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.
Install using pip with:
pip install divexplorer
or, download a wheel or source archive from PyPI.
At the notebook running example - COMPAS dataset you can find a running example of the usage of DivExplorer
For the analysis of the divergent classification behavior in subgroups:
from divexplorer.FP_DivergenceExplorer import FP_DivergenceExplorer
from divexplorer.FP_Divergence import FP_Divergence
# Input: a discretized dataframe with the true class and the predicted class.
# We specify their column names in the dataframe
# The class_map is a dictionary to specify the positive and the negative class (e.g. {"P":1, "N":0})
fp_diver=FP_DivergenceExplorer(df_discretized, true_class_name = "class", predicted_class_name = "predicted", class_map=class_map)
#Extract frequent patterns (FP) and compute divergence
##min_support: minimum support threshold
##metrics: metrics=["d_fpr", "d_fnr"]
# (default metric of interest: False Positive Rate (FPR) d_fpr, False Negative Rate (FNR) d_fnr, Accuracy divergence)
min_sup=0.1
FP_fm=fp_diver.getFrequentPatternDivergence(min_support=min_sup, metrics=["d_fpr", "d_fnr"])The output is a pandas dataframe. Each row indicate a FP with its classification performance and its divergence.
We can then analyze the divergence of FP with respect a metric of interest (e.g. FPR).
fp_divergence_fpr=FP_Divergence(FP_fm, "d_fpr")We can sort the itemset for FPR divergence (and visualize the K=10 most divergent ones)
K=10
FP_fpr_sorted=fp_divergence_fpr.getDivergence(th_redundancy=0).head(K)or directly get the top K divergent patterns
#As a dictionary, where the key are FP and values are their divergence values
topK_dict_fpr=fp_divergence_fpr.getDivergenceTopK(K=10, th_redundancy=0)
#Or as a DataFrame
topK_df_fpr=fp_divergence_fpr.getDivergenceTopKDf(K=10, th_redundancy=0)We can estimate the contribution of each item to the divergence using the notion of Shapley value
#Let be itemset_i a FP of interest, for example the one with highest FP_Divergence
itemset_i=list(topK_dict_fpr.keys())[0]
itemset_shap=fp_divergence_fpr.computeShapleyValue(itemset_i)
#Plot shapley values
fp_divergence_fpr.plotShapleyValue(shapley_values=itemset_shap)
#Alternatively, we can just use as input the itemset itself
fp_divergence_fpr.plotShapleyValue(itemset=itemset_i)The itemset can also be inspected using the lattice graph
#Plot the lattice graph
#Th_divergence: if specified, itemsets of the lattice with divergence greater than specified value are highlighted in magenta/squares
##getLower: if True, corrective patterns are highlighted in light blue/diamonds
fig=fp_divergence_fpr.plotLatticeItemset(itemset_i, Th_divergence=0.15, sizeDot="small", getLower=True)
fig.show()DivExplorer allows to identify peculiar behaviors as corrective phenomena. Corrective items are items that, when added to an itemset, reduce the divergence.
fp_divergence_fpr.getCorrectiveItems()We can then analyze the influence of each item on the divergence of the entire dataset.
We can do with:
- the individual divergence of an item. It is simply the divergence of an item in isolation.
- the global divergence of an item. It a generalization of the Shapley value to the entire set of all items. It captures the role of an item in giving rise to divergence jointly with other attributes. >>> #Compute global shapley value
# Individual divergence
individual_divergence_fpr=fp_divergence_fpr.getFItemsetsDivergence()[1]
fp_divergence_fpr.plotShapleyValue(shapley_values=individual_divergence_fpr,sizeFig=(4,5))
# Global divergence
global_item_divergence_fpr=fp_divergence_fpr.computeGlobalShapleyValue()
fp_divergence_fpr.plotShapleyValue(shapley_values=global_item_divergence_fpr)Note that the evaluation can be perform the analysis of the divergence of a single class of interest:
min_sup=0.1
fp_diver_1class=FP_DivergenceExplorer(X_discretized.drop(columns="predicted"),true_class_name="class", class_map=class_map)
# For example, we analyze the positive rate divergenxe
FP_fm_1class=fp_diver_1class.getFrequentPatternDivergence(min_support=min_sup, metrics=["d_posr", "d_negr"])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.