Transformer-Based Behavioral Representation Learning for Gamer User Engagement Detection in Virtual Reality Environments

1. Introduction

In recent years, online games have progressed into a complex digital ecosystem that facilitates players’ entertainment, social interaction, and economic activities [1]. In these digital environments, player engagement plays a key role in long-term participation, game progression, and monetization outcomes. The continuous engagement of players not only supports the longevity of online games but also influences the decisions associated with content updates [2]. Therefore, analyzing the factors that affect player engagement is an essential focus in game analytics and behavioral data science. Player engagement is a multi-dimensional concept that includes gameplay involvement, playing duration, and progression within the game [3]. Earlier researchers have analyzed player engagements based on various demographic information and game-play related features such as age, gender, region, genre, playing duration, level of difficulty, in-game spending, achievements, and engagement activities [4]. These features are very important for designing a predictive model that can improve the user experience and player retention. The main challenge in the gaming industry is to estimate players’ engagement levels [5]. Player engagement strongly influences the game’s popularity and income generation. However, predicting player engagement level is difficult due to the differences and complexity in player behavior [6].

The advances of Machine Learning (ML), a branch of AI that delivers the capability to learn from datasets using different learning strategies to classify or forecast outcomes by simulating player behavior in advance [7]. The ML model focuses on a system that gathers data, recognizes patterns, and learns from various situations to make decisions. For instance, in digital games, the game behavior can be enhanced autonomously without the need for direct programming [8]. Despite its effectiveness, many traditional ML approaches struggle to capture the non-linear, complex, and high-dimensional nature of gaming data. The advancement of Deep Learning (DL) models is better than the ML models when dealing with complex datasets, as it can learn patterns automatically from the data [9]. In addition, deep learning models use multiple layers that work in parallel, and they can automatically learn relevant features from the dataset. As a result, DL models offer more accurate and reliable outcomes compared to traditional approaches, making them well-suited for classifying and detecting player engagement level in digital gaming settings [10].

1.1. Motivation and Research Contributions

The rapid development of online gaming has created player involvement as an important factor for player retention and platform success. Conversely, engagement is impacted by risky and behavioral patterns that are complex to extract utilizing standard models. Conventional strategy frequently struggles to fully utilize rich user interaction information. This originates a requirement for an intellectual data-based technique that acquires significant patterns from behavioral data. To resolve these conditions, this research presents a Transformer-based Behavior Analysis Framework for Gamer Engagement Detection (TBAF-GED) using behavior data. The core contributions of this study are as follows.

• Introduce a correlation-based feature selection and autoencoder (CFS-AE) for effective elimination of redundant and irrelevant features while learning compact and informative feature representations.
• Develop a Feature Tokenizer Transformer (FT-Transformer) approach to convert gamer behavior features into token embeddings and model their interactions using self-attention. It performs the classification of engagement levels reflecting player retention (High, Medium, Low).
• Integrate a Nadam optimizer to efficiently update model parameters and improve training stability during optimization.
• Validate the performance of the proposed model using the Player Engagement Analysis & Prediction dataset, and demonstrate its effectiveness using various evaluation measures.

The rest of the article is proceeded as follows: Section 2 presents an existing related work on gamer engagement detection. Section 3 offers overview of the proposed method, Section 4 offers the results and discussion, and Section 5 concludes the work with future directions.

2. Related Works

This section presents a concise overview of related works on gamer engagement detection, highlighting identified research gaps. Rashed et al. [11] developed a new architecture for real-time, nonintrusive, and indirect measurement of engagement in multiplayer online games that depends on the flow scheme. The study integrates graph convolutional networks (GCN) for modeling player connections with Transformer systems for temporal processing, empowering indirect measurement of game challenge and player skill, which is utilized to categorize player engagement. Ozyirmidokuz et al. [12] discussed the increasing prevalence of digital game addiction (DGA) by deploying an information scheme for identification and initial involvement, for predicting an individual’s inclination for DGA. A data augmentation approach was employed to improve dataset diversity and improve the method’s generalization ability. Subsequently, wide-ranging preprocessing was trained and assessed six ML strategies. Yin et al. [13] designed a new attention-driven neural point progress system to extract the complexity and richness of customer behavioral dynamics in the current digital context of consumption. The method features a novel multiple-representation, continuous-time attention system that can reliably model dynamic connections among distinct categories of behavior across external effects.

