Metrics roccurve

Development build. This page describes main, not a released package. The latest published Lodestar.Metrics is 0.3.0 — read its documentation.

RocCurve

The receiver operating characteristic as plot data rather than as a number: RocAuc.Score tells you the area, this tells you the shape, and the shape is what says where to put a threshold.

Three parallel arrays of the same length — the false-positive rate, the true-positive rate, and the score at each point. A class rather than a record, for the reason decision 0040 gives.

The first threshold is infinite

Thresholds[0] is +∞, and both rates are 0 there: no sample scores above infinity, so nothing is predicted positive, which is the origin the curve has to start from. The reference prepends that point rather than deriving it, and so does this — a caller iterating the arrays in parallel must expect it.

dropIntermediate defaults to true here and to false on the other two

That asymmetry is scikit-learn's, and it is reproduced rather than normalised so that a caller porting from Python gets the same array lengths without reading a signature. The three curves also drop by two different rules:

• This one drops a point the curve does not bend at — where the second difference of both counts vanishes, so the point is collinear with its neighbours.
• PrecisionRecallCurve and DetCurve drop a point whose true-positive count matches both neighbours, because such points share a recall and stack on one vertical line.

Measured on a ten-sample fixture: this curve goes from 11 points to 5, the precision-recall curve from 11 to 8, and the DET curve from 11 to 8.

Members

Member	What it does
RocCurve.Compute	Draws the curve from labels and scores.