Performance. Learning Classifiers. Instances x i in dataset D mapped to feature space:

Transcription

1 Learning Classifiers Performance Instances x i in dataset D mapped to feature space: Initial Model (Assumptions) Training Process Trained Model Testing Process Tested Model Performance Performance Training Data Test Data decision boundary Classes associated with instances: X, Classification: f(x i ) = c {X, } with x i,j {, }, and f classifier dataset D is a multiset Objective: learn f (supervised) Performance measures: TP: True Positive TN: True Negative FP: False Positive FN: False Negative Note: if N = D, then N = TP+TN+FP+FN Confusion matrix: Predicted class Actual true positive false negative class false positive true negative Performance Performance measures: Success rate σ: TP + TN σ = N Error rate ɛ: ɛ = 1 σ TPR (= recall ρ) True Positive Rate TPR = TP/(TP + FN) FNR False Negative Rate: FNR = 1 TPR FPR False Positive Rate: FPR = FP/(FP + TN) TNR True Negative Rate: TNR = 1 FPR Precision π: π = TP/(TP + FP) F -measure: F = 2 ρ π ρ + π = 2TP 2TP + FP + FN Example: choosing contact lenses Tear production Age prescription Ast rate Lens young myope reduced ne young myope rmal soft young myope reduced ne young myope rmal hard young hypermetrope reduced ne young hypermetrope rmal soft young hypermetrope reduced ne young hypermetrope rmal hard pre-presbyopic myope reduced ne pre-presbyopic myope rmal soft pre-presbyopic myope reduced ne pre-presbyopic myope rmal hard pre-presbyopic hypermetrope reduced ne pre-presbyopic hypermetrope rmal soft pre-presbyopic hypermetrope reduced ne pre-presbyopic hypermetrope rmal ne presbyopic myope reduced ne presbyopic myope rmal ne presbyopic myope reduced ne presbyopic myope rmal hard presbyopic hypermetrope reduced ne presbyopic hypermetrope rmal soft presbyopic hypermetrope reduced ne presbyopic hypermetrope rmal ne Ast = Astigmatism

2 Rule representation and reasoning Rule representation: Logical implication (= rules) LHS RHS (LHS = left-hand side = antecedent; RHS = right-hand side = consequent) Literals in LHS and RHS are of the form: Variable value where {<,, =, >, } Rule-based reasoning: where R F C (or Attribute value) R is a set of rules r R (rule-base) F is a set of facts of the form Variable = value C is a set of conclusions of the same form as facts Basic ideas: ZeroR Construct rule that predicts the majority class Used as baseline performance Example Contact lenses recommendation rule: Lens = ne Total number of instances: 24 Correctly classified instances: 15 (625%) Incorrectly classified instances: 9 (375%) === Detailed Accuracy By Class === TP Rate FP Rate Precision Recall F-Measure Class soft hard ne === Confusion Matrix === a b c <-- classified as a = soft b = hard c = ne OneR OneR: Example Construct a single-condition rule for each variable-value pair Select the rules defined for a single variable (in the condition) which perform best OneR(classvar) { R for each var Vars do for each value Domain(var) do classvarmost-freq-value MostFreq(varvalue, classvar) rule MakeRule(varvalue, classvarmost-freq-value) R R {rule} for each r R do CalculateErrorRate(r) R SelectBestRulesForSingleVar(R) } Rules for contact-lenses recommendation: Tears = reduced Lens = ne Tears = rmal Lens = soft 17/24 instances correct Correctly classified instances: 17 (7083%) Incorrectly classified instances: 7 (2916%) === Detailed Accuracy By Class === TP Rate FP Rate Precision Recall F-Measure Class soft hard ne === Confusion Matrix === a b c <-- classified as a = soft b = hard c = ne

