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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Gains: Fine-grained Federated Domain Adaptation in Open Set

Authors: Zhengyi Zhong, Wenzheng Jiang, Weidong Bao, Ji Wang, Qi (Cheems) Wang, Guanbo Wang, Yongheng Deng, Ju Ren

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on multi-domain datasets across three typical data-shift scenarios demonstrate that Gains significantly outperforms other baselines in performance for both source-domain and target-domain clients. Code is available at: https://github.com/Zhong-Zhengyi/Gains.
Researcher Affiliation Academia 1Laboratory for Big Data and Decision, National University of Defense Technology 2College of Science, National University of Defense Technology 3Department of Computer Science and Technology, Tsinghua University
Pseudocode Yes As shown in Alg. 1, when a new client joins, the server distributes the source global model WS to the target domain for local training, getting WT. Subsequently, the server decomposes WS and WT into an encoder and a classifier and derives the feature using the public dataset. Based on the differences in the feature extracted by ES and ET, as well as the parameter differences between CS and CT, the algorithm discriminates the type of new knowledge and confirms its type.
Open Source Code Yes Code is available at: https://github.com/Zhong-Zhengyi/Gains.
Open Datasets Yes Dataset. The datasets include the Digit Five (i.e., DF) for the digit classification and the Amazon Review (i.e., AR) for the product review. ... 1https://ai.bu.edu/M3SDA 2https://nijianmo.github.io/amazon/index.html
Dataset Splits Yes Under the scenarios of mild data shift and medium data shift, after determining the data classes contained in the source domain clients, we split the data using the Dirichlet distribution with a hyperparameter of 0.1.
Hardware Specification Yes Our experiments are conducted on a single NVIDIA RTX 4090 GPU.
Software Dependencies No The model used for the Digit Five 1 dataset is a CNN model, while the model used for the Amazon Review dataset 2 is an LSTM. The corresponding hyperparameters for the two datasets are as follows: Table 8: Hyperparameter setting. Learning Rate Optimizer Batch Size Digit Five 0.005 SGD 128 Amazon Review 0.5 SGD 64
Experiment Setup Yes Table 8: Hyperparameter setting. Learning Rate Optimizer Batch Size Digit Five 0.005 SGD 128 Amazon Review 0.5 SGD 64