Requirement
"The organization shall define and document its criteria for selecting data preparations and the data preparation methods to be used."[1]
Explanation
Raw data is rarely usable as-is: cleaning, imputation, normalization, labelling and encoding all change what a model learns. This control requires documented criteria for choosing data preparation methods and a record of the methods actually used. Auditors treat undocumented preparation as an unexplained gap between source data and model behavior.
Implementation Guidance
- Data used in an AI system ordinarily needs preparation to make it usable for a given AI task. For example, machine learning algorithms are sometimes intolerant of missing or incorrect entries, non-normal distribution and widely varying scales. Preparation methods and transforms can be used to increase the quality of the data. Failure to properly prepare the data can potentially lead to AI system errors. Common preparation methods and transformations for data used in AI systems include:
- statistical exploration of the data (e.g. distribution, mean, median, standard deviation, range, stratification, sampling) and statistical metadata (e.g. data documentation initiative (DDI) specification[28]);
- cleaning (i.e. correcting entries, dealing with missing entries);
- imputation (i.e. methods for filling in missing entries);
- normalization;
- scaling;
- labelling of the target variables;
- encoding (e.g. converting categorical variables to numbers).
- For a given AI task, the organization should document its criteria for selecting specific data preparation methods and transforms as well as the specific methods and transforms used in the AI task.
- NOTE: For additional information on data preparation specific to machine learning see the ISO/IEC 5259 series and ISO/IEC 23053.
References
- ISO/IEC 42001:2023 — Information technology — Artificial intelligence — Management system, A.7.6
- ISO/IEC 42001:2023 Annex B, B.7.6 (implementation guidance)
