This repository is the official implementation of Huang, R., Wu, W., Yang, J., & Shen, C. (2021). Federated linear contextual bandits. Advances in neural information processing systems, 34, 27057-27068.

• Programming language: Python3,
• Packages Numpy 1.17.4, Scipy 1.3.2, Pandas 0.25.3.

To get a preliminary result, run this command:

python3 main.py

It will plot the regret per client R(T) as function of T of four algorithms specified in the paper, i.e. Fed-PE, Local UCB, Enhanced Fed_PE, and Collaborative. The last algorithm is a modified version for full information exchange.

The feature vectors and arms parameters $\theta$ are generated in SyntheticProblem.py, which is a synthetic dataset.

To run the experiment on the MovieLens-100K, replace

import SyntheticProblem as Construction

by

import MovieLensProblem as Construction

in the file main.py. Then run the command:

python3 main.py

It will plot the regret per client on the MovieLens-100K dataset.

The complete rating matrix is stored in complete_ratings.csv. The file can be get by running this command:

python3 CollaborativeFilteringCopy.py

CollaborativeFilteringCopy.py is directly modified from the Github project: Collaborative-Filtering (https://github.com/kevalmorabia97/Collaborative-Filtering).

MovieLens-100K (Harper, F.M. and Konstan, J.A. (2015) The MovieLens Datasets: History and context. ACM Trans. Intract. Intell. Syst., 5(4).)

To solve multi-client G-optimal design or its equivalent Determinant Maximizaion subject to multi-constraints, use the function OptimalExperiment in minVar.py.