches for the development of customer segmentation models (Sochima et al., 2025; Harbor et al., 2021). This is due to the ability of the algorithms to learn patterns from big data and then make correct decisions. For instance, in Premkumar et al. (2021), multiple MLAs such as Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), Gaussian Naïve Bayes (GNB), and Logistic Regression (LOR) were trained using data collected from Portuguese banks to generate a customer segmentation model for digital marketing. Comparative analysis revealed that LR was the best with accuracy of 92.48%, precision of 70.58%, recall of 54.45, and F1score of 61.53%. In another study by Gu et al. (2020), after comparing Convolutional Neural Network (CNN), Deep Neural Network (DNN), and proposed multiple-filter Convolutional Neural Network (XmCNN), the proposed model reported better prediction performance with accuracy of 0.9018, recall of 0.8664, precision of 0.8366, and F1-score of 0.8502. Similarly, Asif (2018), Elzhan and Yazici (2018), and Fakhri (2022) all took a position with the RF digital marketing prediction model after comparison with other trained MLAs, while Turkmen (2022), who compared Multi-Layered Neural Networks (MLP) and Bayesian Neural Networks (BNN) after training with customer data, reported BNN as the better model.

To this end, this study proposes optimizing digital marketing campaigns using advanced big data analysis and digital marketing techniques. The contribution of the work will be to apply recent and innovate big data analytical frameworks for data analysis and identify patterns in customer behaviour, which will help improve digital marketing. To ensure the reliability of the model when deployed in the field, an algorithm for contact time will be developed and collectively integrated to build the digital marketing software for improved customer service in the banking industry.

RESEARCH METHODOLOGY

The methodology used for this work is Hybrid Metaheuristic-Based Optimization approach. the hybrid combined both the gradient descent optimization and smart regularization approaches to facilitate advance digital marketing. In realizing the methodology, first was to characterize big data for digital marketing campaigns by data collection. Then a big data analytical framework model for improved digital marketing campaign was applied using exploratory data analysis and Python programming language. The new digital marketing software developed was tested and validated experimentally.

The Proposed Big Data Analytical Framework

This section discussed the framework forthe big data analysis. It consists of several components which are big data, big data analysis, exploratory data analysis as shown in Figure 1.

Figure 1:Block Diagram of the Big Data Analytical Framework

Figure 1 presents the block diagram of the of data analytical framework with the several interconnected frameworks. First the big data of customers will be collected, the data analytical steps will be applied such as visualization to provide a physical examination of the data, then imputation will be applied to address possible missing values, outliers and also duplicate values. Normalization is then applied to scale the data into compatible feature sizes, before transformation into a compact feature vector. Exploratory data analysis will be applied to analyze the data and identify patterns such as state of account, demographic, age, etc. The results from the analysis will be presented and then interpreted to make findings.

Data Collection

Dataset used for this work is the bank customer dataset from Kaggle repository with data link [https://www.kaggle.com/datasets/garimam/bank-customer-dataset]. The data contained records of digital marketing campaign carried out from August 2022 to August 2024. The data was obtained through series of phone calls by the customer service department to inform customers about bank products like loans, bonds and then their responses were recorded. The data contained 18 attributes which model bank customer behaviour and were reported in Table 1. The sample size of data collected is 45214 features, from 100 customers. The data attributes characterization was reported in Table 1.

Table 1: Data Characterization of Customer Campaign Records without Contact Time

Table 1 presents the current customer campaign data characterization. The data contain 18 features which model customer behaviour. However, Katamade et al. (2019), Chi and Gang(2022) argued that while this data model behaviour of customers, the reason why it still produced inefficient result in real world test scenario is due to lack of contact time. In the context of this work, contact time is the right time to meet customers, and that right time is determined by economic factors like account statement. To include this variable in the current dataset, a secondary data was collected from the same source. The data contains 7 attributes which model customer account behaviour every quarter of the year.The data contained records of 100 customer’s bank statement from 2022 to 2024.

