IPGAN: Generating Informative Item Pairs by Adversarial Sampling.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Ranking-based recommender models heavily rely on negative sampling.
• Existing methods struggle to generate informative positive-negative item pairs due to limitations in identifying potential positive items and capturing item relationships.

Purpose of the Study:

• To introduce a dynamic sampling strategy for generating informative item pairs in recommender systems.
• To address the limitations of existing negative sampling techniques.

Main Methods:

• Propose an item pair generative adversarial network (IPGAN) with a dynamic sampling strategy.
• Utilize generative models for sampling positive and negative instances, considering correlations.
• Employ a discriminative model to ensure sampled pairs are informative relative to ground truth.
• Introduce a batch-training approach to enhance user and item modeling and accelerate training.

Main Results:

• IPGAN effectively generates informative item pairs by sampling positive instances from observed item features and selecting correlated negative instances.
• The discriminative model ensures the relevance of sampled item pairs.
• The batch-training approach mitigates user-specific biases and speeds up model training.
• Experimental results on three real datasets demonstrate superior recommendation accuracy compared to state-of-the-art methods.

Conclusions:

• The proposed dynamic sampling strategy and IPGAN framework significantly improve the quality of item pairs for recommender systems.
• IPGAN offers enhanced recommendation accuracy and training efficiency.
• This work provides a novel approach to address key challenges in negative sampling for recommender systems.