Machine Learning Approach to Users’ Age Prediction: A Telecom Company Case Study in Saudi Arabia

In the present digital era, telecommunication companies play an instrumental role in ensuring smooth and efficient user experience. These companies have revolutionized information access, automated fundamental operations, and negated the necessity for centralized offices, thereby connecting businesses globally [1]. Among these, the Saudi Telecommunication Company (STC) emerges as a prominent player, providing a comprehensive range of services, including data, Internet, and telephone services [2].

Concurrently, the digital landscape provides innovative marketing and advertising avenues such as online advertising and social network marketing, which significantly contribute to business revenue generation and customer service expansion [3]. Consequently, marketers are continually seeking novel strategies to effectively reach their target audience. In the context of telecommunication services, understanding the age and preferences of users could be instrumental in delivering age-appropriate advertisements [4].

The primary objective of this study is to develop a system capable of determining the users' age. A substantial dataset provided by STC forms the foundation for this system, which, once operational, could enable STC to deliver age-specific advertisements, thereby optimizing their resources for the most relevant user group [5].

Artificial Intelligence (AI) and Machine Learning (ML) systems have become indispensable for businesses, with ML analyzing past behavior data and interpreting them using various algorithms [6]. ML enhances business operations and scalability, ensuring an excellent customer experience [7]. Specifically, telecommunication companies have begun incorporating ML and AI techniques to improve their services. These techniques enable the analysis of substantial customer data and aid in the development of superior user services. Further, ML facilitates reaching the appropriate user demographic with tailored advertisements [8].

Several studies have employed ML models for age prediction with varying degrees of success. For instance, Abu Nada et al. [9] utilized a Convolutional Neural Network (CNN) to predict users' age and gender, achieving 82% accuracy for gender prediction and 57% for age prediction. Similarly, Hassan et al. [10] proposed a CNN model for age prediction from shoe prints, which recorded an impressive accuracy of 86.07% but a considerable error of 40.23% for the prediction of five-year-old individuals. Kim et al. [11] proposed two CNN models using ResNet-152 and DenseNet-169 for age and gender prediction from paranasal sinus images, with the second model achieving a higher accuracy of 77.6 ± 1.5% for age prediction. Nemati et al. [12] proposed an ensemble model using four ML algorithms for gender and age prediction, achieving an accuracy of 60% for age prediction and 83% for gender prediction. Asogwa and Anigbogu [13] developed a ML model for predicting author's age from text using the Naïve Bayes algorithm, attaining a prediction accuracy of 80%.

These studies underscore the potential and challenges of employing ML techniques for age prediction, providing a foundation for this study's endeavor to develop an effective ML-based age prediction system for telecommunication users.

This research is primarily focused on the application of Machine Learning (ML) algorithms for the prediction of users' age. The impetus for this study stems from the company's aim to enhance user experience by appropriately tailoring their services according to the user's age. The multinational population residing in the Kingdom of Saudi Arabia (KSA) provides a diverse demographic wherein age-specific service offerings could significantly improve the quality of service provided. Furthermore, literature research suggests that ML classifiers show promising results in similar age-prediction studies.

Therefore, this study aims to employ and assess the performance of four ML classifiers: Gradient Boosting, Naïve Bayes, Multi-Layer Perceptron, and Logistic Regression. The ultimate goal is to identify the model that provides the highest accuracy using different performance metrics. By predicting user's age, the company could offer more personalized services to its users.

The dataset provided for this study includes several key features and a large number of samples, which will assist in developing a robust ML model for predicting the users' age.

The remainder of the paper is structured as follows: Section 2 presents a review of related work. In Section 3, the preprocessing of the dataset is discussed. Section 4 details the materials and methods employed in the study. Sections 5, 6, and 7 outline the experimental setup, present the results and discussion, and compare this study with previous works, respectively. Finally, Section 8 concludes the paper.

In this section, details about the dataset used are presented as well as some necessary pre-processing techniques that need to be applied to the data set.

3.1 Data description

Table 1. Features description

The study includes data obtained from STC, resulting from an initiative launched under "STC Open Data". The initiative aimed to prepare a technologically aware Saudi generation in line with Vision 2030 through supporting researchers, students, and universities interested in data science. The data used in this study describes the customers' use of uncommon handset devices. The period during which customers' data was collected for STC in Saudi Arabia ranged from 2018 until 2020.The dataset contains over three million rows, 19 features, and one target class called "Age_B" and the possible labels within it are "<18", "18-25", "26-35", "36-55", and ">55" [34]. Table 1 shows a description of each feature:

3.2 Pre-processing techniques

Data quality is considered an essential factor for the success of building machine learning models, as the training phase needs reliable data free of annoying or duplicate values. Data pre-processing acts as a data purifier filter by applying several techniques, including data cleaning, features normalization, and extraction [35]. Table 2 enlists features with missing values.