Exploratory Analysis on the Big Data Using Statistical Descriptive Technique

This section presents the data analysis techniques used to visualize the relationships between variables, identify patterns and draws conclusions. The method used for the big data analysis is exploratory approach while the technique is descriptive analysis. This technique helps in understanding the structure of the dataset, detecting anomalies, and gaining insights into key variables to model customer behaviour for improved digital marketing results. It helps identify customer behaviour, seasonal variations in account balances, and the best contact time based on financial patterns, which aid in strategic decision-making. Figure 2 presents the flow chart for the big data analysis.The diagram began with the loading of the dataset, and then the features are identified and encoded for exploratory analysis, applying different plots such as quarterly balance against customer ID, contact time against customer ID, account type against customer ID, and feature correlations matrix. The results of the data analysis process were reported in the Chapter Five.

Figure 2: Flow Chart for the Big Data Analysis Process

Development of an Improved Digital Marketing Framework Using Machine Learning Algorithm

This section discussed the improved digital marketing model using machine learning algorithm. This was achieved by adopting neural network algorithm and proposing new regularization and optimization algorithm. The neural network type used is Feed Forward Neural Network (FFNN). The FFNN is a model that is inspired by the neurobiological brain shown in Figure 3 (Kekong et al., 2019).

Figure 3: Structure of a Biological Neuron (Source: Openstax, 2018).

From Figure 3, the structure of neurons from the human brain indicates that each neuron has the ability to propagate signals through the neurons, to the rest of the brain. The neuron transmits signals to other neuronswhen the overall potential of the cell body rises above a certain threshold. The individual input to the neuron through the dendrites adjusts the potential of the cell which enables the propagation of the signal. The inputs are groupedin a way that the potential of some cells increaseswhile the potential of others decreases. The FFNN is an interconnected massive parallel computational models, units or nodes, whose functionality mimic the animal neural network in order to process information from the input to the output using the connection strength (weight) obtained by adaptation or learning from a set of training patterns. The architecture of a single-layer network with input  and output  is shown in Figure 4.

Figure 4: Basic Structure of an Artificial Neuron (a) Without Bias (b) With Bias

The mathematical description of ANN process is shown in Figure 5

Figure 5: Mathematical Model Description of an Artificial Neuron

The neuron is a unit of computation that reads the inputs given, processes the input and gives the output in processed form. To get the output of the Artificial Neuron from the activation function, we compute the weighted sum of the inputs as (OduahandOlofin, 2023);

                       (1)

Equation 1: for the weighted sum of the inputs

In Equation 1, x_i is the neuron’s input, w_ki is thecorresponding weightto the input x_i.

The neuron’s output is obtained by sending the weighted sum v_k as the activation function φ input that resolves the output of the specific neuron. y_k=φ(v_k). A step function with threshold t can be used to express a simple activation as shown in Equation 2;

                     (2)

Equation 2: A simple threshold activation function

Any two neurons are connected by a link that has a weight which represents the connection strength between the two neurons. Let w_ij^l denotes the weight for a link between unit j in layer l and unit i in layer l+1. Also b_i^(l )represents the bias of the unit i in layer l+1. For any neural network, the associated parameters inside it are expressed as a function of the weight and the bias of the neurons as shown in Equation 3 (Oduah and Ol of in, 2023);

                                                                        (3)

Equation 3: A function of the weight and the bias of the neurons

The components of equation 3 can be written in the form of w^1∈R^(3×3) and w^1∈R^(1×3).

Let the activation of unit i in layer lbe represented by a_i^l, then the input forthe layer labelled as L_1 we have a_i^1=x_i for the ith input of the whole network. Other layers are given by a_i^l=f(z_i^l ), where z_i^l is the total weighted sum of the inputs to unit i in layer l in addition to the bias term. The activation function of a three-layered FFNN with bias can be computed as Equation 4 (Ogbeta and Luois, 2023);

                                         (4)

Equation 4: the activation function of a three-layered FFNN with bias

Where h_(w,b) (x) is a real number representing the output of the FFNN and the activation function is donated as f(∙) and represented with Equation 5 as a sigmoid function. Figure 4.6 presents the neural network architecture with three hidden layers and a binary classification output which according to the target must be 0 or 1, the algorithm is also presented.

