Data Quality Risks

Data is the foundation of AI systems. Poor data quality leads to unreliable, biased, and potentially harmful AI. This lesson explores data quality dimensions, risks, and quality assurance strategies.

The Centrality of Data Quality

"Garbage In, Garbage Out"

AI models learn patterns from training data:

• Good data → Reliable, fair AI
• Poor data → Unreliable, biased AI

No amount of algorithmic sophistication can overcome fundamentally flawed data.

Data Quality Impact

Performance: Inaccurate data reduces model accuracy and reliability

Fairness: Biased data produces discriminatory outcomes

Safety: Incorrect data can lead to dangerous decisions

Compliance: Poor data quality may violate regulations (GDPR data accuracy requirements)

Trust: Data quality issues erode confidence in AI systems

Data Quality Dimensions

1. Accuracy

Definition: Data correctly represents reality.

Issues:

• Measurement errors
• Transcription mistakes
• Outdated information
• False or fabricated data

Examples:

• Medical records with incorrect patient information
• Financial data with calculation errors
• Sensor data from miscalibrated devices
• Labels from unreliable annotators

Impact on AI:

• Model learns incorrect patterns
• Predictions based on false information
• Systematic errors in outputs
• Cannot distinguish signal from noise

Assessment:

• Validate against ground truth
• Cross-check multiple sources
• Statistical outlier detection
• Expert review samples

Mitigation:

• Source data from reliable providers
• Implement validation rules
• Cross-reference multiple sources
• Regular accuracy audits
• Correct errors when discovered

2. Completeness

Definition: All required data is present; no critical missing values.

Issues:

• Missing features
• Incomplete records
• Null values
• Truncated information

Missing Data Types:

Missing Completely at Random (MCAR): Missingness unrelated to data

• Random sensor failures
• Arbitrary data loss

Missing at Random (MAR): Missingness related to observed data

• Older patients more likely to have missing digital records
• Can be modeled and addressed

Missing Not at Random (MNAR): Missingness related to unobserved data

• People with poor health more likely to skip health surveys
• Creates systematic bias if not addressed

Impact on AI:

• Reduced effective sample size
• Biased estimates if missingness non-random
• Models may learn to exploit missingness patterns
• Cannot detect patterns in missing features

Assessment:

• Calculate missing data rates per feature
• Analyze patterns of missingness
• Test whether missingness is random
• Evaluate impact on subgroups

Mitigation:

• Collect complete data when possible
• Imputation strategies:
  • Simple: mean, median, mode
  • Advanced: regression, k-NN, multiple imputation
• Model missingness explicitly
• Remove features with excessive missingness
• Use algorithms robust to missing data

3. Consistency

Definition: Data is uniform and coherent across sources and time.

Issues:

• Contradictory information across sources
• Format inconsistencies
• Logical contradictions within data
• Evolving definitions or standards

Examples:

• Age calculated from birthdate doesn't match reported age
• Currency in inconsistent formats ($100 vs. 100 USD vs. 100.00)
• Dates in different formats (MM/DD/YYYY vs. DD/MM/YYYY)
• Conflicting information in linked databases

Impact on AI:

• Model confusion from contradictions
• Reduced predictive power
• Errors in data integration
• Difficulty in learning coherent patterns

Assessment:

• Constraint validation (age > 0, date ranges)
• Cross-field consistency checks
• Format standardization assessment
• Duplicate detection

Mitigation:

• Standardize formats and definitions
• Implement consistency constraints
• Resolve conflicts systematically
• Data integration quality assurance
• Version control for evolving standards

4. Timeliness

Definition: Data is current and reflects recent state.

Issues:

• Outdated information
• Stale data from infrequent updates
• Delays in data collection or processing
• Historical data not representative of present

Examples:

• Economic indicators from last year used for current predictions
• User preferences changed but profile not updated
• Seasonal patterns not reflected in older data
• Technology evolution making historical data less relevant

Impact on AI:

• Concept drift (world changes, model doesn't)
• Predictions based on obsolete patterns
• Performance degradation over time
• Misalignment with current reality

Assessment:

• Track data freshness and age
• Monitor performance over time
• Compare predictions to recent outcomes
• Analyze temporal patterns

Mitigation:

• Regular data refresh and updates
• Real-time or near-real-time data pipelines
• Time-windowed training (recent data weighted more)
• Continuous model retraining
• Drift detection and alerts

5. Relevance

Definition: Data is appropriate and applicable for the task.

