Dataset loader utilities

Toy Datasets

from sklearn.datasets import *

# Boston house prices - 506 points, 13 attributes (none missing)
X, y = load_boston(return_X_y=True)
print(X.shape)

(506, 13)

# Iris plants - 150 points (50/class, 3 classes), 4 attributes (none missing)
data = load_iris()
print(data.target[[10, 25, 50]])
print(list(data.target_names))

[0 0 1]
['setosa', 'versicolor', 'virginica']

# diabetes (regression)
# 442 samples, 10 attributes, integer targets
X,y = load_diabetes(return_X_y=True)
print(X.shape)

(442, 10)

# digits (classification)
X,y = load_digits(return_X_y=True)
print(X.shape,y.shape)

(1797, 64) (1797,)

# linnerud (regression)
X,y = load_linnerud(return_X_y=True)
print(X.shape,y.shape)

(20, 3) (20, 3)

# wine (classification)
X,y = load_wine(return_X_y=True)
print(X.shape,y.shape)

(178, 13) (178,)

# breast cancer (classification)
X,y = load_breast_cancer(return_X_y=True)
print(X.shape,y.shape)

(569, 30) (569,)

Real World (larger datasets)

Olivetti Faces

• 400 samples of face images, taken at AT&T 1992-94.
• 4096 (64x64) dimensionality
• feature values = quantized grey levels. loader
• converts unsigned integers to floating point [0.0..1.0].
• targets = integers [0..39] = person's identity
• only 10 examples/class

# olivetti faces (classification)
faces = fetch_olivetti_faces()
print(faces.data)
print(faces.target)

[[0.30991736 0.3677686  0.41735536 ... 0.15289256 0.16115703 0.1570248 ]
 [0.45454547 0.47107437 0.5123967  ... 0.15289256 0.15289256 0.15289256]
 [0.3181818  0.40082645 0.49173555 ... 0.14049587 0.14876033 0.15289256]
 ...
 [0.5        0.53305787 0.607438   ... 0.17768595 0.14876033 0.19008264]
 [0.21487603 0.21900827 0.21900827 ... 0.57438016 0.59090906 0.60330576]
 [0.5165289  0.46280992 0.28099173 ... 0.35950413 0.3553719  0.38429752]]
[ 0  0  0  0  0  0  0  0  0  0  1  1  1  1  1  1  1  1  1  1  2  2  2  2
  2  2  2  2  2  2  3  3  3  3  3  3  3  3  3  3  4  4  4  4  4  4  4  4
  4  4  5  5  5  5  5  5  5  5  5  5  6  6  6  6  6  6  6  6  6  6  7  7
  7  7  7  7  7  7  7  7  8  8  8  8  8  8  8  8  8  8  9  9  9  9  9  9
  9  9  9  9 10 10 10 10 10 10 10 10 10 10 11 11 11 11 11 11 11 11 11 11
 12 12 12 12 12 12 12 12 12 12 13 13 13 13 13 13 13 13 13 13 14 14 14 14
 14 14 14 14 14 14 15 15 15 15 15 15 15 15 15 15 16 16 16 16 16 16 16 16
 16 16 17 17 17 17 17 17 17 17 17 17 18 18 18 18 18 18 18 18 18 18 19 19
 19 19 19 19 19 19 19 19 20 20 20 20 20 20 20 20 20 20 21 21 21 21 21 21
 21 21 21 21 22 22 22 22 22 22 22 22 22 22 23 23 23 23 23 23 23 23 23 23
 24 24 24 24 24 24 24 24 24 24 25 25 25 25 25 25 25 25 25 25 26 26 26 26
 26 26 26 26 26 26 27 27 27 27 27 27 27 27 27 27 28 28 28 28 28 28 28 28
 28 28 29 29 29 29 29 29 29 29 29 29 30 30 30 30 30 30 30 30 30 30 31 31
 31 31 31 31 31 31 31 31 32 32 32 32 32 32 32 32 32 32 33 33 33 33 33 33
 33 33 33 33 34 34 34 34 34 34 34 34 34 34 35 35 35 35 35 35 35 35 35 35
 36 36 36 36 36 36 36 36 36 36 37 37 37 37 37 37 37 37 37 37 38 38 38 38
 38 38 38 38 38 38 39 39 39 39 39 39 39 39 39 39]

20 Newsgroups

• 18K text posts on 20 topics, split into training & testing subsets
• train/test split based on messages posted before/after a specified date
• loader #1: returns list of raw texts for feeding to feature extractors
• loader #2 returns ready-to-use features (no extractor needed)
• Downloads data, extracts to ~/scikit_learn_data/20news_home, then calls load_files on training and/or test data folders

