FedConPE: Efficient Federated Conversational Bandits with Heterogeneous Clients

Authors: Zhuohua Li, Maoli Liu, John C. S. Lui

IJCAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conduct extensive experiments to demonstrate the effectiveness of our algorithm. Specifically, we aim to answer the following research questions: 1. For both the single-client and multi-client settings, does our algorithm Fed Con PE outperform existing state-of-the-art algorithms for conversational contextual bandits? 2. How do the number of clients M and the arm set size K affect the performance of Fed Con PE? 3. Does our algorithm use fewer conversations in practice?
Researcher Affiliation Academia Zhuohua Li , Maoli Liu and John C.S. Lui The Chinese University of Hong Kong {zhli, mlliu, cslui}@cse.cuhk.edu.hk
Pseudocode Yes Algorithm 1: Fed Con PE Algorithm for client i; Algorithm 2: Fed Con PE Algorithm for server
Open Source Code No The paper does not provide a direct link or explicit statement about the release of its source code. It only mentions that details are postponed to an extended version.
Open Datasets Yes The details of data generation and preprocessing are postponed to Appendix G in the extended version of this paper [Li et al., 2024]. Movie Lens-25M [Harper and Konstan, 2015]: A dataset from Movie Lens, a movie recommendation website. Last.fm [Cantador et al., 2011]: A dataset from an online music platform Last.fm. Yelp1: A dataset from Yelp, a website where users post reviews for various businesses.
Dataset Splits No The paper uses standard datasets and evaluates performance, but does not explicitly state the specific train/validation/test splits (e.g., percentages or counts) used for reproducibility.
Hardware Specification No The paper does not specify any hardware details (e.g., CPU, GPU, memory, or cloud instance types) used for the experiments.
Software Dependencies No The paper mentions implementing algorithms and using standard libraries for data processing but does not provide specific software names with version numbers for reproducibility.
Experiment Setup Yes We randomly select 10 users and calculate their cumulative regret over T = 6,000 rounds. We set the arm set size K = 100 and randomly select K arms from |A| for the client. For the baseline conversational algorithms Con UCB, Arm-Con, and Con Lin UCB, we adopt the conversation frequency function b(t) = 5 log(t) , which adheres to their original papers. We set the number of clients M = 10 and independently select K = 100 random arms for each client. Other parameters remain the same as the single-client setting. ...simulations executed for T = 10,000 under different arm set sizes, ranging from 100 to 300 on the synthetic dataset.