SELM-CTR: a stacking ensemble deep learning model with SHAP-based analysis for large-scale click-through rate prediction
Zeeshan Ali, Hassan Ahmed, Abdullah Khan, Shahrzad Saremi, Rania Shibl, Mansooreh Mirzaei, Parvin Rastegari, Mingzhong WangPurpose
Click-through rate (CTR) prediction is a core challenge in digital advertising, as forecasting whether a user will click on an advertisement directly determines ad placement decisions and revenue outcomes. Standard machine learning and deep learning approaches achieve reasonable predictive accuracy but are largely opaque, making it difficult to determine which features drive predictions and why. Deep Neural Networks (DNN), Deep Factorization Machines (DeepFM), and Deep Cross Networks (DCN) each capture different aspects of feature interaction, yet none alone addresses the full complexity of large-scale sparse data.
Design/methodology/approach
This paper proposes a stacking ensemble learning model for click-through rate (SELM-CTR) prediction, a stacking ensemble that combines DNN, DeepFM, and a Gated Deep Cross Network (GDCN) as base models, with XGBoost serving as the meta-model. Rather than treating these architectures as interchangeable alternatives, we exploit their complementary strengths: the DNN learns nonlinear representations, DeepFM captures low- and high-order interactions through factorization, and GDCN applies gated cross-layer interactions. The meta-model is trained on out-of-fold predictions from the base models, allowing it to learn which architecture to trust for different input patterns.
Findings
Evaluation using the publicly available AVAZU dataset indicates that the proposed method achieves an accuracy of 89%, an Area Under the Curve (AUC) of 94%, and a log loss of 0.25. These results represent a measurable improvement over existing baseline approaches. Furthermore, the application of SHAP clarifies how specific features influence the model's predictions, providing practical insights for real-world decision-making.
Originality/value
The primary contribution of this work is the integration of a stacking ensemble architecture (using DNN, DeepFM, GDCN, and XGBoost) with SHAP-based feature analysis. This addresses the common “black-box” limitations of deep learning in advertising, ensuring both high predictive performance and greater transparency regarding feature contributions.