SELM-CTR: a stacking ensemble deep learning model with SHAP-based analysis for large-scale click-through rate prediction

In the rapidly evolving landscape of digital advertising, the ability to accurately predict Click-Through Rate (CTR) is of paramount importance. CTR is a fundamental metric that measures user engagement by determining the likelihood of an individual clicking on an advertisement. This prediction task is crucial for optimizing ad placements, enhancing targeting strategies, and maximizing Return on Investment (ROI) (Wang et al., 2018). With digital advertising expenditure increasing globally, precise CTR prediction has become more critical than ever. In 2020 alone, digital advertising revenue surpassed $139.8 billion, with a significant portion attributed to programmatic advertising and personalized recommendation systems (Mordor Intelligence, 2023). The effectiveness of these systems relies on robust CTR models that can adapt to evolving consumer behavior. Given these trends, the need for scalable and interpretable deep learning-based CTR prediction models is evident (Zheng et al., 2022).

However, the complexity of user behavior and the dynamic nature of online interactions pose significant challenges for traditional Machine Learning (ML) approaches in effectively modeling CTR patterns (Kamal and Bablu, 2022).

To address these challenges, Deep Learning (DL) techniques have emerged as powerful tools capable of capturing intricate feature interactions and nonlinear dependencies (Yang and Zhai, 2022). Traditional models, such as Logistic Regression (LR) and Factorization Machines (FM), have been widely used for CTR prediction due to their computational efficiency (Chen et al., 2016). However, these models fall short in handling high-dimensional sparse data and learning higher-order feature representations (Chapelle et al., 2014). In contrast, DL architectures like Deep Neural Networks (DNN), Deep Factorization Machine (DeepFM), and Gated Deep Cross Networks (GDCN) have demonstrated superior performance in capturing complex user interactions and improving prediction accuracy (Zhang et al., 2016). Nevertheless, while these models achieve high predictive power, they often suffer from a lack of transparency, making it difficult to interpret how different features contribute to the final prediction (Thejas et al., 2021).

This study proposes an ensemble deep learning approach that integrates multiple advanced architectures with a stacking ensemble technique, leveraging XGBoost as a meta-model. Ensemble learning methods, such as stacking, bagging (Breiman, 1996), and boosting (Schapire, 1999), have been widely used to improve predictive accuracy. By combining the strengths of DNN, DeepFM, and GDCN, the proposed method enhances predictive accuracy while maintaining model interpretability through Shapley Additive Explanations (SHAP). The inclusion of SHAP enables a transparent understanding of feature importance, providing insights into user behavior and ad relevance (Krishna et al., 2021).

The contributions of this study are as follows:

1. A stacking ensemble approach integrating DNN, DeepFM, and GDCN to improve CTR prediction performance.

2. Utilization of SHAP analysis to quantify feature contributions, providing actionable insights for advertisers.

3. Extensive evaluation of the proposed model on the AVAZU dataset, demonstrating its effectiveness in improving AUC and reducing log loss.

This study is organized as follows: Section 2 presents related work, highlighting recent advancements in CTR prediction. Section 3 details the proposed methodology, including the DL architecture and ensemble strategy. Section 4 discusses the experimental results and performance evaluation. Finally, Section 5 concludes the study with insights and future research directions.

CTR prediction has evolved through three broad research directions: classical ML models, ensemble-based approaches, and DL architectures. This section reviews key contributions in each direction and identifies the gaps that motivate the proposed method.

Early CTR prediction relied on Logistic Regression (LR) and Factorization Machines (FM). LR is simple and efficient, but it cannot model feature interactions beyond what is manually engineered. FM addressed this limitation by representing the pairwise feature interactions as inner products of latent vectors, making it more effective on sparse, high-dimensional data (Sahllal and Souidi, 2023). However, FM still struggles with class imbalance. Sahllal and Souidi (2023) showed that combining Undersampling with Random Forest improved CTR prediction accuracy by up to 20% by better handling imbalance ratios and feature importance, though the approach was limited to specific industry settings and did not generalize well across domains.

