Chương 5: Thuật toán Khai thác Dữ liệu Cốt lõi
Core Data Mining Algorithms
📚 Mục tiêu học tập (CLO3, CLO4, CLO5, CLO7, CLO8)
Sau khi hoàn thành chương này, sinh viên có thể:
- Implement thuật toán Apriori và FP-Growth cho Association Rules
- Xây dựng Decision Tree (ID3, C4.5) cho bài toán Classification
- Áp dụng Naive Bayes cho phân loại văn bản/dữ liệu
- Thực hiện K-Means và K-Medoids cho Clustering
- Sử dụng DBSCAN cho density-based clustering
- Phân tích outlier bằng phương pháp thống kê và khoảng cách
5.1. Khai thác Luật kết hợp (Association Rules)
Các khái niệm cơ bản
Transaction Database:
TID | Items
─────┼──────────────────────
T1 | {Bread, Milk, Butter}
T2 | {Bread, Milk}
T3 | {Milk, Eggs}
T4 | {Bread, Milk, Eggs}
T5 | {Bread, Butter}
Itemset: Tập các items (VD: {Bread, Milk})
Support(X) = P(X) = count(X) / N
VD: Support({Bread, Milk}) = 3/5 = 0.6 (60%)
Confidence(X → Y) = P(Y|X) = Support(X∪Y) / Support(X)
VD: Confidence({Bread} → {Milk}) = Support({Bread,Milk}) / Support({Bread})
= 3/5 / 4/5 = 0.75 (75%)
Lift(X → Y) = Confidence(X→Y) / Support(Y)
VD: Lift({Bread}→{Milk}) = 0.75 / 0.8 = 0.9375
Lift > 1: X và Y có quan hệ positive
Lift = 1: X và Y independent
Lift < 1: X và Y có quan hệ negative
5.2. Thuật toán Apriori
Nguyên lý Apriori
Nếu một itemset không frequent, thì mọi superset của nó cũng không frequent.
Ngược lại: mọi subset của frequent itemset đều phải frequent.
Apriori Algorithm (min_support = 0.4):
Bước 1: Frequent 1-itemsets (C1 → L1)
┌──────────┬──────────┬─────────┐
│ Itemset │ Support │ Frequent│
├──────────┼──────────┼─────────┤
│ {Bread} │ 4/5=0.80 │ ✅ │
│ {Milk} │ 4/5=0.80 │ ✅ │
│ {Butter} │ 2/5=0.40 │ ✅ │
│ {Eggs} │ 2/5=0.40 │ ✅ │
└──────────┴──────────┴─────────┘
Bước 2: Generate candidates C2 từ L1, scan DB
┌────────────────┬──────────┬─────────┐
│ Itemset │ Support │ Frequent│
├────────────────┼──────────┼─────────┤
│ {Bread, Milk} │ 3/5=0.60 │ ✅ │
│ {Bread,Butter} │ 2/5=0.40 │ ✅ │
│ {Bread, Eggs} │ 1/5=0.20 │ ❌ prune│
│ {Milk, Butter} │ 1/5=0.20 │ ❌ prune│
│ {Milk, Eggs} │ 2/5=0.40 │ ✅ │
│ {Butter, Eggs} │ 0/5=0.00 │ ❌ prune│
└────────────────┴──────────┴─────────┘
Bước 3: C3 từ L2 → {Bread, Milk, Butter}?
Check subsets: {Bread,Butter}✅ {Milk,Butter}❌ → Prune!
→ L3 = ∅ → STOP
Kết quả Frequent Itemsets:
L1: {Bread}, {Milk}, {Butter}, {Eggs}
L2: {Bread,Milk}, {Bread,Butter}, {Milk,Eggs}
FP-Growth (cải tiến Apriori)
FP-Growth vs Apriori:
Apriori: FP-Growth:
- Scan DB nhiều lần - Scan DB chỉ 2 lần
- Generate candidates - KHÔNG generate candidates
- Chậm với large DB - Nhanh hơn nhiều
- Dễ hiểu - Phức tạp hơn
FP-Growth xây dựng FP-Tree (compact representation)
rồi mine frequent patterns trực tiếp từ tree.
5.3. Classification: Decision Trees
Information Gain và Entropy
Entropy(S) = -Σ p_i * log2(p_i)
VD: Dataset có 9 positive, 5 negative
Entropy = -9/14 * log2(9/14) - 5/14 * log2(5/14) = 0.940
Information Gain(S, A) = Entropy(S) - Σ |Sv|/|S| * Entropy(Sv)
Chọn attribute có Information Gain cao nhất để split!
