End-to-End Learning of Probabilistic Hierarchies on Graphs
Authors: Daniel Zügner, Bertrand Charpentier, Morgane Ayle, Sascha Geringer, Stephan Günnemann
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that our model learns rich, high-quality hierarchies present in 11 real world graphs, including a large graph with 2.3M nodes. Our model consistently outperforms recent as well as strong traditional baselines such as average linkage. |
| Researcher Affiliation | Academia | Daniel Zügner, Bertrand Charpentier, Sascha Geringer, Morgane Ayle, Stephan Günnemann Technical University of Munich {zuegnerd, charpent, geringer, ayle, guennemann}@in.tum.de |
| Pseudocode | No | No pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | Yes | Our implementation is available at https://www.daml.in.tum.de/fph |
| Open Datasets | Yes | Datasets. We use 11 real world datasets (Mc Callum et al., 2000; Sen et al., 2008; Aspert et al., 2019; Amunts et al., 2013; Cho et al., 2014; Adamic & Glance, 2005; Patokallio; Wang et al., 2020; Yang & Leskovec, 2015), including the very large ogbn-products dataset with around 2.3M nodes and 62M edges (Hu et al., 2020). |
| Dataset Splits | No | No explicit mention of specific training/validation/test dataset splits (e.g., percentages or sample counts) was found for reproduction purposes. The paper states 'For smaller datasets, we train FPH for 1,000 epochs and restore the best hierarchy after training', implying some form of validation, but without specific split details. |
| Hardware Specification | Yes | We train all models on a single GPU (NVIDIA GTX 1080 Ti or NVIDIA GTX 2080 Ti, 11 GB memory) in our own in-house compute cluster. The machines have 10-core Intel CPUs. |
| Software Dependencies | No | We use Python 3 and Py Torch for all our experiments. To compute the Dasgupta and TSD metrics (as well as to obtain the results for the Louvain algorithm), we use the sknetwork Python library. Specific version numbers for PyTorch and sknetwork are not provided. |
| Experiment Setup | Yes | We use the hyperparameters for models and baselines described in Tab. 7. For smaller datasets, we train FPH for 1,000 epochs and restore the best hierarchy after training. For ogbn-products, ogbnarxiv, ogbl-collab, and DBLP, we train for 2,000 epochs. Similarly, we use different learning rates for A and B for FPH (Das.) on ogbn-arxiv, ogbl-collab, ogbn-products, and DBLP (lr A = 1e-2, lr B = 1e-9). Table 7 further specifies learning rates, batch sizes, epochs, and other parameters for various models. |