Wang, Y. and Lee [14] proposed a methodology to enhance game suggestions and operations by accurately forecasting users’ preferences by integrating the spatiotemporal activity and social ties of game players. Liang et al. [15] investigated a DL-driven user behavior and interaction mechanism optimization approach. To construct a BiLSTM-Attention behavior detection algorithm, it understands model behavior classifier of higher-dimensional operation information, and fine-tunes interaction methodology and response hyperparameters in conjunction with the categorization solution for optimizing the game comments system. Han et al. [16] introduced an advanced churn prediction strategy that exploits user activity and social collaboration information from the huge, multiple-person online role-playing game, Soul and Blade. This approach distinctively predicts persons’ interactions as a graph structure and improves forecasting precision by combining a correct and smooth (C&S) model with a GCN model in social network examination.

Lauer-Schmaltz et al. [17] introduced the model and working of an HDT approach for upper-limb stroke rehabilitation utilizing exoskeletons and critical games, regarding the ETHICA technique. The study validates HDT real-world modifications to therapy problems and exoskeleton support depends on player situations, improve relationship between non-medical and medical stakeholders by decision-support system and data visualizations, and increase player involvement via personalized comments. Marín-Lora et al. [18] examined a mobile Exergame developed particularly for treadmill running, using interface with a smartwatch. The game poses natural settings where smartwatch network, it illuminates the gamers behaviour, converting into running speed and collaborating activities by identifying signals in the game. The major goal is to deliver players with a sustaining gaming involvement tailored to the features of treadmill running. Particularly focus on prioritize the playful element of this game, identifying its relevant context of treadmill running.

For evaluating the success of goals and the presented hypothesis, comparative research was shown between a treadmill running simulator and the presented game.

Existing studies on gamer player engagement recognition mostly utilize statistical examination and conventional ML approaches for analyzing user behavior information, including actions, session time, and interaction occurrence. These techniques, containing Random Forest and SVM, enhance prediction; however, they are highly dependent on feature engineering. Latest research investigates DL strategies like Transformers and LSTM to extract sequential and intricate behavioral patterns. AE methodologies are also used for better feature representation. The proposed model exists to work towards the goal of improving engagement information precision for better personalization and player retention.

3. Material and Methods

This research follows a multi-faceted approach for gamer engagement detection using behavior data. The Player Engagement Analysis & Prediction dataset serves as the foundation for model development and evaluation. Figure 1 showcases the end-to-end architecture of the TBAF-GED approach. The proposed model initially pre-processes the input data through data cleaning, label encoding, outlier handling, and robust scaling to ensure its quality and usability. Additionally, CFS-AE will be performed to eliminate the dimensionality of the dataset and enhance the model efficiency. Besides, the FT-Transformer classifies gamer engagement levels by modeling complex interactions among behavioral features using self-attention. Finally, the Nadam optimizer is leveraged to optimize model parameters.

3.1. Data Acquisition and Preprocessing

Dataset collection involves gathering relevant data from reliable sources for model development and analysis. Data preprocessing involves cleaning, transforming, and preparing the collected data to improve quality and ensure effective model performance.

3.1.1. Dataset Description

The proposed TBAF-GED technique has been conducted by employed Player Engagement Analysis & Prediction dataset [19]. This dataset offers a comprehensive view of demographics, player behavior, and engagement in online gaming surroundings, thus creating it valuable for player retention analysis and engagement prediction. This can be comprised features namely Age, Gender, Location, PlayerLevel, AchievementsUnlocked, and SessionsPerWeek. The target variable, EngagementLevel, classifies players into High, Medium, or Low engagement groups, thereby generating the dataset appropriate for player behavior analysis and engagement prediction. The overall features are 10 and chosen from the original 13 features for evaluation. Table 1 presents three engagement level classes such as High, Medium, and Low with a total of 40,034 samples.

