Local outlier factor gives incorrect results #9874

jeroneandrews-sony · 2017-10-05T19:42:45Z

Steps/Code to Reproduce

import numpy as np
from sklearn.neighbors import LocalOutlierFactor
c=np.array([[0,0],[0,1],[1,1],[3,0]])
k=2 #numNeighbours
clf = LocalOutlierFactor(n_neighbors=k, n_jobs=-1,algorithm='brute',metric='manhattan').fit(c) 
Z = clf._decision_function(c)
print(-Z)
[ 0.875  0.875  0.875  1.5  ]

However, the result should be: [ 0.875 1.333 0.875 2. ] if you work it out by hand (assuming a sample can't be the 1st nearest neighbour to itself). So, I think there is something wrong with the LOF implementation. Could somebody confirm either way?

jeroneandrews-sony · 2017-10-06T08:54:56Z

I have also just compared the scikit result with the R package 'Rlof', which also gives me a result of:

c = matrix(c(0,0,0,1,1,1,3,0),nrow=4, ncol=2,byrow=TRUE)
> lof(c, 2, method="manhattan")
[1] 0.875000 1.333333 0.875000 2.000000

So there is definitely something wrong with the scikit implementation.

albertcthomas · 2017-10-06T14:16:19Z

The LOFs of the training samples are given by clf.negative_outlier_factor_ which gives the opposite of the output you obtain with Rlof.

import numpy as np
from sklearn.neighbors import LocalOutlierFactor
c=np.array([[0,0],[0,1],[1,1],[3,0]])
k=2 #numNeighbours
clf = LocalOutlierFactor(n_neighbors=k, n_jobs=-1,algorithm='brute',metric='manhattan').fit(c) 
Z = clf.negative_outlier_factor_
print(-Z)
[ 0.875  1.33333333  0.875   2. ]

The private _decision_function is meant to be used on test samples.

ngoix · 2017-10-06T14:23:40Z

@jeroneandrews have you tried using fit_predict (public) method? LOF predict (private) method is meant to extend the use of LOF to new data, and thus consider new data as "duplicates" if they already are in the training set.

ngoix · 2017-10-06T14:26:55Z

...predict, or more precisely decision_function private method in your case.
edit: As @albertcthomas pointed out, clf.negative_outlier_factor_ gives the correct answer.

jeroneandrews-sony · 2017-10-07T13:38:14Z

@albertcthomas @ngoix
Thank you for your responses, I missed negative_outlier_factor_ in the documentation 775B .

jeroneandrews-sony closed this as completed Oct 7, 2017

albertcthomas mentioned this issue Oct 9, 2017

[MRG+1] Fix LOF and Isolation benchmarks #9798

Merged

kno10 mentioned this issue Dec 4, 2021

Non-standard "negative_outlier_factor_" #21875

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Local outlier factor gives incorrect results #9874

Local outlier factor gives incorrect results #9874

Local outlier factor gives incorrect results #9874

Local outlier factor gives incorrect results #9874

Comments

Steps/Code to Reproduce