TAFM: A Recommendation Algorithm Based on Text-Attention Factorization Mechanism.

Area of Science:

• Machine Learning
• Recommender Systems
• Natural Language Processing

Background:

• Click-through rate (CTR) prediction is crucial for recommender systems.
• Existing models like DeepFM capture feature interactions but lack user diversity and text utilization.
• DeepFM fails to consider user diversity and text information, limiting its performance.

Purpose of the Study:

• To propose a novel Text-Attention FM (TAFM) model to enhance CTR prediction.
• To address the limitations of DeepFM by incorporating user diversity and text features.
• To improve the accuracy and comprehensiveness of higher-order feature mining in recommender systems.

Main Methods:

• Developed a Text-Attention FM (TAFM) model building upon the DeepFM framework.
• Utilized an attention mechanism to capture diverse user and item representations and identify key features.
• Integrated text components, including a text attention and N-gram extraction, for comprehensive text feature learning.
• Employed a convolutional autoencoder for higher-level feature extraction.

Main Results:

• The TAFM model achieved an AUC score of 0.730 on a public dataset.
• This represents a significant improvement of at least 3% over existing models like DCN, DeepFM, and PNN.
• The accuracy metric also showed an improvement of at least 0.1 percentage points compared to baseline models.

Conclusions:

• The proposed TAFM model effectively enhances CTR prediction by integrating user diversity and rich text features.
• TAFM demonstrates superior performance over state-of-the-art models, highlighting the importance of attention mechanisms and text processing.
• The model offers a more comprehensive approach to feature learning and preference mining in recommender systems.