Databias

Databias is a term used to describe systematic distortions in data that affect analyses, predictions, and decisions derived from that data. It can originate from how data are collected, labeled, stored, or processed, and often reflects real-world inequities or measurement errors. Databias is not only a property of the data but of the entire data pipeline, including sampling methods, recording practices, and preprocessing steps. It can operate alone or amplify existing model biases when used to train or validate algorithms.

Common sources include sampling bias (unrepresentative samples), measurement or labeling bias (inconsistent or subjective labels), historical bias (data reflecting past inequalities), aggregation or feature selection bias, survivorship bias, and recording bias (differences in data capture across groups).

Detection and evaluation involve data audits, exploratory analysis, and fairness metrics; the use of datasets with diverse representations; and documentation such as datasheets for datasets and data provenance. Mitigation strategies include improving data collection to be representative, balancing datasets through resampling or reweighting, removing or adjusting biased features, employing debiasing or fair representation learning techniques, and incorporating monitoring and human oversight during deployment.

Impact and implications include the potential for unfair or inaccurate predictions in hiring, lending, criminal justice, healthcare, and content moderation, as well as the risk of misinforming research and policy decisions. Addressing databias requires explicit goals, stakeholder involvement, and ongoing evaluation.

Governance and standards emphasize risk management, transparency, and auditing. Practices such as dataset documentation, model cards, differential privacy, and responsible data collection are commonly recommended, with continuous updating as data distributions evolve.