• Google Crash Course

Supervised learning

train using qusetion-answer pairs.

Generalized linear model https://www.bilibili.com/video/BV13X4y1R7im

Regression

Linear regression $y=wx+b$$J=\frac{\sum (y^{\prime }-y{)}^2}{n}$

Logistic regression $z=wx+b$$y=\frac{1}{1+e^{-z}}$ Log Loss

Classification

A example usually contains features and a label. Feature indicates question, label indicates answer.

Parameters control predicting, hyperparameters control training.

• Learning rate
• Batch size: the number of examples used to update parameters once.
• Epochs: an epoch means that every example have been used once.

Metric
Accuracy	$\frac{TP+TN}{ALL}$
TPR, Recall	$\frac{TP}{TP+FN}$
FPR	$\frac{FP}{FP+TN}$
Precision	$\frac{TP}{TP+FP}$

Unsupervised learning

train using questions only.

clustering

---

Numerical data

• Clear
• Normalization

Categorical data