Assessing and Remedying Coverage for a Given Dataset

Abstract

Data analysis impacts virtually every aspect of our society today. Often, this analysis is performed on an existing dataset, possibly collected through a process that the data scientists had limited control over. The existing data analyzed may not include the complete universe, but it is expected to cover the diversity of items in the universe. Lack of adequate coverage in the dataset can result in undesirable outcomes such as biased decisions and algorithmic racism, as well as creating vulnerabilities such as opening up room for adversarial attacks.In this paper, we assess the coverage of a given dataset over multiple categorical attributes. We first provide efficient techniques for traversing the combinatorial explosion of value combinations to identify any regions of attribute space not adequately covered by the data. Then, we determine the least amount of additional data that must be obtained to resolve this lack of adequate coverage. We confirm the value of our proposal through both theoretical analyses and comprehensive experiments on real data. c Ai = Π ∀Aj ∈Ai c j to show the number of value combinations for A i .Definition 1 (Pattern). A pattern P is a vector of size d, in which P [i] is either X (meaning that its value is unspecified) or is a value of attribute A i . We name the elements with value X as non-deterministic and the others as deterministic.An item t matches a pattern P (written as M (t, P ) = ), if for all i for which P [i] is deterministic, t[i] is equal to P [i]. Formally:

Managing bias and unfairness in data for decision support: a survey of machine learning and data engineering approaches to identify and mitigate bias and unfairness within data management and analytics systems

“…Asudeh et al [12] first proposed a formalisation of the coverage problem. They also present and evaluate methods both to efficiently evaluate the coverage of a dataset with respect to thresholds set by a practitioner for each dataset attribute, and to identify the type of data samples that are preferable to collect to solve the coverage issue accounting for the cost of data collection.…”

“…Another topic related to the identification of biases within datasets more specific to data management literature is the notion of data coverage. Coverage relates to the idea that data samples in a dataset should sufficiently cover the diversity of items in a universe of discourse [12]. Without adequate coverage, applications using such datasets might be prone to discriminative mistakes.…”

“…After choosing a fairness definition, deciding how to transform age into a categorical attribute can have direct bias consequences. Defining protected classes (male, black, [10][11][12][13][14][15][16][17][18][19][20][21][22][23]) or (male, black, [10][11][12][13][14][15][16][17][18][19][20][21][22][23][24][25]) as protected attributes would both surface and measure different biases. Different mappings of age to its protected class "young" can create different system behaviours: the granularity of the categories chosen would influence both the performance and fairness of the trained inference model.…”

Fairness in rankings and recommendations: an overview

“…A major challenge is that the available data are often limited [7]. This way, any analysis is done with data that has been acquired independently, through a process on which the data scientist has limited control.…”

Investigating and Mitigating Biases in Crowdsourced Data

“…After the data collection process, bias can also be reduced using many other approaches such as data aggregation techniques [19], systematically adding new data points to fix coverage [2], and crowdsourced bias detection [17]. Other approaches to mitigate bias can be employed during feature engineering and model training.…”