Skip to main content

EDA changelog

v.next​

  • GROK-20932: Removed the Playwright tests of the old ML | Models Train view (playwright/MLMethods), which the platform replaced with Forge
  • GROK-20932: ML methods: Fixed training XGBoost, SVM, Softmax and Linear Regression through the API with inputs left out: they now take the defaults of the function's inputs
  • core#1518: Required the Forge plugin for predictive modeling with EDA's methods in the platform's UI; the earlier tool no longer works with this version (needs Datagrok or later)
  • GROK-20932: Added SVM's Max iterations (default 1000000), so a non-converging fit stops and still returns a model
  • GROK-20932: Added meta.mlprobabilities on every classifier's train function and the Class probabilities input of SVM (Platt scaling, set by Forge's Predict probability)
  • GROK-20932: Rewrote the XGBoost hyperparameter tooltips after the XGBoost parameter docs
  • GROK-20932: Added the Row sample and Column sample hyperparameters (subsample / colsample_bytree, default 1, searched in 0.5-1) and a Model seed to XGBoost; the seed matters only when a sample is below 1
  • GROK-20932: Added a Model seed hyperparameter to Linear Regression, where Crux takes one; with the current settings the model does not depend on it
  • GROK-20932: Added meta.mlenabled on train (the inputs Forge disables by the other inputs or the task): SVM gamma / degree / offset by kernel and epsilon for regression only; Linear Regression iterations only with L1 > 0; XGBoost seed only with a sample below 1
  • GROK-20932: Renamed the RealMLP and TabM Validation fraction input to Early stopping fraction
  • GROK-20932: Added a pool of Crux module workers (parallel fits for Forge's search) and the mlparallel / mlspace meta on every train function (search spaces for Forge's Tune)
  • GROK-20932: Added the classificationReport function: Crux (scikit-learn) classification metrics, confusion matrix, per-class values and ROC / precision-recall curves of an actual and a predicted column, used by Forge
  • GROK-20932: Added class probabilities to the apply functions of every classifier but SVM: one proba_ column per class after the prediction
  • GROK-21061: Pareto front: Fixed result columns staying in the table after the viewer is closed
  • GROK-21017: Skip adding MVA viewers when the source table view was closed during the analysis
  • GROK-20932: Added the knnImpute function: headless k-NN imputation of chosen columns, used by Forge
  • GROK-20932: Marked trainSVM with meta.mlhasrows: the model embeds training rows
  • GROK-20932: Added files/legacy-models: one pre-Crux model per ML method with the Legacy models test category
  • GROK-20932: Added the Crux bridge (src/crux): lazy Crux module worker, DataFrame <-> Arrow IPC, cruxFit / cruxPredict
  • GROK-20932: Replaced the PCA, PLS, Linear Regression, Softmax, PLS Regression, XGBoost and SVM kernels with Crux (scikit-learn 1.9.1 and XGBoost 3.4.1 ports); models trained by earlier EDA versions are applied only as described below, the rest must be retrained
  • GROK-20932: Narrowed the ML methods' applicability: features must be INT or FLOAT, regression targets (Linear Regression, PLS Regression, XGBoost and SVM regression) must be INT or FLOAT; BIG_INT and other numerical types are rejected
  • GROK-20932: Removed the gradient-descent Rate hyperparameter of Linear Regression and Softmax; Softmax Penalty (must be > 0) now maps to logistic C, L1 / L2 follow scikit-learn's ridge / elastic net definitions
  • GROK-20932: Changed regression predictions to float columns for integer targets
  • GROK-20932: Changed Linear Regression, Softmax, PLS Regression, XGBoost and SVM training to skip rows with a missing target instead of failing
  • GROK-20932: Removed the sci-comp-ml, XGBoost and LIBSVM wasm modules
