Moth Flame Optimization and CNN-GBM Fusion for Advanced Business Intelligence in Banking

In the era of digital banking, deriving actionable insights from customer complaints is essential to improve service quality, identify problems, and inform decision-making. Conventional methods are ineffective in handling high-dimensional and unstructured data, and their performance in real-world settings is compromised. This research proposes a hybrid deep learning model combining Moth Flame Optimization (MFO) and a CNN-GBM fusion model to overcome these limitations and deliver solid business intelligence from unstructured banking data. Preprocessing was performed on unstructured text data, mainly customer feedback, using common NLP methods such as stopwords removal, tokenization, and lemmatization. The pre-processed text was transformed into contextualized vector representations using BERT embeddings. Moth Flame Optimization was utilized to choose the most descriptive features, which minimized dimensionality and optimized the model's efficiency. A 1D Convolutional Neural Network (CNN) was utilized to obtain high-level features, which were trained using a Gradient Boosting Machine (GBM). The performance was measured with respect to baseline models like SVM and Random Forest. The new framework performed better, with the CNN+GBM model reaching an accuracy of 93.7% and an F1-score of 92.1%. MFO cut down the feature space by 67.5%, highly reducing training time while maintaining the accuracy of models. The value of ROC-AUC at 96.1% validated the high discriminatory capability. Comparative studies proved the new model outperforms traditional classifiers. This work illustrates how to leverage a combination of feature selection through MFO and the fusion of CNN-GBMs for unstructured banking data and business intelligence. Semantic richness comes through the adoption of BERT embeddings, complemented by computationally efficient assistance from MFO. The technique presents an interpretable and scalable banking solution available for institutions requiring inferences of information from bulky amounts of client-authored text.