3 Generalisation: separate-and-cover Covering of classes: Rule-set generation for each class value separately Peeling: box compression instances are peeled off (fall outside the box) one face at the time PRISM algorithm Example: choosing contact lenses Recommended contact lenses: ne, soft, hard General principles: 1 Choose class value, eg hard 2 Construct rule Condition Lens = hard 3 Determine accuracy α = p/t for all possible conditions, where t: total number of instances covered by the rule p: covered instances with the right (positive) class value Condition α = p/t Age = youg 2/8 Age = pre-presbyopic 1/8 Age = presbyopic 1/8 s = myope 3/12 s = hypermetrope 3/12 Astigmatism = 0/12 Astigmatism = 4/12 Tears = reduced 0/12 Tears = rmal 4/12 4 Select best condition (4/12) Separate-and-cover algorithm SC(classvar, D) { R for each val Domain(classvar) do E D while E contains instances with val do rule MakeRule(rhs(classvarval), lhs( )) IR until rule is perfect do for each var Vars, rule IR : var / rule do for each value Domain(var) do inter-rule Add(rule, lhs(varvalue)) IR IR {inter-rule} rule SelectRule(IR) R R {rule} RC InstancesCoveredBy(rule, E) E E\RC } SelectRule: based on accuracy α = p/t; if α = α, for two rules, select the one with highest p Rule: SelectRule example RHS: Lens = hard LHS: Astigmatism =, with α = 4/12 Not very accurate; expanded rule: (Astigmatism = New-condition) Lens = hard Tear product Age prescription Ast rate Lens young myope reduced ne young myope rmal hard young hypermetrope reduced ne young hypermetrope rmal hard pre-presbyopic myope reduced ne pre-presbyopic myope rmal hard pre-presbyopic hypermetrope reduced ne pre-presbyopic hypermetrope rmal ne presbyopic myope reduced ne presbyopic myope rmal hard presbyopic hypermetrope reduced ne presbyopic hypermetrope rmal ne Age = young (2/4); Age = pre-presbyopic (1/4); Age = presbyopic (1/4) s = myope (3/6); s = hypermetrope (1/6) Tears = reduced (0/6); Tears = rmal (4/6)

4 SelectRule example (continued) Rule: (Astigmatism = Tears = rmal) Lens = hard Expanded rule: (Astigmatism = Tears = rmal New-condition) Lens = hard Tear product Age prescription Ast rate Lens young myope rmal hard young hypermetrope rmal hard pre-presbyopic myope rmal hard pre-presbyopic hypermetrope rmal ne presbyopic myope rmal hard presbyopic hypermetrope rmal ne Age = young (2/2); Age = pre-presbyopic (1/2); Age = presbyopic (1/2) s = myope (3/3); s = hypermetrope (1/3) (Astigmatism = Tears = rmal s = myope) Lens = hard SC example (continued) Delete 3 instances from E; new rule: New-condition Lens = hard Tear production Age prescription Ast rate Lens young myope reduced ne young myope rmal soft young myope reduced ne young hypermetrope reduced ne young hypermetrope rmal soft young hypermetrope reduced ne young hypermetrope rmal hard pre-presbyopic myope reduced ne pre-presbyopic myope rmal soft pre-presbyopic myope reduced ne pre-presbyopic hypermetrope reduced ne pre-presbyopic hypermetrope rmal soft pre-presbyopic hypermetrope reduced ne pre-presbyopic hypermetrope rmal ne presbyopic myope reduced ne presbyopic myope rmal ne presbyopic myope reduced ne presbyopic hypermetrope reduced ne presbyopic hypermetrope rmal soft presbyopic hypermetrope reduced ne presbyopic hypermetrope rmal ne Rules: WEKA Results If Astigmatism = and Tears = rmal and s = hypermetrope then Lens = soft If Astigmatism = and Tears = rmal and age = young then Lens = soft If age = pre-presbyopic and Astigmatism = and Tears = rmal then Lens = soft If Astigmatism = and Tears = rmal and s = myope then Lens = hard If age = young and Astigmatism = and Tears = rmal then Lens = hard If Tears = reduced then ne If age = presbyopic and Tears = rmal and s = myope and Astigmatism = then Lens = ne If s = hypermetrope and Astigmatism = and age = pre-presbyopic then Lens = ne If age = presbyopicand s = hypermetrope and Astigmatism = then Lens = ne === Confusion Matrix === a b c <-- classified as a = soft b = hard c = ne Correctly classified instances: 24 (100%) Limitations Adding one condition at the time is greedy search ( optimal state may be missed) Accuracy α = p/t: promotes overfitting: the more correct (higher p compared to t) is, the higher α Resulting rules cover all instances perfectly Example Consider rule r 1 with accuracy α 1 = 1/1 and rule r 2 with accuracy α 2 = 19/20, then r 1 is considered superior to r 2 Alternative 1: information gain Alternative 2: probabilistic measure

5 Information gain where I D (r) = p [ log p t log p ] t α = p/t is the accuracy before adding a condition to r α = p /t is the accuracy after a condition has been added to r Example Consider rule r with α = 1/1 and rule r with accuracy α = 19/20, both modifications of r with α = 20/200 Then is r considered superior to r according to accuracy, but I D (r ) = 1[log(1/1) log(20/200)] = 1 I D (r ) = 19[log(19/20) log(20/200)] 186 hence r is inferior to r according to information gain Comparison accuracy versus information gain Information gain I: Emphasis is on large number of positive instances High coverage cases first, special cases later Resulting rules cover all instances perfectly Accuracy α: Takes number of positive instances only into account if ties break Special cases first, high coverage cases later Resulting rules cover all instances perfectly Probabilistic measure N instance in D Probabilistic measure N instance in D rule r selects t instances M positive concerning the class rule r selects t instances M positive concerning the class p instances concern the class N = D : # instances in dataset D M: # instances in D concerning a class t: # instances in D on which rule r succeeds p: # positive instances Hypergeometric distribution: ( )( ) M N M k t k f(k) = ( ) N t sampling without replacement: probability that k instances out of t belong to the class p instances concern the class Hypergeometric distribution: ( )( ) M N M k t k f(k) = ( ) N t Rule r selects t instances, of which p are positive Probability that a randomly chosen rule r does as well or better than r: ( )( ) min{t,m} P (r min{t,m} M N M k t k ) = f(k) = ( ) N k=p k=p t

