## knn()


Find the approximate k nearest neighbors of every row with NNDescent.


Usage

``` python
knn(
    *features,
    k=15,
    metric="euclidean",
    random_state=None,
    null_policy="error",
    n_threads=None,
    gpu=False,
    verbose=False
)
```


This is the neighbor search UMAP uses internally, a Rust port of PyNNDescent.


## Parameters


`*features: IntoExpr | list[IntoExpr]`  
See [umap()](umap.md#polars_umap.umap).

`k: int = ``15`  
Number of neighbors per row. As in PyNNDescent, the row itself counts as one of them: it is normally the first neighbor, at distance 0.

`metric: Metric = ``"euclidean"`  
Distance metric.

`random_state: int | None = None`  
Seed for the random number generator.

`null_policy: NullPolicy = ``"error"`  
See [umap()](umap.md#polars_umap.umap). Ignored rows get a null result and never appear as neighbors.

`n_threads: int | None = None`  
See [umap()](umap.md#polars_umap.umap).

`gpu: int | None = None`  
See [umap()](umap.md#polars_umap.umap).

`verbose: int | None = None`  
See [umap()](umap.md#polars_umap.umap).


## Returns


`pl.Expr`  
A `Struct` expression with fields `indices` (`Array(UInt32, k)`, row numbers in the input) and `distances` (`Array(Float32, k)`), sorted by ascending distance. When there are fewer than `k` rows, the missing neighbors are null.
