Enhancing sparse data recommendations with self-inspected adaptive SMOTE and hybrid neural networks

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Personalized recommendation systems are crucial for user satisfaction and managing information overload in e-commerce.
• Data sparsity presents a significant challenge for traditional recommendation algorithms.

Purpose of the Study:

• To introduce a novel hybrid framework, LSTM-SC with Self-Inspected Adaptive SMOTE (SASMOTE), for enhanced personalized recommendations.
• To improve data quality and model performance in data-sparse environments.

Main Methods:

• Hybrid framework combining Long Short-Term Memory (LSTM) and modified Split-Convolution (SC) neural networks (LSTM-SC).
• Advanced data sampling technique: Self-Inspected Adaptive SMOTE (SASMOTE) for adaptive neighbor selection and uncertain sample filtering.
• Optimization algorithms: Quokka Swarm Optimization (QSO) for sampling rates and Hybrid Mutation-based White Shark Optimizer (HMWSO) for hyperparameter tuning.

Main Results:

• Demonstrated significant improvements in Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R² metrics.
• Outperformed existing deep learning and collaborative filtering techniques on benchmark datasets (goodbooks-10k, Amazon review).

Conclusions:

• The proposed LSTM-SC framework with SASMOTE offers a superior approach to personalized recommendations, especially in data-sparse e-commerce settings.
• The framework is scalable, interpretable, and applicable to diverse domains like e-commerce and electronic publishing.