Searching to Modulate for Cold-Start Recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Personalized recommendations for cold-start users with limited interaction data are challenging.
• Hypernetworks map user histories to parameters that modulate predictors, achieving state-of-the-art results.
• Current methods lack a general approach for designing the modulation structure.

Purpose of the Study:

• To develop a general, data-driven approach for designing modulation structures in hypernetwork-based recommendation systems.
• To automate the selection of modulation functions and their positions using neural architecture search.

Main Methods:

• Proposed ColdNAS: a symbolic search space for efficient one-shot neural architecture search.
• Proposed ColdNAS+: extended search space with neural networks for modulation functions and a two-stage decoupled stochastic search algorithm.
• Applied methods to recommendation systems and disease-gene association prediction.

Main Results:

• ColdNAS and ColdNAS+ significantly outperform hand-designed cold-start techniques.
• Discovered that optimal modulation functions vary across datasets and metrics, validating the data-driven approach.
• Demonstrated generalization to cold-start tasks beyond recommendations, such as disease-gene prediction.

Conclusions:

• Neural architecture search provides an effective and generalizable method for designing hypernetwork modulation structures.
• Automating modulation structure design is crucial for optimizing performance in cold-start scenarios.
• The proposed ColdNAS and ColdNAS+ frameworks offer robust solutions for personalized recommendations and other cold-start prediction tasks.