Multimodal Deep Neural Framework with Bayesian Optimization for Real-Time E-Commerce user Purchase Behaviour Prediction

This paper introduces a neural network-based method for e-commerce user purchase behaviour prediction, solving the essential problem of personalized recommendation systems. By taking advantage of deep learning architectures, we propose a model that combines multimodal data such as user interactions, product features, and temporal patterns to predict purchase intent with high accuracy. Our developed Deep Neural Network (DNN) performs better than baseline algorithms, with 94.8% accuracy, 95.5% recall, and 94.8% F1-score, showing better capability to identify intricate behavioural patterns. Major innovations are featuring engineering optimized, batch normalization for stable training, and dropout regularization to avoid overfitting. Comparative evaluation against Logistic Regression, Random Forest, XGBoost, and Transformer models validates the DNN's better performance, especially in dealing with sparse and high-dimensional data. The interpretability of the model is enhanced using attention mechanisms, exposing powerful drivers of purchases. Real-world applications involve real-time recommendation systems and targeted marketing, substantially increasing conversion rates. This study highlights the promise of neural networks in revolutionizing e-commerce analytics, providing scalable solutions to dynamic user behaviour prediction.