sensorLeak

sensorLeak provides tools for identifying potential validation and information leakage risks in machine learning workflows using environmental sensor data.

The package focuses on leakage risks that can arise from the way environmental observations are divided into training and testing data, including temporal ordering, shared sensors, and spatial proximity.

Installation

You can install the development version of sensorLeak from GitHub:

# install.packages("pak")
pak::pak("division55/sensorLeak")

Example

Create a simple environmental sensor dataset:

library(sensorLeak)

sensor_data <- data.frame(
  timestamp = as.POSIXct(c(
    "2026-01-01 10:00:00",
    "2026-01-01 12:00:00",
    "2026-01-01 14:00:00",
    "2026-01-01 16:00:00"
  )),
  sensor_id = c("S1", "S1", "S1", "S2"),
  latitude = c(
    13.0827,
    13.0830,
    13.0832,
    13.3000
  ),
  longitude = c(
    80.2707,
    80.2710,
    80.2712,
    80.4000
  ),
  split = c(
    "train",
    "train",
    "test",
    "test"
  )
)

Run the audit:

audit <- sensor_audit(
  sensor_data,
  time = "timestamp",
  sensor = "sensor_id",
  split = "split",
  lat = "latitude",
  lon = "longitude"
)

Print the audit:

audit

Summarize the results:

summary(audit)

Leakage checks

sensorLeak currently provides four diagnostic functions:

Temporal leakage

check_temporal_leakage(
  sensor_data,
  time = "timestamp",
  split = "split"
)

Identifies test observations that occur at or before the end of the training period.

Sensor-group overlap

check_sensor_leakage(
  sensor_data,
  sensor = "sensor_id",
  split = "split"
)

Identifies sensors that occur in both training and testing data.

This should be interpreted according to the intended validation strategy. Sharing a sensor between training and testing data is not automatically invalid; it can be appropriate when the model is intended to predict future observations from known sensors.

Spatial proximity

check_spatial_leakage(
  sensor_data,
  lat = "latitude",
  lon = "longitude",
  split = "split",
  threshold = 1
)

Identifies training/testing observation pairs that are within the specified geographic distance in kilometres.

Spatial proximity may represent a validation risk when the intended goal is geographic generalization.

Overlapping temporal windows

check_window_leakage(
  data,
  start = "window_start",
  end = "window_end",
  split = "split",
  sensor = "sensor_id"
)

Identifies temporal windows that overlap between training and testing data for the same sensor.

Important interpretation

sensorLeak reports potential validation and information leakage risks. A detected condition is not necessarily an error.

Whether a condition represents leakage depends on the scientific question and the intended generalization target of the machine learning workflow.

For example, a model intended to predict future measurements from known sensors may legitimately contain the same sensors in both training and testing data.

Related packages

sensorLeak is designed as a diagnostic tool rather than a replacement for packages that construct spatial or temporal validation schemes.

Users may also consider packages such as blockCV and CAST when constructing spatial or space-time cross-validation schemes.

Development

This package is under active development.

Run the test suite with:

devtools::test()

Run package checks with:

devtools::check()

License

MIT