Leveraging metaheuristics with artificial intelligence for customer churn prediction in telecom industries

1.   Introduction

With an abundance of prototype and in-field industrial applications, machine learning (ML) is a rapidly developing technology ^[1]. Some of the areas that have not covered the potential of the technologies are knowledge acquisition, robotics, manufacturing, execution and control, computer vision, design, medicine, scheduling and planning. Studies utilizing new networks and media have found a niche for ML in the retrieval of information and navigation ^[2,3]. Human-computer interaction (HCI) signifies a vast area of research and applications. HCI has addressed various areas, such as computer-supported cooperative work or user interface design along with the tasks relevant to optimization, user modeling and system adaptation to meet the demand of the user ^[4,5]. HCI professionals and researchers are involved in several fields, including but not limited to manufacturing, aviation, and aerospace. HCI concerns the role of individuals in complicated mechanisms, the model of equipment and facilities for human use and the progression of environments for safety and comfort ^[6]. Figure 1 exhibits the process of visual analysis.

Figure 1.  Process of visual analysis.

DownLoad: Full-Size Img PowerPoint

The progressions of the telecommunication sector and globalization have exponentially increased the number of operators in the market, which increases competition ^[7,8]. In this competitive period, it is required that profits should be maximized periodically, and thus many approaches were framed such as up-selling the existing consumer, expanding the retention time of existing consumers and getting new consumers ^[9]. The important objective of customer churn prediction (CCP) is to assist in accomplishing strategies for retaining the customer. In addition to increasing competition in marketplaces for offering services, the risk of customer churn is also increasing exponentially ^[10,11]. So, creating tactics for tracking loyal consumers (non-churners) turns out to be a necessity. The customer churn methods try to detect early churn signs and also forecast the customers who leave voluntarily ^[12]. Therefore, several companies have understood that their existing customer database was the most valuable asset, and churn prediction was an advantageous tool for predicting consumers at risk. Nowadays, HCI techniques, which are mostly dominated by graphical user interfaces, face the limitations of inadequate capture of emotions, slow processing speeds and low accuracy ^[13]. To overcome the difficulties in these communications, it is required to make breakthroughs in intent understanding, vision, context perception, speech, etc. Deep learning (DL) can find valuable association rules using big data analytics for learning the human cognitive process and building new complicated techniques that lead to accurate and efficient predictions and judgments, thereby making human-computer communications more natural and quicker ^[14]. Hence, there comes a demand for empirical, theoretical, systems and design research on artificial intelligence and evolutionary computation techniques ^[15].

This study presents an artificial intelligence with Jaya optimization algorithm based churn prediction for data exploration (AIJOA-CPDE) technique for HCI application. In the AIJOA-CPDE technique, an initial stage of feature selection using the JOA named the JOA-FS technique is presented to choose feature subsets. For churn prediction, the AIJOA-CPDE technique employs a bidirectional long short-term memory (BDLSTM) model. Lastly, the chicken swarm optimization (CSO) algorithm is enforced as a hyperparameter optimizer of the BDLSTM model. A detailed experimental validation of the AIJOA-CPDE technique ensured its superior performance over other existing approaches.

2.   Related works

Al-Shourbaji et al. ^[16] devised a new FS technique, ACO-RSA, which incorporates two metaheuristic approaches, the reptile search algorithm (RSA) and ant colony optimization (ACO). In their study, the incorporation of ACO and RSA was used for selecting a crucial subset of features for churn prediction. Hegde and Mundada ^[17] aimed to accurately forecast the attrition rate in the banking field with an enhanced deep feed forward neural network (FFNN). In the presented technique, the predictive ML algorithm was used through an optimized deep FFNN which had five hidden layers in it. The Adam optimizer was used for training the model to accomplish the optimum performance. The Banking Churn dataset was fed as input to Deep FFNN. The same dataset was passed as input to other ML algorithms, namely, artificial neural network (ANN), decision tree (DT), logistic regression (LR), and Gaussian Naïve Bayes, for performing the comparison study.

In ^[18], ML and deep learning (DL) algorithms were implemented for predicting telecom customer churn. Ubiquitous models including support vector machines (SVMs) and random forest (RF) classifiers were compared with comparatively new architectures such as XGBoost and DNN to categorize whether a customer will churn or not. Further, the effectiveness of this model was explored by a grid search. Vakeel et al. ^[19] introduced an efficient ML technique that involves a boosting algorithm to recognize Gaussian mixture and customer churn (CC) models for clustering the churned customers. Also, the presented method uses the Light Gradient Boosting (LightGB) mechanism, which outperforms XGBoost and AdaBoost for churn prediction by a factor of 15x.

