REUMass Amherst 2015 Data Science Bootcamp
Day 3: Classification, Regression,
and Model Selection
Prof. Ben Marlin [email protected]
Plan for Day 3: • More on Classifica4on • Introduc4on to Regression • Model Selec4on and Hyperparameters
Classifier Performance
K Nearest Neighbors
• Need to pick number of neighbors K.
• Need to define a distance func8on.
x2 Data
x1
x1<4
0 1 2 3 4 5 6 7
0 1
2 3
4 5
6
Decision Trees
x2 Data
x1
x1<4
x2<2
Y
0 1 2 3 4 5 6 7
0 1
2 3
4 5
6
Decision Trees
x2 Data
x1
x1<4
x2<2
Y
Y N
0 1 2 3 4 5 6 7
0 1
2 3
4 5
6
Decision Trees
x2 Data
x1
x1<4
x2<2 x1<6
Y
Y N
N
0 1 2 3 4 5 6 7
0 1
2 3
4 5
6
Decision Trees
x2 Data
x1
x1<4
x2<2 x1<6
Y
Y N
N
Y N
0 1 2 3 4 5 6 7
0 1
2 3
4 5
6
Decision Trees
x1<4
x2<2 x1<6
Y
Y N
N
Y N
Decision Trees
• Need to choose the maximum depth of the tree.
• Need to choose a spliFng criterion.
Linear Classifiers
Linear Classifiers
Linear Classifiers
Linear Classifiers
Non-‐Linear SVMs and Basis Expansion
i
Non-‐Linear SVMs and Kernels i
Example: Stock Prices
Example: Climate Data
The Regression Learning Problem
KNN Regression (K=1 vs K=9)
KNN Regression (K=1 vs K=9)
Linear Regression
Learning Linear Regression Models
Linear Regression
Non-‐Linear Regression With Kernels
Model Selec4on & Hyperparameters
Hyperparameter and Model Selec4on:
• All of the models we’ve seen have some free parameters (called hyperparameters) that need to be chosen.
• Think of the K in KNN, the depth of a decision tree, the kernel and regulariza8on parameter in an SVM.
• The most common methods for choosing hyperparameters in machine learning are valida8on set methods.
Train
Valida8on
Test
Data Cases
Features
Train models
Pick Best Model
Evaluate Best Model
B3
B2
B1
Train
Val
Learn
Test
Data Cases
Features
B3
B2
B1
Train
Val
Data Cases
Features
Learn
Test
B3
B2
B1 Train
Val Data Cases
Features
Learn
Test