Ad-Click Prediction Enhanced by Nonlinear Dynamics-Inspired Feature Extraction and Ensemble Optimization

The primary objective of this study is to enhance the predictive performance of machine learning models used for estimating Click-Through Rate (CTR), a key metric in digital advertising analytics. Beginning with a baseline Logistic Regression (LR) model applied to the “Click-Through Rate Prediction” dataset from Kaggle, the study systematically incorporates multiple optimization layers to improve forecasting accuracy. Inspired by nonlinear dynamics concepts, new feature representations were derived from temporal patterns and textual fields using TF-IDF and Word2Vec-based embeddings. Hyperparameter optimization techniques were then applied to refine model behavior, followed by the construction of ensemble architectures combining LR, XGBoost, Random Forest (RF), and Support Vector Machine (SVM) classifiers. Experimental results show that the optimized ensemble achieved the highest F1-score of 0.8694, yielding an improvement of approximately 12.7% over the baseline model. Overall, the study provides a comprehensive examination of feature extraction strategies, model optimization procedures, and ensemble fusion techniques, demonstrating the clear advantage of hybrid approaches in complex CTR prediction tasks.