Deep fusion explainable recommendation based on sentiment analysis and feature weighting.

Area of Science:

• Computer Science
• Artificial Intelligence
• Information Retrieval

Background:

• Existing recommendation systems struggle with inaccurate user ratings and lack of explainability.
• User ratings often fail to capture true intentions and sentiment attitudes.
• Personalized information retrieval requires more nuanced approaches than traditional methods.

Purpose of the Study:

• To propose a novel recommendation method, Sentiment Analysis and Feature Weighting (SAFW), to address limitations of current systems.
• To enhance recommendation accuracy by incorporating user sentiment and feature preferences.
• To improve the explainability of recommendations.

Main Methods:

• Utilized BERT pre-trained models for generating review text word vectors.
• Employed Bidirectional Recurrent Neural Network (BiRNN) to extract sentiment features and classify user sentiment towards items.
• Implemented deep fusion of decomposed sentiment and rating matrices to model user-item interactions.
• Calculated feature weights based on review frequency and sentiment scores.

Main Results:

• Reduced deviation between user ratings and true sentiment tendencies.
• Improved accuracy of personalized recommendations through multi-level feature vector fusion.
• Enabled persuasive recommendation explanations by identifying top user-preferred item features.

Conclusions:

• SAFW offers a more accurate and explainable approach to recommendation systems.
• Integrating sentiment analysis and feature weighting significantly enhances personalization.
• The method provides actionable insights into user preferences for improved user experience.