DivExplorer

Machine learning models may perform differently on different data subgroups. We propose the notion of divergence over itemsets (i.e., conjunctions of simple predicates) as a measure of different classification behavior on data subgroups, and the use of frequent pattern mining techniques for their identification. We quantify the contribution of different attribute values to divergence with the notion of Shapley values to identify both critical and peculiar behaviors of attributes. See our paper and our project page for all the details.

Installation

Install using pip with:

pip install divexplorer

or, download a wheel or source archive from PyPI.

Example Notebooks

This notebook gives an example of how to use DivExplorer to find divergent subgroups in datasets and in the predictions of a classifier.

Documentation

For the code details, see the documentation.

The original paper is:

Looking for Trouble: Analyzing Classifier Behavior via Pattern Divergence. Eliana Pastor, Luca de Alfaro, Elena Baralis. In Proceedings of the 2021 ACM SIGMOD Conference, 2021.

You can find more papers and information in the DivExplorer project page.

Quick Start

DivExplorer works on Pandas datasets. Here we load an example one, and discretize in coarser ranges one of its attributes.

import pandas as pd

df_census = pd.read_csv('https://raw.githubusercontent.com/divexplorer/divexplorer/main/datasets/census_income.csv')
df_census["AGE_RANGE"] = df_census.apply(lambda row : 10 * (row["A_AGE"] // 10), axis=1)

We can then find the data subgroups that have highest income divergence, using the DivergenceExplorer class as follows:

from divexplorer import DivergenceExplorer

fp_diver = DivergenceExplorer(df_census)
subgroups = fp_diver.get_pattern_divergence(
    min_support=0.001,
    attributes=["STATE", "SEX", "EDUCATION", "AGE_RANGE"], 
    quantitative_outcomes=["PTOTVAL"])
subgroups.sort_values(by="PTOTVAL_div", ascending=False).head(10)

You can also prune redundant subgroups by specifying:

• a threshold, so that attributes that don't increase the divergence by at least the threshold value are not included in subgroups,
• a minimum t-value, to select only significant subgroups.

from divexplorer import DivergencePatternProcessor

processor = DivergencePatternProcessor(subgroups, "PTOTVAL")
pruned_subgroups = pd.DataFrame(processor.redundancy_pruning(th_redundancy=10000))
pruned_subgroups = pruned_subgroups[pruned_subgroups["PTOTVAL_t"] > 2]
pruned_subgroups.sort_values(by="PTOTVAL_div", ascending=False, ignore_index=True)

Finding subgroups with divergent performance in classifiers

For classifiers, it may be of interest to find the subgroups with the highest (or lowest) divergence in characteristics such as false positive rates, etc. Here is how to do it for the false-positive rate in a COMPAS-derived classifier.

compas_df = pd.read_csv('https://raw.githubusercontent.com/divexplorer/divexplorer/main/datasets/compas_discretized.csv')

We generate an fp column whose average will give the false-positive rate, like so:

from divexplorer.outcomes import get_false_positive_rate_outcome

y_trues = compas_df["class"]
y_preds = compas_df["predicted"]

compas_df['fp'] =  get_false_positive_rate_outcome(y_trues, y_preds)

The fp column has values:

• 1, if the data is a false positive (class is 0 and predicted is 1)
• 0, if the data is a true negative (class is 0 and predicted is 0).
• NaN, if the class is positive (class is 1).

We use Nan for class 1 data, to exclude those data from the average, so that the column average is the false-positive rate. We can then find the most divergent groups as in the previous example, noting that here we use boolean_outcomes rather than quantitative_outcomes because fp is boolean:

fp_diver = DivergenceExplorer(compas_df)

attributes = ['race', '#prior', 'sex', 'age']
FP_fm = fp_diver.get_pattern_divergence(min_support=0.1, attributes=attributes, 
                                        boolean_outcomes=['fp'])
FP_fm.sort_values(by="fp_div", ascending=False).head(10)

Note how we specify the attributes that can be used to define subgroups. In the above code, we use boolean_outcomes because fp is boolean. The following example, from the example notebook, shows how to use quantitative_outcomes for a quantitative outcome.

df_census = pd.read_csv('https://raw.githubusercontent.com/divexplorer/divexplorer/main/datasets/census_income.csv')
explorer = DivergenceExplorer(df_census)
value_subgroups = explorer.get_pattern_divergence(
    min_support=0.001, quantitative_outcomes=["PTOTVAL"])

Analyzing subgroups via Shapley values

Returning to our COMPAS example, if we want to analyze what factors contribute to the divergence of a particular subgroup, we can do so via Shapley values:

fp_details = DivergencePatternProcessor(FP_fm, 'fp')

pattern = fp_details.patterns['itemset'].iloc[37]
fp_details.shapley_value(pattern)

Pruning redundant subgroups

If you get too many subgroups, you can prune redundant ones via redundancy pruning. This prunes a pattern $\beta$ if there is a pattern $\alpha$, subset of $\beta$, with a divergence difference below a threshold.

df_pruned = fp_details.redundancy_pruning(th_redundancy=0.01)
df_pruned.sort_values("fp_div", ascending=False).head(5)

Code Contributors

Project lead:

• Eliana Pastor

Other contributors:

• Luca de Alfaro
• Harsh Dadhich