Colloquium will be held on Friday, October 1st at 2:00PM by Zoom
Tensor Models for Online Decision Making
Abstract: Tensor as a multi-dimensional generalization of matrix has received increasing attentions. In this talk, I will discuss two tensor models that enjoy provable statistical guarantees for online decision making problems. In the first part of the talk, I will discuss a unified framework to extract latent features embedded in multiple sources of tensor data. One typical circumstance is in online advertising, where advertisement information is usually described by both ad click data and ad characteristics data. We conduct cluster analysis of ads based on the extracted latent features from both data sets, which provides interesting insights in linking different ad industries. In theory, the non-asymptotic estimation error of our method sheds some light on when additional data set is helpful for improving estimation accuracy.
In the second part of the talk, I will discuss a tensor bandit model for online interactive recommendation with tensor rewards. Traditional static recommendation system assumes the user’s preference over the item does not change over time. In many recommendation domains such as video recommendation or news recommendation, users constantly interact with the system with dynamic preference, and user feedback is instantly collected for improving recommendation performance. In these settings, it is essential for the recommendation method to adapt to the shifting preference patterns of the users. In theory, the cumulative regret of our model is studied and compared to the benchmark models to justify the importance of exploiting low-rank tensor structure.
Xiwei Tang is inviting you to a scheduled Zoom meeting.
Topic: Statistics Colloquium Fall 2021
Time: This is a recurring meeting Meet anytime
Join Zoom Meeting