Video game addiction is a phenomenon characterized by excessive and compulsive dependence on video games. This addiction can lead to mental health problems and detrimental consequences in other areas of life. Questionnaires and interviews used to identify this dependence can be influenced by biases and errors, requiring active participation from the individuals involved. Moreover, these methods can be ineficient and require significant resources in terms of time and costs. To overcome these limitations, we have leveraged advances in machine learning and deep learning. Utilizing data on gaming habits, playtime durations, demographic data, and other relevant features, our model efectively identifies signs of addiction. The development process followed an incremental approach, beginning with the creation of various models, including decision trees, Support Vector Machines (SVM), Extreme Gradient Boosting (XGBoost), Feed-Forward Neural Networks (FFNN), and random forests. These models were chosen for their popularity and widespread use in the field. Recognizing shortcomings in individual models, we transitioned to an ensemble technique known as "Stacking" to address overfitting issues and enhance overall performance. The final selection of the appropriate model aimed at reducing complexity. The stacking model demonstrated notable accuracy scores Accuracy: 0.95, Recall: 0.89, Precision: 0.87.
Nowadays, video game disorder is a growing problem that raises within our modern society. With
the rising popularity of video games and their easy accessibility through various platforms, young
individuals find themselves struggling with excessive and compulsive addiction to video games, this
addiction can lead to detrimental consequences for their mental health, academic performance, social
relationships, and overall well-being. In this context, the preventive prediction of video game addiction
aims to identify at-risk individuals before the addiction becomes problematic. The goal is to intervene
early by providing tailored support, prevention strategies, and therapeutic practices. The utilization of
Artificial Intelligence (AI) methodologies, such as machine learning and deep learning, has facilitated
significant advancements, it is possible to analyze vast amounts of data related to gaming habits, online
behaviors, and other factors to identify early signs of addiction. In our study, we developed several
individual models, each trained on preprocessed data to predict video game addiction. However, as
we conducted our experiments, we observed a phenomenon of overfitting. This occurs due to the
limitations of the data we have, leading to poor generalization on new data. To overcome the overfitting
issue, we opted for the ensemble learning method known as Stacking. This method allows us to combine
the predictions of several individual models to obtain a more accurate final prediction. The use of
Stacking enabled us to reduce overfitting and biases. Stacking relies on the use of a meta-model that
learns by combining predictions from multiple individual models rather than relying solely on an
individual prediction. The meta-model integrates predictions from each individual model and produces
a more reliable final prediction by leveraging the strengths of each individual model. We were inspired
by the architecture of Stacking ensemble learning proposed in our previous article, which utilized
heterogeneous machine learning and deep learning architectures to predict subsequent substance use
disorder based on initial addiction [
In order to enhance the originality of our current work, we enriched our Stacking-based architecture by adding a significant extension. This extension aims to select the most suitable model from the trained individual models for the given data. By using coeficients for selection, the selected model stood out as the most suitable for the dataset’s characteristics. The decision to select this model is motivated by two factors: first, the increased complexity of the stacking architecture when dealing with large datasets, which results in longer computation times. Second, stacking models are often less interpretable compared to models with simpler architectures. Enriching the architecture allowed us to strike a balance between prediction accuracy and interpretability of the results, which is important for experts in the field of video game addiction to better understand the reasons for video game dependency. The rest of the paper is organized as follows: Section 2 provides an overview of the most recent relevant research. Section 3 provides an overview of the employed methodologies and tools, as well as a detailed explanation of the proposed architectures. Section 4 presents the empirical investigation and an elaborate analysis of the findings. Section 7 ends with conclusion and outlines the intentions for future investigations.
In this section, we present research conducted in the same field as our article, focusing on the application of machine learning and deep learning techniques to study video game addiction. Specifically, the results of these studies and the techniques used will be discussed, along with an attempt to identify their limitations.
