Résultats de recherche pour « Synthetic data »

ObjectivesResearch Data Scotland (RDS) is working to improve the economic, social and environmental wellbeing in Scotland by enabling access to, and linkage of, public-sector data for research in the public good. Our objective was to explore how synthetic data can be used to support this aim. MethodsWe investigated what other, similar, data organisations were doing around synthetic data, both in Scotland and beyond, and the issues to consider. We discussed use cases, tools, level of fidelity of synthesis, disclosure risk, information governance (IG) requirements, where the synthetic data might sit, who could access it, training and accreditation requirements, and barriers to synthesis. The findings were used to draft a synthetic data strategy, create a working group and plan future work. We also held a user workshop to discuss researcher requirements and asked our public panel about their understanding and concerns around synthetic data. ResultsThree workstreams have been set up around disclosure risk & IG, synthesis, and access, promotion & engagement. A test synthesis of education data is planned. From our user workshop one of the main themes that emerged was the usefulness of low fidelity, unlinked datasets for data discovery purposes at the early stages of a project. Other areas where synthetic data can improve data access are for training, where synthetic data can be used to upskill users in the use of administrative data, and in code development, to allow earlier access to data for code generation and to reduce the time needed in secure settings and with access to real data. Future work will bring together data controllers to discuss disclosure risk. ConclusionSynthetic data can potentially be used to improve and speed up data access. RDS will fund and support synthetic data projects to demonstrate value, and work with data controllers to produce synthetic datasets on an ongoing basis, for public sector research datasets in Scotland.

Synthetic data, which is artificially generated and intelligently mimicking or supplementing the real-world data, is increasingly used. The proliferation of AI agents and the adoption of synthetic data create a synthetic mirror that conceptualizes a representation and potential distortion of reality, thus generating trust and accountability deficits. This paper explores the implications for privacy and policymaking stemming from synthetic data generation, and the urgent need for new policy instruments and legal framework adaptation to ensure appropriate levels of trust and accountability for AI agents relying on synthetic data. Rather than creating entirely new policy or legal regimes, the most practical approach involves targeted amendments to existing frameworks, recognizing synthetic data as a distinct regulatory category with unique characteristics.

Abstract Healthcare research increasingly relies on machine learning to guide clinical decisions, yet access to patient data remains restricted due to privacy regulations and institutional policies. Synthetic data generation offers a promising solution, but current approaches face a critical challenge: raw synthetic data from generative models exhibit substantial fidelity and utility gaps compared to real patient data and lack adequate privacy protections, limiting their practical value for clinical applications. We present MAPS (Model-Agnostic Post-hoc Synthetic Data Refinement Framework), a post-hoc approach that enhances synthetic healthcare data quality while maintaining privacy protections. MAPS employs a two-stage refinement process: first, removing synthetic samples that violate privacy through 0-identifiability filtering; second, applying importance weighting via density ratio estimation to resample synthetic data toward the true patient data distribution. We evaluate MAPS across two real-world healthcare datasets—cancer patient records (TCGA-metadata) and cardiovascular ICU data (GOSSIS-1-eICU-cardiovascular)—using four state-of-the-art generative models. MAPS demonstrates substantial improvements in distributional fidelity, with Joint Jensen-Shannon Distance reductions of up to 50%. For privacy protection, MAPS eliminates direct re-identification risk by reducing distance-based identifiability scores from approximately 34% to 0% across all model-dataset combinations. MAPS significantly enhances utility across multiple dimensions. Classification performance improves substantially, with F1 scores increasing from 0.09-0.24 to 0.30-0.38 on cancer type prediction and from 0.13-0.21 to 0.21-0.25 on cardiovascular diagnosis prediction. Notably, MAPS improves uncertainty quantification, an overlooked but essential utility aspect for clinical decision-making. When evaluated using split conformal prediction, MAPS reduces prediction set sizes by 34-77% while maintaining target coverage, enabling more precise and actionable clinical predictions. These improvements demonstrate that post-hoc refinement can simultaneously achieve strong privacy protection, high distributional fidelity, and robust downstream utility, with reliable uncertainty quantification essential for clinical deployment.

The paper is devoted to the problem of generating artificial datasets for data envelopment analysis (DEA), which can be used for testing DEA models and methods. In particular, the papers that applied DEA to big data often used synthetic data generation to obtain large-scale datasets because real datasets of large size, available in the public domain, are extremely rare. This paper proposes the algorithm which takes as input some real dataset and complements it by artificial efficient and inefficient units. The generation process extends the efficient part of the frontier by inserting artificial efficient units, keeping the original efficient frontier unchanged. For this purpose, the algorithm uses the assurance region method and consistently relaxes weight restrictions during the iterations. This approach produces synthetic datasets that are closer to real ones, compared to other algorithms that generate data from scratch. The proposed algorithm is applied to a pair of small real-life datasets. As a result, the datasets were expanded to 50K units. Computational experiments show that artificially generated DMUs preserve isotonicity and do not increase the collinearity of the original data as a whole.