import tensorflow as tf
from tensorflow import keras

# 간단한 CNN 모델
model = keras.Sequential([
    keras.layers.Conv2D(32, (3,3), activation='relu', input_shape=(224,224,3)),
    keras.layers.MaxPooling2D((2,2)),
    keras.layers.Conv2D(64, (3,3), activation='relu'),
    keras.layers.MaxPooling2D((2,2)),
    keras.layers.Flatten(),
    keras.layers.Dense(128, activation='relu'),
    keras.layers.Dense(10, activation='softmax')
])

# 이 모델이 예측을 내릴 때
prediction = model.predict(image)
# 결과: [0.05, 0.02, 0.85, 0.03, 0.01, 0.02, 0.01, 0.00, 0.01, 0.00]
# 3번째 클래스일 확률이 85%

import pandas as pd
import numpy as np
from sklearn.metrics import confusion_matrix
import matplotlib.pyplot as plt

# 데이터 로드
df = pd.read_csv('training_data.csv')

# 1. 클래스 불균형 체크
class_distribution = df['target'].value_counts()
print("클래스 분포:")
print(class_distribution)
print(f"\n불균형 비율: {class_distribution.max() / class_distribution.min():.2f}")

# 2. 보호 속성별 분포 체크 (예: 성별, 연령대 등)
if 'gender' in df.columns:
    gender_target_dist = pd.crosstab(df['gender'], df['target'], normalize='index')
    print("\n성별 타겟 분포:")
    print(gender_target_dist)
    
    # 통계적 유의성 검정
    from scipy.stats import chi2_contingency
    chi2, p_value, dof, expected = chi2_contingency(pd.crosstab(df['gender'], df['target']))
    print(f"\nChi-square test p-value: {p_value:.4f}")
    if p_value < 0.05:
        print("⚠️ 경고: 성별과 타겟 간 통계적으로 유의한 연관성 발견")

# 3. 결측치 패턴 분석
missing_pattern = df.isnull().sum()
print("\n결측치 현황:")
print(missing_pattern[missing_pattern > 0])

# 4. 이상치 탐지
def detect_outliers_iqr(data, column):
    Q1 = data[column].quantile(0.25)
    Q3 = data[column].quantile(0.75)
    IQR = Q3 - Q1
    outliers = data[(data[column] < Q1 - 1.5*IQR) | (data[column] > Q3 + 1.5*IQR)]
    return len(outliers)

numeric_columns = df.select_dtypes(include=[np.number]).columns
print("\n이상치 개수:")
for col in numeric_columns:
    outlier_count = detect_outliers_iqr(df, col)
    if outlier_count > 0:
        print(f"{col}: {outlier_count}개 ({outlier_count/len(df)*100:.2f}%)")

from sklearn.metrics import accuracy_score, precision_score, recall_score
import numpy as np

def evaluate_fairness(y_true, y_pred, sensitive_attribute):
    """
    모델의 공정성을 평가하는 함수
    
    Parameters:
    - y_true: 실제 레이블
    - y_pred: 예측 레이블
    - sensitive_attribute: 보호 속성 (예: 성별, 인종 등)
    """
    
    results = {}
    
    # 전체 성능
    results['overall_accuracy'] = accuracy_score(y_true, y_pred)
    results['overall_precision'] = precision_score(y_true, y_pred, average='weighted')
    results['overall_recall'] = recall_score(y_true, y_pred, average='weighted')
    
    # 그룹별 성능
    unique_groups = np.unique(sensitive_attribute)
    group_metrics = {}
    
    for group in unique_groups:
        mask = sensitive_attribute == group
        group_metrics[group] = {
            'accuracy': accuracy_score(y_true[mask], y_pred[mask]),
            'precision': precision_score(y_true[mask], y_pred[mask], average='weighted'),
            'recall': recall_score(y_true[mask], y_pred[mask], average='weighted'),
            'size': np.sum(mask)
        }
    
    results['group_metrics'] = group_metrics
    
    # 공정성 메트릭 계산
    # 1. Demographic Parity (인구통계학적 동등성)
    positive_rates = {}
    for group in unique_groups:
        mask = sensitive_attribute == group
        positive_rates[group] = np.mean(y_pred[mask])
    
    results['demographic_parity_diff'] = max(positive_rates.values()) - min(positive_rates.values())
    
    # 2. Equal Opportunity (기회 균등)
    true_positive_rates = {}
    for group in unique_groups:
        mask = (sensitive_attribute == group) & (y_true == 1)
        if np.sum(mask) > 0:
            true_positive_rates[group] = np.mean(y_pred[mask])
    
    if len(true_positive_rates) > 1:
        results['equal_opportunity_diff'] = max(true_positive_rates.values()) - min(true_positive_rates.values())
    
    # 3. Disparate Impact (차별적 영향)
    if len(positive_rates) == 2:
        rates_list = list(positive_rates.values())
        results['disparate_impact'] = min(rates_list) / max(rates_list) if max(rates_list) > 0 else 0
    
    return results

# 사용 예시
# results = evaluate_fairness(y_test, y_pred, gender_test)
# 
# print("전체 정확도:", results['overall_accuracy'])
# print("\n그룹별 성능:")
# for group, metrics in results['group_metrics'].items():
#     print(f"\n{group}:")
#     print(f"  정확도: {metrics['accuracy']:.4f}")
#     print(f"  정밀도: {metrics['precision']:.4f}")
#     print(f"  재현율: {metrics['recall']:.4f}")
#     print(f"  샘플 수: {metrics['size']}")
# 
# print(f"\n공정성 지표:")
# print(f"Demographic Parity 차이: {results['demographic_parity_diff']:.4f}")
# if 'equal_opportunity_diff' in results:
#     print(f"Equal Opportunity 차이: {results['equal_opportunity_diff']:.4f}")
# if 'disparate_impact' in results:
#     print(f"Disparate Impact: {results['disparate_impact']:.4f}")
#     if results['disparate_impact'] < 0.8:
#         print("⚠️ 경고: Disparate Impact가 0.8 미만입니다 (차별 가능성)")