v0.7 Release

.github/workflows/ci.yml

@@ -58,6 +58,8 @@ jobs:
           - "--features datasets"
           - "--features ndarray-bindings"
           - "--features std,ndarray-bindings"
+          - "--features lazy-normalization"
+          - "--features lazy-normalization,ndarray-bindings"
           - ""
     steps:
       - uses: actions/checkout@v4
@@ -96,7 +98,7 @@ jobs:
               ;;
           esac
       - name: Tests
-        if: matrix.features == '--features ndarray-bindings' || matrix.features == '--features std,ndarray-bindings' || matrix.features == '--features datasets'
+        if: matrix.features == '--features ndarray-bindings' || matrix.features == '--features std,ndarray-bindings' || contains(matrix.features, 'lazy-normalization') || matrix.features == '--features datasets'
         run: cargo test --no-default-features ${{ matrix.features }}
       - name: Explicit std does not activate ndarray
         if: matrix.features == ''

CHANGELOG.md

@@ -5,12 +5,21 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
+
+## [v0.7]
 ### Added
 - A `Dependency audit` CI workflow runs `cargo deny` and `cargo audit` on each push and pull request, and every week. It fails when a dependency has a known vulnerability, is yanked, uses a license that is not allowed, or comes from an unknown source. The `deny.toml` file holds the rules.
+- Optional `lazy-normalization` support for Lasso and Elastic Net. It applies normalization to a borrowed design and keeps Elastic Net's identity-block augmentation implicit, while retaining the dense Gram workspace. Enable the feature and set `with_lazy_normalization(true)` to select this path. The default remains eager. The new parameter field is available only with the feature and defaults to false during deserialization (#479).
+- `tree`, `ensemble`: `DecisionTreeRegressor`, `RandomForestRegressor` and `ExtraTreesRegressor` get a `compute_feature_importances` method.
 
 ### Changed
 - `Cargo.toml`: optional `ndarray` is now `default-features = false`, and the default feature set is `std` (`ndarray?/std`). Ordinary builds still enable ndarray's standard-library support when `ndarray-bindings` is on, while `std` alone does not activate ndarray. The `std` feature only controls ndarray's standard-library support; smartcore itself is not `no_std` (#421).
   - **Migration:** builds that use `--no-default-features --features ndarray-bindings` (or `default-features = false` with `ndarray-bindings`) no longer get ndarray's `std` feature. Add `std` to the feature list to keep the previous behavior.
+- **Breaking change:** Enabling `lazy-normalization` adds a public `lazy_normalization` field to `LassoParameters`, `LassoSearchParameters`, `ElasticNetParameters`, and `ElasticNetSearchParameters`. Struct literals without `..Default::default()` must add `lazy_normalization: false` to retain the eager path, or `true` to select the lazy path. This also applies when another dependency enables the feature through Cargo feature unification (#479).
+- **Breaking change:** the fitted `DecisionTreeRegressor`, `RandomForestRegressor` and `ExtraTreesRegressor` store new fields that the feature importances use. Models serialized with an earlier version cannot be deserialized. Fit these models again.
+- `tree`: `BaseTreeRegressor` (used by `DecisionTreeRegressor`, `RandomForestRegressor` and `ExtraTreesRegressor`) uses a new internal `Criterion` trait with `MSECriterion`. The MSE split improvement uses a numerically stable formula.  The selected splits do not change when a constant is added to y. Exactly tied splits can be broken in a different way, so fitted trees can differ slightly from earlier versions.
+- `tree`: in the `...Regressor`s, rows with sample weight 0 are removed before the tree grows. They no longer count toward `min_samples_leaf` and `min_samples_split`, and they no longer move a split threshold. A tree fitted with a zero-weight row is the same as a tree fitted without that row.
+- `Cargo.toml`: `ndarray` requires version `0.17.2` or later.
 
 ## [0.7.0]
 ### Added

Cargo.toml

@@ -25,7 +25,8 @@ exclude = [
 [dependencies]
 approx = "0.5.1"
 cfg-if = "1.0.0"
-ndarray = { version = "0.17", optional = true, default-features = false }
+lazymatrix = { version = "0.6.0", optional = true, default-features = false }
+ndarray = { version = "0.17.2", optional = true, default-features = false }
 num-traits = "0.2.12"
 num = "0.4"
 rand = { version = "0.10.1", default-features = false, features = ["alloc"] }
@@ -55,6 +56,9 @@ serde = ["dep:serde", "dep:typetag"]
 # Optional bindings for the ndarray crate (DenseMatrix <-> ndarray).
 ndarray-bindings = ["dep:ndarray"]
 
+# Apply Lasso and Elastic Net normalization without copying the design matrix.
+lazy-normalization = ["dep:lazymatrix"]
+
 # Sample datasets (iris, boston, digits, ...) and dataset generators.
 # Enables serde because datasets ship serialized bundles. Does not enable
 # std_rand: the generators use the crate RNG (SmallRng without std_rand).

README.md

@@ -116,6 +116,29 @@ smartcore organizes algorithms into clear modules with consistent traits:
 
 Recent refactors emphasize reusable components in trees/forests and expanded multiclass SVM capabilities. XGBoost-style regression and single-linkage clustering have been added. See CHANGELOG for API changes and migration notes.
 
+## Optional lazy normalization
+
+The optional `lazy-normalization` feature lets Lasso and Elastic Net borrow the
+design matrix during normalization. Elastic Net also applies its identity-block
+augmentation without storing the augmented design. Enable it with `features =
+["lazy-normalization"]`, then select it with
+`LassoParameters::default().with_lazy_normalization(true)` or
+`ElasticNetParameters::default().with_lazy_normalization(true)`. The default
+remains eager, even when the feature is enabled. The `normalize` option still
+controls whether columns are centered and scaled. Parameter searches use the
+fixed `lazy_normalization` setting. It works with `DenseMatrix` and
+`ndarray-bindings`.
+
+Lazy normalization reduces memory use by computing normalized values when the
+solver needs them. This adds repeated arithmetic and can increase fit time for
+both estimators, including Elastic Net on large inputs. The solver still stores
+a dense Gram matrix with one row and column per feature.
+
+Enabling the feature adds a public `lazy_normalization` field to both estimators'
+parameter and search parameter structs. Code that constructs these structs
+without `..Default::default()` must add the field. Missing fields default to
+`false` during deserialization.
+
 ## Data access and readers
 
 - CSV readers: Read matrices from CSV with configurable delimiter and header rows, with helpful error messages and testing utilities (including non-IO reader abstractions).