Skip to content

Latest commit

 

History

History
328 lines (210 loc) · 16.7 KB

feature-engineering.md

File metadata and controls

328 lines (210 loc) · 16.7 KB
Raw
Support HackTricks and get benefits!

Basic types of possible data

Data can be continuous (infinity values) or categorical (nominal) where the amount of possible values are limited.

Categorical types

Binary

Just 2 possible values: 1 or 0. In case in a dataset the values are in string format (e.g. "True" and "False") you assign numbers to those values with:

dataset["column2"] = dataset.column2.map({"T": 1, "F": 0})

Ordinal

The values follows an order, like in: 1st place, 2nd place... If the categories are strings (like: "starter", "amateur", "professional", "expert") you can map them to numbers as we saw in the binary case.

column2_mapping = {'starter':0,'amateur':1,'professional':2,'expert':3}
dataset['column2'] = dataset.column2.map(column2_mapping)
  • For alphabetic columns you can order them more easily:
# First get all the uniq values alphabetically sorted
possible_values_sorted = dataset.column2.sort_values().unique().tolist()
# Assign each one a value
possible_values_mapping = {value:idx for idx,value in enumerate(possible_values_sorted)}
dataset['column2'] = dataset.column2.map(possible_values_mapping)

Cyclical

Looks like ordinal value because there is an order, but it doesn't mean one is bigger than the other. Also the distance between them depends on the direction you are counting. Example: The days of the week, Sunday isn't "bigger" than Monday.

  • There are different ways to encode cyclical features, ones may work with only just some algorithms. In general, dummy encode can be used
column2_dummies = pd.get_dummies(dataset.column2, drop_first=True)
dataset_joined = pd.concat([dataset[['column2']], column2_dummies], axis=1)

Dates

Date are continuous variables. Can be seen as cyclical (because they repeat) or as ordinal variables (because a time is bigger than a previous one).

  • Usually dates are used as index
# Transform dates to datetime
dataset["column_date"] = pd.to_datetime(dataset.column_date)
# Make the date feature the index
dataset.set_index('column_date', inplace=True)
print(dataset.head())

# Sum usage column per day
daily_sum = dataset.groupby(df_daily_usage.index.date).agg({'usage':['sum']})
# Flatten and rename usage column
daily_sum.columns = daily_sum.columns.get_level_values(0)
daily_sum.columns = ['daily_usage']
print(daily_sum.head())

# Fill days with 0 usage
idx = pd.date_range('2020-01-01', '2020-12-31')
daily_sum.index = pd.DatetimeIndex(daily_sum.index)
df_filled = daily_sum.reindex(idx, fill_value=0) # Fill missing values


# Get day of the week, Monday=0, Sunday=6, and week days names
dataset['DoW'] = dataset.transaction_date.dt.dayofweek
# do the same in a different way
dataset['weekday'] = dataset.transaction_date.dt.weekday
# get day names
dataset['day_name'] = dataset.transaction_date.apply(lambda x: x.day_name())

Multi-category/nominal

More than 2 categories with no related order. Use dataset.describe(include='all') to get information about the categories of each feature.

  • A referring string is a column that identifies an example (like a name of a person). This can be duplicated (because 2 people may have the same name) but most will be unique. This data is useless and should be removed.
  • A key column is used to link data between tables. In this case the elements are unique. his data is useless and should be removed.

To encode multi-category columns into numbers (so the ML algorithm understand them), dummy encoding is used (and not one-hot encoding because it doesn't avoid perfect multicollinearity).

You can get a multi-category column one-hot encoded with pd.get_dummies(dataset.column1). This will transform all the classes in binary features, so this will create one new column per possible class and will assign 1 True value to one column, and the rest will be false.

You can get a multi-category column dummie encoded with pd.get_dummies(dataset.column1, drop_first=True). This will transform all the classes in binary features, so this will create one new column per possible class minus one as the last 2 columns will be reflect as "1" or "0" in the last binary column created. This will avoid perfect multicollinearity, reducing the relations between columns.