Issues:

• Features not predictive of target
• Data from wrong context or population
• Proxy measures that don't capture construct
• Extraneous information adding noise

Examples:

• Using data from one geographic region for another
• Features that worked in past no longer relevant
• Demographic data when predicting objective outcomes
• High-dimensional data with many irrelevant features

Impact on AI:

• Wasted computational resources
• Overfitting to irrelevant patterns
• Spurious correlations
• Poor generalization

Assessment:

• Feature importance analysis
• Correlation with target variable
• Domain expert review
• Ablation studies (removing features)

Mitigation:

• Feature selection based on relevance
• Domain expertise in feature engineering
• Regular review of feature utility
• Remove low-value features
• Context-appropriate data sourcing

6. Representativeness

Definition: Data reflects the diversity of real-world scenarios and populations.

Issues:

• Sampling bias
• Underrepresented subgroups
• Geographic or temporal limitations
• Selection effects

Examples:

• Training data from one demographic applied broadly
• Historical data not representative of current population
• Convenience sampling missing hard-to-reach groups
• Voluntary participation creating selection bias

Impact on AI:

• Poor performance for underrepresented groups
• Bias and discrimination
• Limited generalization
• Safety issues in unrepresented scenarios

Assessment:

• Compare data distribution to target population
• Analyze representation across demographics
• Test performance across subgroups
• Identify gaps in coverage

Mitigation:

• Stratified sampling ensuring representation
• Oversampling underrepresented groups
• Synthetic data generation for minorities
• Multi-source data aggregation
• Explicit diversity goals in data collection

7. Data Lineage and Provenance

Definition: Clear documentation of data origins, transformations, and history.

Issues:

• Unknown data sources
• Undocumented preprocessing
• Lost transformation history
• Inability to trace errors to source

Importance:

• Understanding data biases and limitations
• Reproducing results
• Debugging issues
• Compliance and auditability
• Trust and transparency

Components:

• Origin: Where did data come from?
• Collection: How was data gathered?
• Transformations: What processing was applied?
• Quality: What validation was performed?
• Custody: Who handled data and when?
• Versioning: How has data evolved?

Best Practices:

• Automated lineage tracking
• Metadata documentation
• Version control for datasets
• Transformation logs
• Data catalogs with provenance

Data Quality Risks

1. Training Data Risks

Insufficient Volume:

• Not enough examples to learn patterns
• High variance, poor generalization
• Particularly problematic for rare classes

Mitigation: Data augmentation, transfer learning, few-shot learning techniques

Poor Label Quality:

• Incorrect annotations
• Inconsistent labeling across annotators
• Ambiguous cases mislabeled
• Label noise

Mitigation: Multiple annotators, quality checks, confidence scoring, noise-robust algorithms

Historical Bias:

• Past discrimination encoded in data
• Outdated patterns no longer valid
• Societal inequities reflected

Mitigation: Awareness, data corrections, fairness constraints, stakeholder input

Representation Gaps:

• Edge cases missing
• Minority groups underrepresented
• Limited scenario coverage

Mitigation: Targeted data collection, synthetic data, importance weighting

2. Feature Engineering Risks

Spurious Correlations:

• Features correlate in training but not reality
• Confounding variables
• Coincidental patterns

Example: Ice cream sales correlate with drowning deaths (both caused by summer weather).

Mitigation: Causal reasoning, domain expertise, out-of-distribution testing

Data Leakage:

• Training data contains information from future
• Target variable information in features
• Evaluation data contaminating training

Example: Using total medical costs (which includes treatment costs) to predict need for treatment.

Mitigation: Temporal splits, careful feature review, strict train/test separation

Proxy Discrimination:

• Features correlated with protected attributes
• Seemingly neutral features encode bias

Examples: Zip code proxies for race, first names for gender/ethnicity

Mitigation: Fairness audits, remove high-correlation proxies, include protected attributes explicitly in fairness testing

Poor Feature Selection:

• Irrelevant features add noise
• Overly complex feature spaces
• Missing critical features

Mitigation: Feature importance analysis, domain expertise, iterative refinement

3. Data Drift Risks

Concept Drift:

• Relationship between features and target changes
• Predictive patterns evolve

Example: Consumer preferences shift, making old behavior patterns less predictive.

