Usage

Installation

To use IREX, first install it using pip:

pip install -i https://test.pypi.org/simple/ IREX

or

python -m pip install -i https://test.pypi.org/simple/ IREX

Importation

We recomend to import IREX like this:

from IREX import IREX

Example

First, create an IREX object named IREX, like this:

IREX = IREX()

With this, you can easily see how IREX’s functions works.

Then, we highly recommend to use IREX with the following sequence of functions:

IREX.load_IREX_datasets()

IREX.reset() ##For first iteration or
IREX.iterate() ##For second or greater iteration

IREX.train_model(do_oversample = True, optimize = False, saveModel = True)

IREX.evaluate_model()

IREX.run_ALE()

IREX.apply_Threshold(0.01, -0.01)

IREX.search_PAI("POSITIVE_CLASS")

IREX.run_LIME()

IREX.run_SHAP()

IREX.precompute_Heatmaps()

IREX.run_Feature_Importance_Heatmap()

IREX.run_SHAP_Heatmap()

IREX.run_LIME_Heatmap()

IREX.run_ALE_Heatmap()

IREX.run_Compare_Heatmaps()

IREX.plot_global_process(['gray', 'black', 'red'])

List of functions

Here are all the functions currently available in the latest IREX version.

load_IREX_datasets

Loads examples of dataset ready to work with IREX.

set_source_dataset

-Input parameters:

-main_dataset: the path or name of .csv file which may contain all the questions as columns and all the answers as rows with a final column named “Target” which may contain the final scores of the people who answered the questionnaire.

-target_classes: the number and name of the class to which it corresponds. For example:

{0:"Low",1:"Medium",2:"High"}

set_expected_answers

-Input parameters:

-source: the path or name of .csv file which may contain the expected answers of the questionnaire.

-input_classes: the number and name of the classification which it refers to. For example:

{0:"Low",1:"High"}

reset

Resets the iterative process to start all over again, assigns the local variables for the neural network and also generates the question status and global process dataset.

iterate

Prepares IREX’s local variables for the next iteration.

train_model

Train the prediction model, taking into account these 3 modifiable options:

-do_oversample: True or False parameter wich will apply (or not) the SMOTE oversample technique to the samples by calling another function.

-optimize: True or False parameter wich will optimize (or not) the classification model by calling another function.

-saveModel: True or False parameter wich will save (or not) the classification model by calling another function.

evaluate_model

Generates the confusion matrix of the classification model and prints others metrics values such as accuracy, precision, recal, f1-score, support, etc.

run_ALE

Generates ALE graphs for each of the current questions, this contains a slope per class, which represents the way in which the question influences the final ranking. This also creates a dataset wich save the slope values for a future usage.

apply_Threshold

Use the positive and negative threshold given by the user to identify which items are under these values, tagging them in the dataset which contains the slope values of each question.

-Input parameters:

-positive: assigns the positive threshold to use.

-negative: assigns the negative threshold to use.

search_PAI

Identifies and tags those questions, that match with the selected mode by the user, in the slopes dataset.

-Available modes:

-POSITIVE CLASS: searches for potentially anomalous items (questions) that may lead to a miss classification on the positive class defined by the user.

-NEGATIVE_CLASS: searches for potentially anomalous items (questions) that may lead to a miss classification on the negative class defined by the user.

-ANY_CLASS: searches for potentially anomalous items (questions) that may lead to a miss classification in one class OR another.

-BOTH_CLASSES: searches for potentially anomalous items (questions) that may lead to a miss classification in both classes.

-NO_RELEVANT: searches for potentially anomalous items (questions) that may be between the positive and negative threshold defined by the user.

refine_dataset

Removes from the local dataset the PAIs (Potentially Anomalous Items) defined by the previous function.

run_LIME

Displays the graphics generated by the LIME XAI method.

run_SHAP

Displays the graphics generated by the SHAP XAI method.

precompute_Heatmaps

Prepares the data for the heatmaps generation.

run_Feature_Importance_Heatmap

Displays the heatmaps generated by the Feature Importance XAI method, one per each class and a final graphic whit all heatmaps combined. PAIs are highlighted in red.

run_SHAP_Heatmap

Displays the heatmaps generated by the SHAP XAI method, one per each class and a final graphic whit all heatmaps combined. PAIs are highlighted in red.

run_LIME_Heatmap

Displays the heatmaps generated by the LIME XAI method, one per each class and a final graphic whit all heatmaps combined. PAIs are highlighted in red.

run_ALE_Heatmap

Displays the heatmaps generated by the ALE XAI method, one per each class and a final graphic whit all heatmaps combined. PAIs are highlighted in red.

run_Compare_Heatmaps

Displays all the heatmaps generated in the previous functions by the three XAI methods, separated by classes. For this you will need to run the previous three steps.

plot_global_process

Displays the different graphics generated to represent the iterative process done. These are: items used per iteration, accuracy per iteration, PAIs detected per iteration, accuracy and PAIs per iteration, recall obtained and anomalous questions detected.

-Input parameters:

-plot_colors: list of colors to use to represent the classes in the accuracy and PAIs per iteration and recall obtained and anomalous questions detected plot. For example:

['gray', 'black', 'red']