ID3 Algorithm
ID3 Decision Tree cho Olist (đơn giản hóa):
Target: Is customer satisfied? (review_score >= 4)
┌──────────────┐
│delivery_days │
│ IG = 0.32 │
└──────┬───────┘
│
┌───────────┼───────────┐
│ │ │
≤ 7 days 8-14 days > 14 days
│ │ │
┌────▼────┐ ┌────▼────┐ ┌────▼────┐
│ price │ │Satisfied│ │ NOT │
│IG=0.15 │ │ (68%) │ │Satisfied│
└────┬────┘ └─────────┘ │ (82%) │
│ └─────────┘
┌────┼────┐
│ │
≤ 200 > 200
│ │
┌────▼────┐ ┌──▼──────┐
│Satisfied│ │ Check │
│ (90%) │ │ freight │
└─────────┘ └─────────┘
ID3: Chỉ dùng categorical (discretize nếu cần)
C4.5: Hỗ trợ continuous attributes, handling missing values
C4.5 Improvements over ID3
| Feature | ID3 | C4.5 |
|---|---|---|
| Attributes | Categorical only | Categorical + Continuous |
| Missing values | Not handled | Uses probability weighting |
| Overfitting | Not addressed | Post-pruning (error-based) |
| Splitting criterion | Information Gain | Gain Ratio = IG / SplitInfo |
| Multi-way split | Yes | Yes + Binary split for continuous |
Gain Ratio = Information_Gain(A) / SplitInformation(A)
SplitInfo(A) = -Σ |Sv|/|S| * log2(|Sv|/|S|)
→ Penalize attributes with many values (fix IG bias)
5.4. Naive Bayes Classification
Bayes' Theorem
P(C|X) = P(X|C) * P(C) / P(X)
Posterior = Likelihood × Prior / Evidence
Naive assumption: features are conditionally independent
P(X|C) = P(x1|C) × P(x2|C) × ... × P(xn|C)
Ví dụ: Predict review_score ≥ 4 (satisfied)
Given: delivery_days=5, price=150, state=SP
P(satisfied | features) ∝ P(delivery=5|sat) × P(price=150|sat) × P(SP|sat) × P(sat)
P(unsatisfied | features) ∝ P(delivery=5|unsat) × P(price=150|unsat) × P(SP|unsat) × P(unsat)
Chọn class có probability cao hơn.
Types of Naive Bayes
| Type | Use case | P(x|C) | |------|----------|--------| | Gaussian NB | Continuous features | Normal distribution | | Multinomial NB | Text/Count data | Multinomial distribution | | Bernoulli NB | Binary features | Bernoulli distribution | | Complement NB | Imbalanced text | Complement of class |
5.5. K-Means Clustering
Algorithm
K-Means Algorithm:
Input: K (số clusters), Data points
1. Khởi tạo K centroids (random hoặc K-Means++)
2. REPEAT:
a. ASSIGN: Gán mỗi point vào cluster có centroid gần nhất
b. UPDATE: Tính lại centroid = mean of all points trong cluster
3. UNTIL centroids không đổi (hoặc max iterations)
Iteration 0: Iteration 1: Converged:
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ * │ │ * . . │ │ * . . │
│ . . C1 │ │ . C1 │ │ .C1 . │
│ . . . │ │ . . . │ │ . . . │
│ C2 │ │ │ │ │
│ . . . . │ │ . . C2. . │ │ . . C2. . │
│ . . │ │ . . │ │ . . │
└──────────────┘ └──────────────┘ └──────────────┘
C = Centroid, . = Data point
K-Means++: Khởi tạo thông minh
K-Means++ (Arthur & Vassilvitskii, 2007):
1. Chọn centroid đầu tiên random
2. Cho mỗi centroid tiếp theo:
- Tính D(x) = khoảng cách từ x đến centroid gần nhất
- Chọn point mới với xác suất ∝ D(x)²
→ Centroids xa nhau → converge nhanh hơn
Chọn K tối ưu
Elbow Method: Silhouette Method:
Inertia Silhouette Score
│ │
│\ │ *
│ \ │ * *
│ \ │ * *
│ \ │* *
│ \_____ │
│ ↑ │
│ Elbow │
│ \_____ │
└─────────────── K └─────────────── K
K* = point where adding K* = max silhouette
