Documentation
¶
Overview ¶
Package targetencode is the one reading of FEAT_TARGET_ENCODE's params, shared by the operator (internal/processing/feature) and predict (internal/descriptor) so both refuse the same request with the same code and decide the same streaming verdict. It decides what a request MEANS and never averages an outcome.
The params rules, in the order they are judged (after the params object decoded):
- target is required (the numeric outcome averaged per category);
- smoothing must be >= 0;
- fit_on, when present, must be a non-blank expression (its compile is the field-reference walk's, the one every expression gets);
- folds is 0 (off) or an integer in [2, MaxFolds];
- rng_seed is read only with folds: alone it is refused, as it would seed nothing.
Every refusal is PROCESSING_CONFIG on both sides.
Leaf: stdlib + the public errors / encoding packages + exprident.
Index ¶
Constants ¶
const MaxFolds = 100
MaxFolds caps params.folds: the out-of-fold state is one per-category tally per fold.
Variables ¶
This section is empty.
Functions ¶
func Check ¶
func Check(p Params) *errors.CodedError
Check returns the first params rule p breaks, or nil. The message is the operator's own ("FEAT_TARGET_ENCODE: ..."), raised verbatim by the constructor and by predict.
func Streamable ¶
Streamable reports whether a FEAT_TARGET_ENCODE over field with p can run on the streaming feature drive. Its pre-pass reads field, target and every name fit_on reads from the RAW record, before any feature has emitted, so each must be a stored schema field: a name an earlier feature derives (a split column, a log) is absent there, and the request routes buffered, where features apply in order. A fit_on that does not parse, or reads $env, is judged unstreamable too. A nil schema knows no field.
Types ¶
type Params ¶
type Params struct {
// Target is the numeric outcome averaged per category.
Target string `json:"target"`
// Smoothing is the additive prior weight toward the global mean.
Smoothing float64 `json:"smoothing"`
// FitOn is a boolean expression choosing the training rows: only
// they feed the category and global means; every row is encoded.
// Empty fits on every row (the leakage-prone default).
FitOn string `json:"fit_on"`
// Folds, when >= 2, encodes each training row out-of-fold: from the
// training rows outside its own fold.
Folds int `json:"folds"`
// RngSeed seeds the fold assignment.
RngSeed *int64 `json:"rng_seed"`
}
Params is FEAT_TARGET_ENCODE's params wire shape.
{"target": "default", "smoothing": 10, "fit_on": "split == 0", "folds": 5, "rng_seed": 7}
Source Files