3.1.2. Data Preprocessing

The data preprocessing step encompasses data cleaning, label encoding, outlier handling, and robust scaling to ensure the quality and reliability of the data [20].

Data Cleaning: This method is a primary pre-processing technique, which includes classifying and modifying missing, unreliable or noisy values in the features. Irrelevant entries and duplicate records are eliminated to enhance the quality of data.

Categorical Encoding (Label Encoding): Categorical features are transformed into numerical form employing a label encoder. All single classes are allocated different integer data to generate the data transformation with ML approaches.

Outlier Handling: Outliers are identified utilizing statistical approaches such as the Interquartile Range (IQR). Maximum values that diverge greatly from the normal function are either detached or stopped to minimize their influence.

Robust Scalar: This scaling method is a feature standardization approach that scales values employing statistics that are less sensitive to outliers. It converts values depending on the median and IQR, making it efficient for incorrect data. This assures that maximum values have a minimum effect on the scaler function.

$$x^{\mathrm{\text{'}}}={\displaystyle \frac{x-\mathrm{median}\left(x\right)}{IQR}}$$

(1)

In this Eq. (1), the scalar data is attained for subtracting the median from $x,$ which centers the data. It is separated by the IQR to minimize the impact of outliers and normalize the spread.

3.2. Correlation-Based Feature Selection Strategy

CFS-AE is applied to reduce the dimensionality of the dataset by removing redundant and irrelevant features. It also enhances model efficiency by learning compact and informative feature representations [21].

Autoencoders: AEs are an unsupervised learning method that encodes input values into a compressed hidden representation and decodes it to rebuild an estimation of the real input data. The key AE contains the encoder that processes the bottleneck and the input. The function of the encoding is to map a higher dimension vector $H=(h1,h2,h3,h4,\dots ,hn)$ towards a low-dimension vector $L=(\mathrm{11,12,13,14},13,14,\dots ,\mathrm{l}\mathrm{n})$. This low-dimensional representation takes the most significant behavioral patterns and called the bottleneck.

$$L=f\left(WH+b\right)$$

(2)

Besides, the decoding is the process between the output and hidden layers. Within the context of gamer behavior reconstruction, the decoding captures the compact behavioral representation $L$ and rebuilds a higher-dimensional feature embedding similar to the real input data $H$.

$$X=g\left(W\mathrm{’}L+b\mathrm{’}\right)$$

(3)

In these Eqs. (2) and (3): $f$ and $g$ imply the function of activation maps, $W$ and $b$ signify the encoding’s weights and bias, $W’$ and $b’$ represent the decoding’s weights and bias.

Correlation-based Feature Selection: This technique is used in gamer user behavior classification methods to decide the most important behavioral features by employing a search approach that includes backward reduction, genetic methods, directional search, forward selection, and best-fit approaches. It is a broadly accepted model for improving the performance of the method by choosing the most informative attributes for precise engagement forecast. The FS method is represented as follows:

$$R_{xc}={\displaystyle \frac{n{R}_{xi}}{\sqrt{n+n\left(n-1\right)R_{ii}}}}$$

(4)

Let $R_{xc}$ denote the correlation that occurs among the total subset of attributes and class variables, $n$ implies the subset of attributes $R_{xi}$ represents the average correlation that occurs among the subset of attributes and the class variables, and $R_{ii}$ signifies the inter-correlation average that occurs among subsets of features.

Although the method might still suffer from reconstructed loss owing to noisy higher-dimension behavioral information, redundancy between correlated attributes, overfitting because of extreme model difficulty, and minimized interpretability impacted by dense hidden representations. To solve this problem, the succeeding united loss function is reduced:

$$L=\parallel H-X\parallel 22\mathrm{}+\lambda 1\mathrm{}j\in S\sum \mathrm{}\left(1-Rxj\mathrm{}c\mathrm{}\right)+\lambda 2\mathrm{}\parallel W\parallel F2\mathrm{}+\lambda 3\mathrm{}\parallel L\parallel 1\mathrm{}$$

(5)

Reducing this united loss jointly tackles these challenges by maintaining vital engagement patterns across reconstruction, choosing only relevant features and behavior through correlation limitations, enhancing generalization utilizing weight regularization, and applying sparsity in the bottleneck representation to process more discriminative engagement attributes for precise analysis of user engagement.

