SpeAr: A Spectral Approach for Zero-Shot Node Classification
Authors: Ting Guo, Da Wang, Jiye Liang, Kaihan Zhang, Jianchao Zeng
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, extensive experiments verify the effectiveness of the Spe Ar, which can further alleviate the bias problem. |
| Researcher Affiliation | Academia | Ting Guo1, Da Wang2, Jiye Liang2 , Kaihan Zhang1, Jianchao Zeng1 1. Data Science and Technology, North University of China, Taiyuan, China. 2. School of Computer and Information Technology, Shanxi University, Taiyuan, China. Corresponding author. Email: ljy@sxu.edu.cn. |
| Pseudocode | No | The paper does not contain a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | Justification: We provide the relevant code at github. |
| Open Datasets | Yes | Following the methods DGPN [10], DBi GCN [12], and Graph CEN [13], we seek to substantiate the validity of our proposed Spe Ar through its application to three public datasets: Cora [1], Citeseer [1], C-M10M [35]. |
| Dataset Splits | Yes | The dataset details are shown in Table 1. For Class Split I and II, the evaluation criterion is classification accuracy. In addition to that, we also give the Class Split III to validate the effectiveness of Spe Ar in handling GZNC. The evaluation criterion is H, defined as H = 2 (seen unseen)/(seen + unseen). The seen and unseen are the classification accuracies of seen and unseen classes, respectively. (Table 1 shows Train/Val/Test splits, e.g., Cora Class Split II [2/2/3]) |
| Hardware Specification | Yes | Computation resources: We execute our code on a computer with NVIDIA Ge Force RTX 3090 (GPU) and Intel Xeon Gold 6254 (CPU). |
| Software Dependencies | No | The paper mentions using 'Bert-Tiny' but does not specify any software dependencies with version numbers for their implementation (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | In Spe Ar, we employ a two-stage training strategy. During the first phase, the model is trained by the loss Luscl, which is governed by parameters α and β. In this stage, samples are utilized for updating category prototypes only if the predicted probability of pseudolabels exceeds a predefined threshold q, with q restricted to the range {0.5, 0.6, 0.7, 0.8, 0.9}. To ensure the stability of the update process, the prototype update parameter µ is cautiously set to 0.1 to prevent the introduction of erroneous information that might arise from higher values. In the second phase, the loss function Lscl continues with the parameters α and β established in the first phase. Here, we refine the prototype update process by selecting the top-s nodes with the highest pseudo-label probabilities, with s uniformly set to 1000 across all datasets. Furthermore, to achieve fine-tuning of the prototypes, the update parameter µ is further reduced to 0.01 in this phase. |