  • GROK-20932: Added nine Forge methods on Crux: Random Forest, Extra Trees, Hist Gradient Boosting, CatBoost, Decision Tree, k-Nearest Neighbors, Naive Bayes, LDA and Neural Network (MLP); the stochastic methods have a Model seed hyperparameter; kNN keeps the training rows in the model (meta.mlhasrows)
  • GROK-20932: Marked XGBoost, Random Forest, Extra Trees, Hist Gradient Boosting, CatBoost and Decision Tree with meta.mlnativenans: they take rows with missing feature values (Forge no longer skips or imputes them by default)
  • GROK-20932: Added two Forge methods on Crux's neural tabular models: RealMLP (pytabkit 1.7.3) and TabM (TabM 0.0.3 with rtdl_num_embeddings); both have a Model seed hyperparameter; TabM takes rows with missing feature values (meta.mlnativenans)
  • GROK-20932: Rebuilt the Crux module with crux-fn/realmlp and crux-fn/tabm (same crux revision; saved models keep working)
  • GROK-20932: Restored applying pre-Crux models with the same predictions as before (prediction only): PLS Regression with stored feature names, Linear Regression from 1.8.0, Softmax, XGBoost and SVM from 1.9.0 (src/legacy; XGBoost through Crux's ml.xgboost_predict); older models still require retraining; added files/legacy-models/reference.json and five more fixtures (regression, binary and bool targets)
  • public#4245: T-test, ANOVA, Control comparisons: Fixed the analysis and its checks running on the whole table instead of the rows that pass the filter
  • core#1407: Control comparisons, ANOVA, t-test: Fixed the method radio missing its property type
  • public#4473: MCL: Fixed Markov clustering on number columns only never finishing (the dialog passed the clustering no column); added the package test 'DBSCAN'
  • core#1146: Pareto Front: Fixed a 1 s check that ran for the rest of the session when the table was not applicable
  • Tests: models/share-model claims the second account's side: it finds the shared model under Shared with me and applies it (back in the serial lane, since it signs in as the second account)
  • public#4226: Menu: Reduce Dimensionality, PLS, Multivariate Analysis, T-test, ANOVA, Control Comparisons, Impute Missing Values, Pareto Front and MCL are disabled unless the current view is a table view, instead of failing or warning
  • GROK-20795: Tests: Added the package test of Control comparisons on a filtered table (the women of demog only); analyze/filtered-group-comparison keeps to the dialog
  • Tests: models/train-options moves the Positive class cutoff of Predict probability and reads the re-rendered accuracy (the preview is busy until the re-render ends)
  • Tests: Translated the TestTrack Models cases into BDD in bdd/features/models/ — train-data-checks (the Insights & Tips of Train Model and the remedies that let a model train), train-options (Ignore / Impute missing, Predict probability with its cutoff and ROC curve, One-hot encoding — each saved and applied to new rows) and models-gallery (search, Created by me, Edit..., Run Evaluation, Apply to on a card, Save as Zip, Ctrl+click and Compare), with MISSING listing what is still open
  • #417: ANOVA, t-test, Control comparisons: Fixed a column renamed while the dialog is open disappearing from its column selectors
  • #418: Control comparisons: The Control drop-down stays empty and disabled for an ID-like category instead of listing every value
  • Tests: models/apply-and-delete chooses a model in the Apply dialog by name (the dialog keys its models by id), so it and share-model left the serial lane; the analysis features keep to what the dialog produced, the statistics being the package tests'
  • Tests: Added the BDD feature models/apply-and-delete — PLS and linear regression models saved, applied to new data and deleted from the gallery (from TestTrack General/predictive-models)
  • Tests: Added BDD features for PCA, PLS, multivariate analysis, ANOVA, control comparisons, model training, sharing a model and the Pareto front viewer
  • Tests: Check completed training, rendered charts and fixture cleanup; isolate shared models per run
  • Pareto front: Corrected the viewer help link

