CODEWITHSUNDEEP
pythonknn
user2926523
user2926523

Reputation: 453

Sklearn kNN usage with a user defined metric

Currently I'm doing a project which may require using a kNN algorithm to find the top k nearest neighbors for a given point, say P. im using python, sklearn package to do the job, but our predefined metric is not one of those default metrics. so I have to use the user defined metric, from the documents of sklearn, which can be find here and here.

It seems that the latest version of sklearn kNN support the user defined metric, but i cant find how to use it:

import sklearn
from sklearn.neighbors import NearestNeighbors
import numpy as np
from sklearn.neighbors import DistanceMetric
from sklearn.neighbors.ball_tree import BallTree
BallTree.valid_metrics

say i have defined a metric called mydist=max(x-y), then use DistanceMetric.get_metric to make it a DistanceMetric object:

dt=DistanceMetric.get_metric('pyfunc',func=mydist)

from the document, the line should looks like this

nbrs = NearestNeighbors(n_neighbors=4, algorithm='auto',metric='pyfunc').fit(A)
distances, indices = nbrs.kneighbors(A)

but where can i put the dt in? Thanks

Upvotes: 32

Views: 39493

Answers (3)

Zvi
Zvi

Reputation: 2424

Using KNeighborsRegressor() worked only by setting algorithm='brute' when trying to use a user defined metric.

Otherwise fit() works but predict() fails with error 'returned NULL without setting an error' when using JupyterLab, or 'SystemError: error return without exception set' when using Google Colab

Upvotes: 0

Mahmoud
Mahmoud

Reputation: 539

A small addition to the previous answer. How to use a user defined metric that takes additional arguments. 

>>> def mydist(x, y, **kwargs):
...     return np.sum((x-y)**kwargs["metric_params"]["power"])
...
>>> X = np.array([[-1, -1], [-2, -1], [-3, -2], [1, 1], [2, 1], [3, 2]])
>>> Y = np.array([-1, -1, -2, 1, 1, 2])
>>> nbrs = KNeighborsClassifier(n_neighbors=4, algorithm='ball_tree',
...            metric=mydist, metric_params={"power": 2})
>>> nbrs.fit(X, Y)
KNeighborsClassifier(algorithm='ball_tree', leaf_size=30,                                                                                                                                                          
       metric=<function mydist at 0x7fd259c9cf50>, n_neighbors=4, p=2,
       weights='uniform')
>>> nbrs.kneighbors(X)
(array([[  0.,   1.,   5.,   8.],
       [  0.,   1.,   2.,  13.],
       [  0.,   2.,   5.,  25.],
       [  0.,   1.,   5.,   8.],
       [  0.,   1.,   2.,  13.],
       [  0.,   2.,   5.,  25.]]),
 array([[0, 1, 2, 3],
       [1, 0, 2, 3],
       [2, 1, 0, 3],
       [3, 4, 5, 0],
       [4, 3, 5, 0],
       [5, 4, 3, 0]]))

Upvotes: 22

alko
alko

Reputation: 48357

You pass a metric as metric param, and additional metric arguments as keyword paramethers to NN constructor:

>>> def mydist(x, y):
...     return np.sum((x-y)**2)
...
>>> X = np.array([[-1, -1], [-2, -1], [-3, -2], [1, 1], [2, 1], [3, 2]])

>>> nbrs = NearestNeighbors(n_neighbors=4, algorithm='ball_tree',
...            metric='pyfunc', func=mydist)
>>> nbrs.fit(X)
NearestNeighbors(algorithm='ball_tree', leaf_size=30, metric='pyfunc',
         n_neighbors=4, radius=1.0)
>>> nbrs.kneighbors(X)
(array([[  0.,   1.,   5.,   8.],
       [  0.,   1.,   2.,  13.],
       [  0.,   2.,   5.,  25.],
       [  0.,   1.,   5.,   8.],
       [  0.,   1.,   2.,  13.],
       [  0.,   2.,   5.,  25.]]), array([[0, 1, 2, 3],
       [1, 0, 2, 3],
       [2, 1, 0, 3],
       [3, 4, 5, 0],
       [4, 3, 5, 0],
       [5, 4, 3, 0]]))

Upvotes: 37

Related Questions