Experiment support for a parameter sweep: a deterministic stratified hold-out split, class-ratio resampling of the training half against a fixed test set, a wall-clock stopwatch and label helpers. Compile with -lib ml,ml_lab,csv,json, or with the ai alias group. Design: docs/ML_LAB_SUPPORT_DESIGN.md.
Resample
Changes the class balance of a TRAINING set to a requested positive ratio.
The class-imbalance experiment trains at ratios from 0.001 to 0.5 and must measure every run on the same test set. These functions take only the training half, so the order is fixed by construction: Partition->TrainTest first, then resample what it returned as training data; the test half is never touched.
Example
split := Partition->TrainTest(X, y, 0.3, 42, true);
each(r : ratios) {
train := Resample->ToPositiveRatio(split->GetTrainX(), split->GetTrainY(), ratios[r], 42);
model->Fit(train->GetX(), train->GetY());
preds := model->PredictClass(split->GetTestX()); # the same test set every time
};Operations
CountPositives # function
Number of positive labels.
function : CountPositives(y:Bool[]) ~ IntParameters
| Name | Type | Description |
|---|---|---|
| y | Bool | labels |
Return
| Type | Description |
|---|---|
| Int | count of true values, 0 for Nil |
PositiveRatio # function
Share of positive labels.
function : PositiveRatio(y:Bool[]) ~ FloatParameters
| Name | Type | Description |
|---|---|---|
| y | Bool | labels |
Return
| Type | Description |
|---|---|
| Float | positives over rows, 0.0 for an empty or Nil array |
ToPositiveRatio # function
Down-samples the over-represented class until positives / rows is the requested ratio; the other class is kept whole.
If the current ratio is above the request, positives are cut to Round(ratio * negatives / (1 - ratio)); otherwise negatives are cut to Round(positives * (1 - ratio) / ratio). The rows kept are chosen by a seeded shuffle and returned in ascending source order. The achieved ratio is exact when the arithmetic allows and otherwise within one row; Sample->GetPositiveRatio reports it.
function : ToPositiveRatio(X:Float[,], y:Bool[], ratio:Float, seed:Int) ~ SampleParameters
| Name | Type | Description |
|---|---|---|
| X | Float | training rows x features |
| y | Bool | one training label per row |
| ratio | Float | requested positives / rows, in (0, 1) |
| seed | Int | seeds the choice of rows |
Return
| Type | Description |
|---|---|
| Sample | the resampled training set; raises LabError when a class is empty, the ratio is out of range, or the request would keep zero rows of a class |
ToPositiveRatio # function
Reaches the requested ratio by down-sampling (oversample false, as above) or by duplicating the under-represented class (oversample true): rows of that class are drawn with replacement from a seeded generator and appended after the original rows, and nothing is dropped.
function : ToPositiveRatio(X:Float[,], y:Bool[], ratio:Float, seed:Int, oversample:Bool) ~ SampleParameters
| Name | Type | Description |
|---|---|---|
| X | Float | training rows x features |
| y | Bool | one training label per row |
| ratio | Float | requested positives / rows, in (0, 1) |
| seed | Int | seeds the choice of rows |
| oversample | Bool | true duplicates the minority instead of dropping the majority |
Return
| Type | Description |
|---|---|
| Sample | the resampled training set; raises LabError on bad input |