Re-implementation of the S²GC model described in Zhu and Koniusz, 2021.
The reproduction takes place entirely in SSGC.ipynb. The rest of the files are from the original repository of the paper.
- Download the data-directory from link and add it to "SSGC-Reimplementation"
- Download and unpack from link and add it at "SSGC-Reimplementation/DocumentClassification/data"
This re-implementation produces mostly very similar, if slightly worse, results to the ones described in the paper.
| Datasets | Cora | Citeseer | Pubmed |
|---|---|---|---|
| Test Accuracy - original | 83.5 % | 73.6% | 80.2% |
| Test Accuracy - reproduced | 81.1% | 69.9% | 79.6% |
| Datasets | 20NG | R8 | R52 | Ohsumed | MR |
|---|---|---|---|---|---|
| Test Accuracy - original | 88.6 % | 97.4% | 94.5% | 68.5% | 76.7% |
| Test Accuracy - reproduced | - | 95.6% | 93.3% | 65.8% | 70.0% |
-
The implementation works without problem for all datasets except for the large OGB datasets and one document classification dataset. I simply do not have the memory to load them or generate even sparse matrices.
-
However, in principle the implementation should also work for them, as long as you have enough RAM.
-
Epoch and run count have been reduced by me on some tasks to make the execution of them feasible.