3.3. Gamer Engagement Classification Module

The FT-Transformer is employed as the classification model for gamer engagement detection. It classifies engagement levels by modeling complex interactions among behavioral features using self-attention [22]. The FT-Transformer is applied for classifying player engagement employing behavioral features. The method is intended to manage tabular behavioral data and identify the engagement via a numerical feature tokenizer, an MLP classifier head, and a self-attention-driven feature relationship model. A transformer encoding method encodes global interactions between attributes. The [CLS] outcome is fed to an MLP to forecast engagement level, such as lower, medium, or higher engagement. In this context, the key concept of FT-Transformer is to consider every numerical attribute as an independent token. With the input instances $x=[x_1,x_2,\dots ,x_d]\in R^d$, all numerical attributes $x_i$ has mapped into feature tokenization:

$$t_i=x_i{W}_i+b_i+e_i,i=\mathrm{1,2},\dots ,d$$

(6)

In this Eq. (6), $W_i\in {\mathbb{R}}^k$and $b_i\in {\mathbb{R}}^k$ signify the trainable projection embedding and bias term, $e_i\in {\mathbb{R}}^k$ implies the trainable vector features, and k indicates the token dimension. For achieving each feature token, a trainable classifier token $t_{cls}$ is proposed to build the resulting input sequence of the token:

$$T_0=\left[t_{cls};t_1;t_2;\dots ;t_d\right]$$

(7)

Here, $t_{cls}$ represent the trainable classifier token, ${\mathrm{T}₀\in \mathbb{R}}^{\left(\right(\mathrm{d}+1)\times \mathrm{k})}$signifies the matrix of input tokens, and “;” implies concatenation with sequence size. Let d indicate the count of behavioral data, and k denotes the size of the token. The constructed sequence is then passed to multi-layer transformer encoding to train global dependency between behavioral patterns.

$$\mathrm{A}\mathrm{t}\mathrm{t}\mathrm{n}\left(Q,K,V\right)=softmax\left({\displaystyle \frac{Q{K}^T}{\sqrt{d_k}}}\right)V$$

(8)

The multi-head attention function is represented as:

$$MultiHead\left(Q,K,V\right)=Concat\left(hea{d}_1,hea{d}_2,\dots ,hea{d}_h\right)W^O$$

(9)

Here, Q implies the query matrix, K indicates the key matrix, and V signifies the value matrix; correspondingly, $d_k$ represents the key vectors size$,hea{d}_{\mathrm{i}}$ implies the outcome of the i-th attention head, $W^O$ refers to the matrix of outcome projection, and $h$ implies the count of attention heads. The outcome of the $l_{th}$ transformer encoding layer is provided by:

$$T_l=EncoderLayer\left(T_l-1\right),l=\mathrm{1,2},\dots ,L$$

(10)

Here, $T_l$implies the outcome of the l-th layer, $T_l-1$ denotes the prior layer values, and L signifies the overall count of encoding layers. At last, the performance of the CLS token is utilized as the global behavioral features. The engagement class likelihood has been calculated as:

$$p=softmax\left(MLP\left(h_{cls}\right)\right)$$

(11)

Now, MLP (·) represents the multilayer perceptron classifier head, $h_{cls}$ denotes the resulting latent representation of the CLS token, and p implies the probability function throughout engagement labels. The engagement state detection is attained as:

$$\widehat{y}=argmax\left(p\right)$$

(12)

Therefore, the method achieves engagement identification for all gamer behavior instance that depends on the global representation combined by the CLS token. The challenge of FT-Transformer is its minimized robustness with class imbalance in engagement conditions that might bias detection to the majority categories. To tackle this, a weighted cross-entropy loss has been proposed:

$$\mathcal{L}=-{\displaystyle \sum _{c=1}^C}{w}_c{y}_c\log \left(p_c\right)$$

(13)

In this Eq. (13), $C$ represents the count of engagement classes, $w_c$ implies the class weight, $p_c$ refers to the forecasted probability and $y_c$ implies the actual label. This weighting module allocates high significance to minority engagement categories, assuring balanced training through each class. So, it enhances detection consistency and decreases bias to dominant engagement states in training. Figure 2 shows the architecture of the FT-Transformer-based classification model.

