View
213
Download
0
Tags:
Embed Size (px)
Citation preview
Decision tree
LING 572
Fei Xia
1/10/06
Outline
• Basic concepts
• Main issues
• Advanced topics
Basic concepts
A classification problem
District House type
Income Previous
Customer
Outcome
Suburban Detached High No Nothing
Suburban Semi-detached
High Yes Respond
Rural Semi-detached
Low No Respond
Urban Detached Low Yes Nothing
…
Classification and estimation problems
• Given– a finite set of (input) attributes: features
• Ex: District, House type, Income, Previous customer
– a target attribute: the goal• Ex: Outcome: {Nothing, Respond}
– training data: a set of classified examples in attribute-value representation
• Ex: the previous table
• Predict the value of the goal given the values of input attributes– The goal is a discrete variable classification problem– The goal is a continuous variable estimation problem
DistrictSuburban
(3/5)Rural(4/4)
Urban(3/5)
RespondHouse type
Detac
hed
(2/2
)
Nothing
Previous customer
Yes(3/3) N
o (2/2)
Semi-detached
(3/3)
Respond Nothing Respond
Decision tree
Decision tree representation
• Each internal node is a test:– Theoretically, a node can test multiple attributes– In most systems, a node tests exactly one attribute
• Each branch corresponds to test results– A branch corresponds to an attribute value or a range
of attribute values
• Each leaf node assigns – a class: decision tree– a real value: regression tree
What’s a best decision tree?
• “Best”: You need a bias (e.g., prefer the “smallest” tree): least depth? Fewest nodes? Which trees are the best predictors of unseen data?
• Occam's Razor: we prefer the simplest hypothesis that fits the data.
Find a decision tree that is as small as possible and fits the data
Finding a smallest decision tree
• A decision tree can represent any discrete function of the inputs: y=f(x1, x2, …, xn)
• The space of decision trees is too big for systemic search for a smallest decision tree.
• Solution: greedy algorithm
Basic algorithm: top-down induction
• Find the “best” decision attribute, A, and assign A as decision attribute for node
• For each value of A, create new branch, and divide up training examples
• Repeat the process until the gain is small enough
Major issues
Major issues
Q1: Choosing best attribute: what quality measure to use?
Q2: Determining when to stop splitting: avoid overfitting
Q3: Handling continuous attributesQ4: Handling training data with missing
attribute valuesQ5: Handing attributes with different costsQ6: Dealing with continuous goal attribute
Q1: What quality measure
• Information gain
• Gain Ratio
• …
Entropy of a training set
• S is a sample of training examples
• Entropy is one way of measuring the impurity of S
• Pc is the proportion of examples in S whose target attribute has value c.
cc
c ppSEntropy log)(
Information Gain
• Gain(S,A)=expected reduction in entropy due to sorting on A.
• Choose the A with the max information gain.
(a.k.a. choose the A with the min average entropy)
)(||
||)(),(
)(v
AValuesv
vdef
SEntropyS
SSEntropyASGain
An example
HumidityHigh Normal
S=[9+,5-]E=0.940
[3+,4-] [6+,1-]
InfoGain(S, Humidity)=0.940-(7/14)*0.985-(7/14)0.592=0.151
WindWeak Strong
S=[9+,5-]E=0.940
[6+,2-] [3+,3-]
InfoGain(S, Wind)=0.940-(8/14)*0.811-(6/14)*1.0=0.048
E=0.985 E=0.592 E=0.811 E=1.00
Other quality measures
• Problem of information gain:– Information Gain prefers attributes with many values.
• An alternative: Gain Ratio
Where Si is subset of S for which A has value vi.
||
||log||
||),(
),(
),(),(
21 S
S
S
SASmationSplitInfor
ASmationSplitInfor
ASGainASGainRatio
ic
i
i
Q2: avoiding overfitting
• Overfitting occurs when our decision tree characterizes too much detail, or noise in our training data.
• Consider error of hypothesis h over– Training data: ErrorTrain(h)– Entire distribution D of data: ErrorD(h)
• A hypothesis h overfits training data if there is an alternative hypothesis h’, such that– ErrorTrain(h) < ErrorTrain(h’), and– ErrorD(h) > errorD(h’)
How to avoiding Overfitting
• Stop growing the tree earlier – Ex: InfoGain < threshold– Ex: Size of examples in a node < threshold– …
• Grow full tree, then post-prune
In practice, the latter works better than the former.
