Symmetry-induced Disentanglement on Graphs

Authors: Giangiacomo Mercatali, Andre Freitas, Vikas Garg

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on synthetic and real datasets suggest that these models can learn effective disengaged representations, and improve performance on downstream tasks such as few-shot classification and molecular generation. We conducted extensive experiments that we decribe now. Section 5.1 evaluates the disentanglement capabilities of the models, Section 5.2 provides a compression and few-shots classification experiment, and Section 5.3 assesses the generation capabilities on molecular datasets.
Researcher Affiliation Collaboration Giangiacomo Mercatali University of Manchester André Freitas Idiap Research Institute University of Manchester Vikas Garg Yai Yai Ltd and Aalto University
Pseudocode No The paper describes its architecture and probabilistic formulation in detail, including equations and a diagram (Figure 1), but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] Provided at https://mercatali.gitlab.io/sid/
Open Datasets Yes For metrics, we follow the evaluation protocols from Locatello et al. [36]... We compute 5 metrics, including β-VAE [19]... The task consists in training first without labels to learn the representation, and then using the compressed model as input for training a GCN classifier a labelled subset of the dataset, which involves 100 samples for training and 100 for testing. ...including FRANKENSTEIN [43] (2 classes), COLORS-3 [30] (11 classes), Mutagenicity [51] (2 classes), NCI1 [53] (2 c). ...results on ZINC [23], QM9 [49], and MOSES [47].
Dataset Splits No The paper mentions 100 samples for training and 100 for testing in the few-shot classification task, but it does not specify a separate validation set or its size/proportion for model training or hyperparameter tuning.
Hardware Specification No The paper states, 'We specify the experimental setups in Section 5,' but Section 5 does not contain any specific hardware details such as GPU models, CPU types, or memory specifications.
Software Dependencies No The paper mentions 'Torch Drug2' and 'DIG' as approaches implemented via, and provides a URL for Torch Drug2, but it does not specify version numbers for these or any other software components.
Experiment Setup Yes We set our models to have a group size of 81, a hessian penalty of 40, and a commutative penalty of 5. All models are set to have a pooling rate of 0.8, and a depth of 3 layers. Our models are set with a group size of 81, hessian penalty of 20 and commutative penalty of 5. We train our models for 10 epochs, with a batch size of 32 and the learning rate of 0.001, and compute the metrics over 10K generated molecules.