Detection: Monitor prediction accuracy over time, statistical tests comparing distributions

Mitigation: Regular retraining, online learning, drift-adaptive algorithms

Covariate Shift:

• Input data distribution changes
• Feature statistics evolve

Example: User demographics shift, sensor calibration changes, new product lines.

Detection: Compare feature distributions training vs. production

Mitigation: Domain adaptation, importance weighting, model updates

Label Drift:

• Definition or measurement of target changes
• New classes emerge

Example: New fraud patterns, emerging diseases, novel product categories.

Detection: Review label distributions, analyst feedback

Mitigation: Schema evolution support, regular relabeling, adaptive models

4. Data Collection Risks

Selection Bias:

• Sampling process systematically excludes groups
• Self-selection effects
• Survivorship bias

Example: Survey responses only from those willing to respond (may differ from non-responders).

Mitigation: Random sampling, diverse collection methods, bias modeling

Measurement Bias:

• Instrument bias in data collection
• Different measurement quality across groups
• Systematic measurement errors

Example: Lower resolution images for certain populations, less detailed records for some demographics.

Mitigation: Calibration, standardization, quality metrics by group

Temporal Bias:

• Data collection limited to certain time periods
• Seasonal effects
• Historical events influencing data

Example: Economic data from recession period not representative of normal economy.

Mitigation: Multi-period data collection, seasonal adjustments, temporal modeling

Privacy-Utility Tradeoff:

• Privacy protection (anonymization, aggregation) reduces data utility
• Noise addition for privacy degrades quality
• Access restrictions limit data availability

Mitigation: Privacy-enhancing technologies (differential privacy, federated learning), optimize privacy-utility balance

5. Data Pipeline Risks

ETL Errors:

• Extract, Transform, Load process failures
• Data corruption during transfer
• Integration mismatches

Mitigation: Validation checkpoints, automated testing, data reconciliation

Schema Changes:

• Database schema evolves
• Field definitions change
• Breaking changes to data structure

Mitigation: Version control, backward compatibility, schema validation

Processing Failures:

• Batch job failures
• Data loss during processing
• Incomplete updates

Mitigation: Monitoring and alerting, error handling, rollback capabilities

Scalability Issues:

• Pipeline cannot handle data volume
• Latency increases
• Bottlenecks and failures

Mitigation: Performance testing, distributed processing, capacity planning

Data Quality Assurance

Data Validation Framework

1. Schema Validation:

• Field types and formats correct
• Required fields present
• Constraints satisfied

2. Statistical Validation:

• Ranges and distributions reasonable
• Outliers identified and reviewed
• Consistency with historical data

3. Business Rules Validation:

• Domain-specific constraints met
• Logical consistency
• Referential integrity

4. Cross-Source Validation:

• Agreement across data sources
• Reconciliation of conflicts
• Master data management

Automated Data Quality Checks

Implement automated pipelines checking:

# Example data quality checks
def validate_data(df):
    checks = []

    # Completeness
    missing_rate = df.isnull().sum() / len(df)
    checks.append(("missing_rate", missing_rate < 0.05))  # <5% missing

    # Accuracy
    valid_age = (df['age'] >= 0) & (df['age'] <= 120)
    checks.append(("valid_age", valid_age.all()))

    # Consistency
    age_matches = (df['age'] == (pd.Timestamp.now().year - df['birth_year']))
    checks.append(("age_consistency", age_matches.all()))

    # Representativeness
    gender_balance = df['gender'].value_counts(normalize=True)
    checks.append(("gender_balance", (gender_balance > 0.3).all()))  # No group <30%

    return checks

Alert on failures, block pipeline if critical checks fail.

Data Quality Metrics

Track and Monitor:

Dashboard these metrics, trend over time, alert on degradation.