Post-pruning
• Split data into training and validation set• Do until further pruning is harmful:
– Evaluate impact on validation set of pruning each possible node (plus those below it)
– Greedily remove the ones that don’t improve the performance on validation set
Produces a smaller tree with best performance measure
Performance measure
• Accuracy:– on validation data– K-fold cross validation
• Misclassification cost: Sometimes more accuracy is desired for some classes than others.
• MDL: size(tree) + errors(tree)
Rule post-pruning
• Convert tree to equivalent set of rules
• Prune each rule independently of others
• Sort final rules into desired sequence for use
• Perhaps most frequently used method (e.g., C4.5)
Q3: handling numeric attributes
• Continuous attribute discrete attribute
• Example– Original attribute: Temperature = 82.5– New attribute: (temperature > 72.3) = t, f
Question: how to choose thresholds?
Choosing thresholds for a continuous attribute
• Sort the examples according to the continuous attribute.
• Identify adjacent examples that differ in their target classification a set of candidate thresholds
• Choose the candidate with the highest information gain.
Q4: Unknown attribute values
• Assume an attribute can take the value “blank”.
• Assign most common value of A among training data at node n.
• Assign most common value of A among training data at node n which have the same target class.
• Assign prob pi to each possible value vi of A– Assign a fraction (pi) of example to each descendant in tree– This method is used in C4.5.
Q5: Attributes with cost
• Consider medical diagnosis (e.g., blood test) has a cost
• Question: how to learn a consistent tree with low expected cost?
• One approach: replace gain by – Tan and Schlimmer (1990)
)(
),(2
ACost
ASGain
Q6: Dealing with continuous goal attribute Regression tree
• A variant of decision trees
• Estimation problem: approximate real-valued functions: e.g., the crime rate
• A leaf node is marked with a real value or a linear function: e.g., the mean of the target values of the examples at the node.
• Measure of impurity: e.g., variance, standard deviation, …
Summary of Major issues
Q1: Choosing best attribute: different quality measure.
Q2: Determining when to stop splitting: stop earlier or post-pruning
Q3: Handling continuous attributes: find the breakpoints
Summary of major issues (cont)
Q4: Handling training data with missing attribute values: blank value, most common value, or fractional count
Q5: Handing attributes with different costs: use a quality measure that includes the cost factors.
Q6: Dealing with continuous goal attribute: various ways of building regression trees.
Common algorithms
• ID3
• C4.5
• CART
ID3
• Proposed by Quinlan (so is C4.5)
• Can handle basic cases: discrete attributes, no missing information, etc.
• Information gain as quality measure
C4.5
• An extension of ID3:– Several quality measures– Incomplete information (missing attribute
values)– Numerical (continuous) attributes– Pruning of decision trees– Rule derivation– Random mood and batch mood
CART
• CART (classification and regression tree)
• Proposed by Breiman et. al. (1984)
• Constant numerical values in leaves
• Variance as measure of impurity
Strengths of decision tree methods
• Ability to generate understandable rules
• Ease of calculation at classification time
• Ability to handle both continuous and categorical variables
• Ability to clearly indicate best attributes
The weaknesses of decision tree methods
• Greedy algorithm: no global optimization
• Error-prone with too many classes: numbers of training examples become smaller quickly in a tree with many levels/branches.
• Expensive to train: sorting, combination of attributes, calculating quality measures, etc.
• Trouble with non-rectangular regions: the rectangular classification boxes that may not correspond well with the actual distribution of records in the decision space.
Advanced topics
Combining multiple models
• The inherent instability of top-down decision tree induction: different training datasets from a given problem domain will produce quite different trees.
• Techniques:– Bagging– Boosting
Bagging
• Introduced by Breiman
• It first creates multiple decision trees based on different training sets.
• Then, it compares each tree and incorporates the best features of each.
• This addresses some of the problems inherent in regular ID3.
Boosting
• Introduced by Freund and Schapire
• It examines the trees that incorrectly classify an instance and assign them a weight.
• These weights are used to eliminate hypotheses or refocus the algorithm on the hypotheses that are performing well.
Summary
• Basic case: – Discrete input attributes– Discrete goal attribute– No missing attribute values– Same cost for all tests and all kinds of misclassification.
• Extended cases:– Continuous attributes– Real-valued goal attribute– Some examples miss some attribute values– Some tests are more expensive than others.– Some misclassifications are more serious than others.
Summary (cont)
• Basic algorithm: – greedy algorithm– top-down induction– Bias for small trees
• Major issues:
Uncovered issues
• Incremental decision tree induction?
• How can a decision relate to other decisions? what's the order of making the decisions? (e.g., POS tagging)
• What's the difference between decision tree and decision list?