Improved Bi-LSTM based recommendation system for cold start items with proposed latent factor-based feature

Recommender systems are crucial in various domains, enhancing user experience by recommending items based on historical preferences. However, the challenge of cold start items persists, where limited or no historical data impedes accurate recommendations. To tackle this issue, this paper introduces a novel recommendation model integrating deep learning and latent factor approaches. Using deep learning techniques and latent factor models, the approach begins with Double Sigmoid Normalization for data preprocessing, followed by feature extraction using an Improved Bidirectional Long Short-Term Memory (Bi-LSTM) model to capture item content dynamics effectively. The model integrates seamlessly with latent factors in collaborative filtering, enhancing prediction accuracy by incorporating an Improved Activation function and Fisher Score Ranking. Evaluation on diverse datasets demonstrates the model’s capability to provide robust recommendations even for items with limited interaction history, promising significant advancements in recommendation system performance. The proposed Improved Bi-LSTM method achieves the lowest MSE of 0.050, outperforming all other models including Dense Net, Squeeze Net, CF+DLNN-SADE, DNN, RNN, and Conventional Bi-LSTM thereby demonstrating superior prediction accuracy.