Subsequent work extended FM with richer interaction mechanisms. Raphael et al. (2022) combined K-Means clustering with a Field-aware Neural Factorization Machine (FNFM), achieving 77.9% accuracy on the AVAZU dataset. Yang et al. (2023) proposed the Attention-based Feature Interaction DeepFM (AFI-DeepFM), which uses a Kronecker product to extract fine-grained feature interactions, achieving 75% AUC and 25% log loss. While these methods improve interaction modeling, they do not scale effectively to very high-order feature dependencies or large sparse datasets, motivating the shift toward deep learning.

Ensemble methods improve CTR prediction by combining multiple models to reduce variance and offset individual model weaknesses. He et al. (2020) proposed a GBDT-Stacking model that enhances feature extraction for sparse data, achieving 87.3% accuracy. Despite this strong result, the method was designed for a narrow set of industry contexts and lacks broader applicability. Toğuç and Kuzu (2020) demonstrated that gradient boosting ensembles extend naturally to click fraud detection, achieving 76.1% accuracy on AVAZU with notable gains in precision, recall, and AUC. Their work confirmed the versatility of ensemble methods but did not explore hybrid architectures that combine deep learning with tree-based models.

Biçici (2023a) applied instance weighting across neural networks for CTR prediction on AVAZU, yielding 73.0% accuracy, while Biçici (2024) evaluated Extreme Learning Machines (ELM and ML-ELM) on the same dataset, achieving 71.7% accuracy. Both results fall short of more expressive architectures. Liu et al. (2024) combined ensemble learning with a Reinforcement Learning (RL) strategy for multi-class CTR prediction, reaching 79.2% accuracy. Demirel et al. (2023) explored Bayesian hyperparameter optimization for non-tree ensemble models, achieving 61.0% accuracy on AVAZU while highlighting the critical role of preprocessing and feature elimination. He and Fan (2021) combined tree-based methods with an Inner Product-based Neural Network (IPNN) for default prediction, achieving 71.9% accuracy. Across these studies, a recurring limitation is the absence of explainability: none of the ensemble methods provides a principled mechanism to interpret how features drive individual predictions.

Deep learning models have become the dominant approach for CTR prediction because they can automatically learn high-order feature representations from raw data. Shi et al. (2024) proposed the DDNNFMFM model, which combines a Dense Deep Neural Network with a Field-Matrixed Factorization Machine, achieving 79.2% accuracy on 40 million AVAZU samples. Li et al. (2024) introduced DCNv3, along with Linear Cross Network (LCN) and Exponential Cross Network (ECN) variants, achieving 79.7% AUC by improving the cross-feature interaction mechanism of the original Deep Cross Network. Wang et al. (2023) combined GDCN, Field-level Dimension Optimization (FDO), and Graph Convolutional Networks (GCN) to achieve 79.05% accuracy on the same dataset, demonstrating that graph-based interaction modeling can complement cross-network architectures.

Temporal dynamics and user behavior have also received considerable attention. Gan and Xiao (2019) developed the R-RNN model, which applies attention mechanisms and LSTMs to capture recent user behavior sequences, outperforming DIN-based models with 85.5% accuracy. Kim et al. (2021) proposed DUSIN, a user segmentation interest network that groups users by similar behavioral patterns, achieving 63.5% AUC on AVAZU. Tian et al. (2022) enhanced the AutoInt model by integrating bidirectional LSTMs under a differential privacy constraint, reaching 81.2% accuracy while preserving user data privacy. Wei et al. (2023) combined MLP with Factorization Machines and local attention in the DAIN model, capturing both low- and high-order interactions and achieving 82.9% accuracy across Amazon datasets.