Collinear/Multicollinearity

Collinear appears when 2 features are related to each other. Multicollineratity appears when those are more than 2.

In ML you want that your features are related with the possible results but you don't want them to be related between them. That's why the dummy encoding mix the last two columns of that and is better than one-hot encoding which doesn't do that creating a clear relation between all the new featured from the multi-category column.

VIF is the Variance Inflation Factor which measures the multicollinearity of the features. A value above 5 means that one of the two or more collinear features should be removed.

from statsmodels.stats.outliers_influence import variance_inflation_factor
from statsmodels.tools.tools import add_constant

#dummies_encoded = pd.get_dummies(dataset.column1, drop_first=True)
onehot_encoded = pd.get_dummies(dataset.column1)
X = add_constant(onehot_encoded) # Add previously one-hot encoded data
print(pd.Series([variance_inflation_factor(X.values,i) for i in range(X.shape[1])], index=X.columns))

Categorical Imbalance

This occurs when there is not the same amount of each category in the training data.

# Get statistic of the features
print(dataset.describe(include='all'))
# Get an overview of the features
print(dataset.info())
# Get imbalance information of the target column
print(dataset.target_column.value_counts())

In an imbalance there is always a majority class or classes and a minority class or classes.

There are 2 main ways to fix this problem:

  • Undersampling: Removing randomly selected data from the majority class so it has the same number of samples as the minority class.
from imblearn.under_sampling import RandomUnderSampler
rus = RandomUserSampler(random_state=1337)

X = dataset[['column1', 'column2', 'column3']].copy()
y = dataset.target_column

X_under, y_under = rus.fit_resample(X,y)
print(y_under.value_counts()) #Confirm data isn't imbalanced anymore
  • Oversampling: Generating more data for the minority class until it has as many samples as the majority class.
from imblearn.under_sampling import RandomOverSampler
ros = RandomOverSampler(random_state=1337)

X = dataset[['column1', 'column2', 'column3']].copy()
y = dataset.target_column

X_over, y_over = ros.fit_resample(X,y)
print(y_over.value_counts()) #Confirm data isn't imbalanced anymore

You can use the argument sampling_strategy to indicate the percentage you want to undersample or oversample (by default it's 1 (100%) which means to equal the number of minority classes with majority classes)

{% hint style="info" %} Undersamplig or Oversampling aren't perfect if you get statistics (with .describe()) of the over/under-sampled data and compare them to the original you will see that they changed. Therefore oversampling and undersampling are modifying the training data. {% endhint %}

SMOTE oversampling

SMOTE is usually a more trustable way to oversample the data.

from imblearn.over_sampling import SMOTE

# Form SMOTE the target_column need to be numeric, map it if necessary
smote = SMOTE(random_state=1337)
X_smote, y_smote = smote.fit_resample(dataset[['column1', 'column2', 'column3']], dataset.target_column)
dataset_smote = pd.DataFrame(X_smote, columns=['column1', 'column2', 'column3'])
dataset['target_column'] = y_smote
print(y_smote.value_counts()) #Confirm data isn't imbalanced anymore

Rarely Occurring Categories

Imagine a dataset where one of the target classes occur very little times.

This is like the category imbalance from the previous section, but the rarely occurring category is occurring even less than "minority class" in that case. The raw oversampling and undersampling methods could be also used here, but generally those techniques won't give really good results.

Weights

In some algorithms it's possible to modify the weights of the targeted data so some of them get by default more importance when generating the model.

weights = {0: 10 1:1} #Assign weight 10 to False and 1 to True
model = LogisticRegression(class_weight=weights)

You can mix the weights with over/under-sampling techniques to try to improve the results.

PCA - Principal Component Analysis

Is a method that helps to reduce the dimensionality of the data. It's going to combine different features to reduce the amount of them generating more useful features (less computation is needed).

The resulting features aren't understandable by humans, so it also anonymize the data.

Incongruent Label Categories

Data might have mistakes for unsuccessful transformations or just because human error when writing the data.

Therefore you might find the same label with spelling mistakes, different capitalisation, abbreviations like: BLUE, Blue, b, bule. You need to fix these label errors inside the data before training the model.

