6001 sheet

下面是一份 超级齐全的 Machine Learning Cheat Sheet（中英对照），
覆盖你前面所有 讲义 8 章总结 + 150 道预测题内容 所涉及的全部知识点。

---

1. Machine Learning Taxonomy 机器学习分类

Supervised Learning 监督学习

• Uses labeled data (X, y)
  使用带标签的数据
• Tasks: Classification, Regression
  分类、回归
• Examples: spam detection, churn prediction, house price
  垃圾邮件检测、流失预测、房价预测

Unsupervised Learning 无监督学习

• Uses only X
  只使用特征数据
• Tasks: Clustering, Dimensionality Reduction
  聚类、降维
• Examples: customer segmentation, PCA, anomaly detection
  客户群分、PCA、异常检测

---

2. Classification vs Regression 分类与回归

Classification 分类

• Output is discrete
  输出是离散标签
• Examples: spam/not spam, fraud detection
  垃圾邮件、欺诈检测

Regression 回归

• Output is continuous
  输出连续数值
• Examples: price, temperature
  价格、温度

---

3. Important Algorithms 关键算法

Logistic Regression

• Binary classification
  二分类
• Uses sigmoid → probability (0–1)
  用 sigmoid 输出概率
• Linear decision boundary
  线性决策边界

KNN

• Non-parametric, lazy learner
  非参数、惰性学习
• Sensitive to feature scaling
  对特征缩放敏感
• Predict by nearest neighbors
  通过最近邻预测（投票/平均）

Decision Tree

• Uses Gini / Entropy
  使用 Gini/Entropy
• Easy to interpret
  可解释性强
• Overfits if not pruned
  不剪枝易过拟合

Random Forest

• Bagging of decision trees
  树的 Bagging
• Reduces variance
  降低方差
• More stable than single tree
  比单棵树更稳

K-Means

• Requires K
  必须指定 K
• Minimizes within-cluster distance
  最小化簇内距离
• Sensitive to scale
  对缩放敏感

PCA

• Unsupervised dimensionality reduction
  无监督降维
• Keeps directions with highest variance
  保留最大方差方向
• Removes correlated features
  去相关

---

4. Key Concepts & Terminology 核心概念

Sigmoid

• Output range: 0–1
  范围 0–1

Decision Boundary

• The line/plane separating classes
  分类边界线/面

Centroid (K-Means)

• Cluster center
  聚类中心

False Positive / False Negative

FP = predicted positive but actually negative
预测正类但实际为负类（误报）
FN = predicted negative but actually positive
预测负类但实际为正类（漏报）

---

5. Model Evaluation 模型评估

Confusion Matrix

	Pred +	Pred –
Actual +	TP	FN
Actual –	FP	TN

---

Metrics 指标

Accuracy 准确率

Not reliable under imbalance
类别不平衡时不可靠

Precision 精确率

TP / (TP + FP)
关注预测为正的是否正确

Recall 召回率

TP / (TP + FN)
关注是否漏掉正例

F1-score

Harmonic mean of P & R
精确率与召回率的调和平均

AUC-ROC

Probability model ranks positive > negative
衡量分类器排序能力

RMSE / MSE (Regression)

Measure error for regression
回归误差指标

---

6. Overfitting vs Underfitting 过拟合与欠拟合

Overfitting 过拟合

• High train accuracy
• Low test accuracy
  训练高、测试低
• Model too complex

Underfitting 欠拟合

• Low train
• Low test
  训练和测试都差
• Model too simple

---

7. How to Reduce Overfitting 如何减少过拟合

✓ Regularization (L1 / L2) 正则化
✓ Early stopping 提前停止
✓ More training data 更多数据
✓ Cross-validation 交叉验证
✓ Simpler model 简单模型
✗ Increasing model complexity 会更严重

---

8. Regularization 正则化

L1 (Lasso)

• Shrinks weights to zero → feature selection
  可做特征选择

L2 (Ridge)

• Shrinks weights but not zero
  收缩权重，不变零

Purpose: prevent overfitting
目的：减少过拟合

---

9. Feature Engineering 特征工程

Scaling / Normalization

标准化与归一化

• Needed for KNN, KMeans, Logistic Regression
  KNN/KMeans/逻辑回归需要缩放

One-Hot Encoding

独热编码

• For categorical features
  处理类别特征

Handling Missing Values

处理缺失值

• Mean/median imputation 均值/中位数填补
• KNN imputer
• Drop columns/rows（合理时）

---

10. Cross-Validation 交叉验证

K-Fold CV

• Split into K folds
• Train K times
• Stable evaluation
  稳健评估方式

Train/Validation/Test Split

• Train: fit model
• Validation: tune hyperparameters
• Test: final evaluation
  训练 → 验证 → 测试

---

11. Data Leakage 数据泄漏

Avoid:

• Scaling before split
• Using target-derived features
• Fitting preprocessing on full dataset

避免：

• 在划分前缩放
• 用包含未来/标签的信息
• 在全体数据上 fit 预处理

---

12. Imbalanced Data 不平衡数据

To handle imbalance:
处理方法：

• Class weights 类别权重
• Oversampling (SMOTE) 过采样
• Undersampling 欠采样
• Precision/Recall/PR Curve 专注正类指标

Avoid accuracy as only metric
不要只看准确率

---

13. Gradient Descent 梯度下降

Too large LR → divergence

学习率太大→不收敛

Too small LR → slow

太小→极慢

Used in:

• Logistic regression
• Linear regression
• Neural networks

---

14. Common Exam Patterns 常考题型总结

✔ Identify supervised vs unsupervised
✔ Classification vs regression
✔ Sigmoid range
✔ Interpretation of FP/FN
✔ When to use precision vs recall
✔ KNN sensitivity to scaling
✔ PCA purpose
✔ K-Means requires K
✔ Overfitting signs
✔ Methods to reduce overfitting
✔ Use of cross-validation
✔ Decision tree impurity metrics
✔ Handling missing values
✔ Imbalanced dataset metrics
✔ Train-test split usage

With this cheat sheet, you can answer ALL the predictions questions.

---