Replication code: "Decoupled smoothings on graphs"
This code and data repository accompanies the paper:
- Decoupled smoothing on graphs (WWW 2019) - Alex Chin, Yatong Chen, Kristen M. Altenburger, Johan Ugander.
This repository contains all the correponding code to replicate the figures in "Decoupled smoothing on graphs". We provide links to the datasets (Facebook100) in the data sub-folder.
Reproducing results and figures
This repository set-up assumes that the FB100 (raw
.mat files) have been acquired and are saved the data folder. Follow these steps:
- The Facebook100 (FB100) dataset is publicly available from the Internet Archive at https://archive.org/details/oxford-2005-facebook-matrix and other public repositories. Download the datasets.
- Save raw datasets in placeholder folder data. They should be in the following form:
Go to the code folder, and run code which is briefly described below:
- /soft_smoothing - includes notebooks for code related to simulations for the soft smoothing part (Section 6.1, Figure 2).
- /decouple_smoothing(compared with other methods) - includes all relevant code that compare decoupled smoothing with the other methods - including one hop majority vote, hard smoothing (ZGL) and two hop majority vote (Section 6.2, Figure 3)
- /hard_smoothing_regularization - includes all relevant code that related to iterative hard smoothing and regularization (Section 6.3.1, Figure 4)
- /decoupled_smoothing_regularization - includes all relevant code that related to iterative decoupled smoothing and regularization (Section 6.3.2, Figure 5)
- functions/ - all helper functions that are required by the main code.
All random number generators used in the analysis have been seeded deterministically to produce persistent cross-validation folds and thereby consistent results when re-running the analysis. The code for generating random graphs (sampled from the overdispersed stochastic block model) is not deterministically seeded.
All code was written and tested for Python 3.6 with versions for the following main Python libraries:
scipy(1.1.0). The code has know incompatibilities with Python 2.x and with networkx 1.x.
For questions, please email Yatong at email@example.com.