The contribution proposed by Han, Xu, et al. [
In a similar context, another study conducted in 2019 [
Another study by Aggarwal et al. [
Finally, in the contribution proposed by [
In the context of our article, we attempted to select an appropriate dataset for the purpose of predicting
video game addiction. We found a dearth of datasets on gaming disorders. The datasets used in scientific
articles [
Finally, Our study is the first to utilize this dataset for applying machine learning algorithms to analyze video game addiction, which adds originality to our work.
The data sample used in our study consisted of video game players residing in two diferent countries, namely Gabon and Tunisia. The collected data included socio-demographic information such as age, gender, education, as well as behavioral variables related to players’ habits, mental health, and other relevant factors. The data showed variations based on the measured variables. Variables related to players’ habits included elements such as time spent on games, types of preferred games, frequency of gaming sessions, and other factors. Mental health-related variables encompassed measures of video game addiction, anxiety, depression, and other characteristics. An essential step in this study involves verifying the equivalence of the data used. It is crucial to work with equivalent data, meaning data that has a similar number of instances in both classes of interest: individuals with video game disorders and healthy individuals. Unbalanced data can introduce bias into the results of machine learning models. To assess the data equivalence, we generated a tabular representation of instances for both classes. We observed a disparity in the number of instances between the class of individuals with video game disorders and the class of healthy individuals. This disparity can bias the results and lead to classification errors. To address this imbalance, we employed the ADASYN (Adaptive Synthetic Sampling) method. This method generates new artificial instances of the minority class by extrapolating the characteristics of existing instances. Its objective is to increase the number of samples in the underrepresented class and achieve a balance between the two classes. We chose the ADASYN method among several similar techniques because it is optimized to reduce the risk of overfitting. This step is motivated by the need to balance the number of instances in both classes, allowing the machine learning model to learn patterns from both classes to minimize biases.
The architecture we proposed is based on three complementary modules. The first module, Individual Model Training (IMT), supports the training of individual base models developed earlier, namely Random Forest, Decision Tree, XGBoost, and SVM. This module uses the 5-fold Cross-Validation method. Once all the diferent models are trained, they will be used to make predictions on the validation data.
The second module, Combined Predicted Values (CPV), is responsible for creating a new matrix with new features. The CPV module concatenates the predictions from the base models to form the matrix X of size (n×m), where ’n’ represents the number of instances in the validation dataset for the meta-model, and ’m’ represents the number of base models.
The third module, Meta-Model Training (MMT), handles the training of the meta-model. The Logistic Regression algorithm was chosen as the meta-model in our proposal because it allows for generating coeficients for each base model. To train the meta-model, we used the resulting X matrix from the CPV module and added the corresponding labels as targets. Then, the meta-model makes a final prediction on the test dataset.
The following Fig.1 illustrates the operating principle of the proposed architecture.
The architecture proposed in the previous section ofers significant advantages in terms of prediction. Firstly, by combining the predictions of previously developed individual models, we can capture diferent aspects of the data, enhancing the robustness and performance of the meta-model. Each base model contributes its own expertise in detecting specific patterns or relationships in the data. However, this architecture has certain limitations, particularly regarding the interpretability of results for conducting a causal study. Indeed, video game addiction remains an important area of mental health, so it is essential for practitioners to understand the reasons that motivate the decisions made by the prediction models. By comprehending the underlying features and factors captured by the model, professionals in the field can make informed decisions about therapeutic practices to adopt and even customize these therapies to each individual addict’s profile. In this context, and since we are combining predictions from multiple models, it can be challenging to fully understand and explain the decision-making process of Logistic Regression in the final prediction. This can be a challenge if the goal is to conduct an investigation rather than a predictive study.
Another limitation of stacking-based architectures becomes apparent as the data volume increases. When the dataset becomes very large, using the stacking method can lead to computational time issues. The stacking architecture, therefore, requires training several base models on the training data, and then the models are used for predictions on the validation data. This significantly increases the complexity and computation time because each base model must be trained separately on large amounts of data. Furthermore, when dealing with large datasets, it can also be challenging to store and manage all the underlying model predictions for each instance in the validation dataset. This may require a substantial amount of memory capacity.
