Deep Incomplete Multi-View Clustering via Mining Cluster Complementarity

Authors: Jie Xu, Chao Li, Yazhou Ren, Liang Peng, Yujie Mo, Xiaoshuang Shi, Xiaofeng Zhu8761-8769

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on real-world multi-view datasets demonstrate that our method achieves superior clustering performance over state-of-the-art methods. We evaluate the effectiveness of our proposed DIMVC by comparing it with seven state-of-the-art IMVC methods on real-world multi-view datasets, in terms of three clustering metrics, including clustering accuracy (ACC), normalized mutual information (NMI), and adjusted rand index (ARI).
Researcher Affiliation Academia Jie Xu1, Chao Li1, Yazhou Ren1 , Liang Peng1, Yujie Mo1, Xiaoshuang Shi1 , Xiaofeng Zhu1,2 1School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China 2Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China, Shenzhen 518000, China jiexuwork@outlook.com, lichao.cfm@gmail.com, yazhou.ren@uestc.edu.cn, larrypengliang@gmail.com, moyujie2017@gmail.com, xsshi2013@gmail.com, seanzhuxf@gmail.com
Pseudocode Yes Algorithm 1: Optimization of the proposed DIMVC
Open Source Code Yes The code is provided in the website1. 1https://github.com/SubmissionsIn/DIMVC.
Open Datasets Yes We use four datasets in our experiments, i.e., BDGP (Cai et al. 2012), Caltech (Fei-Fei, Fergus, and Perona 2004), RGB-D (Kong et al. 2014), and Scene (Fei-Fei and Perona 2005). Table 1 presents the description of the used datasets.
Dataset Splits No The paper does not provide specific training/validation/test dataset splits. It describes how incomplete data is constructed from the datasets for evaluation, but not how the datasets themselves are partitioned into reproducible training, validation, and test sets.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments.
Software Dependencies No The paper mentions optimizers like Adam and activation functions like ReLU, but does not specify software platforms (e.g., PyTorch, TensorFlow) or specific library version numbers for reproducibility.
Experiment Setup Yes For our DIMVC, the following settings are adopted for all datasets. Concretely, the autoencoders of all views are implemented by fully connected neural networks with the same structure. For the v-th view, the network structure can be denoted as Xv Fc500 Fc500 Fc2000 Zv Fc2000 Fc500 Fc500 ˆXv, where Fc500 represents the fully connected neural network with 500 neurons. The dimensionality of embeddings Zv is reduced to 10. The activation function is ReLU (Glorot, Bordes, and Bengio 2011). We adopt Adam (Kingma and Ba 2014) to optimize the deep models with a learning rate of 0.001. In the initialization phase, the autoencoders are pre-trained for 500 epochs. The batch size is set to 256. In every iteration of the proposed alternate (EM-like) optimization strategy, the Z-step will train the deep models for 1000 batches after the P-step updates the learning targets. The number of iterations is set to 10.