CPMAI Phase III: Data Preparation

Overview
CPMAI (Cognitive Project Management for AI) is a six-phase, iterative, data-centric methodology for running AI and machine learning projects. The six phases are:
1. Business Understanding
2. Data Understanding
3. Data Preparation
4. Model Development
5. Model Evaluation
6. Model Operationalization

Phase III, Data Preparation, turns raw data that has been identified and assessed into clean, structured, labeled, model-ready data. CPMAI treats AI projects as data projects first. Because of that, Phase III is often the most time-consuming phase and the one most decisive for success.

Why Phase III Is Important
1. Garbage in, garbage out. A model can only be as good as the data it learns from. Poor data quality produces poor models, no matter how sophisticated the algorithm.
2. Most project effort happens here. Industry experience suggests data preparation can consume around 60 to 80 percent of AI project time. Underestimating it is a major cause of schedule overruns and project failure.
3. It is a leading cause of AI project failure. CPMAI was created partly because AI projects fail for data reasons, such as poor quality, insufficient quantity, missing labels, or bias. These fail far more often than for algorithm reasons.
4. It supports trustworthy AI. Bias, privacy, and representativeness are addressed during preparation, before they become embedded in the model.
5. It enables reproducibility. Documented, repeatable data pipelines are required later for model retraining and operationalization in Phase VI.
6. It works as a go/no-go checkpoint. If data cannot be prepared to the needed quality or quantity, the project should loop back to earlier phases rather than push ahead.

What Phase III Is
Data Preparation is the set of activities that transform data assessed in Phase II into datasets ready for model training, validation, and testing.

Key questions answered in Phase III include:
- How do we clean and correct the data?
- How do we label or annotate it for supervised learning?
- How do we transform, normalize, and engineer features?
- How do we merge data from multiple sources?
- How do we split data into training, validation, and test sets?
- How do we handle bias, privacy, and compliance?
- How do we make the pipeline repeatable for production?

CPMAI borrows from CRISP-DM, the Cross-Industry Standard Process for Data Mining. It extends that model with AI-specific concerns such as data labeling, the seven patterns of AI, iteration, and trustworthy AI considerations.

How Phase III Works: Core Activities

1. Data Selection
- Choose the relevant data sources, records, and attributes identified in Phase II.
- Decide what to include and exclude, and document the reasons.
- Make sure the data aligns with the AI pattern chosen in Phase I. The seven patterns are:
  - Recognition
  - Conversational/Human Interaction
  - Predictive Analytics and Decision Support
  - Patterns and Anomalies
  - Hyperpersonalization
  - Autonomous Systems
  - Goal-Driven Systems

2. Data Cleansing
- Handle missing values by deletion or imputation, for example with the mean, median, mode, or a model-based method.
- Remove duplicates and fix inconsistencies such as different formats, units, or spellings.
- Detect and treat outliers. Note that in anomaly-detection projects, outliers may be the signal and should not simply be removed.
- Correct noisy or erroneous entries.

3. Data Labeling and Annotation
- Supervised learning requires labeled data, also called ground truth.
- Labeling options include in-house subject matter experts, crowdsourcing, third-party labeling services, and programmatic or weak labeling.
- Augmented approaches include active learning, semi-supervised learning, and synthetic labeling.
- Quality control matters: clear labeling guidelines, inter-annotator agreement checks, and review cycles.
- Labeling is often the most expensive and slowest step. Plan budget and schedule for it explicitly.

4. Data Transformation
- Normalize or standardize numeric data, for example with min-max scaling or z-score.
- Encode categorical variables, for example with one-hot or label encoding.
- Tokenize and vectorize text, and resize or normalize images.
- Aggregate or reshape data, for example into time windows.

5. Feature Engineering and Selection
- Create new, meaningful attributes from raw data, such as ratios, dates broken into day of week, or rolling averages.
- Remove irrelevant, redundant, or leaking features.
- Use dimensionality reduction where appropriate.

6. Data Integration
- Merge datasets from multiple sources.
- Resolve key mismatches, schema differences, and entity resolution issues.

7. Data Augmentation and Enrichment
- Increase dataset size or diversity, for example by rotating or flipping images, paraphrasing text, or generating synthetic data.
- Enrich data with third-party sources when internal data is insufficient.

8. Handling Class Imbalance
- Techniques include oversampling (for example SMOTE), undersampling, class weighting, and collecting more minority-class data.

9. Data Splitting
- Split data into training, validation, and test sets. Typical ratios are 70/15/15 or 80/10/10.
- Prevent data leakage. Test data must never influence training or preprocessing decisions.
- Use stratified splits for imbalanced classes and time-based splits for temporal data.

10. Bias, Privacy, and Compliance Controls
- Check that the data is representative and look for sampling bias.
- Anonymize, pseudonymize, or mask personally identifiable information (PII).
- Ensure compliance with regulations such as GDPR and HIPAA, and with consent terms.
- Document lineage and provenance for governance and auditability.

11. Building the Data Pipeline
- Automate preparation steps so they are repeatable for retraining and for production inference.
- Apply version control to data and transformations.
- Remember that the same transformations must be applied to new data in production.

