Towards sustainable AI: The Data Science development process
Towards sustainable AI: The Data Science development process
Whilst generally organisation specific, typical data science processes involve the following keys phases: business understanding, data collection, model building, evaluation, and deployment. The “performance metrics” phase introduced here provides a template to the integration of business requirements.
Although performance measurement is generally performed as part of evaluation, it often fails to integrate business requirements not directly related to the AI system’s accuracy. For example, an AI recruitment system satisfying traditional performance metrics may fail to meet an organisations prioritised objectives, such as avoiding gender bias. We refer to such objectives as “soft performance” metrics.
Defining performance metrics as an additional phase allows the business and the AI team to agree on the objectives of the project. Non-functional requirements and expected behaviour of the system are identified and documented, leaving the development team with clear objectives. Based on this, the development team can design a system that meets the organisation’s expectations. It may also result in a project being deemed infeasible, given its dependence on data consistent with these objectives.
