LoginSign Up
← Back to all sparks
M

mlr3fda

ANALYTICS
Velocity2.5

Functional Data Analysis for mlr3

mlr3fda.mlr-org.com ↗

A functional-data feature factory for mlr3, shipping a new extractor almost every month.

mlr3functional-datafeature-extractionpipelinestime-series
◆Current state
mlr3fda adds functional data support to mlr3 pipelines through PipeOps that turn functional columns into tabular features. Five releases since March 2026 have taken it from Fourier features to a catalogue covering wavelets, derivatives, depth, registration, integration and the catch22 time-series feature set. Development tracks the {tf} package closely.
◆Where it's heading
The pattern is one or two new PipeOps per release with fixes to the previous batch alongside — Fourier in 0.4.0, registration in 0.5.0, depth and derivatives in 0.6.0, catch22 and integration in 0.7.0. Performance and parallel-safety work is folded in as it becomes necessary rather than deferred, as with the Fourier speedup and the mlr_reflections registration fix.
◆Prediction
The extraction catalogue is filling out along established functional-data methods, so further tf-backed PipeOps are the likely next additions; the {tf} 0.5.0 compatibility release suggests upstream churn will keep setting the pace.

◆Recent moves

  1. 29d ago

    Compatibility with tf 0.5.0

    A same-fortnight follow-up to 0.7.0 carrying only upstream compatibility with {tf} 0.5.0.

    View source ↗
  2. 1mo ago

    catch22 and integral feature extractors added

    PipeOpFDACatch22 brings the catch22 time-series feature set to functional columns and PipeOpFDAIntegrate computes definite integrals, alongside wavelet filter and one-sided zoom fixes.

    View source ↗
  3. 2mo ago

    Depth and derivative extractors; Fourier made several times faster

    Adds PipeOpFDADepth and PipeOpFDADerive, gives PipeOpFDAExtract a standard-deviation feature, and registers the package in mlr_reflections so it stops erroring under parallelisation. Fourier extraction becomes several times faster.

    View source ↗
  4. 3mo ago

    Curve registration via PipeOpFDARegister

    Functional columns can be aligned to a learned template via PipeOpFDARegister, which is the standard preprocessing step before most functional-data feature extraction.

    View source ↗
  5. 4mo ago

    Fourier feature extraction and package-specific condition classes

    PipeOpFDAFourier introduces FFT-based feature extraction and Mlr3Error/Mlr3Warning condition classes are added; a task column typo is corrected from h20 to h2o.

    View source ↗