Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Teacher-guided Edge Discriminator for Personalized Graph Masked Autoencoder
Authors: Qiqi Zhang, Chao Li, Zhongying Zhao
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on 10 datasets demonstrate the superior performance of TEDMAE on the tasks of node classification and node clustering. ... We conduct extensive experiments to assess the effectiveness of TEDMAE on various node-level tasks, aiming to address the following three key research questions. RQ1: Does the proposed TEDMAE generalize well to multiple downstream tasks? RQ2: To what extent do the various modules in TEDMAE contribute to the classification results? RQ3: How do hyper-parameters affect performance? |
| Researcher Affiliation | Academia | Qiqi Zhang, Chao Li, Zhongying Zhao* College of Computer Science and Engineering, Shandong University of Science and Technology, Qingdao 266590, China EMAIL, EMAIL, EMAIL or EMAIL |
| Pseudocode | Yes | Algorithm 1: Algorithm of TEDMAE Input: adjacency matrix A, homophily / heterophily metrics Ahomo/Aheter, node attributes X, masked feature metrics Xhomo/Xheter, latent representations Hhomo/Hheter, teacher model T1/T2, parameters α, β, γ, training rounds t; Output: node representations Z; # Stage 1. teacher-guided edge discriminator module. 1: for i=1 to max epoch do 2: Calculate the edge weight σ of G (Eqn. (1-2)); 3: Learn the continuous weight of edge ˇ (Eqn. (3)); 4: Generate the teacher embedding h T 1 and h T 2; 5: Calculate the margin ranking loss Lmr (Eqn. (4-6)); 6: Generate matrices Ahomo/heter; 7: end for # Stage 2. personalized graph masked autoencoder module. 8: for j=1 to t do 9: Mask Ahomo and Aheter based on α and β (Eqn. (7)); 10: Generate latent embeddings Hhomo/heter (Eqn. (8)); // Objective 1. reconstruct embedding. 11: Decode the latent embeddings Hhomo and Hheter to the embedding space (Eqn. (9)); 12: Calculate the Info NCE loss Lml (Eqn. (10-11)); // Objective 2. reconstruct feature. 13: Decode the latent embeddings Hhomo and Hheter to the feature space (Eqn. (12)); 14: Calculate the SCE loss Lsce (Eqn. (13)); 15: end for 16: Return Z. |
| Open Source Code | Yes | Code https://github.com/ZZY-Graph Mining Lab/TEDMAE. |
| Open Datasets | Yes | The experiments are conducted on publicly available benchmark datasets, in line with previous studies, including homophily datasets Cora, Wiki CS, Amazon-Computers, Amazon-Photo, Co Author-CS, Co Author-Physics and heterophily datasets Cornell, Texas, Wisconsin, and Actor. |
| Dataset Splits | No | The paper lists benchmark datasets but does not explicitly state the train/test/validation splits used for their experiments (e.g., percentages or specific methodologies). It mentions following hyper-parameter settings from previous studies, which might imply standard splits, but this is not explicitly stated within this paper. |
| Hardware Specification | No | No specific hardware details (e.g., GPU models, CPU specifications, memory) used for running the experiments are mentioned in the paper. |
| Software Dependencies | No | The paper mentions using the Adam optimizer and GAT/MLP as encoder/decoder types, but does not provide specific software names with version numbers (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | As for the proposed TEDMAE, we utilize the Adam optimizer, set the dropout rate to 0.5, and initialize the learning rate to 0.01. Additionally, We also perform an hyper-parameter search for α, β, µ, and ν, with values ranging from 0.1 to 1. |