                                                                                                          (5)

Equation 5: A sigmoid function

Figure 6: Architecture of the FFNN with Three Hidden Layers

Proposed Smart Regularization and Optimization Techniques for the FFNN Training

Regularization techniques were used in the training process to address the issues of over fitting of neurons and generalization of weights, while the optimization techniques is proposed to improve the learning process. In the traditional sense, dropout technique is one of the popular regularization approach used today. The approach monitors neurons and then randomly switches off neurons through probability vector of 0 and 1 to ensure generalization during the training process (Kwubeghari et al., 2024). The Algorithm 1 presents the dropout techniques, However, Salehin and Kang (2021) argued that this approach will indirectly introduce bias in the neurons and may not provide expected generalized models. In addition,Srivastava et al.,(2014) revealed that while dropout is targeted towards addressing the issues of over-fitting, its lack of adaptivity in the neuron selection affects the balance between interdependencies of neurons, learning and prevention of over-fitting. To solve this problem, we propose a strategic approach to neuron selection through smart dropout approach.

Algorithm 1: Dropout algorithm (Kwubeghari et al., 2024)

1. Start
2. Input activation values
3. Parameter initialization (weight, bias, drop rate )
4. Generate probability vector (p) for neurons layers as 0 and 1
5. Choose probability vector as 0 and 1
6. Apply dropout of neurons using probability vector of neurons set to 0
7. Forward propagation with dropout
8. Compute the next layer activated value
9. Compute weight and bias sum of dropout
10. Apply nonlinearity with
11. Repeat for each layer
12. Update neural network training
13. End

In the traditional method of training neural network, popular algorithm like gradients descent back-propagation have been applied over the years to optimize neurons.However, Ogbeta and Louis (2023) argued that the traditional algorithm have fixed learning rate, fixed regularization and robust to noise gradient. The traditional optimization algorithm which used gradient descent technique is presented as algorithm 2.

Algorithm 2: Traditional gradient descent back-propagation [Algorithm 2]

1. Start
2. Initialize weights and bias of neurons
3. Forward pass %% compute activation of neurons
4. Compute loss %% Calculate error rate
5. Back pass %% Compute gradient loss
6. Update weight by applying gradient loss
7. Repeat process
8. Upon meeting stopping criteria
9. End training
10. Stop

The traditional gradient descent back-propagationcomputes the error between predicted and actual outputs and propagates this error backward through the network using the chain rule. Gradients are computed for each layer, and weights are updated using gradient descent, where a fixed learning rate determines the step size. This process continues until the model reaches a local minimum of the loss function.

RESULTS OF BIG DATA ANALYSIS

In the big data analysis, various key attributes were analysed to identify their relationships and how they model customer behaviour by providing insights through feature correlations. For instance, the relationships account balances of customers at different quarters of the day were analysed in Figure 7. From the results, the first and third quarter of the year was recorded to have low account balance for customers, what this mean is that when these customers are visited during this period for marketing a produce, they will likely not subscribe to the product because of economic/financial reasons. The result also revealed that customers recorded the best account balance during the fourth quarter of the years, while the second quarter also recorded high account balance then the first and third quarter. What this implied is that customers when visited during this period will likely subscribe to the product, which is a good revelation and will guide banking sector on the right time of contact to improve digital marketing.

Figure 7: Quarterly Balance Across Customers

Figure 7 presents the quarterly balance across customers showing the quarter customers are mostly financially stable. Figure 8 analyzes the customer categories according to saving and current account, to help identify the best time to meet them.

Figure 8: Quarter Balance Distribution According to Account Type

Figure 8 presents the quarterly balance distribution according to account types. From the results, it was observed that customers of saving and current account types both display similar behavior in their account balance, as both have more balance during the fourth quarter of the years, thus making it the most suitable period to meet them for marketing banking product. Figure 9 presents the correlations matrix heat map. The heat map was used to analyze how the different features in the dataset correlated with each other and how they collectively combined to meet the target.

Figure 9: Correlation Matrix of the Dataset

Figure 9 has demonstrated that across the diagonal matrix, the features perfectly correlated with each other, implying that their arrangement in the dataset is good. Secondly, it was observed that the most features in the dataset exhibit weak correlations, indicating minimal linear relationships between them. The quarterly balances (Q1, Q2, Q3, Q4) and the target variable show near-zero correlation (-0.06 to 0.03), suggesting that a customer’s balance does not significantly influence their suitability for a bank product. Similarly, age and salary are unexpectedly not correlated (~0.01), implying that salary progression is not strongly tied to age in this dataset. Additionally, salary and account balance show a slightly negative correlation (-0.07), which may indicate high expenditures or diverse financial habits among customers. The low correlation between personal and housing loans (0.02) suggests that these financial decisions are independent. Given that the target variable lacks strong correlation with any individual feature, simple linear relationships may not be sufficient to predict digital marketing. This was why neural network was trained and applied for the digital marketing process in the next section.

