Yongbiao Gao

2 previous affiliations
4Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Yongbiao Gao focuses on label distribution learning, reinforcement learning, and their applications in handling real-world ambiguity. His trajectory spans from foundational methods for recovering label distributions to applied reinforcement learning for subjective tasks like video summarization. Most recently, his work addresses sequential decision-making and preference learning in recommender systems.

Research Trajectory

Yongbiao Gao’s research initially concentrated on the challenge of training models when data exhibits inherent ambiguity, contributing to the broader area of Label Distribution Learning for Real World Ambiguity. Because most datasets provide simple logical labels rather than nuanced label distributions, his early work explored how to recover these distributions using prior knowledge. As the first author of Label Enhancement for Label Distribution Learning via Prior Knowledge (2020), Gao formulated label enhancement as a dynamic decision process, utilizing a reinforcement learning agent to adjust label distributions through sequential actions.

In 2021, Gao extended these principles to complex multimedia tasks, specifically targeting the research direction of Embracing Subjectivity in Video Understanding. In Video Summarization via Label Distributions Dual-Reward, he addressed the subjective nature of video summarization—where different viewers prioritize different content—by mapping label distributions from CNN and LSTM-based state representations. This work introduced a dual-reward mechanism in a reinforcement learning framework to better capture human subjectivity and diverse consciousness in video summaries.

Returning to the foundational mechanics of label enhancement, Gao later published Sequential Label Enhancement (2024). This research critiqued the standard practice of using a single objective function to obtain description degrees all at once. Instead, Gao proposed a sequential decision process (Seq_LE) that aligns more closely with the natural, step-by-step process of human annotation, further refining the methodology for recovering label distributions from high-cost or ambiguous data.

Current Directions

The most recent work in the evidence indicates an application of reinforcement learning and distribution modeling to user modeling and recommendation. In Long and Recent Preference Learning With Recent-K Items Distribution for Recommender System (2025), Gao investigates how to better capture the dynamics of user interests over time. This recent research formulates the recommendation task as a Markov decision process and introduces a "recent-k items" distribution branch to explicitly model the correlation between temporally close interactions, thereby improving the agent's ability to learn optimal recommendation policies from interaction trajectories.

Profile updated

See something wrong? Report a mistake