Automated Loss function Search for Class-imbalanced Node Classification

Authors: Xinyu Guo, Kai Wu, Xiaoyu Zhang, Jing Liu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Across 15 combinations of graph neural networks and datasets, our framework achieves a significant improvement in performance compared to state-of-the-art methods. Additionally, we observe that homophily in graph-structured data significantly contributes to the transferability of the proposed framework.
Researcher Affiliation Academia 1School of Artificial Intelligence, Xidian University, Xi an, China 2School of Cyber Engineering, Xidian University, Xi an, China 3Guangzhou Institute of Technology, Xidian University, Guangzhou, China.
Pseudocode Yes Algorithm 1 Auto LINC
Open Source Code No The paper does not provide a direct link to open-source code for the methodology or state that the code is released.
Open Datasets Yes We validate Auto LINC on well-known citation networks (Yang et al., 2016), comprising three datasets: Cora, Cite Seer, and Pub Med, as well as Amazon s co-purchase networks (Mc Auley et al., 2015), which consist of two datasets: Computers and Photo.
Dataset Splits Yes In the case of citation networks, we employ training, validation, and testing set splits as described in (Yang et al., 2016).
Hardware Specification Yes It s important to note that all experiments in this paper are performed using a single NVIDIA Ge Force RTX 3090 GPU.
Software Dependencies No The paper mentions using the Adam optimizer but does not specify versions for any software libraries or dependencies (e.g., Python, PyTorch, TensorFlow).
Experiment Setup Yes All three GNNs consist of a 2-layer neural network with a hidden layer dimension of 256. Other hyperparameters align with those detailed in (Song et al., 2022). This includes employing the Adam (Kingma & Ba, 2014) optimizer, training the model for 2000 epochs, and selecting the optimal model parameters based on the average accuracy and F1 scores on the validation set. The initial learning rate is configured at 0.1 and undergoes halving when there is no improvement in the validation set loss for 100 consecutive generations. Additionally, weight decay is set to 0.0005 and is applied to all learnable parameters except for the last convolutional layer.