Please contact xiawenwen@sjtu.edu.cn for any questions.
Link prediction is a fundamental task of graph analysis and the topic has been studied extensively in the case of static graphs. Recent research interests in predicting links/interactions on temporal graphs. Most prior works employ snapshots to model graph dynamics and formulate the binary classification problem to predict the existence of a future edge. However, the binary formulation ignores the order of interactions and thus fails to capture the fine-grained temporal information in the data. In this paper, we propose a new problem on temporal graphs to predict the interaction order for a given node set (IOD). We develop a Temporal ATtention network (TAT) for the IOD problem. TAT utilizes fine-grained time information by encoding continuous time as fixed-length feature vectors. For each transformation layer of TAT, we adopt attention mechanism to compute adaptive aggregations based on former layer's node representations and encoded time vectors. We also devise a novel training scheme for TAT to address the permutation-sensitive property of IOD. Experiments on several real-world temporal networks reveal that TAT outperforms the state-of-the-art graph neural networks by 55% on average under the AUC metric.
-
python >= 3.7
-
Dependencies are in the requirements.text.
- Install
git clone https://github.com/xiawenwen49/TAT-code.git
cd TAT-code/
pip install -e . # first install the code as a local editable module- Preprocessing a dataset
python -m TAT.preprocessing --dataset CollegeMsg- Run the model
python -m TAT.main --dataset CollegeMsg --model TAT --gpu 0python -m TAT.main --help