An Agnostic Approach to Federated Learning with Class Imbalance
Authors: Zebang Shen, Juan Cervino, Hamed Hassani, Alejandro Ribeiro
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through an extensive empirical study over various data heterogeneity and class imbalance configurations, we showcase that CLIMB considerably improves the performance in the minority class without compromising the overall accuracy of the classifier, which significantly outperforms previous arts. |
| Researcher Affiliation | Academia | Zebang Shen, Juan Cervino, Hamed Hassani, Alejandro Ribeiro Department of Electrical and Systems Engineering University of Pennsylvania Philadelphia, PA 19104, USA {zebang,jcervino,hassani,aribeiro}@seas.upenn.edu |
| Pseudocode | Yes | Algorithm 1 CLIMB: CLass IMBalance Federated Learning |
| Open Source Code | No | The code can be found here. (Note: The provided text does not include an actual link for concrete access to source code.) |
| Open Datasets | Yes | Three benchmark datasets are used in our experiments with the default train/test splits, which are MNIST (Le Cun et al., 1998), CIFAR10 (Krizhevsky et al., 2009) and Fashion-MNIST (Xiao et al., 2017). |
| Dataset Splits | No | The paper mentions "default train/test splits" but does not explicitly describe a validation split or provide specific details on how validation data was used for reproducibility. |
| Hardware Specification | No | The paper does not provide SPECIFIC HARDWARE DETAILS (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper does not provide SPECIFIC ANCILLARY SOFTWARE DETAILS (e.g., library or solver names with version numbers like Python 3.8, CPLEX 12.4) needed to replicate the experiment. |
| Experiment Setup | Yes | We generate heterogeneity in the local data distributions according to the strategy from (Karimireddy et al., 2020; Hsu et al., 2019): Let α [0, 1] be some constant that determines the level of heterogeneity. For a fixed α, we divide the dataset among N = 100 (moderate) or N = 500 (massive) clients as follows: for we allocate to each client a portion of α i.i.d. data and the remaining portion of (1 α) by sorting according to label. ... For the minority class(es), we retain only 1/ρ portion of the corresponding data. Here, ρ 1 the ratio between the numbers of data in the majority class and in the minority class and is termed the imbalance ratio. ... In our experiments, we consider the setting of 1 or 3 minority classes and we take ρ = 5, 10, 20. ... Specifically, we use a 2 hidden layer fully-connected neural network for MNIST, where the numbers of neurons are (128, 128). For CIFAR10, we use a CNN model consisting of 2 convolutional layers with 64 5 5 filters followed by 2 fully connected layers with 394 and 192 neurons. ... The base FL solver is Fed-Avg with partial-participation: 100 devices participate in every communication round. |