How should data be preprocessed for use in machine learning algorithms? How to identify the most predictive attributes of a dataset? What features can generate to improve the accuracy of a model?
Feature Engineering is the process of extracting and selecting, from raw data, features that can be used effectively in predictive models. As the quality of the features greatly influences the quality of the results, knowing the main techniques and pitfalls will help you to succeed in the use of machine learning in your projects.
In this talk, we will present methods and techniques that allow us to extract the maximum potential of the features of a dataset, increasing flexibility, simplicity and accuracy of the models. The analysis of the distribution of features and their correlations, the transformation of numeric attributes (such as scaling, normalization, log-based transformation, binning), categorical attributes (such as one-hot encoding, feature hashing, Temporal (date / time), and free-text attributes (text vectorization, topic modeling).
Python, Python, Scikit-learn, and Spark SQL examples will be presented and how to use domain knowledge and intuition to select and generate features relevant to predictive models.
4. "Feature engineering is the process of
transforming raw data into features that better
represent the underlying problem to the
predictive models, resulting in improved
model accuracy on unseen data."
– Jason Brownlee
5. “Coming up with features is difficult,
time-consuming,
requires expert knowledge.
'Applied machine learning' is basically
feature engineering.”
– Andrew Ng
10. Outbrain Click Prediction - Kaggle competition
Dataset
● Sample of users page views
and clicks during
14 days on June, 2016
● 2 Billion page views
● 17 million click records
● 700 Million unique users
● 560 sites
Can you predict which recommended content each user will click?
11. I got 19th position
from about
1000 competitors
(top 2%),
mostly due to
Feature Engineering
techniques.
More details of my solution
in this post series
12. ● What does the data model look like?
● What is the features distribution?
● What are the features with missing
or inconsistent values?
● What are the most predictive features?
● Conduct a Exploratory Data Analysis (EDA)
First at all … a closer look at your data
15. Data Cleansing
Homogenize missing values and different types of in the same feature, fix input errors, types, etc.
Original data
Cleaned data
16. ML-Ready Dataset
Fields (Features)
Instances
Tabular data (rows and columns)
● Usually denormalized in a single file/dataset
● Each row contains information about one instance
● Each column is a feature that describes a property of the instance
17. Aggregating
Necessary when the entity to model is an aggregation from the provided data.
Original data (list of playbacks)
Aggregated data (list of users)
18. Pivoting
Necessary when the entity to model is an aggregation from the provided data.
Aggregated data with pivoted columns
Original data
# playbacks by device Play duration by device
20. Numerical features
● Usually easy to ingest by mathematical models.
● Can be prices, measurements, counts, ...
● Easier to impute missing data
● Distribution and scale matters to many models
21. Imputation for missing values
● Datasets contain missing values, often encoded as blanks, NaNs or other
placeholders
● Ignoring rows and/or columns with missing values is possible, but at the price of
loosing data which might be valuable
● Better strategy is to infer them from the known part of data
● Strategies
○ Mean: Basic approach
○ Median: More robust to outliers
○ Mode: Most frequent value
○ Using a model: Can expose algorithmic bias
23. Binarization
● Transform discrete or continuous numeric features in binary features
Example: Number of user views of the same document
>>> from sklearn import preprocessing
>>> X = [[ 1., -1., 2.],
... [ 2., 0., 0.],
... [ 0., 1., -1.]]
>>> binarizer =
preprocessing.Binarizer(threshold=1.0)
>>> binarizer.transform(X)
array([[ 1., 0., 1.],
[ 1., 0., 0.],
[ 0., 1., 0.]])
Binarization with scikit-learn
24. Rounding
● Form of lossy compression: retain most significant features of the data.
● Sometimes too much precision is just noise
● Rounded variables can be treated as categorical variables
● Example:
Some models like Association Rules work only with categorical features. It is
possible to convert a percentage into categorial feature this way
Extra slides marker
25. Binning
● Split numerical values into bins and encode with a bin ID
● Can be set arbitrarily or based on distribution
● Fixed-width binning
Does fixed-width binning make sense for this long-tailed distribution?
