Implementation of the Jaccard Timespan Event Score (JTES)

This is a reference implementation of the Jaccard Timespan Event Score (JTES) described here.

Jaccard-Timespan-Event-Score (JTES) is a timespan score based on the Jaccard index also known as Intersection over Union (IoU).

This index is designed to score events that are defined by a timespan defined by two timestamps (t0, t1).

Score Goals:

• Score is in range [0-1], where 0 is lowest and 1 best.
• False-Positives and False-Negatives are equally bad
• If a true event spans over multiple predicted events, the result is averaged
• If the predicted event spans over multiple true events, the result is accounted accordingly
• If len(y_true) == 0 and len(y_pred) > 0, score = 0
• If len(y_true) > 0 and len(y_pred) == 0, score = 0
• If ordered(y_true) == ordered(y_pred), score = 1
• Overlap in y_true is not allowed

Jaccard index in general is defined as |A ∩ B| / |A ∪ B|.

Install

The jtes package is available on pypi

pip install jtes

Install from source (alternative)

python setup.py install

Usage

The jaccard_timespan_event_score function expects two numpy arrays y_true and y_pred. The events are defined as np.datetime64 pairs.

import numpy as np
from jtes import jaccard_timespan_event_score

y_true = np.array([
    (np.datetime64('1900-01-01T00:00:00'), np.datetime64('1900-01-01T01:00:00')),
    (np.datetime64('1900-01-01T03:00:00'), np.datetime64('1900-01-01T04:00:00'))
])

y_pred = np.array([
    (np.datetime64('1900-01-01T00:00:00'), np.datetime64('1900-01-01T01:00:00')),
    (np.datetime64('1900-01-01T03:00:00'), np.datetime64('1900-01-01T05:00:00')),
])

# Returns 0.75
jaccard_timespan_event_score(y_true, y_pred)

License

MIT licensed as found in the LICENSE file.