12. Documentation and Phase Review
- Produce documentation of the prepared dataset, a data quality report, labeling guidelines, and a description of the pipeline.
- Hold a go/no-go review: is the data sufficient for Phase IV, Model Development? If not, iterate back.

Key Outputs and Deliverables of Phase III
- Cleaned, labeled, transformed datasets
- Training, validation, and test splits
- Feature set definitions
- Data quality report and preparation log
- Repeatable data pipeline
- Bias and privacy assessment notes
- Updated risks, schedule, and cost estimates

Relationship to Other Phases
- Phase I (Business Understanding): defines the problem, the AI pattern, and success criteria. These drive what data preparation must achieve.
- Phase II (Data Understanding): identifies data sources and assesses quality and quantity. Phase III acts on those findings.
- Phase IV (Model Development): consumes prepared data. Poor results often send the team back to Phase III.
- Phase V (Model Evaluation): uses the held-out test set created in Phase III.
- Phase VI (Operationalization): reuses the data pipeline and watches for data drift, which may trigger new preparation work.

CPMAI is iterative and agile. Teams are encouraged to work in small iterations and to return to earlier phases whenever new information emerges.

Common Pitfalls in Phase III
- Underestimating time and cost, especially for labeling
- Data leakage, where test or future information enters training
- Ignoring bias and representativeness
- Removing outliers that are actually meaningful
- One-off manual preparation that cannot be reproduced in production
- Starting model building before data is ready
- Poor documentation and missing data lineage
- Failing to involve subject matter experts in labeling and feature decisions

Exam Tips: Answering Questions on CPMAI Phase III: Data Preparation

Tip 1: Think data first. CPMAI is data-centric. When a scenario describes poor model performance, the best answer usually points to data quality, quantity, or labeling, not to switching algorithms.

Tip 2: Know the phase order and boundaries.
- Phase II (Data Understanding): identifying, collecting, exploring, and assessing data. Keywords: assess, explore, inventory.
- Phase III (Data Preparation): cleaning, labeling, transforming, feature engineering, splitting. Keywords: clean, label, transform, engineer, split, augment.
- If a question asks when data is labeled or cleansed, the answer is Phase III.

Tip 3: Expect labeling questions. Know that supervised learning requires labeled data. Know the labeling options: internal, crowdsourced, third-party, and programmatic. Know that labeling quality and cost are major project risks. Answers that include clear guidelines and quality checks are usually best.

Tip 4: Watch for data leakage. Any option that uses test data during training, or fits preprocessing on the full dataset before splitting, is wrong. The correct answer protects the integrity of the holdout set.

Tip 5: Iteration is a correct answer. If data proves insufficient, the CPMAI answer is to loop back to Phase II or Phase I. Proceeding anyway, or forcing a model to work, is not the CPMAI way. Go/no-go decisions are legitimate and encouraged.

Tip 6: Choose sustainable, repeatable options. Prefer answers that build automated, documented, version-controlled pipelines over manual, one-time cleaning. Operationalization depends on consistent preparation.

Tip 7: Integrate trustworthy AI. Choose answers that address bias, representativeness, privacy (PII masking), and compliance during preparation rather than after deployment.

Tip 8: Context matters for outliers and imbalance. In fraud or anomaly scenarios, do not blindly remove outliers. For imbalanced classes, consider resampling, class weighting, or collecting more minority-class data.

Tip 9: Use project manager judgment. CPMAI exam questions often ask what the project manager should do next. Good answers involve:
- Consulting data engineers, data scientists, and subject matter experts
- Updating the schedule and budget for preparation effort
- Documenting decisions
- Raising risks early

Tip 10: Eliminate distractors. Rule out answers that:
- skip phases, for example jumping to deployment
- focus on tuning algorithms when the root cause is data
- ignore stakeholders or governance
- assume more data always fixes everything without checking quality

Tip 11: Remember the effort estimate. If asked which phase typically consumes the most effort, Data Preparation, together with Data Understanding, is the expected answer.

Sample Question Walkthrough
Scenario: A team building an image classifier finds that model accuracy is low. Investigation reveals inconsistent labels from multiple annotators. What should the project manager do?

Best answer: Return to Data Preparation. Establish clear labeling guidelines, measure inter-annotator agreement, and relabel or review the data before retraining.

Wrong answers: switch to a deeper neural network, deploy anyway and monitor, or add more layers. These ignore the data root cause.

Quick Memory Aid
Use C-L-T-F-I-A-S-P: Clean, Label, Transform, Feature-engineer, Integrate, Augment, Split, Pipeline. Wrap all of these in Governance (bias, privacy, documentation) and Iteration (go/no-go).

Summary
Phase III converts assessed data into high-quality, model-ready datasets. It is where most AI project effort and risk concentrates. Master its activities, its boundaries with Phases II and IV, the emphasis on labeling, leakage prevention, repeatable pipelines, and trustworthy AI, and the iterative go/no-go mindset. With these, you will be well prepared to answer CPMAI exam questions on Data Preparation confidently.