Dataset Information

Eliminating unintended bias in personalized policies using bias-eliminating adapted trees (BEAT).

ABSTRACT: SignificanceDecision makers now use algorithmic personalization for resource allocation decisions in many domains (e.g., medical treatments, hiring decisions, product recommendations, or dynamic pricing). An inherent risk of personalization is disproportionate targeting of individuals from certain protected groups. Existing solutions that firms use to avoid this bias often do not eliminate the bias and may even exacerbate it. We propose BEAT (bias-eliminating adapted trees) to ensure balanced allocation of resources across individuals-guaranteeing both group and individual fairness-while still leveraging the value of personalization. We validate our method using simulations as well as an online experiment with N = 3,146 participants. BEAT is easy to implement in practice, has desirable scalability properties, and is applicable to many personalization problems.

SUBMITTER: Ascarza E

PROVIDER: S-EPMC8931211 | biostudies-literature | 2022 Mar

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Eliminating unintended bias in personalized policies using bias-eliminating adapted trees (BEAT).

Ascarza Eva E Israeli Ayelet A

Proceedings of the National Academy of Sciences of the United States of America 20220308 11

SignificanceDecision makers now use algorithmic personalization for resource allocation decisions in many domains (e.g., medical treatments, hiring decisions, product recommendations, or dynamic pricing). An inherent risk of personalization is disproportionate targeting of individuals from certain protected groups. Existing solutions that firms use to avoid this bias often do not eliminate the bias and may even exacerbate it. We propose BEAT (bias-eliminating adapted trees) to ensure balanced al ...[more]

PMID: 35259009

Similar Datasets

Project description:Controlling the spread of COVID-19 - even after a licensed vaccine is available - requires the effective use of non-pharmaceutical interventions, e.g., physical distancing, limits on group sizes, mask wearing, etc1-7. To date, such interventions have not been uniformly and/or systematically implemented across the United States of America (US)8. For example, even when under strict stay-at-home orders, numerous jurisdictions in the US granted exceptions and/or were in close proximity to locations with entirely different regulations in place. Here, we investigate the impact of such geographic inconsistencies in epidemic control policies by coupling high-resolution mobility, search, and COVID case data to a mathematical model of SARS-CoV-2 transmission. Our results show that while stay-at-home orders decrease contacts in most areas of the US, some specific activities and venues often see an increase in attendance. As an example, over the month of March 2020, between 10 and 30% of churches in the US saw increases in attendance; even as the total number of visits to churches declined nationally. This heterogeneity, where certain venues see substantial increases in attendance while others close, suggests that closure can cause individuals to find an open venue, even if that requires longer-distance travel. And, indeed, the average distance travelled to churches in the US rose by 13% over the same period, and over the summer, churches with more than 50 average weekly visitors saw an increase of 81% in distance visitors had to travel to attend. Strikingly, our mathematical model reveals that, across a broad range of model parameters, partial measures can often be worse than no measures at all. In the most severe cases, individuals not complying with policies by traveling to neighboring jurisdictions can create epidemics when the outbreak would otherwise have been contained. Indeed, using county-level COVID-19 data, we show that mobility from high-incidence to low-incidence associated with travel for venues like churches, parks, and gyms consistently precedes rising case numbers in the low-incidence counties. Taken together, our data analysis of nearly 120 million church visitors across 184,677 churches, 14 million grocery visitors across 7,662 grocery stores, 13.5 million gym visitors across 5,483 gyms, 7.7 million cases across 3,195 counties, and modeling results highlight the potential unintended consequences of inconsistent epidemic control policies and stress the importance of balancing the societal needs of a population with the risk of an outbreak growing into a large epidemic, and the urgent need for centralized implementation and enforcement of non-pharmaceutical interventions.

Dataset Information

Eliminating unintended bias in personalized policies using bias-eliminating adapted trees (BEAT).

Publications

Eliminating unintended bias in personalized policies using bias-eliminating adapted trees (BEAT).

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets