Basics of Machine Learning
y -> outcome (position of the planet)
ML GOAL: Predict y as a function of x gives a training set
2 main ways:
- Model based
- Instance based
Model Based
- I choose model , is the parameters of the model.
- Estimate , given a new point x* not previously observed,
- e.g. Regression
Instance Based
- establish a similarity metric
- when a new is given, you compute the corresponding by averaging corresponding to “close” to
- e.g. KNN
Supervised Learning
is given, have labels.
- regression
- KNN
- Decision trees
- SVM
Unsupervised Learning
is given, We want to find the structure of .
- K-means
- PCA
Batch Learning
Online Learning
keep updating data
Least Square
KNN :
How KNN is related to
- expectation is replaced by empirical mean
- the auditioning is relaxed to k points
you can show
Curse of Dimensionality
Big data are data usually defined in very high dimension space. You can show that the median distance from the origin to he closest data point is given by
is an unbiased estimator of
Bias-Variance Decomposition
Let be my model for
since the cross term is zero, use tower property of expectation to prove
Penalization Methods
Ridge Regression
LASSO Regression
Cross Validation
- fix
- divide the T (training set) into K groups of equal size
- group j train the data on the remaining K-1 groups and compute the (mean squared error of group under )
Big will have high variance.
small will have high bias.