Table 2. Count of missing values

3.2.1 Data cleaning

Data cleaning is considered an essential step in preprocessing of machine learning as it deals with missing values and errors in the data set. There are two techniques for dealing with missing values, either by "imputing the missing values" or "Discarding the Missing Values". In this study, one of the types of discarding the missing values technique was applied, "List-wise deletion," which works to delete any row containing a missing value. All missing or unknown values within each column in the data set were removed except for the target class column data [36]. The rows containing unknown values within the target class were transferred to a new set for use in the testing and evaluation phase as a test set. The amount of dataset after the deletion process is 214,730 rows. The table below shows the count of missing values for each feature in the dataset before processing.

Dropping duplication samples. The data used is rows containing the data of customers subscribed to the telephone services provided by STC. Every time a customer made a new phone call through his device, this customer's data was stored in the dataset inside a new row. However, having multiple rows with the same data about one customer is redundant, so duplicated rows have been removed to ensure the quality of data and obtain high predictive accuracy. The number of records in the data set after deletion is 283,278.

3.2.2 Data transformation

Data transformation makes the data homogeneous, organized, and easy to analyze. Normalization and discretization techniques used in this study [37].

Data normalization. Data normalization improves model performance in addition to making its predictive accuracy high. Data normalization is also known as min-max scale or min-max normalization. Also, data normalization restructures the scaling range of the feature and makes it within the range [0, 1] [38]. In this study, feature scaling was performed using data normalization on the only numeric feature present in the dataset called "DEVICE_COUNT" using the formula:

$x^{\prime}=\frac{x-\min }{(x) \max (x)-\min (x)}$        (1)

In the equation, max(x) and min(x) represent the maximum and minimum values of the feature respectively [38].

Data Discretization. The process of converting continuous data into a set of small data periods. This technique is also called the "data reduction mechanism" because it converts a large dataset into categorical data [39]. This technique was applied to the target class by reducing the number of class labels from five to three. The data was classified among the three labels as follows:

The data in ">18" is categorized for a label called "Young".

The data in "<55" is categorized for a label called "Old".

The rest of the data is categorized for a label called "Adult".

Figure 1 and Figure 2 show the number of samples in each label in the target class before and after applying the discretization technique, respectively. As an outcome, it is apparent that before discretization, there are more buckets of the data in terms of age group while after discretization, data has been distributed to relatively a smaller number of age groups. Table 3 contains the selected features list.

21.png

Figure 1. Target class distribution before discretization technique

22.png

Figure 2. Target class distribution after discretization technique

Table 3. Features after the feature selection

3.2.3 Dimensionality reduction

Dimensionality reduction is a technique implemented to improve the accuracy of classifiers [40]. In this study, feature selection was applied, which removes redundant features to reduce the dimensions of the data set by calculating the importance of each feature and its relationship to the target label and only retains the key features with high relation [41]. For the data set in this study, features that have recurring content and are like another feature content within the data set were excluded. For example, the feature content "NATIONALITY_CD" is considered an abbreviation for the feature content "NATIONALITY_NAME", so it is considered a redundant feature. The features that have been removed are "CAL_DT", "MODEL_NAME", "BRAND_FULL_NAME", and "NATIONALITY_CD". The number of features within the data set after the feature selection process has become 15, and the table shows all the remaining features.

3.2.4 Labels binarizing

Label Binarizer is a transformation method that takes categorical data as input and converts it into a Numpy array. It encodes the data into dummy variables that indicate whether a particular label exists or not [41]. This method was performed on all the features whose data type is categorical in the dataset to convert them into numeric data that ML algorithms can deal with them.

3.2.5 Data balancing

A balanced dataset is a set in which the data of each target class has the same number of samples. The most popular techniques for balancing data are oversampling and under sampling [42]. A balancing method was used to make the number of samples in the data of each label in the target class equal to the number of samples in the "Adult" label (containing 220210 samples), as this process is called the "oversampling technique". Figure 3and Figure 4 show the number of samples in each label in the target class before and after applying the balancing method. It is apparent that the data after balancing provides approximately equal number of samples in each class. That potentially can lead to a fairer model building where each class is equally contributing to the model.

23.png

Figure 3. Target class distribution before balancing

24.png

Figure 4. Target class distribution after balancing

This section discusses the analysis and setup for each algorithm for age detection. Including the Artificial Neural Network (ANN), Random Forest (RF), Gradient Boosting (GB) and Decision Tree (DT) model. Also, it involves the optimization strategy.

5.1 Experimental setup and analysis

During the implementation of the experiments, Python programming language was used, and the code was run by the Google Collaboratory platform. The STC dataset contains 19 features with a total of three million instances. Several preprocessing methods were applied to the dataset to build a high-performance model (discussed in detail in the preprocessing section). The STC dataset was divided into two sets (70% for the training set and 30% for the test set). The target prediction class is the Age feature. Also, four machine-learning algorithms were used to predict the age which are Artificial Neural Network (ANN), Random Forest (RF), Gradient Boosting (GB), and Decision Tree (DT) algorithm (discussed in detail in the Experimental Models section). Moreover, the proposed models were trained on the training samples of the dataset then the testing samples were used to determine the performance of the models. Finally, the recall, f1-score, accuracy, confusion matrix, and precision were used as evaluation criteria of the models’ performance.

