Overcoming Catastrophic Forgetting in Graph Neural Networks with Experience Replay
Authors: Fan Zhou, Chengtai Cao4714-4722
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on three benchmark datasets demonstrate the effectiveness of our ER-GNN and shed light on the incremental graph (non-Euclidean) structure learning. |
| Researcher Affiliation | Academia | Fan Zhou, Chengtai Cao* University of Electronic Science and Technology of China fan.zhou@uestc.edu.cn, cct447131988@gmail.com |
| Pseudocode | Yes | Algorithm 1 Framework of our ER-GNN. Input: Continual tasks T : {T1, T2, . . . , Ti, . . . , TM}; Experience buffer: B; Number of examples in each class added to B: e. Output: Model fθ which can mitigate catastrophic forgetting of preceding tasks. 1: Initialize θ at random; 2: while continual task T remains do 3: Obtain training set Dtr i from current task Ti 4: Extract experience nodes B from experience buffer B 5: Compute loss function: L Ti(fθ, Dtr i , B) 6: Compute optimal parameters: θ = arg minθ Θ(L Ti(fθ, Dtr i , B)) 7: Select experience nodes E = Select(Dtr i , e) 8: Add E to experience buffer: B = B E 9: T = T \ {Ti} 10: end while 11: Return model fθ |
| Open Source Code | No | No explicit statement or link indicating the provision of open-source code for the described methodology was found. |
| Open Datasets | Yes | To evaluate the performance of our model on solving the CGL problem, we conduct experiments on three benchmark datasets: Cora (Sen et al. 2008), Citeseer (Sen et al. 2008), and Reddit (Hamilton, Ying, and Leskovec 2017) that are widely used for evaluating the performance of GNN models. |
| Dataset Splits | No | The paper explicitly mentions 'training node set Dtr i' and 'testing node set Dte i' for each task but does not specify a separate validation set or standard train/validation/test splits for the overall datasets. |
| Hardware Specification | No | No specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running experiments were mentioned. |
| Software Dependencies | No | The paper mentions using 'Adma optimizer' but does not specify version numbers for any software libraries or dependencies (e.g., Python, PyTorch, TensorFlow, CUDA versions) that would be needed for reproducibility. |
| Experiment Setup | Yes | The settings of all baselines and the network architecture (i.e., GAT) in our implementation are the same as suggested in the respective original papers. Without otherwise specified, we set the number of experiences stored in the experience buffer from each class as 1 (i.e., e = 1). However, we note that a larger value of e would result in better performance. |