Ensemble CNN with Feature Selection and Soft Voting for Document Classification

Document categorization powers NLP capabilities, including spam filtering, sentiment identification, and document retrieval.Although machine-learning models are faster, they are less interpretable compared to deep learning systems.Deep learning CNNs are better at recognizing complex and non-linear text data associations.These models are difficult to trust in use-case circumstances due to their lack of transparency and noise sensitivity.We're using CNNs, Random Forest, and XGBoost models to address this challenge with a hybrid ensemble framework.Specially built cross-modal feature sanctification achieves our purpose.The traditional feature selection method uses chi-square (χ²) on raw TF-IDF vectors; however, we use a new approach.Using integrated gradients on the TF-IDF matrix, we uncover semantically relevant words and maintain only those that are neurally attentive and statistically discriminative under the chi-square test.These validated features are then fed as structured inputs to RF and XGBoost, while the original CNN continues processing full sequences in parallel.Predictions from four models (CNN, CNN-LSTM, CNN+XGBoost, and CNN+RF) are fused via a confidence-weighted soft voting mechanism, where weights are dynamically assigned based on consensus between neural attention and tree feature importance.The proposed Ensemble (CNN + XGBoost + RF) achieves perfect classification performance with 100% accuracy, precision, recall, and F1score on the BBC News test set, following training on the 70% training split and 30% testing with no overfitting observed, significantly outperforming baseline CNNs (65.25%),CNN-LSTM models (97.48%), and individual hybrids (e.g., CNN + XGBoost: 100%, CNN + RF: 99.21%).Crucially, our framework enhances not only accuracy but also interpretability, enabling traceable decision-making through aligned neural and statistical signals.The proposed architecture demonstrates practical scalability for real-time applications, including news categorization, intelligent content filtering, and semantic information retrieval-bridging the gap between high-performance deep learning and trustworthy, explainable AI.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.