6 P (r ) = Approximation min{t,m} k=p min{t,m} k=p ( )( ) M N M k t k ( ) N t ( t k ) ( M N ) k ( 1 M ) t k N ie hypergeometric distribution approximated by a bimial distribution = I M/N (p, t p + 1) where I x (α, β) is the incomplete beta function: I x (α, β) = 1 x B(α, β) 0 zα 1 (1 z) β 1 dz where B(α, β) is the beta function, defined as 1 B(α, β) = 0 zα 1 (1 z) β 1 dz Reduced-error pruning Danger of overfitting to training set can be reduced by splitting this into: a growing set (GS) (2/3 of training set) a pruning set (PS) (1/3 of training set) REP(classvar, D) { R ; E D (GS, PS) Split(E) while E do IR for each val Domain(classvar) do if GS and PS contain a val-instance then rule BSC(classvarval, GS) while P (rule PS) > P (rule PS) do rule rule IR IR {rule} rule SelectRule(IR); R R {rule} RC InstancesCoveredBy(rule, E) E E\RC (GS, PS) Split(E) } BSC is basic separate-and-cover algorithm, and rule is a rule with last condition removed Divide-and-conquer: decision trees young ne Age Astigmatism soft soft Age hard prescription young pre prebyopic myope hypermetrope presbyopic ne reduced presbyopic Tears pre prebyopic soft rmal Learning decision trees: hypermetrope Lens =? prescription myope hard ne ne R Quinlan: ID3, C45 and C50 L Breiman: CART (Classification and Regression Trees) Which variable/attribute is best? X D X = Dataset D: C X class variable D = D X= D X= Entropy: X DX = C H C (X =x) = P (C =c X =x) ln P (C =c X =x) c Expected entropy: E HC (X) = P (X = x)h C (X = x) x

7 Dataset D: Information gain (again) D = D X= D X= Entropy: H C (X =x) = c Expected entropy: E HC (X) = x P (C =c X =x) ln P (C =c X =x) P (X = x)h C (X = x) Without the split of the dataset D on variable X, the entropy is: H C ( ) = c P (C = c) ln P (C = c) Information gain G C from X: G C (X) = H C ( ) E HC (X) Example: contact lenses recommendation Class variable is Lens: 5/24 if Lens = soft P (Lens) = 4/24 if Lens = hard 15/24 if Lens = ne H( ) = 5 24 ln ln For variable Ast (Astigmatism): ln /12 if Lens = soft P (Lens Ast = ) = 0/12 if Lens = hard 7/12 if Lens = ne Therefore: H(Ast=) = 5 12 ln ln ln Example (continued) For variable Ast (Astigmatism): 0/12 if Lens = soft P (Lens Ast = ) = 4/12 if Lens = hard 8/12 if Lens = ne Therefore: H(Ast=) = 0 12 ln ln ln E H (Ast) = 1 2 H(Ast=) H(Ast=) Information gain: = 1/2( ) = 066 G(Ast) = H( ) E H (Ast) = = 026 Example (continued) For variable Tears: E H (Tears) = 1 2 H(Tears=red) H(Tears=rm) 1/2( ) = 055 Information gain: G(Tears) = H( ) E H (Tears) = = 037 Comparison: G(Tears) > G(Ast) Select Tears as first splitting variable

8 Final remarks I A de with too many branches causes the information gain measure to break down Example Suppose that with each branch of a de a dataset with exactly one instance is associated: E HC (X) = n 1/n (1 log log 0) = 0 if X has n values Hence, G C (X) = H C ( ) 0 = H C ( ) attains a maximum Solution: Gain ratio R C : Split information H X ( ) = x Gain ratio: R C (X) = G C (X)/H X ( ) P (X = x) ln P (X = x) Final remarks II Variable selection is myopic: it does t look beyond the effects of its own values; a resulting decision tree is therefore likely to be suboptimal Decision trees may grow unwieldy, and may need to be pruned (ID3 C45) subtree replacement subtree raising Decision trees can also be used for numerical variables: regression trees