Selective: Feature Selection Library

Selective is a white-box feature selection library that supports supervised and unsupervised selection methods for classification and regression tasks.

Selective also provides optimized item selection based on diversity of text embeddings (via TextWiser) and the coverage of binary labels by solving a multi-objective optimization problem (CPAIOR'21, DSO@IJCAI'22). The approach showed to speed-up online experimentation significantly and boost recommender systems NVIDIA GTC'22.

The library provides:

• Simple to complex selection methods: Variance, Correlation, Statistical, Linear, Tree-based, or Customized.
• Text-based selection to maximize diversity in text embeddings and metadata coverage.
• Interoperable with data frames as the input.
• Automated task detection. No need to know what feature selection method works with what machine learning task.
• Benchmarking multiple selectors using cross-validation with built-in parallelization.
• Inspection of the results and feature importance.

Selective is developed by the Artificial Intelligence Center of Excellence at Fidelity Investments.

Quick Start

# Import Selective and SelectionMethod
from sklearn.datasets import fetch_california_housing
from feature.utils import get_data_label
from feature.selector import Selective, SelectionMethod

# Data
data, label = get_data_label(fetch_california_housing())

# Feature selectors from simple to more complex
selector = Selective(SelectionMethod.Variance(threshold=0.0))
selector = Selective(SelectionMethod.Correlation(threshold=0.5, method="pearson"))
selector = Selective(SelectionMethod.Statistical(num_features=3, method="anova"))
selector = Selective(SelectionMethod.Linear(num_features=3, regularization="none"))
selector = Selective(SelectionMethod.TreeBased(num_features=3))

# Feature reduction
subset = selector.fit_transform(data, label)
print("Reduction:", list(subset.columns))
print("Scores:", list(selector.get_absolute_scores()))

Available Methods

Benchmarking

# Imports
from sklearn.datasets import fetch_california_housing
from feature.utils import get_data_label
from xgboost import XGBClassifier, XGBRegressor
from feature.selector import SelectionMethod, benchmark, calculate_statistics

# Data
data, label = get_data_label(fetch_california_housing())

# Selectors
corr_threshold = 0.5
num_features = 3
tree_params = {"n_estimators": 50, "max_depth": 5, "random_state": 111, "n_jobs": 4}
selectors = {

  # Correlation methods
  "corr_pearson": SelectionMethod.Correlation(corr_threshold, method="pearson"),
  "corr_kendall": SelectionMethod.Correlation(corr_threshold, method="kendall"),
  "corr_spearman": SelectionMethod.Correlation(corr_threshold, method="spearman"),

  # Statistical methods
  "stat_anova": SelectionMethod.Statistical(num_features, method="anova"),
  "stat_chi_square": SelectionMethod.Statistical(num_features, method="chi_square"),
  "stat_mutual_info": SelectionMethod.Statistical(num_features, method="mutual_info"),

  # Linear methods
  "linear": SelectionMethod.Linear(num_features, regularization="none"),
  "lasso": SelectionMethod.Linear(num_features, regularization="lasso", alpha=1000),
  "ridge": SelectionMethod.Linear(num_features, regularization="ridge", alpha=1000),

  # Non-linear tree-based methods
  "random_forest": SelectionMethod.TreeBased(num_features),
  "xgboost_classif": SelectionMethod.TreeBased(num_features, estimator=XGBClassifier(**tree_params)),
  "xgboost_regress": SelectionMethod.TreeBased(num_features, estimator=XGBRegressor(**tree_params))
}

# Benchmark (sequential)
score_df, selected_df, runtime_df = benchmark(selectors, data, label, cv=5)
print(score_df, "\n\n", selected_df, "\n\n", runtime_df)

# Benchmark (in parallel)
score_df, selected_df, runtime_df = benchmark(selectors, data, label, cv=5, n_jobs=4)
print(score_df, "\n\n", selected_df, "\n\n", runtime_df)

# Get benchmark statistics by feature
stats_df = calculate_statistics(score_df, selected_df)
print(stats_df)

Text-based Selection

This example shows how to use text-based selection. In this scenario, we would like to select a subset of articles that is most diverse in the text embedding space and covers a range of topics.

# Import Selective and TextWiser
import pandas as pd
from feature.selector import Selective, SelectionMethod
from textwiser import TextWiser, Embedding, Transformation

# Data with the text content of each article
data = pd.DataFrame({"article_1": ["article text here"],
                     "article_2": ["article text here"],
                     "article_3": ["article text here"],
                     "article_4": ["article text here"],
                     "article_5": ["article text here"]})

# Labels to denote 0/1 coverage metadata for each article 
# across four labels, e.g., sports, international, entertainment, science    
labels = pd.DataFrame({"article_1": [1, 1, 0, 1],
                       "article_2": [0, 1, 0, 0],
                       "article_3": [0, 0, 1, 0],
                       "article_4": [0, 0, 1, 1],
                       "article_5": [1, 1, 1, 0]},
                      index=["label_1", "label_2", "label_3", "label_4"])

# TextWiser featurization method to create text embeddings
textwiser = TextWiser(Embedding.TfIdf(), Transformation.NMF(n_components=20))

# Text-based selection
# The goal is to select a subset of articles 
# that is most diverse in the text embedding space of articles
# and covers the most labels in each topic
selector = Selective(SelectionMethod.TextBased(num_features=2, featurization_method=textwiser))

# Feature reduction
subset = selector.fit_transform(data, labels)
print("Reduction:", list(subset.columns))

Visualization

import pandas as pd
from sklearn.datasets import fetch_california_housing
from feature.utils import get_data_label
from feature.selector import SelectionMethod, Selective, plot_importance

# Data
data, label = get_data_label(fetch_california_housing())

# Feature Selector
selector = Selective(SelectionMethod.Linear(num_features=8, regularization="none"))
subset = selector.fit_transform(data, label)

# Plot Feature Importance
df = pd.DataFrame(selector.get_absolute_scores(), index=data.columns)
plot_importance(df)

Installation

Selective requires Python 3.7+ and can be installed from PyPI using pip install selective.

Source

Alternatively, you can build a wheel package on your platform from scratch using the source code:

git clone https://github.com/fidelity/selective.git
cd selective
pip install setuptools wheel # if wheel is not installed
python setup.py sdist bdist_wheel
pip install dist/selective-X.X.X-py3-none-any.whl

Test your setup

cd selective
python -m unittest discover tests

Support

Please submit bug reports and feature requests as Issues.

License

Selective is licensed under the GNU GPL 3.0.