stickleback 0.1.1

Automated behavioral event detection in bio-logging data.

stickleback

A machine learning pipeline for detecting fine-scale behavioral events in bio-logging data.

Installation

Install with pip.

pip install stickleback

Key concepts

• Behavioral events are brief behaviors that can be represented as a point in time, e.g. feeding or social interactions.
• High-resolution bio-logging data (e.g. from accelerometers and magnetometers) are multi-variate time series. Traditional classifiers struggle with time series data.
• stickleback takes a time series classification approach to detect behavioral events in longitudinal bio-logging data.

Quick start

Load sample data

The included sensor data contains the depth, pitch, roll, and speed of six blue whales at 10 Hz, and the event data contains the times of lunge-feeding behaviors.

import pandas as pd
import sktime.classification.interval_based
import sktime.classification.compose
from stickleback.stickleback import Stickleback
import stickleback.data
import stickleback.util
import stickleback.visualize

# Load sample data
sensors, events = stickleback.data.load_lunges()

# Split into test and train (3 deployments each)
def split_dict(d, ks):
    dict1 = {k: v for k, v in d.items() if k in ks}
    dict2 = {k: v for k, v in d.items() if k not in ks}
    return dict1, dict2

test_deployids = list(sensors.keys())[0:2]
sensors_test, sensors_train = split_dict(sensors, test_deployids)
events_test, events_train = split_dict(events, test_deployids)

sensors[test_deployids[0]]

<style scoped> .dataframe tbody tr th:only-of-type { vertical-align: middle; }

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

</style>

	depth	pitch	roll	speed
datetime
2018-09-05 11:55:52.400	14.911083	-0.059933	-0.012899	4.274450
2018-09-05 11:55:52.500	14.910864	-0.067072	-0.010815	4.044154
2018-09-05 11:55:52.600	14.915853	-0.075173	-0.008335	3.820568
2018-09-05 11:55:52.700	14.923190	-0.085225	-0.005727	3.602702
2018-09-05 11:55:52.800	14.928955	-0.096173	-0.002803	3.432342
...	...	...	...	...
2018-09-05 13:55:51.900	22.552306	-0.010861	0.005441	2.246061
2018-09-05 13:55:52.000	22.571625	-0.010534	0.004674	2.257525
2018-09-05 13:55:52.100	22.588129	-0.010081	0.003841	2.267966
2018-09-05 13:55:52.200	22.603341	-0.009627	0.003042	2.272327
2018-09-05 13:55:52.300	22.619537	-0.009355	0.002164	2.277328

72000 rows × 4 columns

Visualize sensor and event data

plot_sensors_events() produces an interactive figure for exploring bio-logger data.

# Choose one deployment to visualize
deployid = list(sensors.keys())[0]
stickleback.visualize.plot_sensors_events(deployid, sensors, events)

Define model

Initialize a Stickleback model using Supervised Time Series Forests and a 10 s window.

# Supervised Time Series Forests ensembled across the columns of `sensors`
cols = sensors[list(sensors.keys())[0]].columns
tsc = sktime.classification.interval_based.SupervisedTimeSeriesForest(n_estimators=2)
stsf = sktime.classification.compose.ColumnEnsembleClassifier(
    estimators = [('STSF_{}'.format(col),
                   tsc,
                   [i])
                  for i, col in enumerate(cols)]
)

sb = Stickleback(
    local_clf=stsf,
    win_size=50,
    tol=pd.Timedelta("5s"),
    nth=10,
    n_folds=4
)

Fit model

Fit the Stickleback object to the training data.

sb.fit(sensors_train, events_train)

Test model

Use the fitted Stickleback model to predict occurence of lunge-feeding events in the test dataset.

predictions = sb.predict(sensors_test)

Assess results

Use the temporal tolerance (in this example, 5 s) to assess model predictions. Visualize with an outcome table and an interactive visualization. In the figure, blue = true positive, hollow red = false negative, and solid red = false positive.

outcomes = sb.assess(predictions, events_test)
stickleback.visualize.outcome_table(outcomes, sensors_test)

<style scoped> .dataframe tbody tr th:only-of-type { vertical-align: middle; }

.dataframe tbody tr th {
    vertical-align: top;
}

.dataframe thead th {
    text-align: right;
}

</style>

	F1	TP	FP	FN	Duration (hours)
deployid
bw180905-49	0.965517	42	1	2	1.999972
bw180905-53	0.981132	26	1	0	1.999972

deployid = list(events_test.keys())[0]
stickleback.visualize.plot_predictions(deployid, 
                                       sensors_test, 
                                       predictions, 
                                       outcomes)