A Ranking Recommendation Algorithm Based on Dynamic User Preference.

Area of Science:

• Information Retrieval
• Machine Learning
• Recommender Systems

Background:

• Hybrid recommendation techniques based on feature fusion are popular for list ranking.
• Current methods often use implicit feedback, which limits interaction depth and model performance.
• Lack of user satisfaction data hinders capturing the full complexity of user-item interactions.

Purpose of the Study:

• To propose a novel two-stage recommendation model, Attention-based Deep Neural Ranking (A-DNR).
• To address the limitations of existing methods in capturing deep entity interaction features.
• To improve the performance of list ranking recommender systems.

Main Methods:

• A two-stage approach: Stage 1 models user short-term preferences using an attention mechanism and fuses them with long-term preferences into dynamic user preferences.
• Stage 2 models high-order and low-order feature interactions using matrix factorization (MF) and multi-layer perceptron (MLP) models, respectively.
• Feature fusion via a fully connected layer, mapping vectors to scores for final list ranking.

Main Results:

• The A-DNR model demonstrated significant improvements in performance on three real-world datasets (Movielens100K, Movielens1M, Yahoo Movies).
• The proposed attention mechanism effectively captures user short-term preferences.
• The fusion of MF and MLP models successfully models diverse feature interactions.

Conclusions:

• Attention-based Deep Neural Ranking (A-DNR) offers a superior approach to list ranking recommender systems.
• The model's ability to dynamically fuse user preferences and model complex feature interactions leads to enhanced performance.
• A-DNR effectively overcomes the limitations of traditional implicit feedback-based methods.