Bio-logger Ethogram Benchmark: A benchmark for computational analysis of animal behavior, using animal-borne tags

This repository contains the datasets and experiment results presented in our arxiv paper:

B. Hoffman, M. Cusimano, V. Baglione, D. Canestrari, D. Chevallier, D. DeSantis, L. Jeantet, M. Ladds, T. Maekawa, V. Mata-Silva, V. Moreno-González, A. Pagano, E. Trapote, O. Vainio, A. Vehkaoja, K. Yoda, K. Zacarian, A. Friedlaender, "A benchmark for computational analysis of animal behavior, using animal-borne tags," 2023.

Standardized code to implement, train, and evaluate models can be found at https://github.com/earthspecies/BEBE/. 

Please note the licenses in each dataset folder.

Zip folders beginning with "formatted": These are the datasets we used to run the experiments reported in the benchmark paper. 

Zip folders beginning with "raw": These are the unprocessed datasets used in BEBE. Code to process these raw datasets into the formatted ones used by BEBE can be found at https://github.com/earthspecies/BEBE-datasets/.

Zip folders beginning with "experiments": Results of the cross-validation experiments reported in the paper, as well as hyperparameter optimization. Confusion matrices for all experiments can also be found here. Note that dt, rf, and svm refer to the feature set from Nathan et al., 2012.

Results used in Fig. 4 of arxiv paper (deep neural networks vs. classical models)
{dataset}_ harnet_nogyr
{dataset}_CRNN
{dataset}_CNN
{dataset}_dt
{dataset}_rf
{dataset}_svm
{dataset}_wavelet_dt
{dataset}_wavelet_rf
{dataset}_wavelet_svm

Results used in Fig. 5D of arxiv paper (full data setting)
If dataset contains gyroscope (HAR, jeantet_turtles, vehkaoja_dogs):
{dataset}_harnet_nogyr
{dataset}_harnet_random_nogyr
{dataset}_harnet_unfrozen_nogyr
{dataset}_RNN_nogyr
{dataset}_CRNN_nogyr
{dataset}_rf_nogyr

Otherwise:
{dataset}_harnet_nogyr
{dataset}_harnet_unfrozen_nogyr
{dataset}_harnet_random_nogyr
{dataset}_RNN_nogyr
{dataset}_CRNN
{dataset}_rf

Results used in Fig. 5E of arxiv paper (reduced data setting)
If dataset contains gyroscope (HAR, jeantet_turtles, vehkaoja_dogs):
{dataset}_harnet_low_data_nogyr
{dataset}_harnet_random_low_data_nogyr
{dataset}_harnet_unfrozen_low_data_nogyr
{dataset}_RNN_low_data_nogyr
{dataset}_wavelet_RNN_low_data_nogyr
{dataset}_CRNN_low_data_nogyr
{dataset}_rf_low_data_nogyr

Otherwise:
{dataset}_harnet_low_data_nogyr
{dataset}_harnet_random_low_data_nogyr
{dataset}_harnet_unfrozen_low_data_nogyr
{dataset}_RNN_low_data_nogyr
{dataset}_wavelet_RNN_low_data_nogyr
{dataset}_CRNN_low_data
{dataset}_rf_low_data

CSV files: we also include summaries of the experimental results in experiments_summary.csv, experiments_by_fold_individual.csv, experiments_by_fold_behavior.csv. 

experiments_summary.csv - results averaged over individuals and behavior classes
dataset (str): name of dataset
experiment (str): name of model with experiment setting 
fig4 (bool): True if dataset+experiment was used in figure 4 of arxiv paper
fig5d (bool): True if dataset+experiment was used in figure 5d of arxiv paper
fig5e (bool): True if dataset+experiment was used in figure 5e of arxiv paper
f1_mean (float): mean of macro-averaged F1 score, averaged over individuals in test folds
f1_std (float): standard deviation of macro-averaged F1 score, computed over individuals in test folds
prec_mean, prec_std (float): analogous for precision
rec_mean, rec_std (float): analogous for recall

experiments_by_fold_individual.csv - results per individual in the test folds
dataset (str): name of dataset
experiment (str): name of model with experiment setting 
fig4 (bool): True if dataset+experiment was used in figure 4 of arxiv paper
fig5d (bool): True if dataset+experiment was used in figure 5d of arxiv paper
fig5e (bool): True if dataset+experiment was used in figure 5e of arxiv paper
fold (int): test fold index
individual (int): individuals are numbered zero-indexed, starting from fold 1
f1 (float): macro-averaged f1 score for this individual
precision (float): macro-averaged precision for this individual
recall (float): macro-averaged recall for this individual

experiments_by_fold_behavior.csv - results per behavior class, for each test fold
dataset (str): name of dataset
experiment (str): name of model with experiment setting 
fig4 (bool): True if dataset+experiment was used in figure 4 of arxiv paper
fig5d (bool): True if dataset+experiment was used in figure 5d of arxiv paper
fig5e (bool): True if dataset+experiment was used in figure 5e of arxiv paper
fold (int): test fold index
behavior_class (str): name of behavior class
f1 (float): f1 score for this behavior, averaged over individuals in the test fold
precision (float): precision for this behavior, averaged over individuals in the test fold
recall (float): recall for this behavior, averaged over individuals in the test fold
train_ground_truth_label_counts (int): number of timepoints labeled with this behavior class, in the training set