more clusters gives s(i) = (b(i)-a(i)) / max(a(i),b(i))
diminishing returns a(i) = avg distance to same cluster
b(i) = avg distance to nearest other cluster
5.6. DBSCAN (Density-Based Clustering)
Nguyên lý hoạt động
DBSCAN Parameters:
- ε (epsilon): Bán kính neighborhood
- MinPts: Số điểm tối thiểu trong ε-neighborhood
3 loại điểm:
┌──────────────────────────────────────┐
│ │
│ Core Point (≥ MinPts trong ε) │
│ ┌─────┐ │
│ │ ●──┼──. . . (≥5 points) │
│ │ ε │ │
│ └─────┘ │
│ │
│ Border Point (< MinPts nhưng │
│ trong ε của Core Point) │
│ ○ ← border │
│ ↓ │
│ ┌─────┐ │
│ │ ● │ core │
│ └─────┘ │
│ │
│ Noise Point (không thuộc cluster) │
│ ✕ ← noise/outlier │
│ │
└──────────────────────────────────────┘
DBSCAN Algorithm:
1. Cho mỗi point chưa visited:
a. Đánh dấu visited
b. Tìm ε-neighborhood
c. Nếu ≥ MinPts → Core point → Tạo/mở rộng cluster
d. Nếu < MinPts → Tạm đánh dấu noise (có thể thành border sau)
DBSCAN vs K-Means:
┌────────────────┬─────────────┬──────────────┐
│ Feature │ K-Means │ DBSCAN │
├────────────────┼─────────────┼──────────────┤
│ # Clusters │ Phải chỉ K │ Tự phát hiện │
│ Cluster shape │ Spherical │ Arbitrary │
│ Outlier handle │ Không │ Có (noise) │
│ Scalability │ O(nKt) │ O(n²) worst │
│ Parameters │ K │ ε, MinPts │
└────────────────┴─────────────┴──────────────┘
5.7. Phân cụm Phân cấp (Hierarchical Clustering)
Agglomerative (Bottom-up):
Start: Mỗi point = 1 cluster
Step 1: {A}{B}{C}{D}{E} Merge closest: A,B
Step 2: {A,B}{C}{D}{E} Merge closest: D,E
Step 3: {A,B}{C}{D,E} Merge closest: C,{D,E}
Step 4: {A,B}{C,D,E} Merge closest: {A,B},{C,D,E}
Step 5: {A,B,C,D,E} Done!
Dendrogram:
─────────────────────────
│ ┌───────────┤
│ ┌─────┤ │
│ │ │ ┌──────┤
│ ┌──┤ │ │ ┌───┤
─ A B C D E
Linkage Methods:
- Single: min distance between clusters
- Complete: max distance
- Average: mean distance (UPGMA)
- Ward's: minimize variance increase
5.8. Outlier Detection
Phương pháp phát hiện Outlier
| Method | Type | Mô tả |
|---|---|---|
| Z-Score | Statistical | |z| > 3 → outlier |
| Modified Z-Score | Statistical (robust) | Uses MAD instead of std |
| IQR | Statistical | < Q1-1.5IQR or > Q3+1.5IQR |
| Isolation Forest | ML-based | Random splits, outliers isolated faster |
| LOF | Distance-based | Local Outlier Factor, density-based |
| DBSCAN noise | Density-based | Points classified as noise |
| Mahalanobis Distance | Statistical | Accounts for correlation |
🔬 Labs
lab-05-apriori.py— Apriori & FP-Growth trên Olistlab-05-decision-tree.py— Decision Tree Classificationlab-05-naive-bayes.py— Naive Bayes Classificationlab-05-kmeans.py— K-Means Customer Segmentationlab-05-dbscan.py— DBSCAN Anomaly Detection
📝 Câu hỏi ôn tập
- Tính Support, Confidence và Lift cho rule {A} → {B} biết: N=100, count(A)=40, count(B)=50, count(A∪B)=25
- Giải thích nguyên lý Apriori. Tại sao nó giúp giảm số lượng candidates?
- So sánh Information Gain (ID3) và Gain Ratio (C4.5). Tại sao Gain Ratio tốt hơn?
- K-Means có nhược điểm gì? K-Means++ cải thiện như thế nào?
- DBSCAN phát hiện outlier như thế nào? So sánh với K-Means về handling outliers.
- Cho dataset: [1,2,3,100,2,3,1,2,500]. Dùng IQR method, xác định outliers.