Tariq et al. ^[20] aimed to assist an e-business in predicting the churned users through ML. The study focused on performing decision-making and monitoring customer behavior consequently. The proposed technique made use of 2D-CNN, a DL algorithm. The presented method was a layered structure that included 2D-CNN layer, data load layer and pre-processing layer. Furthermore, Apache Spark parallel and distributed architectures have been implemented for data processing in the parallel environment. Mohammad et al. ^[21] aimed to recognize the factors that influencecustomer churn, developed an efficient churn prediction method and provided a better analysis of data visualization outcomes. The presented technique was used for the analysis of churn prediction, which covers multiple stages: evaluation of the classifiers, data pre-processing, analysis, enforcement of ML algorithm and selection of the best one for prediction. Data preprocessing involved three main actions: feature selection, data cleaning and data transformation. ML classifiers selected were RF, LR and ANN.

3.   The proposed model

In the present study, we have established a novel AIJOA-CPDE approach for automated churn prediction in HCI applications. The goal of the AIJOA-CPDE technique exists in the accurate identification of churned or non-churned customers. In the AIJOA-CPDE technique, three subprocesses are used, namely, feature subset election, BDLSTM classification and CSO-based parameter tuning. Figure 2 defines the workflow of the AIJOA-CPDE system.

Figure 2.  Workflow of AIJOA-CPDE approach.

DownLoad: Full-Size Img PowerPoint

3.1.   Feature selection using JOA-FS technique

In this work, the JOA-FS approach was used to produce a set of features to enhance the classification results. Rao ^[22] established the JOA deal with constraint and unconstraint optimized techniques. It is significantly easy to implement this approach because it has only one phase. Jaya means "victory" in Sanskrit. This method exploits a population-oriented metaheuristic that contains evolutionary and SI features. It was identified in the behavior of the "survival of fittest" conception. "The search method of the proposed method aim is to get closer to success by finding the best global solution and evade failure by avoiding the worst choice ^[22]." In this work, properties of swarm‐based intelligence and evolutionary algorithms were incorporated.

Consider that the objective function $\psi \left(x\right)$ must be increased or decreased according to the problems. Assume that $n$ is the number of candidate solutions, and $d$ is the number of design parameters (population size, $k = \mathrm{1, 2}\dots, n)$ for $t$ iterations. Assume the superior candidate obtains the best value of $\psi \left(x\right)$ ($\psi \left(x{)}_{best}\right)$ of all the candidate solutions, and the worst candidate acquires the worst value of $\psi \left(x\right)$ ($\psi (x{)}_{worst}$) of all the candidate solutions ^[23]. Once the values of ${thej}^{th}$ variable for the ${k}^{th}$ candidate at iteration $t$ are ${X}_{(j, k, t)}$, the value is upgraded as follows:

From the expression, the value of the jth parameter for the best candidate represents ${X}_{(j, best, t)}$ while the value of the jth parameter for the worst candidate characterizes ${X}_{(j, worst, t)}.$ ${X}_{(j, k, t)}^{\text{'}}$ shows the upgraded values of ${X}_{(j, k, t)}$, and ${q}_{(1, j, t)}$ and ${q}_{(2, j, t)}$ indicate random numbers within [0, 1] for parameter $j$ at iteration $t$. The term ${q}_{\left(1, j, t\right)}({X}_{\left(j, best, t\right)}-|{x}_{\left(j, k, t\right)}\left|\right)$ demonstrates the solution's tendency to get nearer to the best solution, whereas the term " ${q}_{\left(2, j, t\right)}\left({X}_{\left(j, worst, t\right)}-\left|{x}_{\left(j, k, t\right)}\right|\right)$ " demonstrates the solution's tendency to avoid the worst. Once ${X}_{(j, k, t)}^{\text{'}}$ generates a better function value, it is accepted. In each iteration, every acceptable function value was utilized and kept as the input for the forthcoming iteration.

The fitness function (FF) of JOA considers the classifier accuracy and number of selected features. It reduces the size set of chosen features and maximizes the classifier performances. Thereby, the subsequent FF is exploited for estimating individual solutions, as follows.