5.2 Optimization strategy

Optimization is very important to increase the performance of the model. Optimization plays a big role in machine learning models, where fitting processes the learning model on the training dataset. For optimization in the proposed study, we used three dense layers in ANN model. Also, we used the ‘relu’ function with the number of 36,24, and 16 nodes for each layer respectively. Furthermore, the number of epochs was 60. Moreover, the hyperparameters are applied to the ANN model to increase its accuracy. Hyperparameters have several benefits and positive impacts on the model performance, including optimizer, losses, and metrics. The optimal hyperparameter values used are shown in Table 4:

Table 4. Optimal hyperparameter values

In Random Forest (RF) model we used n_estimators equal to 100 and the criterion was “gini” and the max depth is equal to None. Moreover, for the Gradient Boosting (GB) the loss is ’deviance’, and the learning rate is equal to 0.01 and the n_estimators is equal to 100. Also, for the Decision Tree (DT) algorithm the criterion was “gini”, the splitter was “best” and the max_depth was equal to None.

5.3 Grid search

Grid-searching is the simplest process used to scan the data to find the optimal parameters for each model. Grid-searching can be applied to different model types to find the best parameter for a given machine learning model. Grid-searching is very slow and complex and it's computationally expensive. Basically, the domain of the hyperparameters will be divided into a discrete grid, then we use the grid to try every combination of values and calculate using the cross-validation some performance matrices. In cross-validation, the point of the grid that increases the average value is the best combination of values for the hyperparameter. However, grid-searching is a good method if you want to find the best combination of values for hyperparameters.

5.4 Evaluation matrices

The performance of the Artificial Neural Network (ANN), Random Forest (RF), Gradient Boosting (GB), and Decision Tree (DT) models was evaluated using the confusion matrix, the recall, f score, accuracy, and precision.

5.4.1 Confusion matrix

Confusion matrix can measure the performance of machine learning models. It compares the predicted values with the real values. Figure 5 presents a confusion matrix for a binary problem classification. Where the symbols TP, TN, FP, and FN indicates True Positive, False Positive, False Positive, and False Negative respectively [47-50].

25.png

Figure 5. Confusion matrix for binary classification

5.4.2 Accuracy

Accuracy shows the percentage of the truly predicted samples among all the samples in the testing set.

Accuracy $=(T P+T N) /(T P+T N+F P+F N)$            (2)

5.4.3 Recall

Recall shows the percentage of the positive samples that were correctly predicted among all the real positive samples.

 Recall $=T P /(T P+F N)$             (3)

5.4.4 Precision

Precision presents the percentage of the truly predicted samples of the positive class among all the positive predictions.

precision $=T P /(T P+F P)$               (4)

5.4.5 F score

F-Score shows the average the truly predicted samples of the positive class (Precision) and the positive samples (Recall) that were correctly predicted. This measure used to evaluate the balance of model’s predictions among the two classes.

F-Score=(2×Pr e cision×Recall)/(Pr e cision+Recall)             (5)

This section presents a comparison between the proposed models, Including Artificial Neural Network (ANN), Random Forest (RF), Gradient Boosting (GB) and Decision Tree (DT).

7.1 Comparison between the proposed models

Table 6 presents a comparison for the accuracy between the proposed models. Including Artificial Neural Network (ANN), Random Forest (RF), Gradient Boosting (GB) and Decision Tree (DT) algorithm.

Table 6. Comparison for the accuracy among the proposed models

In this study, several machine learning algorithms were used to predict the users’ age. As a result of implanting these models the accuracy reached is presented in table #. Artificial Neural Network (ANN) reached an accuracy of 60% where the Random Forest (RF) model reached 53%. Also, 53% of accuracy for implanting the Gradient Boosting (GB) and 56% for the Decision Tree model (DT). The highest value of accuracy is 60% for implementing the Artificial Neural Network model (ANN). ANN implemented with Adam optimizer and used three hidden layers and with the value of 60 epochs. Artificial Neural Network model (ANN) achieved good accuracy compared with other models’ results.

7.2 Comparison with the benchmark studies

This section presents a comparison of the proposed model with the benchmark studies. Table 7 shows comparison between the results of the proposed model and other studies.

In this study, ANN, RF, GB, and DT model were used to predict the users’ age. The STC dataset was used to train the model consisting of three million rows divided into three classes which are young, adult, and old features. The number of samples in the dataset in each class is young contains 220210 samples after oversampling.

Table 7. Comparison of the proposed model with benchmark studies

In this section, we compared the proposed model with the previous benchmark studies. The benchmark studies in table# compared with this study were used to predict the users’ age. The first study [9], authors used a CNN model to estimate the users’ age and detect their gender by linking the entered birth information with the users' image. This study reached 57% for accuracy which is lower than the proposed model. So, our proposed model showed a better accuracy performance than this study. Furthermore, the study in [12] used the dataset of 280 samples. Also, this study used the Gradient Boosting Classifier, Multi-Layer Perceptron classifier, multinomial Naive Bayes classifier, and Logistic Regression CV classifier to predict the age and the gender. The study resulted in the highest accuracy of 60%, which is like the proposed model.