Interpretability has emerged as a growing concern in deep CTR models. Jose and Shetty (2022a) addressed this with the Neural Additive Factorization Model (NAFM), which reduces computational cost while maintaining 78.0% accuracy and 37% log loss, providing interpretable feature contributions. Jose and Shetty (2022b) further proposed distilled DNN and DDNN models for real-time CTR systems, achieving 78.4% accuracy with 34.5% log loss by emphasizing scalability. Biçici (2023b) introduced the FFCP model for sparse data, demonstrating competitive AUC and F1 scores without requiring complex feature engineering. Liu et al. (2021) applied Graph Convolutional Networks (GCN-int) to simplify high-order feature interactions, achieving 78.6% accuracy with 37.6% log loss on AVAZU. Feng et al. (2024) proposed GMINN, a gate-enhanced multi-space interaction network that reached 81.33% accuracy. Wang et al. (2024) advanced this line of work with I-DeepFM, which incorporates customizable regularization and attention mechanisms, achieving 87.4% accuracy on both Criteo and AVAZU. Most recently, Zeng et al. (2025) proposed the Adaptive Deep Neural Network (ADNN), achieving 78.9% accuracy on AVAZU through adaptive feature interaction learning. Yilmaz and Aldemir (2024) explored mutual information-based feature interactions for FiBiNET and DCN, achieving 81% AUC on AVAZU, while Mao et al. (2023) proposed FinalMLP, a two-stream MLP architecture with feature gating that achieves 76% AUC on AVAZU. Zhu et al. (2024) proposed Helen combined with SAM optimizer, attaining 81% AUC. Wei et al. (2023) introduced RE-SoRT, a relevance-sorting framework achieving 74.1% AUC. Aksu et al. (2024) proposed NeSHFS, a neural surrogate-based hybrid feature selection method that reaches 74% AUC on AVAZU. Figure 1 shows the chronological progression of CTR prediction research.

Despite these advances, two limitations persist across the literature. First, most methods optimize for predictive performance without providing a transparent mechanism to explain the individual feature contributions, which is critical for advertiser trust and regulatory compliance. Second, while ensemble methods and deep learning models have been explored separately, few studies systematically combine them with feature selection under a unified framework. This study addresses both gaps by proposing a stacking ensemble of DNN, DeepFM, and GDCN with XGBoost as the meta-model, augmented with SHAP-based feature selection and SMOTE for class balancing, achieving superior CTR prediction performance alongside interpretable feature importance analysis.

In this section, proposed DL models and techniques employed to predict CTR are explained. A combination of state-of-the-art models and ensemble strategies have been utilized to effectively capture intricate relationships in the data. Each model is carefully chosen based on its ability to handle high-dimensional sparse data, temporal and categorical feature interactions, and scalability for large datasets. These models include a DNN with residual connections, a DeepFM model, and a GDCN. Additionally, Stacking Ensemble learning techniques are employed to combine predictions from these models.

Figure 2 shows the workflow of the proposed methodology. This study uses the publicly available AVAZU dataset.

First, numerical and categorical features are identified from the dataset. Numerical features are scaled using min-max scaling, while categorical features are encoded using label encoding. Data preprocessing includes handling missing values, removing redundant columns, extracting temporal features, and removing duplicate entries.

The dataset is split into training (80%) and testing (20%) sets. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) (Chawla et al., 2002) is applied only to the training data. SMOTE balances the minority class (Class 1: click) against the majority class (Class 0: non-click). Figure 3 shows the class distribution before and after applying SMOTE. After balancing, three models–DNN, GDCN, and DeepFM–are trained on the balanced training set.

The proposed method uses a stacking ensemble with XGBoost as the meta-model to improve CTR prediction accuracy. Stacking combines predictions from multiple base models using a meta-model, which leverages the strengths of each model to enhance overall performance.

SHAP analysis explains the contribution of each feature to the model's predictions. After stacking, SHAP analysis reveals which features have high, moderate, or low impact on predictions, and which model influences specific features the most.

The models are evaluated using the following metrics: Accuracy, Recall, Precision, AUC, and Log Loss. Figure 2 visually represents the entire process and provides a clear overview of each phase contributing to improved CTR prediction.

The DNN predicts the click-through probability for a given input vector x. It combines embedding layers for categorical features, batch normalization for numerical inputs, and residual connections for robust feature learning. Figure 4 illustrates the model architecture, in which categorical features are embedded, numerical features are batch-normalized, and the resulting concatenated vector passes through three dense layers with a residual connection before the final classification output.