You can clean this issues by lowercasing everything and mapping misspelled labels to the correct ones.

It's very important to check that all the data that you have contains is correctly labeled, because for example, one misspelling error in the data, when dummie encoding the classes, will generate a new column in the final features with bad consequences for the final model. This example can be detected very easily by one-hot encoding a column and checking the names of the columns created.

Missing Data

Some data of the study may be missing.

It might happen that some complete random data is missing for some error. This is kind of da ta is Missing Completely at Random (MCAR).

It could be that some random data is missing but there is something making some specific details more probable to be missing, for example more frequently man will tell their their age but not women. This is call Missing at Random (MAR).

Finally, there could be data Missing Not at Random (MNAR). The vale of the data is directly related with the probability of having the data. For example, if you want to measure something embarrassing, the most embarrassing someone is, the less probable he is going to share it.

The two first categories of missing data can be ignorable. But the third one requires to consider only portions of the data that isn't impacted or to try to model the missing data somehow.

One way to find about missing data is to use .info() function as it will indicate the number of rows but also the number of values per category. If some category has less values than number of rows, then there is some data missing:

# Get info of the dataset
dataset.info()

# Drop all rows where some value is missing
dataset.dropna(how='any', axis=0).info()

It's usually recommended that if a feature is missing in more than the 20% of the dataset, the column should be removed:

# Remove column
dataset.drop('Column_name', axis='columns', inplace=True)
dataset.info()

{% hint style="info" %} Note that not all the missing values are missing in the dataset. It's possible that missing values have been giving the value "Unknown", "n/a", "", -1, 0... You need to check the dataset (using dataset.columnname.valuecounts(dropna=False) to check the possible values). {% endhint %}

If some data is missing in the dataset (in it's not too much) you need to find the category of the missing data. For that you basically need to know if the missing data is at random or not, and for that you need to find if the missing data was correlated with other data of the dataset.

To find if a missing value if correlated with another column, you can create a new column that put 1s and 0s if the data is missing or isn't and then calculate the correlation between them:

# The closer it's to 1 or -1 the more correlated the data is
# Note that columns are always perfectly correlated with themselves.
dataset[['column_name', 'cloumn_missing_data']].corr()

If you decide to ignore the missing data you still need to do what to do with it: You can remove the rows with missing data (the train data for the model will be smaller), you can remove the feature completely, or could model it.

You should check the correlation between the missing feature with the target column to see how important that feature is for the target, if it's really small you can drop it or fill it.

To fill missing continuous data you could use: the mean, the median or use an imputation algorithm. The imputation algorithm can try to use other features to find a value for the missing feature:

from sklearn.impute import KNNImputer

X = dataset[['column1', 'column2', 'column3']]
y = dataset.column_target

# Create the imputer that will fill the data
imputer = KNNImputer(n_neightbors=2, weights='uniform')
X_imp = imputer.fit_transform(X)

# Check new data
dataset_imp = pd.DataFrame(X_imp)
dataset.columns = ['column1', 'column2', 'column3']
dataset.iloc[10:20] # Get some indexes that contained empty data before

To fill categorical data first of all you need to think if there is any reason why the values are missing. If it's by choice of the users (they didn't want to give the data) maybe yo can create a new category indicating that. If it's because of human error you can remove the rows or the feature (check the steps mentioned before) or fill it with the mode, the most used category (not recommended).

Combining Features

If you find two features that are correlated between them, usually you should drop one of them (the one that is less correlated with the target), but you could also try to combine them and create a new feature.

# Create a new feautr combining feature1 and feature2
dataset['new_feature'] = dataset.column1/dataset.column2

# Check correlation with target column
dataset[['new_feature', 'column1', 'column2', 'target']].corr()['target'][:]

# Check for collinearity of the 2 features and the new one
X = add_constant(dataset[['column1', 'column2', 'target']])
# Calculate VIF
pd.Series([variance_inflation_factor(X.values, i) for i in range(X.shape[1])], index=X.columns)
Support HackTricks and get benefits!