Most users (458,234,809 ~ 5*10^8) had only 1 pageview during the period.
27. Log transformation
Compresses the range of large numbers and expand the range of small numbers.
Eg. The larger x is, the slower log(x) increments.
28. Log transformation
Histogram of # views by user Histogram of # views by user
smoothed by log(1+x)
Smoothing long-tailed data with log
29. Scaling
● Models that are smooth functions of input features are sensitive to the scale
of the input (eg. Linear Regression)
● Scale numerical variables into a certain range, dividing values by a
normalization constant (no changes in single-feature distribution)
● Popular techniques
○ MinMax Scaling
○ Standard (Z) Scaling
30. Min-max scaling
● Squeezes (or stretches) all values within the range of [0, 1] to add robustness to
very small standard deviations and preserving zeros for sparse data.
>>> from sklearn import preprocessing
>>> X_train = np.array([[ 1., -1., 2.],
... [ 2., 0., 0.],
... [ 0., 1., -1.]])
...
>>> min_max_scaler =
preprocessing.MinMaxScaler()
>>> X_train_minmax =
min_max_scaler.fit_transform(X_train)
>>> X_train_minmax
array([[ 0.5 , 0. , 1. ],
[ 1. , 0.5 , 0.33333333],
[ 0. , 1. , 0. ]])
Min-max scaling with scikit-learn
31. Standard (Z) Scaling
After Standardization, a feature has mean of 0 and variance of 1 (assumption of
many learning algorithms)
>>> from sklearn import preprocessing
>>> import numpy as np
>>> X = np.array([[ 1., -1., 2.],
... [ 2., 0., 0.],
... [ 0., 1., -1.]])
>>> X_scaled = preprocessing.scale(X)
>>> X_scaled
array([[ 0. ..., -1.22..., 1.33...],
[ 1.22..., 0. ..., -0.26...],
[-1.22..., 1.22..., -1.06...]])
>> X_scaled.mean(axis=0)
array([ 0., 0., 0.])
>>> X_scaled.std(axis=0)
array([ 1., 1., 1.])
Standardization with scikit-learn
32. ● Scales individual samples (rows) to have unit vector, dividing values by
vector’s L2
norm, a.k.a. the Euclidean norm
● Useful for quadratic form (like dot-product) or any other kernel to quantify
similarity of pairs of samples. This assumption is the base of the Vector
Space Model often used in text classification and clustering contexts
Normalization
Normalized vector
Euclidean (L2
) norm
34. Interaction Features
● Simple linear models use a linear combination of the individual input
features, x1
, x2
, ... xn
to predict the outcome y.
y = w1
x1
+ w2
x2
+ ... + wn
xn
● An easy way to increase the complexity of the linear model is to create
feature combinations (nonlinear features).
Area (m2)Example (House Pricing Prediction)
Degree 2 interaction features for vector x = (x1,
x2
)
y = w1
x1
+ w2
x2
+ w3
x1
x2
+ w4
x1
2
+ w4
x2
2
# Rooms
Price
35. Interaction Features
>>> import numpy as np
>>> from sklearn.preprocessing import PolynomialFeatures
>>> X = np.arange(6).reshape(3, 2)
>>> X
array([[0, 1],
[2, 3],
[4, 5]])
>>> poly = poly = PolynomialFeatures(degree=2, interaction_only=False,
include_bias=True)
>>> poly.fit_transform(X)
array([[ 1., 0., 1., 0., 0., 1.],
[ 1., 2., 3., 4., 6., 9.],
[ 1., 4., 5., 16., 20., 25.]])
