My research
The k-max problem with value-index feedback, Work in Progress.
We consider a class of online combinatorial optimization problem where an agent chooses samples sequentially and receives aggregate rewards and index values as feedback. This problem is motivated by real-world applications such as online advertising where users interact with an online system. The goal is to select a set of items with maximum performances according to some stochastic valuation functions.
Sketching stochastic set utility functions, Completed.
We consider the problem of sketching a stochastic valuation function, defined as the expectation of a valuation function of independent random item values. Evaluation of sets of items is important in various applications such as gaming, crowd-sourcing and online advertising. One of the key challenges in these applications is to compute the set value functions accurately and efficiently. We will propose a systematic way of sketching that allows us to approximate the stochastic valuation function everywhere, and have controls over the sketch size at the same time. [link]
MLE for the β-models of random hypergraphs, Completed.
We study a simple model of group outcomes called the β-model, which is known for graphs. We will study a general hypergraph variant where nodes represent items and hyperedges represent relationships. This is motivated by the fact that complete graph data might be expensive to collect and we may only have observations at the set level in real-world applications. The goal is to estimate individual beta values from the group outcomes. We study the inference problem under different settings using maximum likelihood estimation (MLE).
Capturing the Purpose of Websites, Published.
Research on large-scale web-based representation learning at the Alan Turing Institute. We implemented and compared different web embeddings for selected datasets. A local-to-global method is used to enhance the embedding scalability. [link]