Multi-View Representation Learning with Manifold Smoothness

Authors: Shu Li, Wei Wang, Wen-Tao Li, Pan Chen8447-8454

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments conducted on real-world datasets reveal that our Mv DGAT can achieve better performance than state-of-the-art methods.
Researcher Affiliation Academia Shu Li, Wei Wang , Wen-Tao Li, Pan Chen National Key Laboratory for Novel Software Technology Nanjing University, Nanjing 210023, China {lis, wangw, liwt, chenp}@lamda.nju.edu.cn
Pseudocode Yes Algorithm 1 Mv DGAT
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes Course dataset is a webpage classification dataset... (Du et al. 2013; Jing et al. 2017); Advertisement dataset... (Zhao et al. 2018; Zhou, Liu, and Shao 2018); MNIST dataset... adopt the setting used in (Wang et al. 2015).
Dataset Splits Yes For each dataset, we randomly sample 10% data as the validation set, randomly sample γ (γ = 5%, 10%, 15%) of the remaining data as the labeled data, and use the rest data as the unlabeled data.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No The paper mentions software components like 'Adam' and 'linear SVMs' but does not specify their version numbers or the programming environment (e.g., Python, PyTorch versions).
Experiment Setup Yes For Mv DGAT, we first construct k-nearest-neighbor graph for each view with exp( d(xv(i),xv(j)) / σ2 ) for different distance metrics, k {1, 3, 5, 7, 9}, d(xv(i), xv(j)) is set to be Euclidean distance or cosine distance, and σ {10-2, 10-1, 1}, v = 1, 2. ... The final dimension of each output layer is selected from {5, 10, 20, 50, 100} ... In the training process, we use dropout rate p = 0.3 and use Re LU as the activation function... f1 and f2 are trained for a maximum of 200 epochs using Adam (Kingma and Ba 2015) and early stopping with a window size of 5.