In Eq (2), $theErrorRate$ identifies the classifier error rate applying the feature that is chosen. The error rate was calculated as a percentage of inaccurate categories for counting classifications made, specified as values among [0, 1]. $\#SF$ was chosen as feature count, and $\#All\_F$ is the overall amount of features in actual data. $\alpha$ was applied for controlling consequence subset length and classification quality. In the presented method, $\alpha$ is fixed as 0.9.

3.2.   Churn prediction using optimal BDLSTM model

For the churn prediction process, the BDLSTM model was used. LSTM is a variant of recurrent neural networks (RNN) that was presented ^[24]. In every $t$ iteration, ${x}_{t}$ represents the input, and ${h}_{t}$ symbolizes the output of the LSTM cell, and the cell output state of the preceding time step can be represented as ${C}_{t-1}$ are integrated with LSTM cells for updating network parameters in the training procedure. The gates mechanism is presented for controlling the $ceU$ state of LSTM by permitting data to pass through alternatively. Forget, input and output gates are represented by ${f}_{t}$, ${i}_{t}$, and ${0}_{t}$, correspondingly. The value of the cell input state and gate can be computed as follows:

From the equation, ${W}_{f}$, ${W}_{i}$, ${W}_{c}$ and ${W}_{o}$ represent the weight matrices between inputs of the hidden layer, forget gate, input gate and input cell state and the output gate. ${U}_{f}, {U}_{i},$ ${U}_{c}$ and ${U}_{o}$ signify weight matrices between the previous cell output state, forget gate, input gate and input cell state and the output gate. ${b}_{f},$ ${b}_{i},$ ${b}_{c}$ and ${b}_{o}$ denote the bias vectors. BDLSTM is a kind of LSTM variation. In BDLSTM, sequential data is processed with backward and forward LSTM layers, and those 2 hidden layers are closely linked with the same output layer. Based on Eqs (3) to (6), at every iteration $t$, the cell output state ${C}_{t}$ and LSTM layer output ${h}_{t}$ were formulated as follows:

BDLSTM has been efficaciously used in the fields of speech recognition, trajectory prediction, natural language processing, biomedical event analysis and so on. It has been proved that BDLSTM outperforms traditional LSTM in some fields, such as frame automatic speech recognition and understanding and wise phoneme classification. Figure 3 represents the structure of LSTM.

Figure 3.  Architecture of LSTM.

DownLoad: Full-Size Img PowerPoint

To scale up the predictive results of the BDLSTM method, the CSO technique was implemented as a hyperparameter optimizer of the BDLSTM method. In CSO, there exist numerous groups, and each group encompasses some hens, chicks and dominant roosters ^[25]. Hens, chicks and roosters in the groups are determined according to their fitness values. Chicks are chickens with poor fitness values. The rooster (group head) is a chicken that has a better fitness value. Mostly, chickens will be hens, and it randomly selects which group to stay in. Mother-child associations among the hens and the chicks are carried out at random. The dominance and mother‐child relations in the group remain the same and update (G) time step. The chicken movements can be expressed as follows:

1) The rooster location update is shown below:

where

From the expression, $k\in [1, {N}_{r}],$ $k\ne i$ and ${N}_{r}$ denote the number of selected roosters. ${X}_{i, j}$ indicates the position of $\mathrm{i}$ - $\mathrm{t}\mathrm{h}$ rooster in the $\mathrm{j}\mathrm{t}\mathrm{h}$ parameter at $\mathrm{t}$ and $\mathrm{t}+1$ iterations, $randn$ $(O, {\sigma }^{2})$ produces Gaussian random values having mean $0$ and variance ${\sigma }^{2},$ $\epsilon$ signifies an arbitrarily small constant, and ${f}_{i}$ demonstrates fitness values for corresponding rooster $i.$

2) The hen location updating is given in the following:

where

and

Now, ${r}_{1},$ ${r}_{2}\in \left[1, \dots, N\right], r1\ne {r}_{2}$ denote the index of the rooster, $r2$ denotes swarming chicken, eitherrooster or hen and the uniform random number is generated through $randn$.

3) The chick location updating is expressed as follows:

In Eq (13), ${X}_{m, j}^{t}$ denotes the position of the ${i}^{th}$ chick's mother.

The CSO derives a fitness function (FF) for accomplishing improved classifier outcomes. It specifies the positive value to characterize the better accuracy of the candidate solution. The reduction classifier error amount is considered as the FF in the following.

