DyCARS: A dynamic context-aware recommendation system.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Dynamic recommendation systems model evolving user interests using interaction sequences.
• Current methods often struggle with long-term dependencies and delayed interaction patterns.

Purpose of the Study:

• To propose a Dynamic Context-Aware Recommendation System for enhanced dynamic recommendation.
• To improve the modeling of long-term dependencies and extraction of relevant delayed interaction patterns.

Main Methods:

• Utilized a dynamic graph with static embeddings of recent interactions as dynamic context.
• Employed a Gated Multi-Layer Perceptron encoder for long-term dependency structure.
• Implemented an Attention Pooling network with bidirectional attention weights to extract delayed patterns.
• Introduced a Pairwise Cosine Similarity loss function for joint optimization of embeddings.

Main Results:

• The proposed model demonstrated consistent improvements over state-of-the-art baselines.
• Experiments were conducted on the LastFM and Global Terrorism Database datasets.
• The system effectively captured long-term dependency structures and delayed interaction patterns.

Conclusions:

• The Dynamic Context-Aware Recommendation System offers a significant advancement in dynamic recommendation.
• The model's ability to capture complex temporal dynamics enhances recommendation accuracy.
• The approach provides a robust framework for future research in dynamic recommendation systems.