The project for Bayesian Statistics (2019 Spring PKU) by Dinghuai Zhang, Jiaqi Zhang and Feng Zhu.
This project mainly focus on online multi-classification problem under multinomial logit model. We tackle this problem via Thompson Sampling, a bayesian meta-algorithm for contextual bandits problem. We develop two ways, gaussian approximation and Polya-Gamma methods (PG-TS). Details can be seen in our report.pdf. Besides, a simple demo of Thompson Sampling can be found in Code/bernouli_bandit.ipynb.
The code of approximation method can be found in Code/Approximation, while that of PG-TS is available in Code/PG-TS and Code/PGSampler. The real dataset file can be found in Dataset.