4.   Results and discussion

The AIJOA-CPDE procedure was tested by employing the churn dataset ^[26]. It involves 3333 samples with 21 features and two classes. The JOA-FS technique has chosen 12 features out of available 21 features. For experimental validation, the training (TR) and testing (TS) datasets are split in two ways as 90:10 (90% TR samples and 10% TS samples) and 80:20 (80% TR samples and 20% TS samples). Figure 4 demonstrates the churn prediction results of the AIJOA-CPDE model through the confusion matrix under 90:10 and 80:20 of TR/TS datasets.

Figure 4.  Confusion matrices of AIJOA-CPDE system (a-b) TR and TS databases of 90:10 and (c-d) TR and TS databases of 80:20.

DownLoad: Full-Size Img PowerPoint

The outcomes indicated that the AIJOA-CPDE procedure has precisely recognized churn and non-churn customers. As an example, with 90% of the TR dataset, the AIJOA-CPDE model has identified 356 as churn and 2546 as non-churn. Meanwhile, with 10% of the TS database, the AIJOA-CPDE procedure has identified 36 as churn and 288 as non-churn. Moreover, with 80% of the TR dataset, the AIJOA-CPDE method has identified 316 as churn and 2268 as non-churn.

In Table 1, a brief set of CCP outcomes of the AIJOA-CPDE model with 90:10 and 80:20 of TR/TS datasets is shown. With 90% of the TR dataset, the AIJOA-CPDE procedure has attained an average $acc{u}_{bal}$ of 90.20%, $pre{c}_{n}$ of 96.64%, $rec{a}_{l}$ of 90.20%, ${F}_{score}$ of 93.07% and $AU{C}_{score}$ of 90.20%. Meanwhile, with 10% of the TS database, the AIJOA-CPDE approach has acquired an average $acc{u}_{bal}$ of 91.35%, $pre{c}_{n}$ of 94.97%, $rec{a}_{l}$ of 91.35%, ${F}_{score}$ of 93.05% and $AU{C}_{score}$ of 91.35%. Eventually, with 80% of the TR dataset, the AIJOA-CPDE approach has achieved an average $acc{u}_{bal}$ of 90.17%, $pre{c}_{n}$ of 97.45%, $rec{a}_{l}$ of 90.17%, ${F}_{score}$ of 93.37% and $AU{C}_{score}$ of 90.17%. Finally, with 20% of the TS dataset, the AIJOA-CPDE method has reached an average $acc{u}_{bal}$ of 90.02%, $pre{c}_{n}$ of 97.81%, $rec{a}_{l}$ of 90.02%, ${F}_{score}$ of 93.43% and $AU{C}_{score}$ of 90.02%.

Table 1.  CCP outcome of AIJOA-CPDE system under 90:10 and 80:20 of TR/TS datasets.

Class	Accuracybal	Precision	Recall	F-Score	AUC Score
Training Phase (90%)
Churn	80.91	96.48	80.91	88.01	90.20
Non-Churn	99.49	96.81	99.49	98.13	90.20
Average	90.20	96.64	90.20	93.07	90.20
Testing Phase (10%)
Churn	83.72	92.31	83.72	87.80	91.35
Non-Churn	98.97	97.63	98.97	98.29	91.35
Average	91.35	94.97	91.35	93.05	91.35
Training Phase (80%)
Churn	80.61	98.14	80.61	88.52	90.17
Non-Churn	99.74	96.76	99.74	98.22	90.17
Average	90.17	97.45	90.17	93.37	90.17
Testing Phase (20%)
Churn	80.22	98.65	80.22	88.48	90.02
Non-Churn	99.83	96.96	99.83	98.37	90.02
Average	90.02	97.81	90.02	93.43	90.02

---

 | Show Table

DownLoad: CSV

Figure 5 demonstrates the churn prediction results of the AIJOA-CPDE model by means of a confusion matrix under 70:30 and 60:40 of TR/TS databases. The outcomes show the AIJOA-CPDE algorithm has accurately recognized churn and non-churn customers. As an example, with 60% of the TR dataset, the AIJOA-CPDE model has recognized 262 as churn and 1978 as non-churn. In the meantime, with 10% of the TS database, the AIJOA-CPDE methodology has identified 121 as churn and 834 as non-churn. Additionally, with 60% of the TR database, the AIJOA-CPDE approach has detected 232 as churn and 1713 as non-churn.

Figure 5.  Confusion matrices of AIJOA-CPDE system (a-b) TR and TS databases of 70:30 and (c-d) TR and TS databases of 60:40.

