FSRLeaning - Python Library

FSRLeaning is a Python library for feature selection using reinforcement learning. It's designed to be easy to use and efficient, particularly for selecting the most relevant features from a very large set.

Installation

Install FSRLearning using pip:

pip install FSRLearning

Example usage

Data Pre-processing

The Dataset

In this example, we're using the Australian credit approval dataset. It has 14 features that have been intentionally anonymized. The goal is to predict whether the label is 0 or 1. We're using this dataset to demonstrate how to use the library, but the model can work with any dataset. You can find more details about the dataset here.

The process

The first step is a pre-processing of the data. You need to give as input to the method for feature selection a X and y pandas DataFrame. X is the dataset with all the features that we want to evaluate and y the label to be predicted. It is highly recommended to create a mapping between features and a list of number. For example each feature is associated with a number. Here is an example of the data pre-processing step on a data set with 14 features including 1 label.

import pandas as pd

# Get the pandas DataFrame
australian_data = pd.read_csv('australian_data.csv', header=None)

# Get the dataset with the features
X = australian_data.drop(14, axis=1)

# Get the dataset with the label values
y = australian_data[14]

After this step we can simply run a feature selection and ranking process that maximises a metric.

from FSRLearning import FeatureSelectorRL

# Create the object of feature selection with RL
fsrl_obj = FeatureSelectorRL(14, nb_iter=200)

# Returns the results of the selection and the ranking
results = fsrl_obj.fit_predict(X, y)
results

The feature_Selector_RL has several parameters that can be tuned. Here is all of them and the values that they can take.

• feature_number (integer) : number of features in the DataFrame X

• feature_structure (dictionary, optional) : dictionary for the graph implementation

• eps (float [0; 1], optional) : probability of choosing a random next state, 0 is an only greedy algorithm and 1 only random

• alpha (float [0; 1], optional): control the rate of updates, 0 is a very not updating state and 1 a very updated

• gamma (float [0, 1], optional): factor of moderation of the observation of the next state, 0 is a shortsighted condition and 1 it exhibits farsighted behavior

• nb_iter (int, optional): number of sequences to go through the graph

• starting_state ("empty" or "random", optional) : if "empty" the algorithm starts from the empty state and if "random" the algorithm starts from a random state in the graph

The output of the selection process is a 5-tuple object.

• Index of the features that have been sorted

• Number of times that each feature has been chosen

• Mean reward brought by each feature

• Ranking of the features from the less important to the most important

• Number of states visited

Existing methods

• Compare the performance of the FSRLearning library with RFE from Sickit-Learn :

fsrl_obj.compare_with_benchmark(X, y, results)

Returns some comparisons and plot a graph with the metric for each set of features selected. It is useful for parameters tuning.

• Get the evolution of the number of the visited states for the first time and the already visited states :

fsrl_obj.get_plot_ratio_exploration()

Returns a plot. It is useful to get an overview of how the graph is browse and to tune the epsilon parameter (exploration parameter).

• Get an overview of the relative impact of each feature on the model :

fsrl_obj.get_feature_strengh(results)

Returns a bar plot.

• Get an overview of the action of the stop conditions :

fsrl_obj.get_depth_of_visited_states()

Returns a plot. It is useful to see how deep the Markovian Decision Process goes in the graph.

Your contribution is welcomed !

• Automatise the data processing step and generalize the input data format and type
• Distribute the computation of each reward for making the algorithm faster
• Add more vizualization and feedback methods

References

This library has been implemented with the help of these two articles :

• Sali Rasoul, Sodiq Adewole and Alphonse Akakpo, FEATURE SELECTION USING REINFORCEMENT LEARNING (2021)
• Seyed Mehdin Hazrati Fard, Ali Hamzeh and Sattar Hashemi, USING REINFORCEMENT LEARNING TO FIND AN OPTIMAL SET OF FEATURES (2013)