Handling Missing Data with Graph Representation Learning

Authors: Jiaxuan You, Xiaobai Ma, Yi Ding, Mykel J. Kochenderfer, Jure Leskovec

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on nine benchmark datasets show that GRAPE yields 20% lower mean absolute error for imputation tasks and 10% lower for label prediction tasks, compared with existing state-of-the-art methods.
Researcher Affiliation Academia Jiaxuan You1 Xiaobai Ma2 Daisy Yi Ding3 Mykel Kochenderfer2 Jure Leskovec1 1Department of Computer Science, 2Department of Aeronautics and Astronautics, and 3Department of Biomedical Data Science, Stanford University {jiaxuan, jure}@cs.stanford.edu {maxiaoba, dingd, mykel}@stanford.edu
Pseudocode Yes Algorithm 1 GRAPE forward computation
Open Source Code Yes 1Project website with data and code: http://snap.stanford.edu/grape
Open Datasets Yes We conduct experiments on 9 datasets from the UCI Machine Learning Repository [1].
Dataset Splits Yes We randomly split the labels Y into 70/30% training and test sets, Ytrain and Ytest respectively.
Hardware Specification No The paper does not provide specific details about the hardware used for experiments, such as GPU models, CPU types, or memory specifications. It mentions running '5 trials' but not on what hardware.
Software Dependencies No The paper mentions 'Adam optimizer [28]', 'RELU activation', and 'MLP' but does not specify any software libraries or frameworks with version numbers (e.g., TensorFlow, PyTorch, scikit-learn versions).
Experiment Setup Yes For all experiments, we train GRAPE for 20,000 epochs using the Adam optimizer [28] with a learning rate at 0.001. For all feature imputation tasks, we use a 3-layer GNN with 64 hidden units and RELU activation. The AGGl is implemented as a mean pooling function MEAN( ) and Oedge as a multi-layer perceptron (MLP) with 64 hidden units. For label prediction tasks, we use two GNN layers with 16 hidden units. Oedge and Onode are implemented as linear layers. The edge dropout rate is set to rdrop = 0.3.