release v0.3.3

azmyrajab · Apr 26, 2024 · 49ae93c · 49ae93c
1 parent 1746bf2
commit 49ae93c
Show file tree

Hide file tree

Showing 6 changed files with 41 additions and 49 deletions.
diff --git a/Cargo.lock b/Cargo.lock
diff --git a/Cargo.toml b/Cargo.toml
@@ -2,7 +2,7 @@ cargo-features = ["profile-rustflags"]
 
 [package]
 name = "polars_ols"
-version = "0.3.2"
+version = "0.3.3"
 edition = "2021"
 
 [lib]

diff --git a/README.md b/README.md
@@ -54,7 +54,7 @@ df = pl.DataFrame({"y": [1.16, -2.16, -1.57, 0.21, 0.22, 1.6, -2.11, -2.92, -0.8
                    "weights": [0.34, 0.97, 0.39, 0.8, 0.57, 0.41, 0.19, 0.87, 0.06, 0.34],
                    })
 
-lasso_expr = pl.col("y").least_squares.lasso(pl.col("x1"), pl.col("x2"), alpha=0.0001, add_intercept=True).over("group")
+lasso_expr = pl.col("y").least_squares.lasso("x1", "x2", alpha=0.0001, add_intercept=True).over("group")
 wls_expr = pls.compute_least_squares_from_formula("y ~ x1 + x2 -1", sample_weights=pl.col("weights"))
 
 predictions = df.with_columns(lasso_expr.round(2).alias("predictions_lasso"),
@@ -154,6 +154,7 @@ Currently, this extension package supports the following variants:
 - Weighted Least Squares: ```least_squares.wls```
 - Regularized Least Squares (Lasso / Ridge / Elastic Net) ```least_squares.{lasso, ridge, elastic_net}```
 - Non-negative Least Squares: ```least_squares.nnls```
+- Multi-target Least Squares:  ```least_squares.multi_target_ols```
 
 As well as efficient implementations of moving window models:
 - Recursive Least Squares: ```least_squares.rls```

diff --git a/polars_ols/least_squares.py b/polars_ols/least_squares.py
@@ -62,6 +62,7 @@
 _VALID_NULL_POLICIES: Set[NullPolicy] = set(get_args(NullPolicy))
 _VALID_OUTPUT_MODES: Set[OutputMode] = set(get_args(OutputMode))
 _VALID_SOLVE_METHODS: Set[SolveMethod] = set(get_args(SolveMethod)).union({None})
+_EPSILON: float = 1.0e-12
 
 
 @dataclass
@@ -190,7 +191,8 @@ def _pre_process_data(
     # handle sample weights
     sqrt_w: Optional[pl.Expr] = None
     if sample_weights is not None:
-        sqrt_w = parse_into_expr(sample_weights).sqrt()
+        # missing sample weights are assumed to be minimal 'epsilon' weight
+        sqrt_w = parse_into_expr(sample_weights).sqrt().fill_null(_EPSILON)
         target *= sqrt_w
         features = [(expr * sqrt_w) for expr in features]
     return target, features, sqrt_w
@@ -211,22 +213,17 @@ def _register_least_squares_plugin(
 
     # register either coefficient or prediction plugin functions
     if mode == "coefficients":
-        # TODO: remove 'rename_fields' after https://github.com/pola-rs/pyo3-polars/issues/79
-        #  it currently breaks input_wildcard_expansion=True correctly returning a struct.
-        return (
-            register_plugin_function(
-                plugin_path=Path(__file__).parent,
-                function_name=f"{function_name}_coefficients",
-                args=[target_fit, *features_fit],
-                kwargs=ols_kwargs.to_dict(),
-                is_elementwise=False,
-                changes_length=returns_scalar_coefficients,
-                returns_scalar=returns_scalar_coefficients,
-                input_wildcard_expansion=True,
-            )
-            .alias("coefficients")
-            .struct.rename_fields([f.meta.output_name() for f in features_fit])
-        )
+        return register_plugin_function(
+            plugin_path=Path(__file__).parent,
+            function_name=f"{function_name}_coefficients",
+            args=[target_fit, *features_fit],
+            kwargs=ols_kwargs.to_dict(),
+            is_elementwise=False,
+            changes_length=returns_scalar_coefficients,
+            returns_scalar=returns_scalar_coefficients,
+            input_wildcard_expansion=True,
+            pass_name_to_apply=True,
+        ).alias("coefficients")
     else:
         predictions = register_plugin_function(
             plugin_path=Path(__file__).parent,
@@ -237,7 +234,7 @@ def _register_least_squares_plugin(
             input_wildcard_expansion=True,
         )
         if sqrt_w is not None:
-            predictions /= sqrt_w  # undo the scaling implicit in WLS weighting
+            predictions *= 1.0 / sqrt_w  # undo the scaling implicit in WLS weighting
         if mode == "predictions":
             return predictions
         else:

diff --git a/src/least_squares.rs b/src/least_squares.rs
@@ -252,6 +252,11 @@ pub fn solve_multi_target(
     alpha: Option<f64>,
     rcond: Option<f64>,
 ) -> Array2<f64> {
+    // handle degenerate case of no data
+    if x.is_empty() {
+        return Array2::zeros((x.ncols(), y.ncols()));  // n_features x n_targets
+    }
+    // Choose SVD implementation based on L2 regularization
     let alpha = alpha.unwrap_or(0.0);
     if alpha > 0.0 {
         solve_ridge_svd(y, x, alpha, rcond)

diff --git a/tests/test_ols.py b/tests/test_ols.py
@@ -90,6 +90,7 @@ def test_multi_target_regression(alpha, mode, null_policy):
         n_features=3,
         add_missing=null_policy not in {"zero", "ignore"},
         missing_columns=("x1",),
+        n_groups=3,
     )
     df = df.with_columns(
         pl.struct(
@@ -104,24 +105,24 @@ def test_multi_target_regression(alpha, mode, null_policy):
         multi_target = df.select(
             compute_multi_target_least_squares(
                 "y",
-                "x1",
-                "x2",
-                "x3",
+                pl.col("^x.*$"),
                 mode=mode,
                 ols_kwargs=ols_kwargs,
-            ).alias(mode)
+            )
+            .over("group")
+            .alias(mode)
         )
 
     with timer("compute multiple linear regressions"):
         expected = df.unnest("y").select(
             compute_least_squares(
                 target,
-                "x1",
-                "x2",
-                "x3",
+                pl.col("^x.*$"),
                 mode=mode,
                 ols_kwargs=ols_kwargs,
-            ).alias(target)
+            )
+            .over("group")
+            .alias(target)
             for target in ("y1", "y2", "y3")
         )
 
@@ -131,25 +132,15 @@ def test_multi_target_regression(alpha, mode, null_policy):
 def test_fit_missing_data_coefficients():
     df = _make_data(add_missing=True)
 
-    # # in presence of unhandled nulls assert the rust library raises ComputeError
-    # with pytest.raises(pl.exceptions.ComputeError):
-    #     df.select(
-    #         pl.col("y").least_squares.ols(
-    #             pl.col("x1"), pl.col("x2"), null_policy="ignore", mode="coefficients"
-    #         )
-    #     )
-
     # test rust zero policy is sane
     assert np.allclose(
         df.select(
-            pl.col("y").least_squares.ols(
-                pl.col("x1"), pl.col("x2"), null_policy="zero", mode="coefficients"
-            )
+            pl.col("y").least_squares.ols(pl.col("^x.*$"), null_policy="zero", mode="coefficients")
         ).unnest("coefficients"),
         df.fill_null(0.0)
         .select(
             pl.col("y").least_squares.ols(
-                pl.col("x1"), pl.col("x2"), null_policy="ignore", mode="coefficients"
+                pl.col("^x.*$"), null_policy="ignore", mode="coefficients"
             )
         )
         .unnest("coefficients"),
@@ -158,14 +149,12 @@ def test_fit_missing_data_coefficients():
     # test rust drop (any) policy is sane
     assert np.allclose(
         df.select(
-            pl.col("y").least_squares.ols(
-                pl.col("x1"), pl.col("x2"), null_policy="drop", mode="coefficients"
-            )
+            pl.col("y").least_squares.ols(pl.col("^x.*$"), null_policy="drop", mode="coefficients")
         ).unnest("coefficients"),
         df.drop_nulls()
         .select(
             pl.col("y").least_squares.ols(
-                pl.col("x1"), pl.col("x2"), null_policy="ignore", mode="coefficients"
+                pl.col("^x.*$"), null_policy="ignore", mode="coefficients"
             )
         )
         .unnest("coefficients"),
@@ -175,14 +164,14 @@ def test_fit_missing_data_coefficients():
     assert np.allclose(
         df.select(
             pl.col("y").least_squares.ols(
-                pl.col("x1"), pl.col("x2"), null_policy="drop_y_zero_x", mode="coefficients"
+                pl.col("^x.*$"), null_policy="drop_y_zero_x", mode="coefficients"
             )
         ).unnest("coefficients"),
         df.drop_nulls(subset=["y"])
         .fill_null(0.0)
         .select(
             pl.col("y").least_squares.ols(
-                pl.col("x1"), pl.col("x2"), null_policy="ignore", mode="coefficients"
+                pl.col("^x.*$"), null_policy="ignore", mode="coefficients"
             )
         )
         .unnest("coefficients"),
@@ -393,7 +382,7 @@ def test_coefficients_ols_groups():
         df.select(
             "group",
             pl.col("y")
-            .least_squares.from_formula("x1 + x2 -1", mode="coefficients")
+            .least_squares.ols(pl.col("^x.*$"), mode="coefficients")
             .over("group")
             .alias("coefficients"),
         )
@@ -406,7 +395,7 @@ def test_coefficients_ols_groups():
         df.filter(pl.col("group") == 1)
         .select(
             pl.col("y")
-            .least_squares.from_formula("x1 + x2 -1", mode="coefficients")
+            .least_squares.ols(pl.col("^x.*$"), mode="coefficients")
             .alias("coefficients")
         )
         .unnest("coefficients")