3.4. Parameter Optimization

The NAdam optimizer is an advanced optimization algorithm that combines Adam with Nesterov momentum. It is used to accelerate convergence and improve the stability and efficiency of model training [23]. Nadam optimizer algorithm is used to improve the training effectiveness and prediction solution of the engagement classifier tasks. This optimizer integrates the adaptive learning ability of the Adam with Nesterov momentum (NAG) to enable stable convergence when learning difficult gaming patterns from behavioral data. The key concept of NAdam is it applies NAG’s look-ahead update.

$${\theta }_{temp}={\theta }_{t-1}-{\beta }_1\mathrm{*}{\displaystyle \frac{m_{t-1}}{1-{\beta }_1^{t-1}}}$$

(14)

Afterward, the gradient has been measured at the point instead of the temporary value to prevent the upgrade of redundancy computation.

NAdam performs look-ahead through mathematical evaluation. This momentum approximation $\widehat{m}$ is computed as

$${\widehat{m}}_t^{nesterov}={\beta }_1\cdot {\displaystyle \frac{m_t}{1-{\beta }_1^t}}+\left(1-{\beta }_1\right)\cdot {\displaystyle \frac{g_t}{1-{\beta }_1^t}}$$

(15)

This formula estimates the momentum condition at the look-ahead value ${\theta }_{temp}$ for weighting the recent bias-corrected momentum ${\widehat{m}}_t^{nesterov}$ with ${\beta }_1^{t+1}$, which implies the decayed coefficient of another stage, then $g_t$ denotes the recent gradient, and this approximation of the look-ahead momentum substitutes the Adam momentum estimation. The resulting parameter upgrade as follows:

$${\theta }_t={\theta }_{t-1}-\alpha \cdot {\displaystyle \frac{{\widehat{m}}_t^{nesterov}}{\sqrt{V_t}+e^{\mathrm{\text{'}}}}}$$

(16)

In this Eq. (16), $v_t$ signifies the 2nd-moment estimation, and its biased correction follows the same expression as Adam. NAdam incorporates adaptive learning rates with NAG, making it efficient for heterogeneous and higher-dimensional gaming behavior information.

Table 2 portrays the hyperparameter settings employed across various phases. The model utilizes CFS with the correlation threshold of 0.85 to eliminate extremely correlated features and decrease redundancy. An autoencoder with dropout of 0.2, a latent dimension of 64, and 2 hidden layers (256 neurons each) has been employed for nonlinear feature compression. A Feature Tokenizer Transformer with dropout 0.2, embedding dimension 128, feed-forward size 5124 layers, and 4 attention heads is utilized for feature representation learning. Training is accomplished by implementing the Nadam optimizer with a learning rate of 0.001 (β₁=0.9, β₂=0.999). The model is trained with a batch size of 32 for 25 epochs, and early stopping with a patience of 10 is employed to avoid overfitting.

3.5. Model Evaluation Metrics

In this research, we used several performance metrics to assess our presented method, such as accuracy, recall, F-score, precision, and Matthew’s correlation coefficient (MCC), which were all discussed in the examination. A confusion matrix is a table that defines how well a categorization approach executes, and it summarizes and visualizes the prediction outcomes for a classification issue. In a confusion matrix, true negative $\left(TN\right)$ signifies a value that is correctly predicted as negative, true positive $\left(TP\right)$ denotes a value that is correctly predicted as positive, false negative $\left(FN\right)$ illustrates a value incorrectly predicted as negative, and false positive $\left(FP\right)$ represents a value incorrectly predicted as positive. Mathematically, all these evaluation metrics can be represented as:

$$Accuracy=\left({\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}\right)\times 100\mathrm{\%}$$

(17)

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(18)