Polynomial features with scikit-learn
36. Interaction Features - Vowpal Wabbit
vw --loss_function logistic --link=logistic --ftrl --ftrl_alpha 0.005 --ftrl_beta 0.1
-q cc -q zc -q zm
-l 0.01 --l1 1.0 --l2 1.0 -b 28 --hash all
--compressed -d data/train_fv.vw -f output.model
Feature interactions with VW
Interacting (quadratic) features of some namespaces
vw_line = '{} |i {} |m {} |z {} |c {}n'.format(
label,
' '.join(integer_features),
' '.join(ctr_features),
' '.join(similarity_features),
' '.join(categorical_features))
Separating features in namespaces in Vowpal Wabbit (VW) sparse format
1 |i 12:5 18:126 |m 2:0.015 45:0.123 |z 32:0.576 17:0.121 |c 16:1 295:1 3554:1
Sample data point (line in VW format file)
38. Categorical Features
● Nearly always need some treatment to be suitable for models
● Examples:
Platform: [“desktop”, “tablet”, “mobile”]
Document_ID or User_ID: [121545, 64845, 121545]
● High cardinality can create very sparse data
● Difficult to impute missing
39. One-Hot Encoding (OHE)
● Transform a categorical feature with m possible values into m binary features.
● If the variable cannot be multiple categories at once, then only one bit in the
group can be on.
● Sparse format is memory-friendly
● Example: “platform=tablet” can be sparsely encoded as “2:1”
41. Large Categorical Variables
● Common in applications like targeted advertising and fraud detection
● Example:
Some large categorical features from Outbrain Click Prediction competition
42. Feature hashing
● Hashes categorical values into vectors with fixed-length.
● Lower sparsity and higher compression compared to OHE
● Deals with new and rare categorical values (eg: new user-agents)
● May introduce collisions
100 hashed columns
43. Feature hashing
import hashlib
def hashstr(s, nr_bins):
return int(hashlib.md5(s.encode('utf8')).hexdigest(), 16)%(nr_bins-1)+1
CATEGORICAL_VALUE='ad_id=354424'
MAX_BINS=100000
>>> hashstr(CATEGORICAL_VALUE, MAX_BINS)
49389
Feature hashing with pure Python
Original category
Hashed category
import tensorflow as tf
ad_id_hashed = tf.contrib.layers.sparse_column_with_hash_bucket('ad_id',
hash_bucket_size=250000, dtype=tf.int64, combiner="sum")
Feature hashing with TensorFlow
44. Feature hashing
vw --loss_function logistic --link=logistic --ftrl --ftrl_alpha 0.005 --ftrl_beta 0.1
-q cc -q zc -q zm -l 0.01 --l1 1.0 --l2 1.0
-b 18 --hash all
--compressed -d data/train_fv.vw -f output.model
Feature hashing with Vowpal Wabbit
Hashes values to a feature space of 218
positions (columns)
45. Bin-counting
● Instead of using the actual categorical value, use a global statistic of this
category on historical data.
● Useful for both linear and non-linear algorithms
● May give collisions (same encoding for different categories)
● Be careful about leakage
47. LabelCount encoding
● Rank categorical variables by count in train set
● Useful for both linear and non-linear algorithms (eg: decision trees)
● Not sensitive to outliers
● Won’t give same encoding to different variables
48. Category Embedding
● Use a Neural Network to create dense embeddings from categorical
variables.
● Map categorical variables in a function approximation problem into Euclidean
spaces
● Faster model training.
● Less memory overhead.
● Can give better accuracy than 1-hot encoded.
51. Time Zone conversion
Factors to consider:
● Multiple time zones in some countries
● Daylight Saving Time (DST)
○ Start and end DST dates
52. ● Apply binning on time data to make it categorial and more general.
● Binning a time in hours or periods of day, like below.
● Extraction: weekday/weekend, weeks, months, quarters, years...
Hour range Bin ID Bin Description
[5, 8) 1 Early Morning
[8, 11) 2 Morning
[11, 14) 3 Midday
[14, 19) 4 Afternoon
[19, 22) 5 Evening
[22-24) and (00-05] 6 Night
Time binning
53. ● Instead of encoding: total spend, encode things like:
Spend in last week, spend in last month, spend in last
year.
● Gives a trend to the algorithm: two customers with equal
spend, can have wildly different behavior — one
customer may be starting to spend more, while the other
is starting to decline spending.