Data Quality Improvement Process

1. Assessment: Measure current quality across dimensions

2. Root Cause Analysis: Identify sources of quality issues

3. Remediation: Fix data and underlying processes

4. Prevention: Implement controls to maintain quality

5. Monitoring: Continuous quality tracking

6. Iteration: Regular review and improvement

Data Governance for AI

Data Governance Framework

Data Ownership:

• Clear accountability for data quality
• Domain stewards responsible for their data
• Cross-functional data governance committee

Data Standards:

• Standardized definitions and formats
• Data dictionaries and catalogs
• Metadata management

Data Quality Rules:

• Defined quality criteria
• Validation procedures
• Escalation processes for issues

Access Controls:

• Role-based access
• Data classification (public, internal, confidential)
• Audit logging

Lifecycle Management:

• Retention policies
• Archival procedures
• Deletion protocols

Data Documentation

Datasheets for Datasets:

Comprehensive documentation including:

• Motivation: Why created, who created, who funded
• Composition: What data, how many instances, missing data, relationships
• Collection: How collected, sampling strategy, who was involved
• Preprocessing: Cleaning, labeling, raw data availability
• Uses: Prior uses, repository, other uses
• Distribution: How distributed, when, license
• Maintenance: Who maintains, how to contribute, updates

Benefits:

• Transparency about data characteristics
• Understanding biases and limitations
• Informed use of data
• Reproducibility

Responsible Data Collection

Ethical Considerations:

• Informed consent for data use
• Privacy protection
• Respect for data subjects
• Benefit sharing with data sources

Inclusive Collection:

• Diverse, representative sampling
• Accessibility accommodations
• Multiple languages and formats
• Reaching underrepresented groups

Quality by Design:

• Clear data requirements upfront
• Quality criteria in collection protocols
• Training for data collectors
• Validation during collection, not just after

Case Study: Financial Fraud Detection Data Quality

Context: Building AI for detecting credit card fraud.

Data Quality Challenges:

1. Severe Class Imbalance:

• Fraud rate: 0.1% (1 in 1000 transactions)
• Model would achieve 99.9% accuracy by predicting "not fraud" for everything
• Needs to detect rare fraud without excessive false positives

Solutions:

• Undersampling majority class
• Oversampling/SMOTE for minority class
• Stratified train/test split preserving fraud rate
• Evaluation metrics: precision-recall, F1, not just accuracy

2. Rapid Drift:

• Fraud patterns evolve quickly (weeks/months)
• New fraud schemes emerge constantly
• Historical data becomes less relevant

Solutions:

• Continuous monitoring and retraining (weekly)
• Focus on recent data (sliding window)
• Online learning to adapt quickly
• Fraud analyst feedback loop

3. Delayed Labels:

• Fraud not detected immediately
• Labels updated days/weeks after transaction
• Creates label drift and evaluation challenges

Solutions:

• Model actual-at-time-of-transaction risk
• Delayed label handling in training
• Continuous model evaluation with updated labels
• Short-term proxy metrics (customer disputes)

4. Privacy Constraints:

• Cannot store sensitive transaction details
• Aggregation reduces signal
• Privacy regulations limit data use

Solutions:

• Feature engineering creating privacy-preserving aggregates
• Federated learning across banks
• Differential privacy for model training
• Minimize retention of raw data

5. Data Quality Variability:

• Different merchants provide varying data quality
• International transactions have different formats
• Missing fields common

Solutions:

• Robust to missing data (ensemble methods, special handling)
• Standardization layer for ingestion
• Quality scores per data source
• Explicit missing data features

Results:

• 95% fraud detection rate (up from 70% with poor data quality)
• 0.5% false positive rate (acceptable for business)
• Model performance maintained over 6-month periods
• Continuous monitoring catches drift early

Lessons:

• Data quality directly impacts fraud detection capability
• Multiple quality issues require multiple solutions
• Continuous attention needed for evolving patterns
• Privacy and quality must be balanced

Summary

Data is Foundation: AI quality cannot exceed data quality.

Multiple Dimensions: Accuracy, completeness, consistency, timeliness, relevance, representativeness all matter.

Systematic Risks: Training data, feature engineering, drift, collection, and pipeline issues.

Proactive Quality Assurance: Validation, monitoring, governance, and documentation.

Continuous Process: Data quality requires ongoing attention and improvement.

Ethical Imperative: Poor data quality leads to unfair and harmful AI.

ISO 42001 Integration: Data governance controls (Annex A) address quality systematically.

Next Lesson: Security and adversarial risks - protecting AI from attacks.