프로젝트 범위 설정: 문제와 성공 기준 정의

# Scoping checklist — ML task type
questions = [
    'What is the exact output the model should produce?',
    'Is the output a number (regression) or a category (classification)?',
    'Do we need probabilities, or just labels?',
    'Is there a natural ordering in the labels (ordinal vs nominal)?',
    'Is this supervised (labelled data exists) or unsupervised?',
    'Is time ordering important (time-series)?'
]
for q in questions:
    print('•', q)

# Example: formalising target definition in code
import pandas as pd

def label_churn(df, observation_date, prediction_window_days=90):
    '''
    Label a customer as churned (1) if they made no purchase
    in the prediction_window_days after observation_date.
    '''
    cutoff = pd.Timestamp(observation_date)
    window_end = cutoff + pd.Timedelta(days=prediction_window_days)
    churned = (
        df.groupby('customer_id')['purchase_date']
        .apply(lambda dates: not any((dates > cutoff) & (dates <= window_end)))
        .reset_index(name='churned')
    )
    return churned

# Data audit template
data_sources = [
    {'feature': 'days_since_last_purchase', 'source': 'orders DB', 'latency_hours': 1, 'leakage_risk': False},
    {'feature': 'total_spend_90d', 'source': 'orders DB', 'latency_hours': 1, 'leakage_risk': False},
    {'feature': 'support_tickets_opened', 'source': 'CRM', 'latency_hours': 24, 'leakage_risk': False},
    {'feature': 'refund_requested', 'source': 'finance DB', 'latency_hours': 1, 'leakage_risk': True}
]
import pandas as pd
audit = pd.DataFrame(data_sources)
print(audit)
print('\nFeatures with leakage risk:')
print(audit[audit['leakage_risk'] == True]['feature'].tolist())

# Formalised success criteria
success_criteria = {
    'primary_metric': {
        'name': 'recall (churn class)',
        'threshold': 0.80,
        'rationale': 'Missing a churner costs more than a false alert'
    },
    'secondary_metric': {
        'name': 'precision (churn class)',
        'threshold': 0.60,
        'rationale': 'Intervention budget: can only contact 40% of flagged customers'
    },
    'business_metric': {
        'name': 'retained revenue lift',
        'threshold': '+15% over no-model baseline',
        'measurement': 'A/B test over 90 days post-launch'
    }
}
for metric, spec in success_criteria.items():
    print(f'{metric}: {spec}')

from sklearn.dummy import DummyClassifier
from sklearn.metrics import classification_report
import numpy as np

# Simulate imbalanced dataset (10% churn rate)
np.random.seed(42)
y = np.random.choice([0, 1], size=1000, p=[0.9, 0.1])

# Majority-class baseline
dummy = DummyClassifier(strategy='most_frequent')
dummy.fit(np.zeros((1000, 1)), y)  # features irrelevant for dummy
y_dummy = dummy.predict(np.zeros((1000, 1)))
print(classification_report(y, y_dummy, target_names=['retained', 'churned']))

# Deployment constraint spec
deployment = {
    'serving_mode': 'batch (nightly scoring of all active customers)',
    'max_latency_ms': None,   # batch: no real-time constraint
    'throughput_rps': None,
    'memory_budget_mb': 512,  # must run on a single EC2 t3.medium
    'retraining_frequency': 'weekly',
    'max_model_size_mb': 100,
    'explainability_required': True,
    'regulatory_audit_trail': True
}
for k, v in deployment.items():
    print(f'{k}: {v}')

# Ethics and legal checklist
ethics_checklist = [
    ('High-stakes decisions on individuals?', True,   'churn model triggers retention offers, low risk'),
    ('Protected attributes in data?',          True,   'age and region present — must audit for proxy bias'),
    ('GDPR data minimisation required?',       True,   'only collect features needed for prediction'),
    ('Right to explanation required?',         True,   'EU AI Act high-risk? Check with legal team'),
    ('Regulatory filing required?',            False,  'B2C retention model, not financial services')
]
for question, answer, note in ethics_checklist:
    print(f'[{"YES" if answer else "NO"}] {question}  -> {note}')

charter_template = '''
PROJECT CHARTER
===============
Problem: Identify customers at risk of churning before they cancel.
ML Task: Binary classification (churned=1, retained=0)
Target: No purchase in 90 days after monthly observation date
Data Sources: orders DB, CRM (support tickets), product analytics
Primary Metric: Recall >= 0.80 on held-out test set
Secondary Metric: Precision >= 0.60
Baseline: DummyClassifier majority-class (recall = 0, precision = NaN)
Deployment: Batch scoring, weekly retraining, 100 MB model limit
Ethical Flags: Audit for age/region proxy bias; GDPR minimisation
Timeline: MVP in 6 weeks; A/B test for 90 days; review in month 5
'''
print(charter_template)

# Charter revision log (tracked in version control)
revision_log = [
    {'version': 'v1.0', 'date': '2026-06-01',
     'change': 'Initial scope: 30-day churn window'},
    {'version': 'v1.1', 'date': '2026-06-05',
     'change': 'EDA: 30-day window has only 4% churn rate (too imbalanced); widened to 90 days'},
    {'version': 'v1.2', 'date': '2026-06-10',
     'change': 'Baseline run: dummy recall=0, real-time serving infeasible (10s latency); switched to batch'}
]
for rev in revision_log:
    print(f"{rev['version']} ({rev['date']}): {rev['change']}")

project_phases = [
    ('Phase 1', 'EDA + data quality audit', '2026-06-01', '2026-06-07'),
    ('Phase 2', 'Feature engineering + baseline model', '2026-06-08', '2026-06-14'),
    ('Phase 3', 'Model selection + hyperparameter tuning', '2026-06-15', '2026-06-21'),
    ('Phase 4', 'Evaluation, fairness audit, model card', '2026-06-22', '2026-06-25'),
    ('Phase 5', 'Deploy + A/B test + monitoring setup', '2026-06-26', '2026-07-31')
]
print(f'{'Phase':8} | {'Deliverable':42} | Start      | End')
print('-' * 80)
for phase, desc, start, end in project_phases:
    print(f'{phase:8} | {desc:42} | {start} | {end}')