Given m categorical features x_cat = [x_cat,1, …, x_cat,m], each feature x_cat,i is mapped to a d-dimensional dense vector through an embedding layer:

The m embedding vectors are concatenated and flattened into a single vector:

where ‖ denotes vector concatenation.

The n numerical features $xnum∈Rn$ are normalized via batch normalization:

The concatenated input vector is then:

The model applies the following transformation order uniformly across all hidden layers: Linear → ReLU → Batch Normalization → Dropout. This ordering follows the practice described by Ioffe and Szegedy (2015) and subsequently evaluated by Bjorck et al. (2018), where batch normalization is applied after the activation to stabilize the output distribution, and dropout is applied last to prevent overfitting.

• Layer 1:

• Layer 2 (residual connection):

The residual term requires dimensional compatibility between the dense transformation output $(R256)$ and the skip connection from h₁ $(R512)$⁠. A linear projection $Ws∈R256×512$ is applied to h₁ before addition:

• Layer 3:

Parameters are updated using Adam with learning rate η = 0.001:

Batch normalization normalizes the output of each activation using the batch statistics:

where $μB$ and $σB2$ are the mean and variance computed over the current mini-batch $B$⁠, ϵ is a small constant for numerical stability, and γ, β are learnable scale and shift parameters.

Table 1 summarizes the layered architecture of the DNN, and Table 2 lists its training hyperparameters.

The GDCN predicts click-through probability by routing the input through two parallel branches: a Gated Cross Network (GCN) branch that captures explicit high-order feature interactions, and a DNN branch that learns implicit nonlinear representations. The outputs of both branches are concatenated before the final prediction. Figure 5 demonstrates core steps of GDCN model.

Note on terminology: Throughout this section, GCN refers exclusively to the Gated Cross Network component within the GDCN architecture. It is not related to Graph Convolutional Networks, which share the same abbreviation in other contexts.

Each categorical feature x_cat,i is mapped to a d-dimensional embedding:

The embeddings are concatenated into a flat vector:

Numerical features are batch-normalized:

The shared input vector fed to both branches is:

Let $x(0)=v∈RD$⁠. The cross network applies K stacked cross layers. At each layer ℓ ∈ {1, …, K}:

A layer-specific sigmoid gate controls the contribution of each cross interaction:

The gated cross output for layer ℓ is:

The output of the Gated Cross Network branch after K layers is $xgcn=x(K)∈RD$⁠.

where LReLU(z) = max(αz, z) with α = 0.1.

The model minimizes the same binary cross-entropy loss as Eq. (10), optimized using Adam with learning rate η = 0.0001.

Table 3 summarizes the GDCN architecture, and Table 4 lists its hyperparameters.

DeepFM combines a Factorization Machine (FM) component and a DNN component that share the same input embeddings (Rendle, 2010). The FM component captures low-order (order-1 and order-2) feature interactions, while the DNN component learns high-order feature correlations. Both components are trained jointly in an end-to-end manner without FM pre-training. Figure 6 shows the overall architecture.

Let $f∈RNf$ denote the concatenated flattened embedding vector. The final CTR prediction is:

Both the FM and DNN components operate on the same embedding representation:

The DNN output scalar is:

Table 5 summarizes the DeepFM architecture, and Table 6 lists its hyperparameters.

The proposed SELM follows a two-level stacking architecture. At Level 0, the three base models (DNN, DeepFM, and GDCN) are trained on the training set. Their prediction outputs on held-out data are stored as new features to form the meta-training set. At Level 1, a meta-model (XGBoost) trains on this meta-training set to produce the final prediction.

