Repository for training a Temporal Graph Neural Network model for predicting frequency of keywords (node regression)
- Based on .csv data containing keywords appearing together in scientific journals for specific months of an 8-year period (2014-2021), train model to predict the future frequency (e.g. for the next year) of the same keywords
- The main steps involve:
- Creating data windows of one year (8 data windows in total)
- Creating graphs for each data window in order to leverage structural dependencies between keywords (e.g. co-occurence)
- Transforming graphs into proper format in order to be ingested by Temporal GNN model
- Training the model
- First run
python generate_graphs.py
to generate the graph data- This will create a graphs.pkl file
- Then run
python training.py
to train the model using the generated graphs- This will save the best model parameters based on evaluation loss
You can evaluate the model and generate predictions using the evalutation.ipynb
notebook