A Contextual Bandit Algorithm for Linear Mixed Effects Models
-
SeriesResearch Master Defense
-
SpeakerHong Deng
-
LocationOnline
-
Date and time
August 28, 2020
15:00 - 16:00
The thesis generalizes the linear contextual bandit problems for potentially individual-clustered data. Upper confidence bound-typed bandit algorithms are widely used for contextually dependent decisions, such as customized recommender systems; however, the correlations of observations within individuals are rarely discussed in prior work. To allow for the presence of individual heterogeneity, linear mixed effects models are imposed for the reward generation, and a learning algorithm taking into account individual heterogeneity, called LIME-UCB, is proposed. The algorithm constructs the confidence interval by combing information across and within individuals, and achieves efficient learning for data with high level of individual heterogeneity.