# 20 newsgroups text corpus (20 classes, 18.8K samples)
data = fetch_20newsgroups(subset='train')
from pprint import pprint
pprint(list(data.target_names))

['alt.atheism',
 'comp.graphics',
 'comp.os.ms-windows.misc',
 'comp.sys.ibm.pc.hardware',
 'comp.sys.mac.hardware',
 'comp.windows.x',
 'misc.forsale',
 'rec.autos',
 'rec.motorcycles',
 'rec.sport.baseball',
 'rec.sport.hockey',
 'sci.crypt',
 'sci.electronics',
 'sci.med',
 'sci.space',
 'soc.religion.christian',
 'talk.politics.guns',
 'talk.politics.mideast',
 'talk.politics.misc',
 'talk.religion.misc']

print(data.filenames.shape)
print(data.target.shape)
print(data.target[:10])

(11314,)
(11314,)
[ 7  4  4  1 14 16 13  3  2  4]

• converting text to TF-IDF vectors of unigram tokens
• vectors should be very sparse: ~159 non-zero values in a 30K dimensional space.

from sklearn.feature_extraction.text import TfidfVectorizer
categories = ['alt.atheism', 'talk.religion.misc',
              'comp.graphics', 'sci.space']

newsgroups_train = fetch_20newsgroups(subset='train',
                                      categories=categories)
vectorizer = TfidfVectorizer()
vectors = vectorizer.fit_transform(newsgroups_train.data)
print(vectors.shape)
print(vectors.nnz/float(vectors.shape[0]))

(2034, 34118)
159.0132743362832

• It's easy to overfit a classifier on irrelevant Newsgroup data such as headers.
• Example: Multinomial Naive Bayes - fast to train, decent F-score:

from sklearn.naive_bayes import MultinomialNB
from sklearn import metrics

newsgroups_test = fetch_20newsgroups(subset='test', categories=categories)
vectors_test    = vectorizer.transform(newsgroups_test.data)
clf             = MultinomialNB(alpha=.01).fit(vectors, 
                                               newsgroups_train.target)

pred = clf.predict(vectors_test)
print(metrics.f1_score(newsgroups_test.target, 
                       pred, 
                       average='macro'))

0.8821359240272957

• Most informative features?

import numpy as np

def show_top10(classifier, vectorizer, categories):
    feature_names = np.asarray(vectorizer.get_feature_names())
    
    for i, category in enumerate(categories):
        top10 = np.argsort(classifier.coef_[i])[-10:]
        print("%s: %s" % (category, " ".join(feature_names[top10])))

show_top10(clf, vectorizer, newsgroups_train.target_names)

alt.atheism: edu it and in you that is of to the
comp.graphics: edu in graphics it is for and of to the
sci.space: edu it that is in and space to of the
talk.religion.misc: not it you in is that and to of the

/home/bjpcjp/.local/lib/python3.8/site-packages/sklearn/utils/deprecation.py:101: FutureWarning: Attribute coef_ was deprecated in version 0.24 and will be removed in 1.1 (renaming of 0.26).
  warnings.warn(msg, category=FutureWarning)

• Things that cause overfit:
  • Headers: "NNTP-Posting-Host:", "Distribution:"
  • Whether the sender is affiliated with a university (headers, signatures)
  • The word "article"

• These clues make newsgroup classification much easier. To make the job harder, use remove to strip out information. (params can be a combination of 'headers','footers','quotes').

newsgroups_test_nh = fetch_20newsgroups(subset='test', remove=('headers'),
                                        categories=categories)
newsgroups_test_nf = fetch_20newsgroups(subset='test', remove=('footers'),
                                        categories=categories)
newsgroups_test_nq = fetch_20newsgroups(subset='test', remove=('quotes'),
                                        categories=categories)
newsgroups_test_all = fetch_20newsgroups(subset='test', remove=('headers',
                                                                'footers',
                                                                'quotes'),
                                        categories=categories)

vectors_test_nh = vectorizer.transform(newsgroups_test_nh.data)
vectors_test_nf = vectorizer.transform(newsgroups_test_nf.data)
vectors_test_nq = vectorizer.transform(newsgroups_test_nq.data)
vectors_test_all = vectorizer.transform(newsgroups_test_all.data)

pred_nh = clf.predict(vectors_test_nh)
pred_nf = clf.predict(vectors_test_nf)
pred_nq = clf.predict(vectors_test_nq)
pred_all = clf.predict(vectors_test_all)

print(metrics.f1_score(pred_nh, newsgroups_test_nh.target, average='macro'))
print(metrics.f1_score(pred_nf, newsgroups_test_nf.target, average='macro'))
print(metrics.f1_score(pred_nq, newsgroups_test_nq.target, average='macro'))
print(metrics.f1_score(pred_all, newsgroups_test_all.target, average='macro'))