DownLoad: Full-Size Img PowerPoint

In Table 2, a brief set of CCP outcomes of the AIJOA-CPDE model with 70:30 and 60:40 of TR/TS datasets is shown. With 70% of the TR dataset, the AIJOA-CPDE method has attained an average $acc{u}_{bal}$ of 89.36%, $pre{c}_{n}$ of 93.58%, $rec{a}_{l}$ of 89.36%, ${F}_{score}$ of 91.32% and $AU{C}_{score}$ of 89.36%. In the meantime, with 30% of the TS dataset, the AIJOA-CPDE procedure has obtained an average $acc{u}_{bal}$ of 88.19%, $pre{c}_{n}$ of 94.17%, $rec{a}_{l}$ of 88.19%, ${F}_{score}$ of 90.85% and an $AU{C}_{score}$ of 88.19%. In parallel, with 60% of the TR dataset, the AIJOA-CPDE method has achieved an average $acc{u}_{bal}$ of 91.74%, $pre{c}_{n}$ of 96.68%, $rec{a}_{l}$ of 91.74%, ${F}_{score}$ of 94.01% and $AU{C}_{score}$ of 91.74%. Lastly, with 40% of the TS dataset, the AIJOA-CPDE procedure has gained an average $acc{u}_{bal}$ of 91.41%, $pre{c}_{n}$ of 97.63%, $rec{a}_{l}$ of 91.41%, ${F}_{score}$ of 94.20% and $AU{C}_{score}$ of 91.41%.

Table 2.  CCP outcome of AIJOA-CPDE system under 70:30 and 60:40 of TR/TS datasets.

Class	$\boldsymbol{A}\boldsymbol{c}\boldsymbol{c}{\boldsymbol{u}}_{\boldsymbol{b}\boldsymbol{a}\boldsymbol{l}}$	$\boldsymbol{P}\boldsymbol{r}\boldsymbol{e}{\boldsymbol{c}}_{\boldsymbol{n}}$	$\boldsymbol{R}\boldsymbol{e}\boldsymbol{c}{\boldsymbol{a}}_{\boldsymbol{l}}$	${\boldsymbol{F}}_{\boldsymbol{s}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$	$\boldsymbol{A}\boldsymbol{U}{\boldsymbol{C}}_{\boldsymbol{s}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$
Training Phase (70%)
Churn	80.12	90.34	80.12	84.93	89.36
Non-Churn	98.60	96.82	98.60	97.70	89.36
Average	89.36	93.58	89.36	91.32	89.36
Testing Phase (30%)
Churn	77.56	92.37	77.56	84.32	88.19
Non-Churn	98.82	95.97	98.82	97.37	88.19
Average	88.19	94.17	88.19	90.85	88.19
Training Phase (60%)
Churn	84.06	95.87	84.06	89.58	91.74
Non-Churn	99.42	97.50	99.42	98.45	91.74
Average	91.74	96.68	91.74	94.01	91.74
Testing Phase (40%)
Churn	83.09	98.29	83.09	90.05	91.41
Non-Churn	99.73	96.98	99.73	98.34	91.41
Average	91.41	97.63	91.41	94.20	91.41

---

 | Show Table

DownLoad: CSV

The TACC and VACC of the AIJOA-CPDE algorithm are examined on CCP achievement in Figure 6. The results exhibited that the AIJOA-CPDE method has demonstrated improved accomplishment with improved values of TACC and VACC. Particularly, the AIJOA-CPDE technique has attained the greatest TACC results.

Figure 6.  TACC and VACC analysis of AIJOA-CPDE approach.

DownLoad: Full-Size Img PowerPoint

The TLS and VLS of the AIJOA-CPDE method are verified on CCP performance in Figure 7. The figure showed that the AIJOA-CPDE procedure has revealed improved accomplishment with the least values of TLS and VLS. Apparently, the AIJOA-CPDE method has resulted in decreased VLS results.

Figure 7.  TLS and VLS analysis of AIJOA-CPDE approach.

DownLoad: Full-Size Img PowerPoint

A precision-recall examination of the AIJOA-CPDE procedure under the test dataset is shown in Figure 8. The figure depicted the AIJOA-CPDE approach has improved values precision-recall values under all classes.

Figure 8.  Precision-recall analysis of the AIJOA-CPDE approach.

DownLoad: Full-Size Img PowerPoint

An elaborated ROC research of the AIJOA-CPDE method under the test dataset is shown in Figure 9. The figure exhibited the AIJOA-CPDE method has exposed its capability to categorize different classes.