To overcome the limitations of the architecture presented in section 4.1, we have added a fourth module that collaborates with the MMT module from the previous architecture, aiming to select the best model suited to the dataset from the base models. This new architecture ofers significant advantages in scenarios with the previously mentioned limitations, which are execution time and interpretability. The added module, named Best Model Selection (BMS) (see Fig.2, facilitates the choice of the most performant model among the base models using the meta-model. This selection is made by using coeficients assigned by the Logistic Regression model to each base model based on their contributions to the final prediction. During the meta-model training process, coeficients are generated, which are values associated with the features in the X matrix (the predictions of the base models) in a metamodel (Logistic Regression). The initial values of these coeficients in the meta-model are typically set randomly during training. These coeficients determine the importance of each underlying model in the meta-model’s final prediction. These coeficients are then adjusted to minimize the error between the meta-model predictions and the corresponding labels. Therefore, the meta-model seeks to find a balance between the contributions of the diferent base models. This process is performed automatically with the goal of optimizing the coeficients to obtain a meta-model capable of accurately predicting the target label based on the base model’s predictions. In summary, the coeficients play a significant role in this process as they determine the relative weight of each underlying model in the final prediction. Base models that provide more information are assigned higher coeficients. Once the coeficients have been generated by the Logistic Regression model, we used the ’Argmax’ function, which selects the index corresponding to the maximum value in a set of values to choose the base model with the highest coeficients. This means it has a significant impact on the final predictions of the meta-model. By adding the BMS module, we can reduce computation time by identifying the most performant model among the individual models. This allows us to focus resources on a single model rather than distributing them across multiple models, which can significantly accelerate the prediction process. Additionally, the selected model is considered a more straightforward alternative to interpret.
The evaluation metrics that were used in our study are as follows: Recall: It is defined as the rate of True Positives (TP) or sensitivity, representing the proportion of correctly identified positive examples: However, it is straightforward to achieve a high recall by predicting that all examples are positive . Therefore, this criterion cannot be used alone and is often associated with precision.
Precision: refers to the proportion of correct predictions among positive predictions: Recall =
+
.
Precision =
+ .
Accuracy: Also known as the rate of correct classification, it is an evaluation criterion used in classification that measures the proportion of examples correctly classified among all examples: Accuracy =
+ + + + .
The results of the machine learning models we developed individually are summarized in Table 1:
By observing the performance metrics mentioned in this table, we found that SVM, XGBoost, and Decision Tree models seem to have relatively high scores with high precision and recall performances. However, it is important to note that in some cases, perfect scores may indicate the presence of overfitting issues.
The Random Forest model achieved perfect scores for all evaluated metrics. This could be an indicator of overfitting. Indeed, when a model achieves precision, recall, and accuracy of 1.00, it may reveal that it has overly adapted to the specific training data, making it challenging to generalize well to new data.
The Logistic Regression model shows slightly lower precision at 0.926 compared to other models, suggesting some improvement in terms of generalization.
We have incorporated confusion matrices for the models, and the results are presented in the figure below. The scores from these confusion matrices show that models based on Decision Tree and XGBoost algorithms performed well by successfully classifying positive (True Positive, TP = 23) and negative (True Negative, TN = 66) instances. However, they produced two false positives (False Positive, FP = 2), indicating a failure to correctly classify two non-addict individuals as addicts. No false negatives (False Negative, FN = 0) were observed, and an increase in FN values is unfavorable in our study, as it means the model is not correctly identifying truly addictive individuals.
On the other hand, the model based on the Random Forest algorithm demonstrated excellent performance compared to the previous models. This model stands out with a high number of true positives (TP = 25) and the same value for true negatives (TN = 66). No false positives (FP = 0) or false negatives (FN = 0) were observed, demonstrating its high ability to correctly classify instances. However, it is important to note that this perfect performance could be an indicator of overfitting tendencies. Therefore, these results require a validation study to assess the model’s generalization on new, unseen data.
