2.3 KiB
Executable File
2.3 KiB
Executable File
Logarithms
log_2X used for generating decision trees
- Power to which we have to raise 2 to get X
- When using, X will be probability between 0 and 1
- log of probability is always negative
Decision Tree for Contact Lenses
- Upside down
- Ellipse at top = root
- Edges = branches
- Rectangles = leaves
- Leaves assign classification
Strategy
- Grow trees from root
- Top-Down
- More specific as grown, described as general-to-specific
- Divide and Conquer
- Stop if all examples have same class
- How is attribute or root node selected?
- Consider how to generate decision tree for weather dataset with nominal values only.
Weather Dataset
Criterion for Attribute Selection
- Which is best?
- Smallest tree
- Heuristic: choose attribute which produces purest nodes
- Information gain popular criteria for measuring impurity
- Increases with average purity of subsets
- Choose attribute that gives greatest information gain
Information
- Expected amount of information needed to specify whether new example should be classified as yes or no, given it reached that node.
Computing Information
- Measure information in bits
- Given probability distribution, info required to predict an event is the distributions entropy
- Entropy gives the information required in bits
I(p_1,p_2,...,p_n)=-p_{1}\log_{2}p_1 -p_{2}\log_{2}p_2 ... -p_{n}\log_{2}p_n
Where n = number of classes, and p_1 + p_2 + ... p_{n} = 1
Minus signs included since output must be positive
Expected Information for Outlook
- Outlook = Sunny
info([2,3]) = I(\frac{2}{5},\frac{3}{5}) = -\frac{2}{5}\log_2(\frac{2}{5}) - \frac{3}{5}\log_2(\frac{3}{5}) = 0.971 bits
- Outlook = Overcast
info([4,0]) = I(\frac{4}{4},\frac{0}{4}) = -1\log_2(1) -0\log_2(0) = 0 bits
- Outlook = Rainy
info([3,2]) = I(\frac{3}{5},\frac{2}{5}) = -\frac{3}{5}\log_2(\frac{3}{5}) - \frac{2}{5}\log_2(\frac{2}{5}) = 0.693 bits
Computing Information Gain
Information before splitting - information after splitting
gain(outlook) = info([9,5])-E(Outlook) = 0.940 - 0.693 = 0.247
Information Gain for Attributes of Weather Data
gain(Outlook) = 0.247 gain(Temperature) = 0.029 gain(Humidity) = 0.152 gain(Windy) = 0.048
Outlook Selected for root because gains most information