Open-source agentic AI for transaction monitoring. Just ask Jooby.What's Jube? →
jube
  • Documentation
  • Analytical Methodology
  • Other Model Types
  • K-Nearest-Neighbours Regressor

K-Nearest-Neighbours Regressor

The regression counterpart predicts the plain mean of the k nearest neighbours’ targets — no vote fraction to calibrate, since that mean is already on the target’s own scale.

R equivalent

library(FNN)

set.seed(1)
n <- 200
account_age_days <- runif(n, 1, 2000)
avg_monthly_txn <- runif(n, 1, 60)
settlement_days <- 12 - 0.004 * account_age_days + 0.05 * avg_monthly_txn +
  rnorm(n, 0, 1.5)

train_x <- scale(data.frame(account_age_days, avg_monthly_txn))
new_cases <- data.frame(account_age_days = c(30, 1500), avg_monthly_txn = c(5, 45))
new_x <- scale(new_cases, center = attr(train_x, "scaled:center"),
               scale = attr(train_x, "scaled:scale"))

result <- knn.reg(train = train_x, test = new_x, y = settlement_days, k = 15)
print(result$pred)

Console output:

[1] 11.455776  8.206563

Tool equivalent: Jube.Service.Repository.AnalyticsTrainingRun.AdvancedAnalyticsKNearestNeighboursRegressorTrainingService / tool SubmitKNearestNeighboursRegressorTraining — Kind = Regressor, recall class KNearestNeighboursRegressorRecall. Same NeighbourCount/Standardise shape as the classifier, no calibration fields, same stored-rows recall model.

Code audit

AdvancedAnalyticsKNearestNeighboursRegressorTrainingService delegates to KNearestNeighboursRegressorTrainer.Train (Jube.Engine/AdvancedAnalytics/Regression/KNearestNeighboursRegressorTrainer.cs:33-55), the regression sibling of the classifier trainer, and it shares the same hand-written implementation: no Accord.MachineLearning.* type is instantiated anywhere in this family either.

  • Distance metric and search structure: identical to the classifier — squared Euclidean distance summed per variable, brute-force linear scan of every stored row (KNearestNeighboursRegressorPipeline.Explain, Jube.Engine/AdvancedAnalytics/Regression/KNearestNeighboursRegressorPipeline.cs:84-107; square root taken only for the reported distances at line 118, not for ranking).
  • Tie-breaking: the same stable-sort rule as the classifier — Enumerable.Range(0, _rows.Length).OrderBy(i => distancesSquared[i]).Take(NeighbourCount) (lines 104-107) — so equidistant neighbours are resolved by ascending storage index.
  • k validated against dataset size: KNearestNeighboursRegressorTrainer.TrainWithSplit throws if trainX.Length < options.NeighbourCount before fitting (lines 82-85), and KNearestNeighboursRegressorPipeline’s constructor repeats the check (lines 40-43). As with the classifier, only the training split size is checked, not validation or test.
  • Prediction: the plain arithmetic mean of the k neighbours’ stored targets (neighbourTargets.Average(), KNearestNeighboursRegressorPipeline.Explain line 135) — an unweighted mean, not a distance-weighted (inverse-distance) average; every one of the k neighbours contributes equally regardless of how much closer some are than others.
  • No calibration surface: KNearestNeighboursRegressorOptions (Jube.Engine/AdvancedAnalytics/Regression/KNearestNeighboursRegressorOptions.cs:19-38) carries no Calibrate or DecisionThreshold field at all — confirmed by reading the options class directly, not inferred — since a regression target has nothing analogous to calibrate.
  • Snapshot cost: as with the classifier, every standardised, imputed training row and its target value is stored in the model version (AdvancedAnalyticsKNearestNeighboursRegressorPipelineSnapshot, KNearestNeighboursRegressorTrainer.cs:187-200), and recall repeats the same brute-force search, so recall cost scales with the training split size; the same LargeSnapshotWarningThreshold (200,000 rows, lines 31, 181-185) documents this directly in a training warning.