The accuracy table metric is a multi-use non-scalar metric that can be used to produce multiple types of line charts that vary continuously over the space of predicted probabilities. Examples of these charts are ROC, precision-recall, and lift curves.
log_accuracy_table_to_run(name, value, description = "", run = NULL)
A string of the name of the metric.
A named list containing name, version, and data properties.
(Optional) A string of the metric description.
The calculation of the accuracy table is similar to the calculation of an ROC curve. An ROC curve stores true positive rates and false positive rates at many different probability thresholds. The accuracy table stores the raw number of true positives, false positives, true negatives, and false negatives at many probability thresholds.
There are two methods used for selecting thresholds: "probability" and "percentile." They differ in how they sample from the space of predicted probabilities.
Probability thresholds are uniformly spaced thresholds between 0 and 1. If NUM_POINTS were 5 the probability thresholds would be c(0.0, 0.25, 0.5, 0.75, 1.0).
Percentile thresholds are spaced according to the distribution of predicted probabilities. Each threshold corresponds to the percentile of the data at a probability threshold. For example, if NUM_POINTS were 5, then the first threshold would be at the 0th percentile, the second at the 25th percentile, the third at the 50th, and so on.
The probability tables and percentile tables are both 3D lists where the first dimension represents the class label, the second dimension represents the sample at one threshold (scales with NUM_POINTS), and the third dimension always has 4 values: TP, FP, TN, FN, and always in that order.
The confusion values (TP, FP, TN, FN) are computed with the one vs. rest strategy. See the following link for more details: https://en.wikipedia.org/wiki/Multiclass_classification.
N = # of samples in validation dataset (200 in example), M = # thresholds = # samples taken from the probability space (5 in example), C = # classes in full dataset (3 in example)
Some invariants of the accuracy table:
TP + FP + TN + FN = N for all thresholds for all classes
TP + FN is the same at all thresholds for any class
TN + FP is the same at all thresholds for any class
Probability tables and percentile tables have shape (C, M, 4)
Note: M can be any value and controls the resolution of the charts. This is independent of the dataset, is defined when calculating metrics, and trades off storage space, computation time, and resolution.
Class labels should be strings, confusion values should be integers, and thresholds should be doubles.