GraphPulse: Topological representations for temporal graph property prediction

Authors: Kiarash Shamsi, Farimah Poursafaei, Shenyang Huang, Bao Tran Gia Ngo, Baris Coskunuzer, Cuneyt Gurcan Akcora

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experimentation, we demonstrate that our model enhances the ROC-AUC metric by 10.2% in comparison to the top-performing state-of-the-art method across various temporal networks.
Researcher Affiliation Collaboration 1Department of computer science, University of Manitoba, 2Mila Quebec AI Institute, 3School of Computer Science, Mc Gill University, 4University of Texas at Dallas, 5AI Institute University of Central Florida
Pseudocode No The paper describes the methodology in prose and with flowcharts (Figure 3) but does not include any explicit pseudocode blocks or algorithm listings.
Open Source Code Yes We provide the implementation of Graph Pulse at https://github.com/kiarashamsi/Graph Pulse.
Open Datasets Yes We perform experiments on Math Overflow (Paranjape et al., 2017) and Reddit-Body (Kumar et al., 2018) datasets, and seven ERC20 token networks that we have extracted from the Ethereum blockchain. The datasets used in this study are publicly available
Dataset Splits Yes Based on the chronological order, the graphs are divided into 80% training and 20% testing data... For all methods, we utilized a chronological %80 %20 split of the graph snapshot sequence as our train-validation and test data, respectively.
Hardware Specification Yes We ran all experiments on a Dell Power Edge R630, featuring an Intel Xeon E5-2650 v3 Processor (10-cores, 2.30 GHz, 20MB Cache), and 192GB of RAM (DDR42133MHz).
Software Dependencies No The paper mentions software components like 'Adam optimizer' and 'LSTM+GRU model' but does not specify their versions or the versions of underlying deep learning frameworks (e.g., PyTorch, TensorFlow) or programming languages.
Experiment Setup Yes We set the Mapper hyperparameters as cls = 5, n cubes = 2, and perc overlap = 0.4. GIN and TDA-GIN models use a Graph Isomorphism Network with 64 hidden units followed by a target output dimension of two. Raw RNN and TDA RNN models utilize LSTM and GRU layers with an Adam optimizer and a learning rate of 1e-4. A hybrid LSTM-GRU model processes sequences in a (7,3) and (7,5) format for input, respectively. We set the final embedding dimension as 16. For HTGN, the number of historical windows in the HTA module is set to 5.