Meta-Model Fusion Logic. The XGBoost meta-model combines the three base-model probability outputs through a nonlinear, tree-based ensemble rather than a fixed weighted sum. Specifically, at Level 1, the input to XGBoost is a 3-dimensional feature vector $[pˆDNN,pˆDeepFM,pˆGDCN]$ for each instance, where each $pˆ$ is the base model's predicted click probability. XGBoost learns a gradient-boosted ensemble of decision trees that discovers nonlinear combinations of these base predictions, adaptively up-weighting base models that are more discriminative on different subspaces of the feature distribution. This differs from both fixed-weight averaging (which treats all base models equally) and logistic regression blending (which is constrained to linear combinations), allowing the meta-model to exploit complementary strengths among the base models.

Figure 7 illustrates the complete stacking pipeline. Algorithm 1 summarizes the complete training and prediction pipeline of the proposed SELM framework.

Compared to bagging and boosting, stacking leverages the complementary strengths of diverse architectures through a learned meta-model, allowing SELM to overcome the individual limitations of each base model.

Each base model q_k produces a prediction for a given instance d. The set of base model outputs is:

The meta-model g combines these predictions to produce the final output:

The base model output matrix is defined as:

The final class label is assigned by:

All models are implemented in TensorFlow and trained on Google Colab Pro + TPUs (v2-8). This section presents evaluation metrics, dataset description, SHAP feature analysis, and results across three experimental configurations.

Model performance is evaluated using five standard metrics. Accuracy (Acc) measures the overall fraction of correctly classified instances. Precision (PR) measures the proportion of predicted clicks that are true clicks. Recall (RE) measures the proportion of actual clicks that are correctly identified. The Area Under the Receiver Operating Characteristic Curve (AUC-ROC) quantifies the model's ability to discriminate between clicked and non-clicked samples; a value of 1.0 indicates perfect discrimination and 0.5 indicates random performance. Log Loss measures the quality of predicted probabilities, with lower values indicating better calibration. Table 7 provides the formal definitions.

The AVAZU dataset is a widely used benchmark for CTR prediction. It contains approximately 40 million ad impression records spanning 10 days of training data and one day of test data. The training set is approximately 6 GB and the test set approximately 2 GB. Most features are categorical and are encoded using label encoding before model training.

The dataset includes a binary target variable click (1 = clicked, 0 = not clicked) and 24 features organized into four groups: (1) site and app attributes, (2) device attributes, (3) ad placement attributes, and (4) anonymized contextual attributes. Four engineered temporal features–Day, Hour_of_day, Day_of_week, and Hour_day_interaction–are additionally derived during preprocessing. Table 8 summarizes all 28 features.

SHAP (Lundberg and Lee, 2017) assigns each feature a value that quantifies its contribution to the model's output. After training the SELM with XGBoost as the meta-model, SHAP analysis is applied to the out-of-fold meta-training set to identify the most influential features.

Figure 8 shows the SHAP summary plot. Each point represents one data instance. Red points (positive SHAP values) indicate that a high feature value increases the predicted click probability; blue points (negative SHAP values) indicate a decrease. Features are ranked top-to-bottom by their overall influence on the model output.

Figure 9 shows the top 10 features ranked by mean absolute SHAP value. Feature C19 has the highest importance, followed by App_id and C21. Site-related features–Site_domain and Site_category–rank highly, reflecting the strong influence of the advertising environment on user click behavior. Anonymous features C18 and C14 also contribute substantially despite being privacy-masked. Features Site_id, App_domain, and Device_model contribute additionally to predictive performance. Overall, the SHAP analysis confirms that both site/app context and anonymized behavioral signals are the primary drivers of CTR in the AVAZU dataset (Wang et al., 2022).

Three experiments are conducted to systematically assess the impact of class-balancing and feature selection on model performance.

1. Experiment 1 – Under-Sampling: The majority class is down-sampled to balance the training set. All 28 features are used.

2. Experiment 2 – Full Features without SHAP Selection: SMOTE is applied for class balancing, and all 28 features are used without SHAP-guided selection.

3. Experiment 3 – SHAP-Selected Features with SMOTE: SMOTE is applied, and only the top SHAP-ranked features are retained for training.

Each experiment evaluates all four models using confusion matrices, ROC curves, and the metrics defined in Section 4.1.

