CPMAI’s glossary defines data splitting as “dividing a data set into subsets (e.g., training, validation, test) for model development and evaluation,” typically achieved via random subsampling to ensure each subset is representative of the underlying distribution and to prevent sampling bias.

Under Required AI Explainability Considerations, CPMAI mandates that when a chosen model is a “black-box” with limited native interpretability, teams must implement post-hoc interpretability techniques (e.g., feature‐importance plots, surrogate models) to “interpret AI results so that cause and effect can be represented,” ensuring stakeholders understand why the model makes its predictions.

=========

Phase V: Model Evaluation is dedicated to validating a trained model’s performance against technical metrics and the business success criteria defined earlier. This phase encompasses tasks such as generating performance results, measuring KPIs, and deciding whether to retrain or proceed to operationalization.

=========

The CPMAI Glossary’s entry for Generative AI highlights its use in creating new content (text, images, or code), enhancing training datasets via data augmentation, powering virtual avatars/characters, and serving as an Augmented Intelligence tool to boost human productivity . It also underpins programmatic content generation across multiple media types. Generative AI is not designed primarily for explainable decision-support interfaces.

=========

Phase IV: Model Development is explicitly where “Model Training / Model Building” occurs. This phase includes tasks for selecting modeling techniques, conducting hyperparameter optimization, and executing the actual training runs on the prepared datasets.

Neural Networks—especially deep architectures—typically require extensive computational resources, longer training times, and higher infrastructure costs compared to simpler methods. In contrast, algorithms like Naive Bayes train very quickly on large datasets, and Gaussian Mixture Models or SVMs have more moderate training complexity and infrastructure demands. Therefore, given strict constraints on training time, cost, and complexity, Neural Networks are the least suitable choice.

=========

The Data Labeling task in Phase III: Data Preparation specifies that teams should identify labeling methods such as using internal staff, contracting third-party labelers, leveraging pre-existing labeled datasets, or combining those modes. Soliciting end-users to label data falls outside these recommended approaches and introduces uncontrolled variability and quality issues .

=========