Graph Data Mining - Handling sparse data inputs with missing values using graph representation learning
A tweaked version of GRAPE for handling missing values found in sparse data
The challenge of dealing with missing data attributes has been a real-world problem for decades. The same becomes even more challenging when the data is sparse in nature. We have different approaches to deal with issue one of them being the use of neural networks where data has representations in a high dimensional space. One such model that leverage this idea is GRAPE, which is a graph-based neural network framework designed for performing feature imputation as well as label prediction for the missing data. The motivation behind this project is to see how well GRAPE works in case of sparse data input and also to suggest some potential adjustments that can be made to the GRAPE model for handling sparse data inputs.
I used the mean absolute error (MAE) and Root Mean Squared Error (RMSE) for experimental evaluations. The following visualizations show the behaviour of the MAE and MSE when both the models i.e. original GRAPE and tweaked version of GRAPE for sparse data is run for 1000 epochs.
Please see the original work at - http://snap.stanford.edu/grape/
Official GitHub (GRAPE) - https://github.com/maxiaoba/GRAPE
You can cite their wonderful work-
@article{you2020handling,
title={Handling Missing Data with Graph Representation Learning},
author={You, Jiaxuan and Ma, Xiaobai and Ding, Daisy and Kochenderfer, Mykel and Leskovec, Jure},
journal={NeurIPS},
year={2020}
}