The Avazu dataset is split into 80% training and 20% test sets using a stratified split to preserve the original class ratio. A held-out validation set of 10% is drawn from the training portion and used exclusively for hyperparameter tuning and early stopping; it is never used for final performance reporting. The test set is held out entirely until final evaluation.

The raw Avazu dataset is heavily imbalanced, with approximately 83% non-click (Class 0) and 17% click (Class 1) samples. Three strategies are evaluated systematically. Experiment 1 applies random undersampling to the majority class. Experiment 2 applies SMOTE (Chawla et al., 2002) to the training set only. Experiment 3 applies SMOTE combined with SHAP-guided feature selection, retaining the top K = 10 features identified from the preliminary model trained on the training data.

Critically, SMOTE is applied only to the training data after the train/test split, ensuring that no synthetic samples appear in the test set.

To prevent data leakage in the stacking pipeline, the following strict protocol is enforced:

1. The dataset is split into training (80%) and test (20%) sets before any processing.

2. SMOTE is applied only to the training portion.

3. The SELM meta-model is trained using 5-fold out-of-fold predictions from the base models. In each fold, base models are trained on 4 folds and generate predictions on the held-out fold. This ensures the meta-model never trains on predictions from instances the base models were trained on.

4. SHAP analysis (Phase 3 of Algorithm 1) is conducted exclusively on the out-of-fold meta-training set. The test set is completely invisible at this stage.

5. The top-K = 10 features are selected based solely on training-data SHAP values. Only after feature selection is determined are the corresponding test set columns appended to the meta-test matrix.

6. The final meta-model is retrained from scratch on the selected features using only training data; the test set is used only for final evaluation.

This strict separation guarantees no data leakage at any stage of the SHAP-based feature selection pipeline.

All hyperparameters reported in Tables 2, 4, and 6 are selected via grid search on the validation set. The XGBoost meta-model hyperparameters (learning rate, max depth, number of estimators) are tuned using the same validation set with 5-fold cross-validation.

Under-sampling reduces the majority class to match the minority class size. Table 9 reports the results for all models. All four models achieve moderate performance, with accuracy between 0.68 and 0.70 and AUC between 0.73 and 0.76. The SELM achieves the lowest log loss (0.56) and the highest precision (0.67), though recall drops slightly compared to the individual base models. Figures 10 and 11 show the corresponding confusion matrices and ROC curves.

In this experiment, SMOTE balances the training set, and all 28 features are used without SHAP-guided selection. Table 10 presents the verified results from the experimental notebooks. The SELM achieves 0.89 accuracy, 0.97 precision, 0.81 recall, 0.95 AUC, and 0.25 log loss. DeepFM shows notably lower recall (0.76), suggesting it struggles to identify true clicks when noisy features are included. Figures 12 and 13 present the corresponding confusion matrices and ROC curves.

This experiment applies SMOTE for class balancing and uses only the top K = 10 features identified by SHAP analysis. Removing low-importance features reduces noise and allows each model to focus on the most discriminative signals. Table 11 presents the verified results from the experimental notebooks. The DNN, GDCN, and SELM each achieve 0.89 accuracy and 0.95 AUC. The SELM achieves 0.97 precision, 0.81 recall, and 0.25 log loss, matching the best base model performance while providing more stable and interpretable aggregation through the nonlinear XGBoost meta-model. DeepFM slightly underperforms the other models (0.87 accuracy, 0.93 AUC, 0.29 log loss). Figures 14 and 15 show the confusion matrices and ROC curves.

Table 12 compares the SELM meta-model across all three experiments. Moving from under-sampling to SMOTE with full features improves accuracy by 19% points (0.70–0.89) and AUC by 19 points (0.76–0.95). The SHAP-guided feature selection in Experiment 3 yields comparable performance to Experiment 2, confirming that the top-K = 10 features capture the most discriminative information. Importantly, Experiment 3 achieves the same overall accuracy (0.89) and AUC (0.95) as Experiment 2 while operating on a reduced feature set, demonstrating more compact and interpretable models without sacrificing predictive performance. Log loss drops to 0.25, the lowest across all three experiments. Figure 16 visualizes these trends.

