Predicting Dynamic Embedding Trajectory in Temporal Interaction Networks.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Modeling sequential user-item interactions is vital for e-commerce, social networks, and education.
• Dynamic embedding methods capture evolving user/item representations but often lack future trajectory modeling.

Purpose of the Study:

• To introduce JODIE, a novel coupled recurrent neural network model for learning user and item embedding trajectories.
• To explicitly model and predict future user/item embeddings and interactions.

Main Methods:

• JODIE utilizes two recurrent neural networks to update embeddings at each interaction.
• A novel projection operator estimates future embeddings, enabling prediction of future user-item interactions.
• A t-Batch algorithm ensures time-consistent batches for scalable and efficient training (9x faster).

Main Results:

• JODIE demonstrated superior performance on future interaction and state change prediction tasks across four real-world datasets.
• Outperformed six state-of-the-art algorithms by at least 20% in future interaction prediction.
• Achieved over 12% improvement in state change prediction accuracy.

Conclusions:

• JODIE effectively models dynamic user-item embeddings and their future trajectories.
• The proposed method offers significant improvements in predictive accuracy for sequential interaction data.
• JODIE provides a scalable and efficient solution for dynamic representation learning.