Non-Conformity in Aquaculture Social Audits - A Study on Labour Rights using Predictive Models
Author
Moustsen, Theis
Term
4. term
Education
Publication year
2025
Abstract
Dette speciale undersøger sandsynligheden for overholdelse af arbejdsmæssige rettigheder på akvakulturfarme ved hjælp af auditdata fra Aquaculture Stewardship Council (ASC) og prædiktive modeller. Et datasæt af ASC-audits danner grundlag for ti XGBoost-modeller, der for hver af ILO’s fem grundlæggende rettigheder forudsiger, om en audit vil registrere en ikke-overensstemmelse. Med SHAP-analyse vurderes, hvilke variable der driver forudsigelserne, og landet, hvor farmen ligger, fremstår som den afgørende faktor. Seks lande med flest audits (Ecuador, Chile, Vietnam, Indien, Norge og Storbritannien) analyseres nærmere: På tværs af modellerne indikerer auditdata oftest ikke-overensstemmelse i Storbritannien, derefter Indien og Ecuador; undtagelsen er modellerne for foreningsfrihed og kollektiv forhandling, hvor alle lande indikerer overensstemmelse. Norge er generelt forbundet med overensstemmelse, Chile er aldrig indikator for ikke-overensstemmelse, og Vietnam er indikator for ikke-overensstemmelse i modellerne om tvangsarbejde. Sammenligninger med eksisterende forskning viser ingen klar overensstemmelse og i nogle tilfælde direkte modstrid, især for foreningsfrihed. Mulige forklaringer er, at ASC-certificerede farme repræsenterer bedste fald, eller at audits har begrænset evne til at finde krænkelser. Et sociologisk rammeværk med Standings prekariat og Grimshaw et al.’s protective gaps anvendes til at fortolke resultaterne; rammerne forklarer delvist mønstrene og flugter mere med den øvrige forskning end med modellerne. Modellerne har lav præcision og recall, hvilket peger på behov for mere data og yderligere forskning i arbejdsvilkår i akvakultur samt auditters evne til at afdække krænkelser.
This thesis examines the likelihood that aquaculture farms comply with labour rights standards using Aquaculture Stewardship Council (ASC) audit data and predictive models. A dataset of ASC audits underpins ten XGBoost models that, for each of the ILO’s five fundamental rights, predict whether an audit will record a non-conformity. SHAP analysis is used to understand which variables drive predictions, with the country of operation emerging as the dominant factor. The six most-audited countries (Ecuador, Chile, Vietnam, India, Norway, and the UK) are examined in detail: Across models, the UK most strongly indicates non-conformity, followed by India and Ecuador; the exception is the freedom of association and collective bargaining models, where all countries indicate compliance. Norway is generally associated with compliance, Chile is never an indicator of non-conformity, and Vietnam is an indicator of non-conformity in the forced labour models. Comparing the models’ indications with existing research shows no clear alignment and, in some cases, contradictions, especially regarding freedom of association. Possible explanations include that ASC-certified farms represent best-case sites or that audits struggle to detect violations. A sociological framework drawing on Standing’s precariat and Grimshaw et al.’s protective gaps is used to interpret the findings; it partially explains the patterns and aligns more with other research than with the models. The models have low precision and recall, underscoring the need for more data and further research on labour rights in aquaculture and on how effectively audits detect abuses.
[This summary has been generated with the help of AI directly from the project (PDF)]
Documents