Documentation
¶
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func F1 ¶
F1 computes the balanced multiclass F1 score.
It is computed with the following formula:
F1 = 2 * (precision * recall) / (precision + recall)
func R1 ¶
R1 is the minimization inner objective function (equation 7) from "Optimizing Classifier Performance via an Approximation
to the Wilcoxon-Mann-Whitney Statistic"
by Yan, Dodier, Mozer and Wolniewicz, Proceedings of ICML 2003.
It is designed to be differentiable without too much difficulty.
pred1 = a prediction for a known target of 1 pred0 = a prediction for a known target of 0
The two meta-parameters for the optimization are margin and power: a) 0 < margin <= 1 improves generalization; and b) power > 1 controls how strongly the margin minus the difference between pred1 and pred0 is amplified; the margin-difference is raised to power before being returned.
In examples they try margin = 0.3 and power = 2; or margin = 0.2 and power = 3; but these should be optimized in an outer loop.
Yan: "In general, we can choose a value between 0.1 and 0.7 for margin. Also, we have found that power = 2 or 3 achieves similar, and generally the best results."
INVAR: power > 1.0 INVAR: 0 < margin <= 1 These invariants should be checked by the caller for speed and to allow potential inlning. They are not checked within R1().
Types ¶
type AUCroc ¶
type AUCroc struct {
TargetRange TargetRange
Auc1 float64
Auc0 float64
U1 float64
U0 float64
N1 float64
N0 float64
// Objective function value, set if TargetRrange.ComputeObjective is true.
Obj float64
}
AUCroc is returned by AreaUnderROC() call.
func AreaUnderROC ¶
func AreaUnderROC(predictor, target []float64, targetRanges []*TargetRange) (res []*AUCroc, err error)
AreaUnderROC() is for classifying a target real values as <= targetRanges[i].Threshold versus target > targetRanges[i].Threshold, for each i over the supplied targetRanges. The AUC for ROC is equivalent to the Wilcoxon or Mann-Whitney U test statistic with the relation:
AUC = U/(n0 * n1)
No NaN handling at present so handle those prior. This could be added without too much difficulty. VposSlice will sort them to the end, but another pass would be needed to keep pair-wise complete target/predictor pairs.
returns: auc1 is the area under the curve for classifying target > targetRanges[i].Threshold (typically what one wants). auc0 is the area under the curve for classifying target <= targetRanges[i].Threshold.
type ConfusionMatrix ¶
type ConfusionMatrix struct {
*tensor.Dense
Labels []string // optional Labels
Iter [][]float64 // a native iterator of the confusion matrix. This allows for quick access.
// contains filtered or unexported fields
}
ConfusionMatrix represents a confusion matrix.
It is ordered in the following shape (predicted, actual)
func Confusion ¶
func Confusion(pred, correct qol.Classes) *ConfusionMatrix
Confusion creates a Confusion Matrix.
func (*ConfusionMatrix) Heatmap ¶
func (m *ConfusionMatrix) Heatmap() (*plot.Plot, error)
Heatmap is a convenience method to create a heatmap. It creates a *gonum.org/v1/plot.Plot, which contains a heatmap structure. The Palette, X, Y are all accessible for customization.
Not safe to be run concurrently.
type TargetRange ¶
type TargetRange struct {
// We just have just one threshold for now, but could be a range
// or one range vs. another for ordinal/multiclass disrimination.
Thresh float64
// should we compute the U_R objective function, a differentiable
// approximation to the Area under ROC, but designed to be minimized.
//
// Not implemented: actual optimization by using the derivative
// of the objective with respect to inputs X. Hence think of
// this simply as a proof of concept to allow comparison of
// the actual ROC and the approximation function to check for
// applicability on a given data context.
//
ComputeObjective bool
// If ComputeObjective is true, then Margin and Power must be set.
Margin float64 // Margin must be > 0, and Margin must <= 1. Starting guess might be 0.2
Power float64 // Power must be > 1. Starting guess might be 3.0
}
TargetRange defines how to map a real target feature to a category/class.
Source Files