Generalize model verification to unsupervised learning methods, custom output fields

[KapoorHitesh]

For kmeans algorithm, not getting ModelVerification section/tag in generated pmml file.

iris_pipeline = PMMLPipeline([
	("pca", PCA(n_components = 3)),
	("clusterer", KMeans(n_clusters = 3))
])

x_columns = ["Sepal.Length", "Sepal.Width", "Petal.Length","Petal.Width"]
iris_pipeline.fit(iris_df[x_columns])
from sklearn2pmml import sklearn2pmml
iris_pipeline.verify(iris_df[x_columns])
sklearn2pmml(iris_pipeline, "kmeanstry.pmml", with_repr = True)

[vruusmann]

Model verification is most useful with supervised learning methods such as regression and classification, where there is a "clear" target field to be checked.

Clustering is an unsupervised learning method. As the target field is missing, then something else needs to be checked, such as the distance of the verification data record to a (sub)set of clusters (in PMML speak, "cluster affinities").

[KapoorHitesh]

Thank you for the quick response. Do you have any example PMML file which uses this?
Are you planning to implement this feature in sklearn2pmml?

[vruusmann]

Do you have any example PMML file which uses this?

Integration tests for the KMeans estimator type use cluster affinities for "external" validation:
https://github.com/jpmml/jpmml-sklearn/blob/master/src/test/resources/main.py#L112-L118

So, instead of saving cluster affinities to a CSV file, they should be bundled into the PKL file so that the JPMML-SkLearn library could see them.

Are you planning to implement this feature in sklearn2pmml?

Not a priority for me.

But lets keep this issue open, as a means to track ideas/progress on this topic (and related topics).

[pradeeppadmarajaiah]

Hi,
Can you please post an example which explains on how to build PMML pipeline for K-means clustering using sklearn2pmml?

Thanks,