Trendlines
54. ● Hardcode categorical features from dates
● Example: Factors that might have major influence on spending behavior
● Proximity to major events (holidays, major sports events)
○ Eg. date_X_days_before_holidays
● Proximity to wages payment date (monthly seasonality)
○ Eg. first_saturday_of_the_month
Closeness to major events
55. ● Differences between dates might be relevant
● Examples:
○ user_interaction_date - published_doc_date
To model how recent was the ad when the user viewed it.
Hypothesis: user interests on a topic may decay over time
○ last_user_interaction_date - user_interaction_date
To model how old was a given user interaction compared to his last
interaction
Time differences
57. Spatial Variables
● Spatial variables encode a location in space, like:
○ GPS-coordinates (lat. / long.) - sometimes require projection to a different
coordinate system
○ Street Addresses - require geocoding
○ ZipCodes, Cities, States, Countries - usually enriched with the centroid
coordinate of the polygon (from external GIS data)
● Derived features
○ Distance between a user location and searched hotels (Expedia competition)
○ Impossible travel speed (fraud detection)
58. Spatial Enrichment
Usually useful to enrich with external geographic data (eg. Census demographics)
Beverage Containers Redemption Fraud Detection: Usage of # containers redeemed (red circles) by
store and Census households median income by Census Tracts
60. Natural Language Processing
Cleaning
• Lowercasing
• Convert accented characters
• Removing non-alphanumeric
• Repairing
Tokenizing
• Encode punctuation marks
• Tokenize
• N-Grams
• Skip-grams
• Char-grams
• Affixes
Removing
• Stopwords
• Rare words
• Common words
Roots
• Spelling correction
• Chop
• Stem
• Lemmatize
Enrich
• Entity Insertion / Extraction
• Parse Trees
• Reading Level
61. Represent each document as a feature vector in the vector space, where each
position represents a word (token) and the contained value is its relevance in the
document.
● BoW (Bag of words)
● TF-IDF (Term Frequency - Inverse Document Frequency)
● Embeddings (eg. Word2Vec, Glove)
● Topic models (e.g LDA)
Document Term Matrix - Bag of Words
Text vectorization
63. from sklearn.feature_extraction.text import TfidfVectorizer
vectorizer = TfidfVectorizer(max_df=0.5, max_features=1000,
min_df=2, stop_words='english')
tfidf_corpus = vectorizer.fit_transform(text_corpus)
face person guide lock cat dog sleep micro pool gym
0 1 2 3 4 5 6 7 8 9
D1 0.05 0.25
D2 0.02 0.32 0.45
...
...
tokens
documents
TF-IDF sparse matrix example
Text vectorization - TF-IDF
TF-IDF with scikit-learn
64. Similarity metric between two vectors is cosine among the angle between them
from sklearn.metrics.pairwise import cosine_similarity
cosine_similarity(tfidf_matrix[0:1], tfidf_matrix)
Cosine Similarity with scikit-learn
Cosine Similarity
65. Textual Similarities
• Token similarity: Count number of tokens that appear in
two texts.
• Levenshtein/Hamming/Jaccard Distance: Check
similarity between two strings, by looking at number of
operations needed to transform one in the other.
• Word2Vec / Glove: Check cosine similarity between two
word embedding vectors
69. Feature Selection
Reduces model complexity and training time
● Filtering - Eg. Correlation our Mutual Information between
each feature and the response variable
● Wrapper methods - Expensive, trying to optimize the best
subset of features (eg. Stepwise Regression)
● Embedded methods - Feature selection as part of model
training process (eg. Feature Importances of Decision Trees or
Trees Ensembles)
70. Diverse set of Features and Models leads to different results!
Outbrain Click Prediction - Leaderboard score of my approaches
72. “More data beats clever algorithms,
but better data beats more data.”
– Peter Norvig
73. “...some machine learning projects
succeed and some fail.
Where is the difference?
Easily the most important factor is the
features used.”
– Pedro Domingos