thanhtai435's picture
Add chapter-04-data-preprocessing/README.md
a993a7b verified

Chương 4: Tiền xử lý Dữ liệu

Data Preprocessing for Knowledge Discovery


📚 Mục tiêu học tập (CLO1, CLO2, CLO3, CLO4)

Sau khi hoàn thành chương này, sinh viên có thể:

  • Đánh giá chất lượng dữ liệu đa chiều (accuracy, completeness, consistency, timeliness)
  • Thực hiện làm sạch dữ liệu: xử lý nhiễu, outliers, missing values
  • Áp dụng các kỹ thuật biến đổi: smoothing, normalization, discretization
  • Giảm số chiều dữ liệu bằng PCA, feature selection
  • Feature engineering cho mô hình ML

4.1. Đa chiều Chất lượng Dữ liệu

    ┌──────────────────────────────────────────────────────┐
    │              DATA QUALITY DIMENSIONS                  │
    │                                                      │
    │  ┌────────────┐  ┌─────────────┐  ┌──────────────┐ │
    │  │ ACCURACY   │  │COMPLETENESS │  │ CONSISTENCY  │ │
    │  │            │  │             │  │              │ │
    │  │ Du lieu co │  │ Bao nhieu   │  │ Du lieu co   │ │
    │  │ dung khong?│  │ % bi thieu? │  │ mau thuan?   │ │
    │  │            │  │             │  │              │ │
    │  │ VD: Gia    │  │ VD: 15%     │  │ VD: Age=5    │ │
    │  │ am (-$50)  │  │ thieu email │  │ nhung Married│ │
    │  └────────────┘  └─────────────┘  └──────────────┘ │
    │                                                      │
    │  ┌────────────┐  ┌─────────────┐  ┌──────────────┐ │
    │  │ TIMELINESS │  │  VALIDITY   │  │  UNIQUENESS  │ │
    │  │            │  │             │  │              │ │
    │  │ Du lieu co │  │ Dung format │  │ Co bi trung  │ │
    │  │ cap nhat?  │  │ va range?   │  │ lap khong?   │ │
    │  │            │  │             │  │              │ │
    │  │ VD: Data   │  │ VD: Email   │  │ VD: 2 rows   │ │
    │  │ tu 2019    │  │ khong co @  │  │ cung 1 order │ │
    │  └────────────┘  └─────────────┘  └──────────────┘ │
    └──────────────────────────────────────────────────────┘

4.2. Kỹ thuật Làm sạch Dữ liệu

Xử lý Missing Values

Phương pháp Mô tả Khi nào dùng
Loại bỏ Xóa row/column có missing Missing random, tỷ lệ thấp (<5%)
Mean/Median Thay bằng giá trị trung bình/trung vị Numerical, phân phối chuẩn/lệch
Mode Thay bằng giá trị phổ biến nhất Categorical variables
Forward/Backward Fill Dùng giá trị trước/sau Time series data
KNN Imputation Dùng K neighbors gần nhất Dữ liệu có correlation
Regression Predict missing từ các biến khác Strong linear relationships
MICE Multiple Imputation by Chained Equations Complex missing patterns

Xử lý Outliers

    Phương pháp phát hiện Outlier:
    
    1. IQR Method:
       Q1 = 25th percentile
       Q3 = 75th percentile
       IQR = Q3 - Q1
       Outlier nếu: x < Q1 - 1.5*IQR hoặc x > Q3 + 1.5*IQR
    
    2. Z-Score Method:
       z = (x - mean) / std
       Outlier nếu: |z| > 3
    
    3. Modified Z-Score (robust):
       MAD = median(|xi - median(x)|)
       Modified_Z = 0.6745 * (xi - median(x)) / MAD
       Outlier nếu: |Modified_Z| > 3.5
    
    Xử lý:
    - Capping/Winsorizing: Thay outlier bằng percentile 1%/99%
    - Log transform: Giảm skewness
    - Loại bỏ: Nếu chắc chắn là lỗi dữ liệu
    - Giữ nguyên: Nếu outlier có ý nghĩa business (VD: đơn hàng lớn)

4.3. Data Integration & Deduplication

Entity Resolution

    Record A: "Sao Paulo", "SP", "sao_paulo@email.com"
    Record B: "São Paulo", "SP", "saopaulo@email.com"  
    Record C: "S. Paulo",  "SP", "sao_paulo@email.com"
    
    → Tất cả là cùng 1 entity!
    
