Adaptation to sklearn

[MuellerSeb]

sklearn provides a lot of functionality, we could use to simplify our code.

Distance matrix calculation:
https://scikit-learn.org/stable/modules/generated/sklearn.metrics.pairwise_distances.html

KDtree:
https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.KDTree.html

Search radius to optimize moving window kriging (See: #57):
https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.KDTree.html#sklearn.neighbors.KDTree.query_radius

Nearest neighbors:
https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.KDTree.html#sklearn.neighbors.KDTree.query

Beside that, one can specify the metric in use (see: #120):
https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.DistanceMetric.html

• “euclidean” | EuclideanDistance |   | sqrt(sum((x - y)^2))
• “manhattan” | ManhattanDistance |   | sum(|x - y|)
• “chebyshev” | ChebyshevDistance |   | max(|x - y|)
• “minkowski” | MinkowskiDistance | p | sum(|x - y|^p)^(1/p)
• “wminkowski” | WMinkowskiDistance | p, w | sum(|w * (x - y)|^p)^(1/p)
• “seuclidean” | SEuclideanDistance | V | sqrt(sum((x - y)^2 / V))
• “mahalanobis” | MahalanobisDistance | V or VI | sqrt((x - y)' V^-1 (x - y))

Or for geo-coordinates (see: #121):

• “haversine” | HaversineDistance | 2 arcsin(sqrt(sin^2(0.5dx) + cos(x1)cos(x2)sin^2(0.5dy)))
  (np.deg2rad needed here)

So this could solve a lot of issues.

[rth]

Yes, I think that would be a good idea. I'm all for better integration with scikit-learn in general #31

When improving the API for 2.0, making one that's compatible with scikit-learn if possible would be useful https://scikit-learn.org/stable/developers/develop.html#rolling-your-own-estimator

[MuellerSeb]

@rth Could you draft a general workflow for the integration with sklearn? I have to learn a bit more about sklearn to get to know the dos and don'ts..

[rth]

Essentially one could start with minimal functionality along the lines of,

from sklearn.base import BaseEstimator

class OrdinaryKrigging(BaseEstimator):
    def __init__(self, ...):
        ...
 
    def fit(self, X, y):
        ...
    
    def predict(self, X, return_std=False):
        ...

where X is (n_samples, n_features) array (with for now n_features = 2 or 3 being supported only).

Looking at the usage of GaussianProcessRegressor in scikit-learn could be a start. Then one would progressively add features, and form time to time run check_estimator to make sure the API is compatible.

We already have a scikit-learn wrapper class BTW

PyKrige/pykrige/rk.py

Line 27 in 40a8140

class Krige(RegressorMixin, BaseEstimator):

, but the idea could be to make this more of a default API.

[MuellerSeb]

So we should have a look at:
https://github.com/scikit-learn-contrib/project-template

[rth]

So we should have a look at: scikit-learn-contrib/project-template

Worth looking into but there are also things outdated there.

[MuellerSeb]

OK. In the end I am thinking of a single class for kriging where the dimension is just a parameter. And since Ordinary kriging is just universal kriging with a constant drift (or simply an unbiased estimator), we could combine all 4 kriging classes that are present at the moment.
When providing these in the proposed sklearn way, we can even get rid of the RK class. Which sounds nice to me.

You once mentioned that anisotropy and rotation could be provided by pipelines (#138).
How does that work?