Patterns Count-Based Labels for Datasets

TL;DR

This paper introduces a method for creating compact labels that estimate counts of attribute-value combinations in datasets, aiding in profiling and bias detection without storing all counts explicitly.

Contribution

It proposes a novel label-based approach and heuristic algorithm for efficiently estimating pattern counts in datasets with limited label size.

Findings

The label-based estimation achieves high accuracy in count approximation.

The heuristic algorithm efficiently generates near-optimal labels.

Experimental results demonstrate the method's practicality and effectiveness.

Abstract

Counts of attribute-value combinations are central to the profiling of a dataset, particularly in determining fitness for use and in eliminating bias and unfairness. While counts of individual attribute values may be stored in some dataset profiles, there are too many combinations of attributes for it to be practical to store counts for each combination. In this paper, we develop the notion of storing a "label" of limited size that can be used to obtain good estimates for these counts. A label, in this paper, contains information regarding the count of selected patterns--attributes values combinations--in the data. We define an estimation function, that uses this label to estimate the count of every pattern. We present the problem of finding the optimal label given a bound on its size and propose a heuristic algorithm for generating optimal labels. We experimentally show the accuracy of count estimates derived from the resulting labels and the efficiency of our algorithm.