Factorizing time-heterogeneous Markov transition for temporal recommendation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Temporal recommendation systems are crucial but challenged by large event spaces and sparse user data.
• Understanding time-heterogeneous user behavior is key to improving recommendation accuracy.
• Existing methods struggle to model the dynamic and evolving nature of user preferences over time.

Purpose of the Study:

• To propose a novel model, Neural-based Time-heterogenous Markov Transition (NeuralTMT), to address challenges in temporal recommendation.
• To effectively capture time-heterogeneous temporal patterns in user behavior.
• To enhance recommendation quality by considering temporal dynamics.

Main Methods:

• Users' temporal behaviors are modeled using third-order Markov transition tensors.
• A linear co-factorization model is developed to learn time-evolving user and item factors.
• The model is extended to a neural-based framework (NeuralTMT) using nonlinear mappings and attention mechanisms.

Main Results:

• NeuralTMT demonstrates significantly superior performance compared to state-of-the-art baseline methods.
• The model effectively captures complex, time-varying user behavior patterns.
• Experiments on four diverse datasets validate the effectiveness of NeuralTMT.

Conclusions:

• NeuralTMT offers a robust solution for temporal recommendation by handling time-heterogeneous user behaviors.
• The proposed neural-based approach provides flexibility and improved accuracy in sequential recommendation.
• The work bridges tensor factorization and neural sequential recommendation, offering insights for future research.