Finally, the model based on the Logistic Regression algorithm obtained a lower number of true negatives (TN = 64) compared to SVM (TN = 66). This indicates that SVM is more efective in correctly predicting non-addictive individuals in the context of video game addiction. On the other hand, the results obtained with the deep learning model, represented in Fig.4 , show the curves of Recall, Precision, and Accuracy measures. The confusion matrix allows the analysis of the model’s predictions based on actual categories. The scores of this matrix are depicted in Fig.5 : The corresponding confusion matrix reveals that the model correctly predicted 64 individuals not afected by video game addiction and 21 individuals afected by addiction. However, the confusion matrix also indicates that there were 2 cases of individuals with addiction incorrectly classified as non-afected and 4 cases of non-afected individuals wrongly classified as having addiction.
We would like to emphasize that the performances of the individual base models within the stacking architecture have shown promising results in terms of accuracy, especially for XGBoost (0.944) and Random Forest (0.944). However, as highlighted in section 6.1, their tendency to overfit becomes apparent when data is limited. Figure 6 illustrates the performance of predictions from the basic models within the stacking architecture.
The meta-model achieved competitive performance compared to the base models (Accuracy: 0.95, Recall: 0.89, Precision: 0.87). Although the base models obtained highly accurate results, the meta-model combined their predictions in an optimal manner, thus producing a more realistic and robust prediction. For a better understanding of the performance of our meta-model (stacking-based architecture with three modules), Figure 7 presents the corresponding confusion matrix. Examining the values of this matrix, we observed that the model demonstrated a remarkable ability to predict non-addiction cases (True Negatives, TN = 237) as well as addiction cases (True Positives, TP = 51). This confirms the robustness of the stacking model in identifying players genuinely afected by video game addiction. On the other hand, the absence of null or high values for False Negatives (FN) and False Positives (FP) indicates that the model is not prone to bias. In other words, it does not tend to underestimate or overestimate the presence or absence of addiction in individuals. This characteristic is crucial for ensuring realistic outcomes from this model.
Furthermore, during the process of selecting the most suitable model for the dataset among the base models, the meta-model chose the Decision Tree model, even though it was not the best in terms of Accuracy score. The selection was based on the importance coeficients generated by the Logistic Regression model. These coeficients highlighted the significant contribution of the Decision Tree model in the final prediction of the stacking model. Despite other models, such as XGBoost, having a high Accuracy score, the Decision Tree model was considered the most contributing according to the meta-model. These results emphasize the importance of not solely focusing on performance metrics. Figure 8 illustrates the selection of the best model among base models by the meta-model based on their coeficients.
In conclusion, this study employed a sophisticated approach, utilizing stacking ensemble learning, to predict gaming disorder. By comparing the results of individual models and the stacking ensemble, we aimed to identify the most suitable predictive model for our dataset. The comprehensive evaluation allowed us to discern the strengths and weaknesses of ensemble learning model and individuals models, in addition, the selection of an optimal model that strikes a balance between predictive accuracy and reduced complexity. Our findings underscore the importance of methodological considerations in predictive modeling for gaming disorder. The tailored selection of a model based on dataset characteristics and performance metrics is crucial for achieving accurate predictions while mitigating unnecessary complexity. This study contributes to the evolving field of predictive modeling for behavioral disorders and highlights the potential of stacking ensemble learning as an efective strategy for optimizing predictive outcomes.
During the preparation of this work, the author(s) used ChatGPT for rephrasing and improving clarity of certain paragraphs, as well as Grammarly for grammar and spelling checks. All content generated or suggested by these tools was critically reviewed and edited by the authors. The author(s) afirm full responsibility for the accuracy, originality, and integrity of the final manuscript.