Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Harnessing the Power of Federated Learning in Federated Contextual Bandits
Authors: Chengshuai Shi, Ruida Zhou, Kun Yang, Cong Shen
TMLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We substantiate these claims through rigorous theoretical analyses and empirical evaluations. ... Experimental results using real-world data with several different FL choices corroborate the practicability and flexibility of Fed IGW. ... In this section, we report the empirical performances of Fed IGW on two distinct real-world multi-label classification datasets, Bibtex (Katakis et al., 2008) and Delicious (Tsoumakas et al., 2008), which are also used in other practical CB investigations such as Cortes (2018). |
| Researcher Affiliation | Academia | Chengshuai Shi EMAIL Department of Electrical and Computer Engineering University of Virginia Ruida Zhou EMAIL Department of Electrical and Computer Engineering University of California, Los Angeles Kun Yang EMAIL Department of Electrical and Computer Engineering University of Virginia Cong Shen EMAIL Department of Electrical and Computer Engineering University of Virginia |
| Pseudocode | Yes | Algorithm 1 Fed IGW (Agent m) ... Algorithm 2 The FL component commonly adopted in existing studies on federated linear bandits: one-shot aggregation of compressed local data ... Algorithm 3 The (simplified) Fed Avg algorithm as an example of the canonical FL framework: multi-round aggregation of local model parameters |
| Open Source Code | Yes | Additional experimental details and results are discussed in Appendix G, while the codes for the experiments can be found at https://github.com/Shen Group/Fed IGW. |
| Open Datasets | Yes | In this section, we report the empirical performances of Fed IGW on two distinct real-world multi-label classification datasets, Bibtex (Katakis et al., 2008) and Delicious (Tsoumakas et al., 2008), which are also used in other practical CB investigations such as Cortes (2018). |
| Dataset Splits | No | at each time step, a context is randomly sampled from the dataset while the true labels are concealed from the agents. |
| Hardware Specification | No | The paper mentions two-layer multi-layer perceptrons (MLPs) are used to approximate reward functions but provides no specific details about the hardware used to run the experiments, such as CPU or GPU models. |
| Software Dependencies | No | The paper states that codes for the experiments are available but does not provide specific version numbers for any software libraries, programming languages, or other dependencies in the text. |
| Experiment Setup | Yes | For practical conveniences, instead of selecting a theoretically sound but sophisticated choice of Ξ³ for Fed IGW as in Theorem 4.1, we set it as a constant hyper-parameter and perform some preliminary manual selections with the final adopted values reported in Table 5. We believe this approach is more practically appealing as it does not need to scale Ξ³ consistently; a similar choice of using constant Ξ³ s is also adopted in Agarwal et al. (2023). Also, the temperature parameter ΞΆ used in softmax can be found in Table 5. ... During each FL process, the local batch size, the number of communications, and the local learning rate are specified in Table 5. Moreover, the epoch length is designed to be growing exponentially as in Corollaries 4.2, D.8 and E.2, i.e., Ο l = 2l, while culminating at an upper limit of 4096 to maintain timely updates. |