Furkan Kocayusufoglu successfully defended his PhD on the topic of "Learning with Richly Structured Data".

Committee:  Ambuj Singh (Chair), Xifeng Yan, Francesco Bullo, Yu-Xiang Wang

Learning problems often involve richly structured data, coupled with a diverse set of learning objectives depending on the application domain. The structure of the data inevitably shapes how the learning algorithms are derived for satisfactory results on the downstream task. In this talk, I will present an overview of my thesis, which includes various learning tasks (e.g., prediction, generation, recommendation) on spatial and temporal structures from multiple problem domains such as user modeling, representation learning, and learning with complex networks. We will first discuss the benefits of incorporating formatting information (structure) along with linguistic content of richly formatted emails to learn improved representations. Next, I will present our work on search personalization, which leverages temporal structure in modeling dynamic user interests to better capture their search intent while interacting with online e-commerce and content-sharing platforms. The second part of my talk will focus on problems about a complex family of graphs called flow graphs. We will cover two studies aiming to (i) estimate missing flows based on graph topology and physics, and (ii) generate realistic flow graphs, providing a data-driven alternative to domain-specific simulations. Lastly, I will present an ongoing work about transferring knowledge across flow graphs for improved flow estimation, a problem that is motivated by the findings of our previous studies on flow graphs. In the end, we will summarize the contributions, implications, and potential weaknesses of our work, and discuss future research directions.