✨ 1. enhance MLP 🛠 2. bug fixes.
- AR and ARMA for time series modeling ⚡️ 2. Optimize interpolation package ⚡️ 3. Optimize matrix decomposition memory usage. 🛠 4. Bug fixes.
- Generalized ridge regression
- ✨ Enhance JSON library
- ⚡️ Do NOT transparently include OpenBLAS library to save space. If the users need optimized matrix computation, they should add the dependency based on their platform. See README for details.
- New matrix design
- New formula design
- Generalized linear models (GLM) 📜 4. Sparse logistic regression 🛠 5. Bug fixes
- All new declarative data visualization
- TreeSHAP (contributed by Ray Ma @rayeaster)
- UMAP (contributed by Ray Ma @rayeaster)
- Levenberg-Marquardt algorithm
- 📦 The packages smile-cas and smile-vega are merged into scala-scala package
- Spark integration in smile-spark
- NLP in Kotlin
- Grid search and random search for hyperparameter tuning
- 🐛 Bug fixes
- Smile Shell is based on Scala REPL (2.13.2) again
- DataFrame and Tuple -> JSON
- Kotlin and Clojure notebooks
Kudos to Ray Ma @rayeaster for great contributions!
Various minor improvements
💻 The CAS module is a computer algebra system that has the ability to manipulate mathematical
expressions in a way similar to the traditional manual computations of
mathematicians and scientists.
👍 The symbolic manipulations supported include:
simplification to a smaller expression or some standard form,
including automatic simplification with assumptions and
simplification with constraints
substitution of symbols or numeric values for certain expressions
🔄 change of form of expressions: expanding products and powers, partial
and full factorization, rewriting as partial fractions, constraint
satisfaction, rewriting trigonometric functions as exponentials,
transforming logic expressions, etc.
partial and total differentiation
matrix operations including products, inverses, etc.
- Vega-lite based plot
- Jupyter notebook examples 🛠 3. Bug fixes
Smile has been fully rewritten with more than 150,000 lines change.
- Fully redesigned API. It is leaner, simpler and even more friendly.
- Faster implementation and memory optimization. Many algorithms are fully reimplemented. RandomForest is 8X faster than XGBoost on large benchmark data (10MM samples).
- 🆕 New parallelism mechanism
- All new DataFrame and Formula
- 🆕 New algorithms such as ICA, error reduction prune, quantile loss, TWCNB, etc.
- 👌 Support arbitrary class labels.
- ✨ Enhancement and harden numeric computations.
- 👌 Support Parquet, SAS, Arrow, Avro, etc.
- 🐛 Bug fixes.