IML reliability: empirical study on the reliability of interpretations of interpretable machine learning methods

imlreliability provides a unified framework for the reliability test based on repeated data perturbation, as specified in paper... Users are free to implement their own IML methodology, data set or consistency metrics. The output can be directly uploaed and compare to our reliability experiments in the dashboard.

Motivation

IML Tasks

• Feature importance
  • Regression
  • Classification
• Clustering
• Dimension Reduction

Reliability Test

Criteria

imlreliability Quickstart

Installation

IML_reliability 0.1 and later require Python 3.7 or Python 3.8.

Clone this repo and run

python setup.py install

The dependencies of this package will be automatically installed into your environment.

Usage

Detailed working examples and tutorials for each task can be found in tutorial folder of the repository. Here we give a simple illustration on the usage of the package. To measure interpretation reliability for any of the IML tasks, imlreliability consists of two steps:

1. Obtain a set of intepretation by repeatly applying the IML algorithm to perturbed data sets. This step is acheived by initializing the model model with:
   • feature_importance.feature_impoClass() for feature importance (classification)
   • feature_importance.feature_impoReg() for feature importance (regression)
   • clustering.clustering() for clustering
   • dimension_reduction.dimension_reduction() for dimension reduction

and perform the model fitting via model.fit().

    # Pseudo-code 
    ## Load example data 
    from sklearn.preprocessing import scale, normalize
    communities_data = pd.read_csv('http://archive.ics.uci.edu/ml/machine-learning-databases/communities/communities.data').to_numpy()
    communities_data = np.delete(communities_data, np.arange(5), 1)
# remove predictors with missing values
    communities_data = np.delete(communities_data,
                         np.argwhere((communities_data == '?').sum(0) > 0).reshape(-1), 1)
    communities_data = communities_data.astype(float)
    x = communities_data[:, :-1]
    y = communities_data[:, -1]


    ### scale and normalize data 
    x = normalize(scale(x))
    y = (scale(y))
    ## The input data includes both features and response: data: (X, Y), X: array of shape (N, M), Y: array of shape (N,)
    data_reg=(x,y)

    ## Here we aim to evaluate the interpretation reliability of Ridge regression with cross validation, using the ``feature_impoReg``function. We use ``RidgeCV()`` from ``sklearn`` as our estimator. By setting ``importance_func=None``, the magnitude of coefficients will be used to evaluate feature importance. 

    from sklearn.linear_model import RidgeCV
    estimator=RidgeCV()
    importance_func=None

    ## We initialize the model with the ``mlreliability.feature_importance.feature_impoReg`` function.
    import imlreliability
    model_reg = imlreliability.feature_importance.feature_impoReg(data,estimator, importance_func) # < construct the model
    model_reg.fit() # < train the model

Details of arguments are listed below.

Parameter name	Type	Task	Description
data	numpy array, required	All	Input samples (and response) used to be fed to data for reliability test.
estimator	object, required	All	scikit-learn object of the IML model to run.Either estimator needs to provide a score function or scoring must be passed.
n_repeat	int, optional	All	Number of repeats in data perturbation, default as 100.
split_proportion	float in (0,1), optional	All	The proportion of training set in data splitting, default as 0.7.
importance_func	str, callable, list, tuple or dict, optional	Feature importance	Strategy to evaluate feature importance score.
K_max	int, optional	Feature importance	Number of top features of interest, default as 30.
evaluate_fun	str, callable, list, tuple or dict	Feature importance	Function to evaluate prediction performance. Default as mean_squared_error in regression,
get_prediction_consistency	{True,False}, optional	Feature importance	Controls whether to calculate prediction consistency, default as True.
K	int, required	Clustering	Number of clusters.
label	array of shape (N,1), optional	Clustering	True cluster labels, default as None.
rank	int, optional	Dimension reduction	The number of reduced dimensions, default as 2.
perturbation	{'noise','split'}, optional	Clustering, dimension reduction	Controls the way of perturbation, default as 'noise'.
sigma	float,optinal	Clustering, dimension reduction	Controls level of noise. need to specify if perturbation=='noise', default as 1.
stratify	{True,False}, optional	All	Controls whether to conduct stratified sampling, default as True.
norm	{True,False}, optional	All	Controls whether to conduct data normalization, default as True.
rand_index	RandomState instance, optional	All	Make sure the sampling uses the same RandomState instance for all iterations.
verbose	{True,False}, optional	All	Controls the verbosity, default as True.

2. Calculate interpretation consistency of the obtained interpretations via given reliability metrics. The consistency results is achieved via

   • model.get_consistency() for feature importance

   • model.get_consistency_clustering() for clustering

   • model.get_consistency_knn() for neiborhood consistency in dimension reduction

   • model.get_consistency_clustering() for clustering consistency in dimension reduction

The .get_consistency function results in three pandas dataframe: accuracy: prediction accuracy on test set; consistency: interpretation consistency measured by RBO, Jaccard score, or user-defined metrics; and prediction_consistency measured by prediction entropy and purity if get_prediction_consistency ==True.

# Pseudo-code 
model_reg.get_consistency(data_name='communities', estimator_name='Ridge',impotance_func_name='Coef')
print(model_reg.accuracy)
print(model_reg.consistency)
print(model_reg.prediction_consistency)

The consistency pandas dataframe can be downloaded and upload to the dashboard.

Details of arguments are listed below.

Parameter name	Type	Task	Description
data_name	str, required	All	Name of the data set.
method_name	str, required	All	Name of IML method.
impotance_func_name	str, optional	Feature importance	Name of funtion to measure feature importance, default as None.
Kranges	callable, optional	Dimension reduction (KNN)	list of numbers of nearest neighbors to measure, default as None.
cluster_func	callable, optional	Dimension reduction (clustering)	clustering object applied to the reduced dimension, default as AgglomerativeClustering().
cluster_func_name	str, optional	Dimension reduction (clustering)	Name of clustering function, default as 'HC (ward)'.
user_metric	callable, optional	All	User-defined evaluation metric of consistency, default as None.
user_metric_name	callable, optional	All	User-defined evaluation metric of consistency, default as 'user_metric'.

API Documentations

The detailed API documentation for this package can be found at [https://DataSlingers.github.io/IML_reliability]