Research
My research examines how structure supports representation learning and decision-making: learning structure from experience in reinforcement learning, learning over graph-structured data, and applying graph-based methods to real-world problems.
Learning Laplacian representations of environment dynamics when the graph structure is not given, and using them for value function approximation and exploration in deep reinforcement learning.
Explore reinforcement learning →Learning from graph-structured data through multi-hop propagation, local structure, and multiple relations, alongside systematic benchmarking and analysis of graph representations.
Explore graph learning →Using relational structure for dialogue response generation, dynamic scene understanding, vehicle behavior classification, and influence maximization in unknown social networks.
Explore applications →A connecting perspective
In reinforcement learning, the graph of environment dynamics is not given explicitly. I study learning Laplacian representations from experience and using them for value function approximation and temporally extended exploration.
In graph-based learning, relational structure is part of the problem. My work examines how models combine information across neighborhoods and relations, preserve node information, and perform under systematic evaluation.
Applications connect these methods to concrete tasks, including dialogue, scene understanding, and decision-making in unknown social networks.