Learning Deep Tree-Based Retriever for Efficient Recommendation: Theory and Method.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning significantly enhances recommendation accuracy but faces efficiency challenges due to large item sets and high computation costs.
• Existing tree-based deep recommendation models learn tree structures for efficiency but struggle to fully satisfy the max-heap assumption using one-versus-all classification.

Purpose of the Study:

• To propose a Deep Tree-based Retriever (DTR) for efficient and accurate recommendations.
• To enhance the max-heap assumption adherence and recommendation performance in tree-based models.

Main Methods:

• DTR frames training as a softmax-based multi-class classification over tree nodes at the same level, promoting competition and mimicking beam search.
• A rectification method is introduced for the loss function to better align with the max-heap assumption.
• Sampled softmax and a tree-based sampling method are employed to improve efficiency and reduce bias.

Main Results:

• DTR demonstrates improved recommendation efficiency and accuracy.
• Theoretical analysis confirms DTR's generalization capability.
• Experiments on four real-world datasets validate the effectiveness of DTR, its rectification method, and tree-based sampling.

Conclusions:

• DTR offers a more efficient and accurate approach to deep recommendation systems.
• The proposed methods effectively address the limitations of previous tree-based models.
• DTR shows strong generalization and practical applicability in real-world scenarios.