Persistent Local Homology in Graph Learning
Authors: Minghua Wang, Yan HU, Ziyun Huang, Di Wang, Jinhui Xu
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In the experiment section of our study, we comprehensively evaluate the performance enhanced by PLH in both synthetic and real-world scenarios, demonstrating their superiority over conventional PH. Specifically, in the synthetic experiments, we meticulously design tasks to illustrate the enhanced capabilities of PLH compared to PH. Complementing this, our real-world experiments provide empirical evidence of PLH s overall superiority in practical applications. Furthermore, an ablation study reveals that integrating PLH into a SOTA model yields approximately a 1% increase in performance. The ablation study subsection also provides a more detailed comparative analysis of the computational efficiency between the r PLH and PLH algorithms. Notably, throughout our experiments, we employ various metrics to ensure a robust and fair comparison with existing baselines, reflecting the diverse tasks and datasets on which PLH is implemented. |
| Researcher Affiliation | Academia | Minghua Wang EMAIL Department of Computer Science and Engineering State University of New York at Buffalo, NY, USA Yan Hu EMAIL Computer, Electrical and Mathematical Sciences and Engineering Division King Abdullah University of Science and Technology, Saudi Arabia Ziyun Huang EMAIL Department of Computer Science and Software Engineering Penn State Erie, The Behrend College, PA, USA Di Wang EMAIL Computer, Electrical and Mathematical Sciences and Engineering Division King Abdullah University of Science and Technology, Saudi Arabia Jinhui Xu EMAIL Department of Computer Science and Engineering State University of New York at Buffalo, NY, USA |
| Pseudocode | Yes | Algorithm 1 Reduced annular local subgraph vertex set building algorithm. 1: V0 {v} Initialize with the given vertex v 2: for i 1 to s do Iterate over s layers 3: Vi 4: for all u Vi 1 do 5: Vi Vi top-D i-th edge layer vertices with highest degree connected to vertex u 6: return Vr+1 . . . Vs Output the vertex set that induces annular local subgraph gs r |
| Open Source Code | No | The paper does not explicitly state that the source code for the methodology described in this paper is available, nor does it provide a link to a code repository. It mentions "Additional dependency requirements are specified within the source code" in the Experimental Environment section, but this is ambiguous and does not confirm the release of their own implementation code. |
| Open Datasets | Yes | Real-world Datasets include ogbn-arxiv, ogbl-ddi, and Protein-Protein Interaction networks (PPI). Both the ogbn-arxiv and ogbl-ddi datasets are part of the Open Graph Benchmark (OGB) (Hu et al., 2020). The PPI dataset is derived from the research paper by Zitnik & Leskovec (2017). |
| Dataset Splits | No | The paper mentions using OGB datasets and PPI datasets, which often have standard splits, but it does not explicitly state the specific train/test/validation split percentages, sample counts, or the methodology used for splitting in its text. It does not provide citations to predefined splits for these experiments directly within the experimental details, nor does it specify any custom splitting strategy with seeds. |
| Hardware Specification | Yes | The experiments are conducted on a server, which is powered by an Intel(R) Xeon(R) W-2123 CPU clocked at 3.60GHz and is backed by 64 GB of RAM. The system is further enhanced with four Ge Force RTX 2080 Ti GPUs, each with 11 GB of memory. |
| Software Dependencies | Yes | On the software side, we utilize key packages for deep learning, with the main package versions being Py Torch 1.12.1, CUDA 11.3, and cu DNN 8. |
| Experiment Setup | Yes | Hyperparameters remained consistent across all experimental setups and can be found in the Appendix. Table 8: Hyperparameters. In synthetic experiments, the threshold parameter D is set to adhering to the standard PLH algorithm s configuration. Experiment H D Resolution #Ephocs #Runs synthetic 2 8 100 20 ogbn-arxiv 2 8 16 500 10 ogbl-ddi 2 6 16 200 10 PPI 2 8 16 2500 20 |