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(19)

$$Fscore={\displaystyle \frac{2\times (Precision\times Recall)}{Precision+Recall}}$$

(20)

$$MCC={\displaystyle \frac{(TP\times TN)-(PP\times PN)}{\sqrt{(TP+PP)\times (TP+PN)\times (TN+PP)\times (TN+PN)}}}$$

(21)

4. Findings and Discussions

The performance analysis of proposed TBAF-GED model is inspected using Player Engagement Analysis & Prediction dataset. The correlation heatmap shown in Figure 3 illustrate the features used for gamer engagement level classification. The heatmap demonstrate the relation amongst player activity, progression, and achievement related attributes. Strong positive correlation is observed between all selected features, which exhibits the meaningful information for accurate gamer engagement level classification.

Figure 4 illustrate the exploratory data analysis of video game player engagement through multiple visualization. The dataset includes players from various age group with both genders, which supporting balanced analysis. The scatter plot highlights the interaction amongst players level and achievement unlocked, which reflects players exhibiting high participation. Furthermore, the analysis of weekly session reveals highly engaged players maintains regular participations in gaming activities compared to low and medium level engagement players.

Figure 5 shows the classifier result of the proposed model using 80% of TRAIP. Fig. (5a) displays the confusion matrix; the majority of classes predicted correctly with minimal classification error which indicates the effectiveness of the proposed model. Fig (5b) represents the PR curve analysis exhibits strong predictive performance.

Table 3 summarize the player engagement level result with 80% of TRAIP. The training phase indicates strong classification performance across all metrics and achieved average accuracy of 92.23% with balanced precision, recall, and F-Score values, which indicating the consistent class discrimination. The average MCC value of 81.49% further validates the stable and reliable performance of the proposed model during training phase.

Figure 6 present the classifier result of the TBAF-GED method under confusion matrix and ROC curve of the 20% of TRAIP. Figure 6a, reveal the confusion matrix of the TBAF-GED method that correctly identified engagement level with 1707 samples as lower, 3554 samples as medium and 1778 samples as higher. Likewise, Figure 6b exhibit the ROC inspection of the TBAF-GED system offers optimal classification through each three categorizes.

Table 4 displays the player engagement level classifier outcome of the proposed model using 20% TESTP. It is evident that the proposed model reaches excellent performance by achieving average accuracy of 91.94% and average F-Score value of 87.07%, which highlight the effective classification capability of model. The close alignment between precision and recall further validate the balanced classification across all classes. Moreover, the average value of MCC with 80.77% confirms the stability and robustness of the proposed model during testing phase.

Figure 7 displays the classifier outcome of the TBAF-GED method with confusion matrix and PR curve of the 70% of TRAIP. Figure 7a, shows the confusion matric of the TBAF-GED technique, which precisely categorized engagement level with 6530 instances as low, 12847 instances as medium and 6335 instances as high. Also, Figure 7b portrays the PR analysis of the TBAF-GED method, thus representing maximal performance at all classes.

Table 5 represent the engagement level classifier outcome of the TBAF-GED approach with 70% of TRAIP. The proposed TBAF-GED technique accomplished an average value with $accu{r}_y$ of 94.50% and MCC of 86.85%, representing consistent and strong classification performance.

Figure 8 shows the classifier outcome of the TBAF-GED approach with confusion matrix and ROC curve of the 30% of TRAIP. Figure 8a, present the confusion matric of the TBAF-GED model, which accurately classified engagement level with 2826 instances as low, 5463 instances as medium and 2687 instances as high. Also, Figure 8b shows the displays the ROC analysis of the TBAF-GED method, demonstrating that the system provides effective classification across all three classes.

Table 6 shows the engagement level classifier outcome of the TBAF-GED system with 30% of TESTP. The proposed TBAF-GED technique succeeded an average value with $accu{r}_y$ of 94.26% and MCC of 86.29%, presenting constant and high-quality classification outcomes.

