Anomaly-Aware Supervised Learning — tl_anomaly

Detect outliers using DBSCAN or other methods, then optionally remove them or down-weight them before supervised learning.

Usage

tl_anomaly_aware(
  data,
  formula,
  response,
  anomaly_method = "dbscan",
  action = "flag",
  supervised_method = "tree",
  ...
)

Arguments

data: A data frame
formula: Model formula
response: Response variable name, left out of the detection
anomaly_method: Method for anomaly detection. Only "dbscan" is implemented; its noise points are the anomalies.
action: Action to take: "remove", "flag", "downweight". "downweight" gives anomalies a case weight of 0.1, and needs a supervised_method that takes case weights: "linear", "polynomial", "logistic", "tree", "ridge", "lasso", "elastic_net", "forest", "boost", "nn" or "xgboost". A forest reads them as sampling weights. "svm" and "deep" take none and are refused.
supervised_method: Supervised learning method (default: "tree", which handles both regression and classification with any number of classes). "logistic" is binary-only and errors on a response with more than two levels.
...: Additional arguments for DBSCAN, such as eps and minPts

Value

A tidylearn model object with additional class "tidylearn_anomaly_aware". The model includes an anomaly_info element with anomaly_model, is_anomaly (logical vector), n_anomalies, and action.

Details

DBSCAN runs on the predictors the formula names, on their own scale: its eps and minPts (defaults 0.5 and 5, passed through ...) are a distance and a count in those units. If every row comes out as noise the call stops, since no normal data would be left to model.

Examples

# \donttest{
model <- tl_anomaly_aware(iris, Species ~ ., response = "Species",
                           anomaly_method = "dbscan", action = "flag")
# }