1.9.0 (2026-08-26)​

  • Added SVM (classification and regression)
  • PLS scores plot: replaced fixed-radius circles with Hotelling's T-squared confidence ellipses (95% and 99%)
  • Fixed saved Softmax and XGBoost prediction on reordered feature columns

1.8.0 (2026-08-14)​

Improved ML features:

  • Added variables importance in projection the PLS regression and multivariate analysis
  • Updated the multivariate analysis demo application
  • Fixes:
    • Serialization of PLS and OLS models;
    • The loss of the analysis results in saved projects.

1.7.0 (2026-07-17)​

Upgrade XGBoost

  • XGBoost v3.3.0, minimal build
  • Fixed classification objectives
  • Training 1.3 - 2.2× faster
  • Prediction 1.4 - 3.4× faster
  • Input validation in TypeScript
  • Breaking: models saved by earlier EDA versions (v1 container) are no longer loadable — retrain and re-save them
  • Moved the EDA Playwright E2E suite into the package (playwright/); helpers from @datagrok-libraries/test/src/playwright

1.6.2 (2026-06-23)​

Fix the package build

1.6.1 (2026-06-22)​

Update Datagrok tools

1.6.0 (2026-06-17)​

Migrate ML kernels to Rust + WebAssembly; remove LS-SVM

  • PCA, PLS, softmax and linear regression now run on the Rust + WASM sci-comp-ml backend (XGBoost unchanged).
  • Linear regression adds Elastic Net (L1/L2) and gradient-descent hyperparameters; defaults remain ordinary least squares.
  • Removed LS-SVM and the legacy C++/Emscripten backend.

1.5.6 (2026-06-01)​

  • Add two-sample t-test (Welch and Student)
  • Add Control comparisons (Dunnett's test and Holm-corrected Welch's t-tests)

1.5.5 (2026-05-26)​

Implement Welch ANOVA

1.5.4 (2026-05-16)​

Updated the Multivariate Analysis demo

1.5.3 (2026-05-14)​

Fixed hints in the multivariate analysis demo

1.5.2 (2026-04-24)​

Fix the column_list issue in PCA

1.5.1 (2026-03-27)​

Add formula for prediction in multivariate analysis

1.5.0 (2026-03-13)​

Probabilistic Multi-Parameter Optimization (pMPO) - Enhancements

  • Added support for both numeric and categorical desirability columns
  • Introduced automatic parameter tuning for streamlined configuration
  • Added option to run pMPO without sigmoidal correction
  • New ROC curve and confusion matrix visualizations for model evaluation
  • Expanded correctness test coverage

1.4.12 (2026-01-16)​

Implemented

  • Interactive app for training probabilistic multi-parameter optimization (pMPO) models
  • Export pMPO models to MPO desirability profiles

1.4.11 (2025-11-21)​

Implemented

  • Pareto front viewer
  • Pareto optimization app

1.4.10 (2025-11-05)​

Correct MCL worker termination

1.4.9 (2025-11-04)​

Updated ML lib

1.4.8 (2025-10-15)​

Updated tutorials

1.4.7 (2025-10-01)​

Updated utils and js-api

1.4.6 (2025-09-23)​

Improved ANOVA:

  • Updated the F-test description
  • Implemented the tab control for the results

1.4.5 (2025-09-19)​

Dependencies update

1.4.4 (2025-09-17)​

Improved the PCA dialog

1.4.3 (2025-03-28)​

Updated grok tools

1.4.2 (2025-03-28)​

Updated Multivariate Analysis:

  • Implemented the Quadratic PLS regression
  • Improved MVA and PLS UI/UX

1.3.5 (2025-03-07)​

Vector distance functions support for dimensionality reduction on webGPU and without

1.3.4 (2025-02-28)​

MCL: Updated ML library to support large MCL matrices

1.3.3 (2025-02-20)​

  • Updated the MVA demo

1.3.2 (2025-02-03)​

  • Add History support in MCL Dialog

1.3.1 (2025-01-06)​

  • Rework MCL Viewer

1.2.6 (2024-11-05)​

Fixed:

  • Samples naming in Multivariate Analysis

1.2.5 (2024-11-05)​

Fixed:

  • Labels in Multivariate Analysis

1.2.4 (2024-10-31)​

Fixed:

  • Labels and help links in the PLS tools
  • Fixed k-NN imputer description
  • PCA computations

1.2.3 (2024-10-25)​

Fixed:

  • k-NN imputer input form
  • console output

1.2.2 (2024-09-12)​

Updated ANOVA:

