Decoupled Contrastive Multi-View Clustering with High-Order Random Walks

Authors: Yiding Lu, Yijie Lin, Mouxing Yang, Dezhong Peng, Peng Hu, Xi Peng

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

Reproducibility Variable Result LLM Response
Research Type Experimental To verify the efficacy of DIVIDE, we carry out extensive experiments on four benchmark datasets comparing with nine state-of-the-art Mv C methods in both complete and incomplete Mv C settings.
Researcher Affiliation Academia College of Computer Science, Sichuan University
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes The code is released on https://github.com/XLearning SCU/2024-AAAI-DIVIDE.
Open Datasets Yes Scene-15 (Li and Perona 2005) contains 4,485 images of 15 scene categories. Caltech-101 (Li et al. 2015) consists of 8,677 images of objects from 101 classes. Reuters (Amini, Usunier, and Goutte 2009) is a multilingual news dataset with 18,758 samples from various languages. Land Use-21 (Yang and Newsam 2010) contains 2,100 satellite images from 21 classes.
Dataset Splits Yes Specifically, we randomly select m = η n samples and remove one view from each to simulate incomplete data, where η is the missing rate and n is the total number of samples. To warm up, we set the target T in Eq. (1) as an identity matrix In for the first 100 epochs and then adopt the rectified target Eq. (6) in the remaining epochs.
Hardware Specification Yes We implement our method in Py Torch 1.13.0 and run all experiments on NVIDIA 3090 GPUs in Ubuntu 20.04 OS.
Software Dependencies Yes We implement our method in Py Torch 1.13.0 and run all experiments on NVIDIA 3090 GPUs in Ubuntu 20.04 OS.
Experiment Setup Yes We train our model for 200 epochs using the Adam optimizer without weight decay. The initial learning rate is set to 2 10 3, and the batch size is fixed to 1024. To warm up, we set the target T in Eq. (1) as an identity matrix In for the first 100 epochs and then adopt the rectified target Eq. (6) in the remaining epochs. For the other hyper-parameters, we fix the contrastive temperature τ = 0.5, the temperature of the kernel function σ = 0.1, the walking step t = 5, and the rectified weight α = 0.5 throughout experiments.