Result of Neural Network Experimental Training

This section discussed the result of the neural network training process, while considering the AGOA and the traditional back-propagation algorithm with normal dropout.  The performance evaluation was carried out with RMSE, MSE and MAE. The MAE result with back-propagation was reported in Figure 10, considering training and validation set.

Figure 10: MAE with Back-propagation

The MAE measured the absolute difference between the actual and predicted values. From the results, the MAE is for the training and validation is 0.003 which is very good as it is acceptable error and implied that the absolute difference in prediction success of customers which will likely subscribe to the banking product, when compared to the actual customers that subscribe is marginal and tolerable. To measure the MSE, which is the actual difference between true customers and predicted customer attributes is reported in Figure 11, while Figure12 recorded the RMSE result.

Figure 11: The MSE of the Digital Marketing Model with Back-propagation

Figure 12: The RMSE Result of the Digital Marketing Model with Back-propagation

The MSE in Figure 11, measures the deviation from the true and actual predicted values during the training process. From the results, it was observed that averagely the MSE for the training which recorded 0.00176and validation set recorded 0.0013 are both tolerable and implied minimal error, which suggested the model was able to predict the true value with high success rate.

Figure 12 presents the RMSE, which is the square root of MSE and provides an error metric in the same unit as the original data, making it more interpretable. The recorded training RMSE of 0.005 and test RMSE of 0.000492 further confirm the model’s effectiveness in minimizing prediction errors. The significantly lower RMSE on the test set suggests that the model performs exceptionally well on new data, ensuring reliable digital marketing. This predictive accuracy is essential for digital marketing in the banking industry, as it allows banks to effectively target potential customers for financial products.

CONCLUSION

This study explored the architecture, training, and performance evaluation of a Feedforward Neural Network (FFNN) for digital marketing enhancement in the banking sector by predicting customer subscription behaviour. The artificial neuron model was mathematically described using weighted inputs, bias terms and activation function (Sigmoid Function) and the FFNN architecture included multiple hidden layers and incorporated bias parameters to improve learning efficiency.To enhance generalization and reduce overfitting in the algorithm, regularization technique was introduced which includes the traditional dropout algorithm.

Furthermore, the conventional gradient descent back-propagation algorithm was discussed and applied to optimize the network and the fourth quarter emerged as the most financially favourabletime for customer outreach, aligning marketing efforts with periods of higher account balances. The correlation matrix revealed weak linear relationships between most features and the target variable, justifying the adoption of a neural network capable of capturing complex and non-linear patterns.Experimental training results of the neural network in this study demonstrated high predictive accuracy and minimal errors, with MAE (0.003), MSE (0.0013) and RMSE (0.000492) values which indicate strong model generalization and low deviation from true values. In conclusion, this work successfully integrates intelligent neural network design, smart regularization and adaptive optimization to support decision-making in the financial sector and the work underscores the potential of deep learning approaches to revolutionize customer engagement and resource targeting in digital banking.

