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..
Improving Evolutionary Multi-View Classification via Eliminating Individual Fitness Bias
Authors: Xinyan Liang, Shuai Li, Qian Guo, Yuhua Qian, Bingbing Jiang, Tingjin Luo, Liang Du
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 Experiments In this section, we aim to verify the effectiveness of EFB-EMVC method from four aspects: (1) Comparison with SOTA methods; (2) Impact analysis of the evolutionary navigator (EN) on EFBEMVC; (3) Impact analysis of the distillation loss function ENL on EFB-EMVC; (4) Generality of evolutionary navigator. In addition, we conducted an in-depth analysis of the hyperparameters of EFB-EMVC and the experimental results are presented in Appendix A.4. |
| Researcher Affiliation | Academia | Xinyan Liang1,2, Shuai Li1, Qian Guo3, Yuhua Qian1 , Bingbing Jiang4, Tingjin Luo5, Liang Du1 1 Institute of Big Data Science and Industry, Shanxi University 2 State Key Laboratory of AI Safety, Beijing, 100086 3 School of Computer Science and Technology, Taiyuan University of Science and Technology 4 School of Information Science and Technology, Hangzhou Normal University 5 College of Science, National University of Defense Technology EMAIL, EMAIL EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 EFB-EMVC Framework 1: Input: D = (X, Y ): training dataset; ˆD = ( ˆX, ˆY ): test dataset; F: a set of basic fusion operators; T: iteration number; evolutionary navigator set; 2: Output: The satisfactory MVM and its corresponding accuracy; 3: Population initialization: Generate an initial population P0; 4: Fitness evaluation: Obtain the unbiased fitness values of all individuals in P0; 5: for t = 1 to T do 6: Generate offspring Qt using the crossover operator; 7: Conduct mutation on each individuals in Qt; 8: Obtain the unbiased fitness values of all individuals in Qt; 9: Select next generation population Pt+1 from Qt Pt using a selection operator; 10: end for 11: return The satisfactory MVM and its corresponding accuracy. |
| Open Source Code | Yes | (The code is available at https://github.com/Li Shuailzn/Neurips2025-EFB-EMVC) |
| Open Datasets | Yes | All used datasets in our paper can be downloaded from the https://github.com/Li Shuailzn/ Neurips-2025-EFB-EMVC. Datasets. In the experiments, nine multi-view datasets are used and they are MVox Celeb [16], Youtube Face [19], NUS-WIDE-128 (NUS) [20], Reuters [21], CB [7], MM-IMDB [22], NTU RGBD [23], and Ego Gesture [24]. |
| Dataset Splits | Yes | In the experimental process, to effectively avoid the randomness interference caused by data partitioning and network initialization operations, a 5-fold cross-validation strategy was adopted for the MVox Celeb [16], Youtube Face [19], NUS-WIDE-128 (NUS) [20], Reuters5 [16] and Reuters3 [16] datasets, dividing each dataset into a training set and a test set. The specific partitioning ratio is as follows: 80% of the samples are used for model training, and 20% are used for model testing. For the remaining datasets, since their original authors have already completed the partitioning of the training and test sets, the experiments were independently repeated five times. |
| Hardware Specification | Yes | The computing environment includes Ubuntu 24.04.2 LTS as the operating system, equipped with an AMD EPYC processor with 160 physical cores (320 logical threads), 566 GB of DDR4 memory, and 8 NVIDIA Ge Force RTX 5090 GPUs, each with 32 GB of VRAM. |
| Software Dependencies | Yes | In our experiments, all methods are implemented using Tensor Flow 2.10.0. The experimental setup is based on Python 3.9.23 and CUDA 11.2. |
| Experiment Setup | Yes | a) Training of multi-view model (MVM): All MVM are trained using the Adam algorithm. The learning rate is set to 0.001, with a first-moment exponential decay rate of 0.9 and a second-moment exponential decay rate of 0.999. Each network undergoes training for 200 epochs. To prevent overfitting, if the performance of a MVM does not improve after 10 epochs, the training process will be halted. b) EFB-EMVC Algorithm: Inspired by the work of EDF[7], we set the population size to 28, the number of iteration rounds to 20, the crossover rate to 0.9, and the mutation rate to 0.2. ... The weight of the fusion output, γ1, is set to 1.0 V, where V represents the number of views of the current individual. For datasets using the ℓENL loss function, the output weight of each view, γ2, is set to 1.0. For the cross-entropy loss between EN and the view branch on the target category, the parameter α is set to 1.0; for the Wasserstein distance between their output distributions on non-target categories, the parameter β is set to 0.3. In addition, for datasets using KL divergence, the weight of each view is directly set to 1.0. The temperature coefficient is uniformly set to t=2. |