FedNE: Surrogate-Assisted Federated Neighbor Embedding for Dimensionality Reduction

Authors: Ziwei Li, Xiaoqi Wang, Hong-You Chen, Han Wei Shen, Wei-Lun (Harry) Chao

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct comprehensive experiments on both synthetic and real-world datasets. The results demonstrate that our FEDNE can effectively preserve the neighborhood data structures and enhance the alignment in the global embedding space compared to several baseline methods.
Researcher Affiliation Academia Ziwei Li, Xiaoqi Wang, Hong-You Chen, Han-Wei Shen, Wei-Lun Chao The Ohio State University {li.5326, wang.5502, chen.9301, shen.94, chao.209}@osu.edu
Pseudocode No The paper does not contain explicitly labeled pseudocode or algorithm blocks. It describes the overall framework with a diagram and accompanying text.
Open Source Code Yes The main implementations will be available soon under https://github.com/Ziwei-0129/Fed NE.
Open Datasets Yes We conduct experimental studies on four benchmark datasets that have been widely used in the field of dimensionality reduction [35, 55]: MNIST [26], Fashion-MNIST [48], mouse retina single-cell transcriptomes [32], and CIFAR-10 [25].
Dataset Splits Yes We assess the quality of data embeddings by analyzing the input high-dimensional data points and their corresponding 2D positions [14]. First, to evaluate the preservation of neighborhood structures, we compute trustworthiness and continuity scores.
Hardware Specification Yes We conducted experiments using MNIST with 20 clients on a server with 4 NVIDIA Ge Force RTX 2080 Ti GPUs.
Software Dependencies No The paper mentions 'Adam optimizer' but does not specify its version or the version of any other libraries or programming languages used.
Experiment Setup Yes In the experiments of MNIST and Fashion-MNIST datasets, we use Adam optimizer with a learning rate of 0.001 and a batch size of 512 (i.e., the number of edges in the k NN graphs not the number of data instances). The learning rate was decreased by 0.1 at 30% and 60% of the total rounds. For the experiments with the single-cell RNA-Seq and CIFAR-10 dataset, the learning rate was initially set to 1 10 4. For negative sampling, we fix the number of non-neighboring data points sampled per edge to be 5 (b = 5).