0.870270086666271
0.8819288688799596
0.8401551094938573
0.7731035068127478

Labeled Faces in the Wild

• Collection of .jpegs; each picture is centered on a single face.
• 5.7K classes, 13.2K samples, 5.8K dimensions, 0-255 data
• Dataset size >200MB
• Loader downloads to ~/scikit_learn_data/lfw_home/ using joblib, parses metadata, decodes jpegs, converts slices into memmapped numpy arrays.
• The first loader is for face ID (a multiclass classification task).
• Default slice = rectangular shape around face (most background removed).
• Each face assigned to single person in target.
• Second loader is for face verification - each sample is a 2-pic pair, belonging (or not) to the same person.
• Both loaders can add RGB color info with color=True.
• Datasets are divided into train,test and 10_folds evaluation subsets.

data = fetch_lfw_people(min_faces_per_person=70, resize=0.4)
print(data.target_names.size,"\t",data.target_names[0])
print(data.data.dtype)
print(data.data.shape)
print(data.images.shape)

7 	 Ariel Sharon
float32
(1288, 1850)
(1288, 50, 37)

pairs = fetch_lfw_pairs(subset='train')
print(pairs.target_names)
print(pairs.pairs.shape)
print(pairs.data.shape)
print(pairs.target.shape)

['Different persons' 'Same person']
(2200, 2, 62, 47)
(2200, 5828)
(2200,)

Forest Covertypes

• Describes dominant tree species for 30x30m patches of US forestry.
• 581K samples.
• Seven possible values (multiclass)
• 54 features (some boolean, some discrete, some continuous.
• Default: returns data (a dict-like 'Bunch' object) with feature matrix & target.
• as_frame=True: returns as a Pandas dataframe instead.

data = fetch_covtype()
print(data.data.shape)
print(data.target.shape)
print(data.data[0])

(581012, 54)
(581012,)
[2.596e+03 5.100e+01 3.000e+00 2.580e+02 0.000e+00 5.100e+02 2.210e+02
 2.320e+02 1.480e+02 6.279e+03 1.000e+00 0.000e+00 0.000e+00 0.000e+00
 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
 1.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00]

Reuters Newswire Corpus

• 103 classes
• 804K samples with 47K dimensions, [0..1] values. Compressed dataset size ~656MB.
• Returns a dict-like object with:
  • data is a scipy sparse CSR matrix. Non-zero values are cosine-normalized TF-IDF vectors; first 23K samples = training, last 781K samples = test. Array should have 0.16% non-zero values.
  • target is a scipy sparse CSR matrix. 804K samples, 103 categories. Each sample has 1 in its category, 0 in the others. 3.1% non-zero.
  • each sample can be identified by its ID#.
  • target_names are the topics of each sample (1 to 17). 103 topics (strings). Corpus frequencies range from 5 ('GMIL') to 381K ('CCAT')

data = fetch_rcv1()
print(data.data.shape)
print(data.target.shape)
print(data.sample_id[:3])
print(data.target_names[:10].tolist())

(804414, 47236)
(804414, 103)
[2286 2287 2288]
['C11', 'C12', 'C13', 'C14', 'C15', 'C151', 'C1511', 'C152', 'C16', 'C17']

KDD CUP (1998 DARPA intrusion detection eval dataset0

• Artificial data with hand-injected attacks.

• Original dataset has ~80% of abnormal data = unrealistic for anomaly detection, so dataset is splite into SA && SF.

  • SA selects all normal data + ~1% of abnormal data.
  • SF selects all data with logged_in = positive (focuses on intrusion attack ~ 0.3%)
  • http and smtp are subsets of SF.

• 4.89M samples, 41 dimensions, discrete/continuous data

• targets: 'normal' or anomaly name
• SA: 976K samples, 41 dimensions;
• SF: 699K samples, 4 dimensions
• http: 619K samples, 3 dimensions

• smtp: 95K samples, 3 dimensions

• returns data feature matrix, target values. as_frame=True returns data as a Pandas dataframe and target as a Pandas series.

data = fetch_kddcup99(subset='smtp')
print(data.data.shape)
print(data.target.shape)
print(data.data[0])

(9571, 3)
(9571,)
[-2.3025850929940455 8.12151008316269 5.796361655949294]

California Housing

• 20K samples, 8 attributes, no missing attribute values
• Measures median house value for California districts, 1990 census

data = fetch_california_housing()
print(data.data.shape)
print(data.target.shape)
print(data.data[0])

(20640, 8)
(20640,)
[   8.3252       41.            6.98412698    1.02380952  322.
    2.55555556   37.88       -122.23      ]