kscorer 2.2.0

KScorer: Auto-select optimal K-means clusters with advanced scoring

Basic Usage

LIVE demo-notebook is here

Load Modules

In [1]: import numpy as np
   ...: import pandas as pd
   ...: from sklearn import datasets
   ...: from sklearn.metrics import balanced_accuracy_score
   ...: from sklearn.model_selection import train_test_split
   ...: from kscorer.kscorer import KScorer

Init KScorer

In [2]: ks = KScorer()

Get Data

In [3]: X, y = datasets.load_digits(return_X_y=True)
   ...: X.shape
Out[3]: (1797, 64)

Train/Test Split

In [4]: X_train, X_test, y_train, y_test = train_test_split(
   ...:     X, y, test_size=0.2, random_state=1234)

Fit KScorer (i.e. Perform Unsupervised Clustering)

In [5]: labels, centroids, _ = ks.fit_predict(X_train, retall=True)
100%|██████████| 13/13 [00:09<00:00,  1.39it/s]

Optimal Clusters

In [6]: ks.show()

In [7]: ks.optimal_
Out[7]: 10

Confusion Matrix

In [8]: labels_mtx = (pd.Series(y_train)
   ...:               .groupby([labels, y_train])
   ...:               .count()
   ...:               .unstack()
   ...:               .fillna(0))
   ...: # match arbitrary labels to ground-truth labels
   ...: order = []
   ...: 
   ...: for i, r in labels_mtx.iterrows():
   ...:     left = [x for x in np.unique(y_train) if x not in order]
   ...:     order.append(r.iloc[left].idxmax())
   ...: 
   ...: confusion_mtx = labels_mtx[order]
   ...: confusion_mtx
Out[8]:

Cluster Unseen Data (you would prefer to build classifier instead)

In [9]: labels_unseen = ks.predict(X_test, init=centroids)

Evaluate Accuracy

In [10]: y_clustd = pd.Series(labels).replace(dict(enumerate(order)))
    ...: y_unseen = pd.Series(labels_unseen).replace(dict(enumerate(order)))

In [11]: balanced_accuracy_score(y_train, y_clustd)  # train data
Out[11]: 0.6940733254455871

In [12]: balanced_accuracy_score(y_test, y_unseen)  # unseen data
Out[12]: 0.646615365026082

ToDo:

• consider applying power-transform before initial scaling
• consider pyckmeans
• consider pyxmeans
• consider spherecluster
• consider benchmark testing