Figure 9 illustrate the classifier performance of the TBAF-GED model by offering the accuracy and loss analyses. The $accu{r}_y$ curves specify that the TBAF-GED method realizes gradually higher values as the number of epochs rises, considering efficient learning and convergence. Likewise, the validation $accu{r}_y$ closely followed the training accuracy, signifying that the model generalizes better to unnoticed data. The loss curves exhibit that the training and validation losses converge to significantly correlated values. This closer arrangement amongst training and validation loss further ensures that the TBAF-GED system gains efficiently and demonstrate stronger generalization ability with lower overfitting.

and (b) Loss curves.

Table 7 depict the comparison outcome of the TBAF-GED approach with other methodologies using evaluation metrics [24,25,26]. The findings shown that the EEGNet, ShallowConvNet, and CSP + LDA methods have attained minimal performance, while the LSTM and ResNet18+NN models have got slightly greater value. Moreover, the Decision Tree, and Random Forest methodologies have gained moderate performance. At the same time, the proposed TBAF-GED system has achieved maximal solutions with $accu{r}_y$ of 94.50%, $prec{i}_n$ of 91.43%, $reca{l}_l$ of 90.92%, and $F_{score}$ of 91.17%. These results confirm the superiority and efficiency of the proposed system over the other methodologies.

An ablation study calculates the contribution of various modules by selectively eliminating or modifying them to notice performance variations. It supports to identify the significance of every component and gives insights into optimization and model behavior. Table 8 present the ablation analysis of the TBAF-GED method. The outcome implied that the Without Feature Selection, Correlation-Based FS Only + FTT, Autoencoder Only + FTT, and Correlation-Based FS + Autoencoder + FTT (without Nadam) approaches obtain lower performance with various metrics, whereas the TBAF-GED (Preprocessing + Correlation FS + Autoencoder + FTT + Nadam) method achieve higher performance with $accu{r}_y$ of 94.50%, $prec{i}_n$ of 91.43%, $reca{l}_l$ of 90.92%, and $F_{score}$ of 91.17%. This noticeably show the efficacy of integrating all proposed components into integrated structure.

Table 9 classifies game difficulty levels and their associated genres, mapping low difficulty to RPG and Simulation, medium to Strategy, and high to Action and Sports. This classification supports to comprehend how game genres associate with modifying difficulty stages.

The proposed technique for gamer player engagement identification combines a stronger pipeline comprising data preprocessing, CFS-AE-driven feature extraction, FT-Transformer for classifier, and Nadam optimization for effective learning. It efficiently operates noisy, higher-dimension behavioral data by integrating association-based filtering with higher autoencoder extraction learning. The FT-Transformer improves the outcome by extracting risk feature interactions and sequential dependency in users’ behavior. In comparison with standard ML techniques, the proposed architecture delivers enhanced precision and superior generalization on datasets, namely the Kaggle user engagement dataset.

5. Conclusion and Future Work

This research has introduced a TBAF-GED using behaviour data. The proposed model employs a CFS-AE to reduce redundancy and irrelevant features in the feature space. For gamer engagement detection, the FT-transformer provides a novel tokenization model that transforms behavioral features into embeddings and captures their complex interactions through self-attention. Additionally, the Nadam optimizer is utilized to update model parameters and improve convergence stability during training. Experimental evaluation demonstrates that the proposed TBAF-GED framework attains promising performance, with an accuracy of 94.50% compared with existing approaches. The final outcome categorizes gamer engagement levels based on game difficulty and genre as Low (RPG, Simulation), Medium (Strategy), and High (Action, Sports). The proposed model supports game analytics by enabling proactive identification of gamer engagement levels from behavioral data. It helps improve game design decisions and enhances player retention through intelligent engagement analysis. Although the proposed technique may face difficulties in scalability on extremely huge datasets because of higher computational needs, and may demonstrate minimized results in hugely different or unobserved gaming settings without appropriate field adaptation. In the future, it can concentrate on building lighter transformer techniques to decrease computation and assist real-world involvement recognition. XAI approaches can be further added to create predictions that are most understandable for developers. Superior strategies, such as contrasting learning or data reinforcement, can enhance outcomes on unnecessary information. The system can be encompassed for cross-gaming and real-world adaptative engagement prediction to enhance practical usage.