Figure 9.  ROC curve analysis of AIJOA-CPDE approach.

DownLoad: Full-Size Img PowerPoint

Figure 10.  $Acc{u}_{y}$ analysis of AIJOA-CPDE approach with recent methods.

DownLoad: Full-Size Img PowerPoint

A comparative analysis of the AIJOA-CPDE technique with current procedures on churn prediction is reported in Table 3 ^[27,28,29]. A brief analysis of the AIJOA-CPDE method with current methods in terms of $acc{u}_{y}$ is depicted. Based on $acc{u}_{y}$, the outcomes exhibited that the DT method has reduced $acc{u}_{y}$ by 76.78%. Then, the LR, SVM, SGD, and RMSProp models have reported slightly improved $acc{u}_{y}$ of 80.65, 84.41, 84.54 and 87.48% respectively. Although the ISMOTE-OWELM models have reached reasonable outcomes with $anacc{u}_{y}$ of 90.59%, the AIJOA-CPDE model has shown a maximum $acc{u}_{y}$ of 91.41%.

Table 3.  Comparative analysis of AIJOA-CPDE technique with recent approaches ^[27,28,29].

Methods	$\boldsymbol{A}\boldsymbol{c}\boldsymbol{c}{\boldsymbol{u}}_{\boldsymbol{y}}$	$\boldsymbol{P}\boldsymbol{r}\boldsymbol{e}{\boldsymbol{c}}_{\boldsymbol{n}}$	$\boldsymbol{R}\boldsymbol{e}\boldsymbol{c}{\boldsymbol{a}}_{\boldsymbol{l}}$	${\boldsymbol{F}}_{\boldsymbol{s}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$	$\boldsymbol{A}\boldsymbol{U}{\boldsymbol{C}}_{\boldsymbol{s}\boldsymbol{c}\boldsymbol{o}\boldsymbol{r}\boldsymbol{e}}$
AIJOA-CPDE	91.41	97.63	91.41	94.20	91.41
Logistic Regression [27]	80.65	79.44	80.56	79.17	82.30
Decision Tree [27]	76.78	56.90	75.81	65.09	78.37
ISMOTE-OWELM [28]	90.59	91.78	89.51	89.77	89.96
SVM [28]	84.41	84.66	84.11	85.71	84.10
SGD-NN [29]	84.54	86.22	85.92	84.43	84.91
RMSProp-NN [29]	87.48	85.29	85.30	85.20	86.38

---

 | Show Table

DownLoad: CSV

Although the ISMOTE-OWELM approaches have reached reasonable outcomes with a $rec{a}_{l}$ of 89.51%, the AIJOA-CPDE technique has shown maximum $rec{a}_{l}$ of 91.41%. Although the ISMOTE-OWELM techniques have reached reasonable outcomes with ${aF}_{score}$ of 89.77%, the AIJOA-CPDE model has shown a maximum ${F}_{score}$ of 94.20%. In addition, the AIJOA-CPDE technique has shown a maximum $AU{C}_{score}$ of 91.41%. Thus, the AIJOA-CPDE method can be employed for an accurate churn prediction process.

5.   Conclusions

In this study, we have advanced a novel AIJOA-CPDE technique for automated churn prediction in HCI applications. The goal of the AIJOA-CPDE technique exists in the accurate identification of churned or non-churned customers. In the AIJOA-CPDE technique, an early stage of feature selection through JOA named the JOA-FS technique is presented to choose feature subsets. To identify churn customers, the AIJOA-CPDE technique employed the BDLSTM model. Lastly, the CSO algorithm is implemented as a hyperparameter optimizer of the BDLSTM method. A detailed experimental validation of the AIJOA-CPDE technique ensured its superior performance over other existing approaches. In the future, the outlier detection process can be employed to boost the prediction performance of the AIJOA-CPDE technique.

Use of AI tools declaration

The authors declare they have not used Artificial Intelligence (AI) tools in the creation of this article.

Acknowledgments

This work was supported by the Ulsan City & Electronics and Telecommunications Research Institute (ETRI) grant funded by the Ulsan City [23AS1600, the development of intelligentization technology for the main industry for manufacturing innovation and Human-mobile-space autonomous collaboration intelligence technology development in industrial sites].

Conflict of interest

The authors declare that they have no conflict of interest. The manuscript was written through contributions of all authors. All authors have given approval to the final version of the manuscript.