Using NAMs for eye-tracking data

Get PoTeC data

git clone git@github.com:dili-lab/PoTeC PoTeC-data
cd PoTeC-data

python download_data_files.py

took ~7 minutes

Training and evaluating the NAM model

The model can be trained using three different labels for the three different tasks which are: expert_cls_label, all_bq_correct and all_tq_correct. An example call for one label including hyperparemeter tuning is shown below:

python nam_train.py --hp_tuning --label "expert_cls_label" --dataset_folder PoTeC-data

Training and evaluating the baselines

The configurations for each setting are stored in a separate .json file. See below for an example call.

python evaluation.py --config "evaluation_configs/config_baseline_hp_tuning_2_labels_new-reader-split_label_expert_cls.json" --hp-tuning

Analyse features

In order to extract the most important features, extract_feature_contribution.py can be run. Note that the log directory to the trained model and the label need to be provided manually at the top of the file (lines 30 and 33). The Jupyter Notebook in the folder "feature_analysis" can then be run to analyse the files and create the plots.