An Explainable Deep Learning Based Data Mining Framework for Automated Data Loading Optimization and Pipeline Evaluation Using Sentiment Analysis

Abstract

Most contemporary deep learning deployments are plagued by inefficiencies in pipeline setup and data mining, especially with large-scale, semi-structured, and unstructured text-based datasets. Non-adaptive nature of traditional systems means they need frequent manual interventions to sync data preprocessing with changing model needs. The limitations are addressed through this research by proposing AutoDL-Pipeline, an intelligent, modular system that combines adaptive pipeline training and automated data integration. It uses the Improved Rüppell's Fox Optimizer (IRFO) to dynamically manage data flow operations such as load scheduling, batch optimization, and schema flexibility. A hybrid storage architecture is used in order to support both archival requirements and real-time analytical processing. In the center of the pipeline, a Multilayer Interactive RoBERTa based Bidirectional Convolutional Gated Recurrent Unit (MIRBCGRU) model is used to analyze the effect of data quality on the classification result. The system integrates explainability modules such as Local interpretable model-agnostic explanations (LIME) to provide complete transparency to both loading and prediction stages. Experiments across IMDb, Yelp, Amazon Reviews, and Sentiment140 datasets show statistically significant improvements (p < 0.05) in accuracy, time cost, and stability, with Kendall’s W values between 0.74 and 0.78 confirming consistent superiority.Finally, the AutoDL-Pipeline not only streamlines end-to-end deep learning pipelines but also offers transparent and scalable solutions to practical data pipeline problems—filling an essential gap in automated model life cycle management.