Table 13 compares the proposed SELM against recent CTR prediction methods on the AVAZU dataset. The proposed method outperforms all listed baselines on AUC and achieves competitive log loss. We have expanded the comparison table to include additional recent models (2024–2025) to provide a more complete picture of the state-of-the-art landscape.

To disentangle the contribution of individual components, we conduct a systematic ablation study across two dimensions: (1) the effectiveness of SHAP-based feature selection relative to alternative selection methods, and (2) the contribution of each base model to ensemble performance.

To verify that SHAP-based selection is superior to other approaches, we evaluate three alternative feature selection strategies under the same SMOTE + stacking framework: (1) Random selection of K = 10 features, (2) Variance-based selection retaining the 10 highest-variance features, and (3) Mutual information (MI)-based selection retaining the 10 features with highest MI with the target label. Table 14 summarizes results.

SHAP-based selection matches full-feature performance (0.89 accuracy, 0.95 AUC) while using only 10 of 28 features, outperforming all alternative selection strategies. This confirms that SHAP identifies the most informative subset rather than simply selecting any 10 features.

To quantify each base model's contribution, we evaluate three degraded SELM variants: SELM without DNN (meta-model trained on DeepFM and GDCN predictions only), SELM without DeepFM, and SELM without GDCN. Table 15 presents results on Experiment 3 (SHAP + SMOTE).

Removing GDCN causes the largest performance drop (AUC from 0.95 to 0.93, log loss from 0.25 to 0.30), confirming GDCN's dominant role in capturing high-order feature interactions via its gated cross layers. Removing DNN causes a modest drop (AUC 0.95 to 0.94), while removing DeepFM has minimal impact on the GDCN + DNN combination. The full three-model ensemble achieves the best overall balance, validating the complementarity assumption underlying the stacking design.

This study proposed an approach to tackle the challenges of enhancing both predictive accuracy and interpretability in CTR prediction models. The proposed method utilizes DNN, DeepFM, and Gated Cross Networks as base models within a stacking ensemble framework, with XGBoost serving as the meta-model to refine predictions. This ensemble technique combines the strengths of models, significantly improving the overall performance and robustness of the CTR prediction system. Furthermore, the application of SHAP for feature importance analysis provides valuable insights into how individual features influence the model's predictions, offering much-needed transparency and interpretability in complex machine learning models. The results of the proposed ensemble method outperform existing CTR prediction techniques in terms of accuracy of 89% (0.89), AUC 0.95, and log loss 0.25. This study not only contributes to the advancement of CTR prediction but also lays the groundwork for more transparent and interpretable machine learning systems, which are critical for applications in digital advertising and beyond. Future research could focus on further optimizing the ensemble approach, exploring real-time prediction capabilities, and incorporating additional features to improve model performance. To enhance transparency, further research could incorporate explainable AI techniques, such as LIME and Counterfactual Explanations, to provide deeper insights into feature contributions and model behavior. The evaluation of the proposed model on the Criteo Dataset (Criteo: Criteo Display Advertising Challenge, 2014) is also a part of our future work. Additionally, integrating more contextual features, user behavior, and exploring quantum machine learning (QML) could further improve model performance and efficiency for large-scale, dynamic CTR prediction tasks.

1. Zeeshan Ali: Conceptualization, methodology, software, formal analysis, writing – original draft.

2. Hassan Ahmed: Methodology, writing – review and editing.

3. Abdullah Khan: Data curation, software, visualization.

4. Shahrzad Saremi: Supervision, writing – review and editing, project administration.

5. Rania Shibl: Writing – review and editing.

6. Mansooreh Mirzaei: Supervision, project administration, writing – review and editing.

7. Parvin Rastegari: Writing – review and editing, validation.

8. Mingzhong Wang: Supervision, writing – review and editing.

This study does not involve human participants, human data, or animal subjects. The dataset used in this work (AVAZU) is publicly available and does not contain any personally identifiable information.

All authors have read and approved the final version of the manuscript and consent to its publication.