    Kỹ thuật:
    1. Exact matching (= )
    2. Fuzzy matching (Levenshtein distance, Jaro-Winkler)
    3. Phonetic matching (Soundex, Metaphone)
    4. ML-based matching (trained classifier)

4.4. Biến đổi Dữ liệu (Transformation)

Smoothing Methods

Method Công thức Use case
Moving Average MA(k) = (1/k) * Σ x_i Time series noise reduction
Weighted MA WMA = Σ w_i * x_i / Σ w_i Recent data more important
Exponential Smoothing S_t = α*x_t + (1-α)*S_{t-1} Adaptive smoothing
Binning Group into bins, replace with bin mean/median Discretization

4.5. Data Reduction: Giảm số chiều

PCA (Principal Component Analysis)

    Original: 10 features          PCA: 3 components (95% variance)
    ┌──────────────────┐          ┌──────────────────┐
    │ x1 x2 x3 ... x10│    →     │ PC1  PC2  PC3    │
    │                  │  PCA     │                  │
    │ 1000 x 10        │          │ 1000 x 3         │
    └──────────────────┘          └──────────────────┘
    
    Bước thực hiện:
    1. Standardize dữ liệu (mean=0, std=1)
    2. Tính Covariance Matrix
    3. Tính Eigenvalues & Eigenvectors
    4. Chọn top-k components (giải thích 95%+ variance)
    5. Transform dữ liệu sang không gian mới

Feature Selection Methods

Method Type Mô tả
Variance Threshold Filter Loại features có variance thấp
Correlation Filter Loại features có correlation cao với nhau
Chi-squared Filter Đánh giá independence giữa feature và target
Mutual Information Filter Đo lượng thông tin chung
Forward Selection Wrapper Thêm feature từng cái, chọn cái tốt nhất
Backward Elimination Wrapper Bỏ feature từng cái, bỏ cái kém nhất
L1 Regularization Embedded Lasso tự động zero-out features
Tree Importance Embedded Feature importance từ Random Forest

4.6. Discretization

Chuyển continuous → categorical

    Binning Methods:
    
    1. Equal-width (equal interval):
       price: [0-100], [100-200], [200-300], [300-400], [400-500]
       → Đơn giản nhưng sensitive với outliers
    
    2. Equal-frequency (equal depth):
       price: Mỗi bin có 20% dữ liệu
       → Cân bằng hơn, nhưng mất ý nghĩa range
    
    3. Entropy-based (supervised):
       Chọn split points tối ưu cho classification target
       → Tốt nhất cho ML, nhưng cần label
    
    4. Clustering-based:
       Dùng K-Means để tìm natural groups
       → Data-driven, phù hợp cho khám phá

4.7. Chuẩn hóa Dữ liệu (Normalization)

    1. Min-Max Normalization:
       x' = (x - min) / (max - min)
       Range: [0, 1]
       Ưu: Giữ nguyên distribution shape
       Nhược: Sensitive với outliers
    
    2. Z-Score Standardization:
       x' = (x - μ) / σ
       Range: unbounded (typically -3 to +3)
       Ưu: Robust hơn với outliers
       Nhược: Không bounded
    
    3. Decimal Scaling:
       x' = x / 10^j (j = smallest int so that max(|x'|) < 1)
       VD: 345 → 0.345 (j=3)
    
    4. Robust Scaling:
       x' = (x - median) / IQR
       Ưu: Rất robust với outliers
       Nhược: Ít phổ biến

4.8. Feature Engineering

Tạo features mới từ dữ liệu Olist

    Raw Data                         Engineered Features
    ┌──────────────────────┐        ┌──────────────────────────────┐
    │ order_purchase_ts    │   →    │ hour_of_day                  │
    │                      │        │ day_of_week                  │
    │                      │        │ is_weekend                   │
    │                      │        │ month                        │
    │                      │        │ days_since_first_order       │
    ├──────────────────────┤        ├──────────────────────────────┤
    │ price, freight       │   →    │ freight_ratio (freight/price)│
    │                      │        │ price_category (binned)      │
    │                      │        │ is_free_shipping             │
    ├──────────────────────┤        ├──────────────────────────────┤
    │ review_comment       │   →    │ comment_length               │
    │                      │        │ has_comment (bool)           │
    │                      │        │ sentiment_score              │
    ├──────────────────────┤        ├──────────────────────────────┤
    │ customer_city,       │   →    │ distance_to_seller           │
    │ seller_city          │        │ same_state (bool)            │
    │                      │        │ region_pair                  │
    ├──────────────────────┤        ├──────────────────────────────┤
    │ product dimensions   │   →    │ volume                       │
    │ (l, w, h, weight)   │        │ density (weight/volume)      │
    │                      │        │ is_heavy (>5kg)              │
    └──────────────────────┘        └──────────────────────────────┘

🔬 Labs


📝 Câu hỏi ôn tập

  1. Liệt kê 6 chiều chất lượng dữ liệu. Cho ví dụ vi phạm với dataset Olist.
  2. So sánh 5 phương pháp xử lý Missing Values. Khi nào dùng phương pháp nào?
  3. IQR method phát hiện outlier như thế nào? Tính toán cho cột price có Q1=50, Q3=200.
  4. PCA giảm chiều như thế nào? Giải thích "explained variance ratio".
  5. Cho 3 ví dụ Feature Engineering cho dataset Olist mà có ý nghĩa business.