This is the official repository of the paper, "You're Not Alone in Battle: Combat Threat Analysis Using Attention Networks and a New Open Benchmark," published in CIKM 2023.

Dataset, benchmark, and training details are provided in the Appendix.pdf.

This repository contains the proposed (1) combat simulation dataset, (2) model SAFETY, and (3) online Appendix.

Note: In coming weeks, we will make more details about the dataset and codes available.

./train folder has codes to run all deep learning models.

./dataset folder has dataset zip files.

./run folder has codes to reproduce the experimental outcomes in the paper.

./best_hyperparam folder has txt files with best hyperparameters for each deep learning model.

statistics_eval.ipynb file has codes to check raw data statistics.

The raw datasets are the 1.zip, 2.zip, 3.zip, 4.zip files in the dataset folder, where each number indicates the assigned tactic.

Preprocessed data for training and evaluation is data_dict file.

cd ./SAFETY/
unzip ./dataset/1.zip ./dataset/2.zip ./dataset/3.zip ./dataset/4.zip

This will return four folders within ./dataset. Each folder has four csv files, each with features and labels.

Run codes in table2dict.ipynb to generate data_dict.pkl. (Note that the zip file of data_dict is already provided in ./dataset.)

Run codes in statistics_eval.ipynb to check unzipped raw data statistics.

To run the model code, refer to ./run/run.sh file.

python ./train/main.py --pred y_int --stress-type none --model SAFETY # train to predict y_int
python ./train/main.py --pred y_atk --stress-type none --model SAFETY # train to predict y_atk

Running the code will automatically load the tuned best hyperparameters from ./best_hyperparam. If the model does not have tuned hyperparameters in ./best_hyperparam, grid search the model's best hyperparameters

python ./train/main.py --pred y_int --stress-type none --model SAFETY --optimize hyperparam # search best hyperparameters for y_int prediction
python ./train/main.py --pred y_atk --stress-type none --model SAFETY --optimize hyperparam # search best hyperparameters for y_atk prediction

As described in the Appendix.pdf, the maximum observed time (t_max) is 300. Only 20 randomly sampled timestamps (t_sample_size) serve as model input. To change the setting, modify --t-max and --t-sample-size in the command.