SMONAC: Supervised Multiobjective Negative Actor-Critic for Sequential Recommendation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Traditional recommendation systems often rely solely on accuracy metrics, leading to repetitive suggestions and reduced user engagement.
• Multiobjective reinforcement learning (RL) offers a promising approach to balance multiple recommendation objectives like accuracy, diversity, and novelty.
• Existing RL methods face challenges including neglecting negative action values and limited integration with supervised learning models.

Purpose of the Study:

• To develop a novel algorithm that addresses the deficiencies in existing multiobjective RL for recommendation systems.
• To improve recommendation quality by simultaneously optimizing for accuracy, diversity, and novelty.
• To enhance long-term user engagement through more varied and novel recommendations.

Main Methods:

• Introduced the supervised multiobjective negative actor-critic (SMONAC) algorithm.
• Implemented a negative action update mechanism using offline RL to learn values of sampled negative actions.
• Developed a multiobjective actor-critic mechanism that integrates accuracy, diversity, and novelty into a scalarized value for supervised learning network criticism.

Main Results:

• SMONAC demonstrated significant performance improvements on two real-world datasets.
• The algorithm particularly excelled in enhancing recommendation diversity and novelty.
• Comparative experiments validated the effectiveness of the proposed negative action update and multiobjective actor-critic mechanisms.

Conclusions:

• The SMONAC algorithm effectively balances multiple recommendation objectives, outperforming existing methods.
• Addressing negative action values and integrating RL with supervised learning leads to superior recommendation quality.
• The findings suggest SMONAC is a viable approach for improving user engagement in recommendation systems.