Effect of Choosing Loss Function when Using T-batching for Representation Learning on Dynamic Networks
This repository contains the code for the research paper titled "Effect of Choosing Loss Function when Using T-batching for Representation Learning on Dynamic Networks." The code is based on the JODIE project, and alternative loss functions have been added to investigate their impact on dynamic network representation learning.
To run the code with different loss functions, you need to specify the desired model using command-line arguments.
- For the original loss function, use
--model jodie. - For
$loss_{item-sum}$ , use--model jodie-sum. - For
$loss_{full-sum}$ , use--model jodie-full-sum.
Both the jodie.py script for training and the evaluate_interaction_prediction.py script for validation and testing accept these command-line arguments.
This paper introduces a new dataset related to android application install interactions in the Myket android application market. The dataset can be accessed from the following link: Myket Android Application Install Dataset.
In addition to the new dataset, we also use three other datasets from the JODIE project:
- LastFM
- Wikipedia
The dataset .csv files should be placed in the data/ directory. For the Myket dataset for instance, you should put the myket.csv file under the path data/myket.csv.
If you use this code in your research or work, please cite the following preprint:
@article{loghmani2023effect,
author = {Erfan Loghmani and MohammadAmin Fazli},
title = {Effect of Choosing Loss Function when Using T-batching for Representation Learning on Dynamic Networks},
journal = {CoRR},
volume = {abs/2308.06862},
year = {2023},
url = {https://doi.org/10.48550/arXiv.2308.06862},
doi = {10.48550/ARXIV.2308.06862},
eprinttype = {arXiv},
eprint = {2308.06862},
timestamp = {Wed, 23 Aug 2023 14:43:32 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2308-06862.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}