JadeML is a Windows desktop app for learning machine learning through data tables, workflow diagrams, and interactive model visualizations. Open independent prediction and clustering tasks to prepare data, fit models, and explore how their results are produced.
This source tree is prepared for version 3.0.1; see the release notes. Published installers are available on the releases page.
- Use Windows x64 and install the .NET 9 Desktop Runtime if it is not included with your chosen distribution.
- Download and run an installer from the releases page, or build from source below.
- Install 64-bit Python 3.11, NumPy 1.26.4, and scikit-learn 1.5.1. Python is required by every model except MLP, and by PCA preprocessing.
To select a Python installation or virtual environment explicitly, run the following in PowerShell, then launch JadeML from the same session:
$env:JADEML_PYTHON = 'C:\path\to\python.exe'
& $env:JADEML_PYTHON -m pip install 'numpy==1.26.4' 'scikit-learn==1.5.1'Install packages into that interpreter without --user. JadeML also checks the standard per-user Python 3.11 installation and python.exe on PATH when JADEML_PYTHON is unset. Restart JadeML after changing Python or its packages. See Python setup for discovery rules and compatibility details.
Try the CSV files in SampleDatasets. Numeric CSV values use decimal points.
| Task | Open from the File menu | Workflow |
|---|---|---|
| Supervised prediction | New prediction task — Ctrl+N | Load data → choose inputs and a target → prepare data → split training/test rows → train → evaluate and predict |
| Unsupervised clustering | New clustering task — Ctrl+R | Load data → select numeric features → prepare data → fit clusters → explore and export assignments |
Each task keeps its own data and model. Standardization, min-max scaling, constant-feature removal, and optional PCA with whitening help prepare inputs. The prediction workflow refits preprocessing on training rows after splitting; clustering fits preprocessing on all rows being grouped. Changing data or settings clears dependent results.
For clustering, exclude identifiers and known class labels, including numeric labels. No target column or train/test split is needed. Selected features must be finite numeric values in every row; clustering import preserves incomplete rows so assignments stay aligned with original records. Correct missing values or exclude the affected feature.
K-Means and Gaussian mixtures can assign new samples using their fitted preprocessing and model. Hierarchical clustering and DBSCAN need a new fit to include new samples. CSVs with headers match selected inputs by exact name; files without headers use their original selected-column order.
All regression models accept one or multiple numeric inputs and one numeric target.
| Order | Models |
|---|---|
| 1 | Dummy Regressor |
| 2 | Linear Regression |
| 3 | Ridge Regression; Lasso Regression; Elastic Net Regression |
| 4 | Bayesian Ridge Regression; Huber Regression; RANSAC Regression; Quantile Regression |
| 5 | Poisson Regression; Gamma Regression; Tweedie Regression |
| 6 | Polynomial Regression; Partial Least Squares Regression |
| 7 | KNN Regression |
| 8 | Decision-Tree Regression; Random-Forest Regression; Extra-Trees Regression |
| 9 | Support Vector Machine; Gaussian Process Regression |
| 10 | AdaBoost Regression; Gradient Boosting; Histogram Gradient Boosting |
| 11 | Voting Regressor; Stacking Regressor |
| 12 | Multilayer Perceptron (MLP) |
Models support binary and multiclass tasks unless stated otherwise.
| Order | Models |
|---|---|
| 1 | Dummy Classifier |
| 2 | Nearest-Centroid Classifier; K-Nearest Neighbors |
| 3 | Naïve Bayes (Gaussian); Bernoulli Naive Bayes; Multinomial Naive Bayes; Complement Naive Bayes |
| 4 | Logistic Regression (binary); Multinomial Logistic Regression (multiclass task) |
| 5 | Linear Discriminant Analysis; Quadratic Discriminant Analysis |
| 6 | Decision Tree; Random Forest; Extra Trees |
| 7 | Support Vector Machine |
| 8 | AdaBoost (binary only); Gradient Boosting; Histogram Gradient Boosting |
| 9 | Voting Classifier; Stacking Classifier |
| 10 | Multilayer Perceptron (MLP) |
| Order | Model |
|---|---|
| 1 | K-Means |
| 2 | Hierarchical (agglomerative) clustering |
| 3 | DBSCAN |
| 4 | Gaussian mixture |
The ScikitLearn 1.1.0 C# wrapper connects these models to scikit-learn. MLP uses TorchSharp. See the model guide for input requirements, settings, and limits.
Supervised explorers use familiar tables, TreeViews, diagrams, and graphs. Follow a decision-tree path, compare forest predictions, inspect linear coefficients or class evidence, trace boosting stages, or explore an MLP's layers, neurons, and weights. Editable examples show how inputs contribute to a fitted model's prediction.
The clustering Results tab provides:
- Scatterplots: a PCA overview with explained variance and selectable feature-pair plots.
- Cluster charts: group-size bars, feature-distribution box plots, and mean profiles.
- Explore model: center distances, hierarchical dendrograms, DBSCAN neighborhoods, or Gaussian mixture probabilities and weights.
These views preserve the fitted model and assignments. Cluster numbers are arbitrary group IDs; projected plots can hide separation in other dimensions. Silhouette measures separation, not classification accuracy, and excludes DBSCAN noise. Graphs identify sampling and display-only transformations. More interpretation notes are in the visualization guide.
Use Windows x64 and a .NET SDK supporting net9.0-windows8.0; SDK 10.0.401 is the verified setup. Configure Python as described above, then run:
dotnet restore JadeML.sln
dotnet build JadeML.sln -c Release --no-restore
dotnet test JadeML.sln -c Release --no-build --no-restoreThe xUnit suite in Tests/JadeML.Tests covers model parity with scikit-learn, preprocessing, evaluation, CSV handling, workflow state, and visual explorers. WinForms tests require a Windows desktop session and run sequentially because rendering and native runtimes are shared. TorchSharp tests use the CPU; a CUDA GPU is not required. The CUDA package has large native dependencies, so the first restore can take time.
Set JADEML_VISUAL_QA_DIRECTORY to an absolute directory before testing to save visual previews. See validation and practical limits for further notes and the release preparation guide for packaging version 3.0.1.
