Skip to content

Releases: ccao-data/lightsnip

Release list

0.0.7

Choose a tag to compare

@jeancochrane jeancochrane released this 15 Jul 16:54
Immutable release. Only release title and notes can be modified.
855d364

What's changed

Full Changelog: 0.0.6...0.0.7

New features

  • Add support for save_tree_error train argument to preserve record_evals by @jeancochrane in #10

    • Summary: train_lightgbm() now supports an optional boolean parameter save_tree_error, which allows the caller to specify that the model should use its training data as a validation set in order to save a record_evals attribute that records the error for each tree in the model. This option supports the experimental comps algorithm described in the vignette Finding comparables with LightGBM.
  • Add MSE + covariance penalty term custom objective function by @wagnerlmichael in #21

    • Summary: train_lightgbm() now supports an optional objective = "mse_cov" parameter value and an optional mse_cov_rho parameter to enable a new objective function that takes mean squared error and adds an additional penalty on the squared covariance between prediction residuals and centered sale prices. This method is based on forthcoming research by @nicacevedo and Haihao Lu at MIT that aims to reduce regressivity in mass appraisal. See their code at nicacevedo/soft-vertical-equity-constrained-mass-appraissal for details.
      • New parameter values:
        • objective = "mse_cov": Enables the new objective function
      • New parameters:
        • mse_cov_rho (optional integer): Rho value for the covariance penalty term; controls how much to penalize covariance between prediction residuals and centered sale prices

Bug fixes

  • Update lgbm_load() to only load record_evals if they exist by @jeancochrane in #13

Documentation

  • Create a vignette outlining how to extract comparables from LightGBM by @dfsnow in #3
  • Update comps vignette copy and tables by @dfsnow in #6
  • Update comps vignette with latest version of the similarity scoring algorithm by @jeancochrane in #19
  • Enable llms.txt for docs by @Damonamajor in #24

Maintenance

Dependencies

New Contributors

Special thanks to @nicacevedo and Haihao Lu for developing the covariance penalty objective function added in #21.

0.0.6

Choose a tag to compare

@dfsnow dfsnow released this 05 Jul 02:02
  • Update supporting files for transition to GitHub