Release 22.7.2.0

Release 22.7.2.0

Benchmarks

Features

• Experimental N-HITS forecasting mixer (#886)
• Differencing blocks for time series tasks (#903)
• Linear tree for LightGBMArray mixer (#902)
• STL decomposition blocks for time series tasks (#907)

Fixes

• Restored statsforecast as default backend for ARIMA models (#904)

Other

N/A

Release 22.6.1.2

Changelog

Features

• #894

Fixes

• #889
• #895
• #896
• #897

Other

• #789

Benchmarks:

• #8b07e93
• #b92c8fd

Release v22.5.1.0

Changelog:

Features

• #865 - ProphetMixer to forecast

Bug fixes

• #866 & #872 - Fix TS float cleaner can't deal with malformed input
• #873 - Fix non-null TS delta analysis

Other

Benchmarks

Release v22.4.1.0

Lightwood 22.4.1.0 changelog:

Features: -

Bug fixes:

• #848 - Log runtime per mixer
• #855 - Fix dimension error in neural mixer

Other

• #850 - Restore Windows CI

Full Changelog:

b05e8b4...v22.4.1.0

Release 22.2.1.0

Lightwood 22.2.1.0 changelog:

Features:

• Simpler and better Json AI #826
• Compute & log time per phase #828

Bug fixes: -

Other:

• Remove anomaly_error_rate arg in favor of fixed_confidence #825

Full Changelog: v22.1.4.0...v22.2.1.0

Release 22.1.4.0

Lightwood 22.1.4.0 changelog:

Moving forward, our release versioning schema will follow this format:
[year's last two digits].[month].[week].[patch]

(note: other MindsDB repositories will also switch to this)

Features:

• ConfStats block (#800): provides calibration insights for the lightwood predictor
• Temperature scaling block (#795, experimental & non-default): alternative to ICP block for confidence estimation
• Improved documentation pages (#806)
• Replaced default encoder for time series forecasting tasks (from RNN to simple MA features, #805)
• Explicit detrend and deseasonalize options for sktime mixer (#812)

Bug fixes:

• Updated update model tutorial (#774)
• Fix forecast horizon lower bound (#801)
• Handle empty input when predicting (#811)

Other

• Rename nr_predictions parameter to horizon (#803)
• Set allow_incomplete_history to True by default (#818)

Full Changelog

v1.9.0...v22.1.4.0

Release 1.9.0

Lightwood 1.9.0 changelog:

Features:

• Improved T+N forecast bounds (#788)
• Optimized classifier ICP block for confidence estimation (#798)

Bug fixes:

• Fixed initialization issues in confidence normalizer (#788)
• Fixed no analysis mode (+ parameter to specify this in a problem definition, #791)
• Fixed temporal delta estimation for ungrouped series (#792)

Other

• Add original query index column in output (used internally in MindsDB, #794)
• Streamlined explain() arg passing #797

Release 1.8.0

Lightwood 1.8.0 changelog:

Features:

• SkTime mixer 2.0 (#758, #767)
• Improve time aim feature (#763)
• Improved OHE and binary encoders, standardized a few more (#755, #785)
• Streamlined predictor.adjust signature (#762)
• Add precision, recall, f1 (#776)

Bug fixes:

• Do not drop single-group-by column (#761, #756)
• OH and Binary Encoders weighting fix (#769)
• LGBM array mixer does not modify the datasource (#771)
• Fixes missing torchvision import (#784)

Other

• Make image encoder optional (#778)
• Revamp notebooks test docs (#764)

Release 1.7.0

Lightwood 1.7.0 changelog:

Features:

• Simplified type mapping in Json AI (#724)
• Setter for neural mixer # epochs (#737)
• Improved nan handling (#720)
• Drop columns with no information (#736)
• LightGBM mixer supports weights (#749)
• Improved OneHot and Binary encoders' logic around weights (#749)
• New accuracy function lookup hierarchy (#754)
• Better warning logs when nan or inf values are encountered (#754)

Bug fixes:

• Fixed LightGBM error on CPU (#726)
• Cast TS group by values to string to avoid TypeError (#727)
• Check target values when transforming time series if task requires them (#747)
• Streamline encode/decode in TsArrayNumericEncoder (#748)
• target_weights argument is now used properly (#749)
• Use custom R2 accuracy to account for edge cases (#754)
• Fixed target dropping behavior (#754)

Other

• Update README.md example (#731)
• Separate branch for docs (#740)
• Docs for image and audio encoders; LightGBM and LinearRegression mixers (#721, #722)

Release 1.6.0

Lightwood 1.6.0 changelog:

Many thanks to our community contributors for this release!
@MichaelLantz @mrandri19 @ongspxm @vaithak

Features:

• SHAP analysis block (#679, @mrandri19)
• Disable GlobalFeatureImportance when we have too many columns (#681, @ongspxm; #698)
• Added cleaner support for file path data types (image, audio, video) (#675)
• Add partial_fit() to sktime mixer (#689)
• Add ModeEnsemble (#692, @mrandri19)
• Add weighted MeanEnsembler (#680, @vaithak)

Bug fixes:

• Normalized column importance range (#690)
• Fix ensemble supports_proba in calibrate.py (#694, @mrandri19)
• Remove self-referential import (#696)
• Make a integration test for time_aim (#685, @MichaelLantz)
• Fix for various datasets (#700)

Other

• Improve logging for analysis blocks (#677; @MichaelLantz)
• Custom block example: LabelEncoder (#663)
• Implement ShapleyValues analysis (#679)
• Move array/TS normalizers to generic helpers (#702)