  • Add tests & benchmarks
  • Fixed the incorrect feature column type bug

1.2.1 (2024-09-12)​

  • Add tests & benchmarks for missing values imputation using KNN

1.2.0 (2024-09-10)​

  • 1.21.1 release

1.1.35 (2024-08-28)​

  • Add History support and enable gpu by default if available for dim-red

1.1.34 (2024-08-09)​

  • Updated multivariate analysis
  • Reduced bias in the PLS regression and Multivariate Analysis
  • Fixed linear regression fails

1.1.33 (2024-08-14)​

Bump ml version to 6.6.21

1.1.32 (2024-08-13)​

Add XGBoost to predictive modeling tools

1.1.31 (2024-08-06)​

Add MCL layout with force directed graph and atlas

1.1.30 (2024-07-29)​

Add PLS regression to predictive modeling tools

1.1.29 (2024-07-29)​

Add the softmax classifier

1.1.28 (2024-07-23)​

Bump dependencies versions, datagrok-api to 1.20.0

1.1.27 (2024-06-17)​

Add linear regression

1.1.26 (2024-05-27)​

Add inflation factor to MCL

1.1.25 (2024-05-16)​

Fix inconsistent KNN size in webGPU.

1.1.24 (2024-05-09)​

  • Fix GPU description nullish value.

1.1.23 (2024-05-09)​

  • Disable webGPU input in case if none is available.
  • Improvements to GPU device handling

1.1.22 (2024-04-25)​

Add support for webGPU MCL in full (sparse matrix and expansion/normalization operators)

1.1.21 (2024-04-23)​

Add GPU information to clustering algorithms.

1.1.20 (2024-04-16)​

Improve dimensionality reduction dialog.

1.1.19 (2024-04-15)​

  • PLS components computation
  • Update UI for Multivariate Analysis - main feature & demo app

1.1.18 (2024-04-15)​

Fixed webGPU numeric distance with 0 range

1.1.17 (2024-04-14)​

Features​

  • Add webGPU UMAP implementation.
  • Add webGPU Sparse matrix calculation.
  • Add webGPU option for MCL.

1.1.16 (2024-04-05)​

Add webGPU KNN calculation option to UMAP.

1.1.15 (2024-02-27)​

Add support for post processing functions in dimensionality reduction. The method Dimensionality Reduction (see Top Menu > ML > Dimensionality Reduction) now allows to apply post processing functions to the resulting embeddings.

1.1.14 (2024-02-20)​

Fix NW distance function

1.1.13 (2024-02-12)​

Fix PCA & PLS applying to table from local file

1.1.12 (2024-02-12)​

Add MCL clustering. The method is available in Top Menu > ML > Cluster > MCL.

1.1.11 (2023-12-28)​

Add missing values imputation using the KNN method.

1.1.10 (2023-12-28)​

Add option to pass random seed to dimensionality reduction methods. This allows to reproduce results of dimensionality reduction.

1.1.9 (2023-12-22)​

Improvements to multi column dimensionality reduction.

1.1.8 (2023-12-18)​

Removed separate methods of dimensionality reduction and substituted with a single method Dimensionality Reduction (see Top Menu > ML > Dimensionality Reduction) that supports t-SNE and UMAP. The method allows to use multiple columns (like number, string, Molecule, Macromolecule, etc.) as features with different distance functions. The method also allows to cluster resulting embeddings using DBSCAN algorithm and color resulting scatterplot according to clusters.

1.1.7 (2023-12-14)​

Add DBSCAN clustering to Top Menu > ML > Cluster > DBSCAN.

1.1.6 (2023-10-31)​

Methods descriptions are updated.

1.1.5 (2023-10-23)​

Anova top-menu item is improved.

1.1.4 (2023-10-11)​

Add one-way ANOVA (see Top Menu > ML > Analysis of Variances (ANOVA)...)

1.1.3 (2023-08-24)​

  • The method PCA is replaced to Top Menu > ML > Dimensionality Reduction
  • The method UMPA is added (see Top Menu > ML > Dimensionality Reduction)
  • The method t-SNE is added (see Top Menu > ML > Dimensionality Reduction)
  • The method SPE is added (see Top Menu > ML > Dimensionality Reduction)

1.1.2 (2023-07-27)​

  • Move Multivariate Analysis using partial least squares (PLS) regression to Top Menu | ML
  • In PLS and PCA, just numerical columns can be selected as features