FASTENER

FeAture SelecTion ENabled by EntRopy for Python

FASTENER is a state-of-the-art feature selecton algorithm for remote sensing, but performs well also on several other data sets. It is most suitable for large datasets with several hundreds of features. It has been develped for the use case of crop/land-cover classification based on Sentinel-II data.

Prerequisites

• Python 3.6+
• scikit-learn (0.22.2+)
• mypy(0.761)
• For example:
  • Pandas

For further details see requirements.txt.

Installation

Install using pip:

pip install fastener

Users' Manual

Basic documentation is available within the code.

A simple workflow is described below.

0. Includes

# import dataset
from sklearn.datasets import load_breast_cancer

# import preprocessing tools
from sklearn import preprocessing
import numpy as np
import pandas as pd

# import learning/evaluation
from sklearn.tree import DecisionTreeClassifier
from sklearn.metrics import f1_score
from sklearn.model_selection import cross_val_score

# typing
from typing import Dict, List, Callable, Any, Tuple, Optional, \
    Counter as CounterType, Set

# FASTENER specific imports
from fastener.random_utils import shuffle
from fastener.item import Result, Genes, RandomFlipMutationStrategy, RandomEveryoneWithEveryone, \
    IntersectionMating, UnionMating, IntersectionMatingWithInformationGain, \
    IntersectionMatingWithWeightedRandomInformationGain
from fastener import fastener

1. Prepare data

# loading breast cancer dataset
# scikit-learn 0.22+ is needed
cancer = load_breast_cancer(as_frame=True)
X_df = cancer.data
y_df = cancer.target

# basic dataset split
n_sample = X_df.shape[0]
n_test = int(n_sample * 0.8)

labels_train = y_df.to_numpy()[:n_test]
labels_test = y_df.to_numpy()[n_test:]

XX_train = X_df.to_numpy()[:n_test, :]
XX_test = X_df.to_numpy()[n_test:, :]

2. Define feature evaluation function

def eval_fun(model: Any, genes: "Genes", shuffle_indices: Optional[List[int]] = None) -> "Result":
    test_data = XX_test[:, genes]
    if shuffle_indices:
        test_data = test_data.copy()
        for j in shuffle_indices:
            shuffle(test_data[:, j])
    pred = model.predict(test_data)
    res = Result(f1_score(labels_test, pred))
    return res

3. Configure the FASTENER

By default fastener runs for 1000 iterations. The number of iterations can be adjusted with number_of_rounds parameter in the fastener.Config.

number_of_genes = XX_train.shape[1]
general_model = DecisionTreeClassifier
# output folder name must be changed every time the algorithm is run
output_folder_name = "output"

# to start the algorithm initial_genes or initial_population must be provided
initial_genes = [
    [0]
]

# Select mating selection strategie (RandomEveryoneWithEveryone, NoMating) and mating strategy
# (UnionMating, IntersectionMating, IntersectionMatingWithInformationGain, 
# IntersectionMatingWithWeightedRandomInformationGain) 
# If regression model is used IntersectionMatingWithInformationGain, IntersectionMatingWithWeightedRandomInformationGain 
# must have regression=True set (eg. IntersectionMatingWithInformationGain(regression=True))
mating = RandomEveryoneWithEveryone(pool_size=3, mating_strategy=IntersectionMatingWithWeightedRandomInformationGain())

# Random mutation (probability of gene mutating: 1 / number_of_genes)
mutation = RandomFlipMutationStrategy(1 / number_of_genes)

entropy_optimizer = fastener.EntropyOptimizer(
    general_model, XX_train, labels_train, eval_fun,
    number_of_genes, mating, mutation, initial_genes=initial_genes,
    config=fastener.Config(output_folder=output_folder_name, random_seed=2020, reset_to_pareto_rounds=5)
)

4. Run FASTENER loop

entropy_optimizer.mainloop()

5. Check evaluation of the 1000th iteration

# read log from last generation
object = pd.read_pickle(f'log/{output_folder_name}/generation_1000.pickle')

# list of best-scoring EvalItem objects for each number of features
best = list(object.front.values())

for item in best:
    # names of best features
    selected_features = X_df.iloc[:, item.genes].columns.tolist()

    X = X_df[selected_features].values.astype(float)
    y = y_df.values.astype(float)

    # evaluates each set of features with cross validation
    model = DecisionTreeClassifier()
    cvs = cross_val_score(model, X, y, cv=10)
    print("Features:", selected_features)
    print("Accuracy: ", cvs.mean(), " stdev: ", cvs.std(), "\n")

For detailed workflow check Example.ipynb.

Mating strategy

The following mating strategies are available:

• Union mating: If either (or both) of the parents have the feature selected the descendent will have it too.

mating_strategy = UnionMating()

• Intersection mating: If both of the parents have the feature the descendent will have it too.

mating_strategy = IntersectionMating()

• Intersection mating with information gain: If both of the parents have the feature the descendent will have it too. Additionally, some features from either one of the parents, that have the highest information gain are added.

mating_strategy = IntersectionMatingWithInformationGain()

• Intersection mating with weighted random information gain: If both of the parents have the feature the descendent will have it too. Additionally, some features from either one of the parents, will be added. The probability of a feature being selected is proportional to scaling function applied to it's calculated information gain.

mating_strategy = IntersectionMatingWithWeightedRandomInformationGain()

Note: If regression model is used with Intersection mating with information gain or Intersection mating with weighted random information gain, the regression flag must be set to True (eg. IntersectionMatingWithWeightedRandomInformationGain(regression=True)). However, if the dataset is large this can cause errors so intersection mating or union mating is a better choice.

Future Work

• Update documentation
• Prepare example notebooks
• Create unit tests

Publications

If you use the algorithm, please cite the following paper:

• Koprivec, F.; Kenda, K.; Šircelj, B., FASTENER Feature Selection for Inference from Earth Observation Data. Entropy 2020, 22, 1198 (link).

Abstract:

In this paper, FASTENER feature selection algorithm is presented. The algorithm exploits entropy-based measures such as mutual information in the crossover phase of the genetic algorithm approach. FASTENER converges to an (near) optimal subset of features faster than previous state-of-the-art algorithms and achieves better classification accuracy than similarity-based methods such as KBest or ReliefF or wrapper methods such as POSS. The approach was evaluated using the Earth Observation dataset for land-cover classification from ESA's Sentinel-2 mission, the digital elevation model and the ground truth data of the Land Parcel Identification System from Slovenia. The algorithm can be used in any statistical learning scenario.

Acknowledgements

This research was funded by European Union's Horizon 2020 programme project PerceptiveSentinel (Research and Innovation) grant number 776115, project NAIADES (Innovation Action) grant number 820985.