REFERENCES

1. Asif, M. (2018). Predicting the success of bank digital-marketing using various classification algorithms. Journal of Risk and Financial Management, 15(6). DOI:10.3390/jrfm15060269. https//www.researchgate.net.
2. Chi, X., & Gang, W. (2022). How to improve the success of bank digital-marketing? Prediction and interpretability analysis based on machine learning. Computers and Industrial Engineering, Article 1088874.
3. Elzhan Z. K., & Adnan Y. (2018). Comparative study of the classification models for prediction of bank digital-marketing. Nazarbayev University, Computer Science, Astana, Kazakhstan. https://doi.org/10.1109/ICAICT.2018.8747086
4. Fakhri A. (2022). Predicting call success of bank digital-marketing campaign with machine learning—A Pythonista with a passion for data analytics, data science, QA automation, and network DevOps. https://www.linkedin.com/in/fakhri-azhar
5. Gu, J., Na, J., Park, J., & Kim, H. (2020). Predicting success of outbound digital-marketing in insurance policy loans using an explainable multiple-filter convolutional neural network. Applied Sciences, 11(21). https://doi.org/10.3390/app11157147
6. Harbor M.C, Eneh I.I., Ebere U.C. (2021). Nonlinear dynamic control of autonomous vehicle under slip using improved back-propagation algorithm. International Journal of Research and Innovation in Applied Science (IJRIAS); Vol. 6; Issue 9; https://rsisinternational.org/journals/ijrias/DigitalLibrary/volume-6-issue-9/62-68.pdf
7. https://openstax.org/books/biology-2e/pages/35-1-neurons-glail-cells.
8. Katamade, D., Aduselidze, G., Katamadze, G., &Slobodianyk, A. (2019). Bank marketing problems and some aspect of their management. Economics and Management, 16, 120–130.
9. Kekong P.E, Ajah I.A.,   Ebere U.C. (2019). Real-time drowsy driver monitoring and detection system using deep learning based behavioural approach. International Journal of Computer Sciences and Engineering 9 (1), 11-21; http://www.ijcseonline.isroset.org/pub_paper/2-IJCSE-08441-18.pdf
10. Kwubeghari, A., Ezigbo, L. I., & Okoye, F. A. (2024). Modelling of cyber attack detection and response system for 5G network using machine learning technique. African Journal of Engineering Research and Development (AJERD), 7, 297–307.
11. Muneeb, A. (2018). Predicting the success of bank digital-marketing using various classification algorithms. Örebro University School of Business. Available at https://www.diva-portal.org
12. Namulia, M. (2011). Effect of selected marketing communication tools on student enrolment in private universities in Kenya. European Journal of Business and Management, 3(3), 172–205.
13. Oduah, O., &Olofin, B. B. (2023). Development of multi-level intrusion detection system for cloud-based log management using machine learning technique. International Journal of Real Time Applications and Computing System (IJORTACS), 1, 101–114.
14. Ogbeta, L. K., & Lois, N. (2023). Neuro-based strategy for real-time protection of wireless network ecosystem against DDOS attack. I1SRED, 5, 79–98.
15. Openstax, (2018). Structure of a neuron [figure], in Biology 2e.openstax, Rice University.
16. Premkumar, B., Prabhakar, N., &Madhavaiah, C. (2021). Predicting the success of bank digital-marketing for selling long-term deposits: An application of machine learning algorithms. Theresa Journal of Humanities and Social Sciences. https://www.researechgate.net
17. Salehin, I., & Kang, D. K. (2023). A review on dropout regularization approaches for deep neural networks within the scholarly domain. Electronics, 12, 3106. https://doi.org/10.3390/electronics12143106
18. Salman, M. (2020a). Predicting the success of bank digital-marketing using artificial neural network. International Journal of Economics and Management Engineering, 14, 1–4.
19. Salman, M. (2020b). Predicting the success of bank digital-marketing using artificial neural network. International Journal of Economics and Management Engineering, 14, 1–4.
20. Sergio M., Paulo C., & Paulo R. (2017). A Data-driven approach to predict the success of bank telemarketing. – University institute of Lisbon, 1649-026 Lisboa, Portugal, and Research Centre, Univ. of Minho, 4800-058 Guimaraes, Portugal. https://www.core.ac.uk.com.
21. Sochima V.E. Asogwa T.C., Lois O.N. Onuigbo C.M., Frank E.O., Ozor G.O., Ebere U.C. (2025)”; Comparing multi-control algorithms for complex nonlinear system: An embedded programmable logic control applications; DOI: http://doi.org/10.11591/ijpeds.v16.i1.pp212-224
22. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., &Salakhutdinov, R. (2014). Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15, 1929–1958.
23. Turkmen, B. (2022). Customer segmentation with machine learning for online retail industry. European Journal of Social and Behavioural Sciences, 31, 111–136. https://doi.org/10.15405/ejsbs.316
24. Turkmen, E. (2021). Deep learning-based methods for processing data in digital-marketing-success prediction. In 2021 Third International Conference on Intelligent Communication Technologies and Virtual Mobile Networks (ICICV) (Vol. 6, pp. 1161–1166). Tirunelveli, India.
25. Vafeiadis, T., Diamantaras, K. I., Sarigiannidis, G., &Chatzisavvas, K. C. (2015). A comparison of machine learning techniques for customer churn prediction. Simulation Modelling Practice and Theory, 55, 1–9. https://doi.org/10.1016/j.simpat.2015.03.003
26. Yufeng, S. (2022). Bank digital-marketing analysis prediction: Customers’ response to future marketing campaigns [Article]. GitHub. https://www.yfsui.github.io