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 [1].
Automatically Identify and Rectify: Robust Deep Contrastive Multi-view Clustering in Noisy Scenarios
Authors: Xihong Yang, Siwei Wang, Fangdi Wang, Jiaqi Jin, Suyuan Liu, Yue Liu, En Zhu, Xinwang Liu, Yueming Jin
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we perform a series of experiments to evaluate the effectiveness and advantages of our proposed method. Specifically, we aim to address the following research questions (RQs): RQ1: How does AIRMVC compare with other leading deep multi-view clustering techniques in terms of performance? RQ2: What art the impacts of the components of AIRMVC to enhance multi-view clustering results? RQ3: What clustering structures are identified by AIRMVC? RQ4: What is the effect of hyper-parameters on the efficacy of AIRMVC? |
| Researcher Affiliation | Collaboration | 1School of Computer, National University of Defense Technology, Changsha, Hunan, China 2National University of Singapore, Singapore. 3Intelligent Game and Decision Lab, Beijing, China, xihong EMAIL. Correspondence to: Xinwang Liu <EMAIL>. |
| Pseudocode | Yes | Due to the space limited, we present the training algorithm in Alg. 1 in Appendix. Algorithm 1 Training Algorithm of our designed AIRMVC |
| Open Source Code | Yes | The code of AIRMVC are available at https://github.com/xihongyang1999/AIRMVC on Github. |
| Open Datasets | Yes | To evaluate the effectiveness of the proposed AIRMVC framework, we implement comprehensive experiments on six widely used datasets: BBCSport, Reuters, Caltech101, UCI-digit, Web KB, SUNRGB-D, and STL10. Tab. 1 provides detailed statistical information for all datasets. BBCSport1: The BBCSport dataset includes 544 sports news articles categorized into five classes: athletics, cricket, football, rugby, and tennis. Each document is described from three different views with dimensions of 2582, 2544, and 2465. Reuters2: Reuters is a collection of documents. It contains 1200 samples, which can be divided into 5 classes. UCI-digit3: The UCI-digit Dataset, also known as the Optical Recognition of Handwritten Digits, contains handwritten digit images (0-9) represented as 8x8 grayscale matrices, resulting in 64 features per image. Web KB4: The Web KB dataset features web pages from four Wisconsin universities, focusing on both the content and the links between them. SUNRGBD5: The SUNRGBD dataset, essential for indoor scene understanding, contains 10,335 samples across 45 categories, including RGB images, depth images, and 3D point cloud data. STL106: The STL10 dataset is a key benchmark for unsupervised feature learning, deep learning, and self-taught learning. It contains 13,000 labeled images across 10 classes and 100,000 unlabeled images, each at 96x96 pixels with four views per image. Caltech1017: The Caltech101 dataset is a widely used benchmark in machine learning and computer vision, consisting of 9,144 samples categorized into 102 distinct clusters. It is characterized by its diversity and complexity, featuring five distinct views that provide complementary information for each sample. 1http://mlg.ucd.ie/datasets/bbc.html 2http://archive.ics.uci.edu/dataset/137/reuters+21578+text+categorization+collection 3https://cs.nyu.edu/roweis/data.html 4https://lig-membres.imag.fr/grimal/data.html 5https://rgbd.cs.princeton.edu/ 6https://cs.stanford.edu/acoates/stl10/ 7https://data.caltech.edu/records/mzrjq-6wc02 |
| Dataset Splits | No | The paper mentions datasets and noise introduction but does not specify any training, validation, or test splits. For example, it mentions "To simulate a noisy input environment, we randomly introduce noise into the multi-view input data x at varying proportions of {10%, 30%, 50%, 70%, 90%}" but not how the original datasets are partitioned into train/test sets. |
| Hardware Specification | Yes | In this study, all experiments are conducted on the Py Torch (Imambi et al., 2021) platform using an NVIDIA A6000 GPU. |
| Software Dependencies | No | The paper mentions "Py Torch (Imambi et al., 2021)" but does not provide a specific version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | Our proposed AIRMVC is trained using the Adam optimizer (Kingma & Ba, 2014) with its default settings. To enhance the discriminative power of the network and obtain reliable soft predictions in an unsupervised setting, we pre-train the model with an auto-encoder module for 100 epochs. The trade-off hyperparameters α and β are consistently set to 1.0. The maximal training epoch is set to 400 for all datasets. A detailed overview of the hyperparameter settings is provided in Tab. 7 in the Appendix. Table 7. Hyper-parameters Statistics in AIRMVC. BBCSport Web KB Reuters UCI-digit Caltech101 STL10 Views 2 2 5 3 5 4 Samples 544 1051 1200 2000 9144 13000 Clusters 5 2 6 10 102 10 Learning Rate 1e-4 1e-4 1e-5 1e-4 1e-4 1e-4 τ 0.8 0.8 0.8 0.8 0.8 0.8 α 1.0 1.0 1.0 1.0 1.0 1.0 β 1.0 1.0 1.0 1.0 1.0 1.0 |