230 likes | 433 Views
Advisor : Prof. Sing Ling Lee Student : Chao Chih Wang Date : 2012.10.11. Collective Classification for Network Data With Noise. Outline. Introduction Network data Collective Classification ICA Problem Algorithm For Collective Inference With Noise Experiments Conclusions.
E N D
Advisor : Prof. Sing Ling Lee Student : Chao Chih Wang Date : 2012.10.11 Collective Classification for Network Data With Noise
Outline • Introduction • Network data • Collective Classification • ICA • Problem • Algorithm For Collective Inference With Noise • Experiments • Conclusions
Introduction – Network data • traditional data: • instances are independent of each other • network data: • instances are maybe related of each other • application: • email • web pages • papers citation independent related
Introduction – Collective Classification • classify interrelated instances using relational features. • Related instances: • The instances to be classified are related. • Classifier: • The base classifier uses contentfeatures and relational features. • Collective inference: • update the class labels • recompute the relational feature values
Introduction – ICA • ICA : Iterative Collective Algorithm Initial : Training local classifier use content features predict unlabel instances Iterative{ for predict each unlabel instance { set unlabel instance ’s relational feature use local classifier predict unlabel instance } } step1 step2
Introduction – ICA Example unlabel data: Class label : 1 2 3 training data: • Initial : • use content features predict unlabel instances 3 1 H C 2 A 3 3 1 B 2 1 Iterative 1: 1. set unlabel instance ’s relational feature 2. use local classifier predict unlabel instances Iterative 2: 1. set unlabel instance ’s relational feature 2. use local classifier predict unlabel instances D 2 E 1 F 1 G 2 1
Problem – Noise C A • label the wrong class • make a mistake • difficult to judge B G D 2 E F 2 1 1 1 1 or 2 ? 2 1
Problem – use ICA True label: unlabel data: C B A ICA: training data: 2 2 1 • Initial : • use content features predict unlabel instances 1 A 1 1 1 C 1 Iterative 1: 1. set unlabel instance ’s relational feature 2. use local classifier predict unlabel instances Iterative 2: 1. set unlabel instance ’s relational feature 2. use local classifier predict unlabel instances 2 D 1 • noise B 2 A : 1 1 Iteration 1 2/3 1/3 0 Iteration 2 2/3 1/3 0
ACIN • ACIN : Algorithm For Collective Inference With Noise Initial : Training local classifier use content features predict unlabel instances Iterative{ for predict each unlabel instance { for nbunlabel instance ’s neighbors{ if(need to predict again) (class label, probability ) = local classifier(nb) } set unlabel instance ’s relational feature (class label, probability ) = local classifier(A) } retrain local classifier } step1 step2 step3 step4 step5
ACIN -Example True label: unlabel data: C B A ACIN: training data: 2 2 1 • Initial : • use content features predict unlabel instances 1 ( 1 , 60%) ( 1 , 60%) A 1 2 1 1 predict again C ( 1 , 70%) 1 Iterative 2: 1. repredictunlabel instance ’s neighbors 2. set unlabel instance ’s relational feature 3. use localclassifier predict unlabel instances Iterative 1: 1. predict unlabel instance ’s neighbors 2. set unlabel instance ’s relational feature 3. use local classifier predict unlabel instances ( 2 , 60%) ( 1 , 60%) ( 1 , 70%) 2 ( 2 , 60%) ( 2 , 60%) ( 2 ,60%) D 1 • noise B ( 1 , 90%) 2 A : 1 1 Iteration 1 70/130 60/130 0 Iteration 2 60/120 60/120 0
ACIN -Analysis • Compare with ICA • 1. different method for compute relational feature • 2. predictunlabel instance ’s neighbors again • 3.retrain local classifier
ACIN –Analysis #1 • compute relational feature • use probability General method : Class 1 : 1/3 Class 2 : 1/3 Class 3 : 1/3 A 1 3 2 ( 1 , 80%) ( 3 , 70%) Our method: Class 1 : 80/(80+60+70) Class 2 : 60/(80+60+70) Class 3 : 70/(80+60+70) ( 2, 60%)
ACIN –Analysis #2 • predictunlabel instance ’s neighbors again • first iteration need to predict again • different between originaland predict label : • Next iterationneed predict again • This iterationnot to adopt • similarbetween originaland predict label : • Next iterationnot need predict again • Average the probability A Example: predict again ( 2, 60%) ( 2, 80%) 1 2 1 C B ( 2, 70%) ( 2, 60%) ( 2, 60%) ( 1 , 80%)
ACIN –Analysis #2 B’s True label : 2 2 A - noise B D 1 3 C ( 1 , 60%) ( 3 , 60%) ( 2 , 70%) ( 2 , 70%) predict again ( 3 , 60%) ( 2 , 80%) B is noise: Method 2 >Method 3> Method 1 ( ? , ??%) ( 3 , 60%) ( 2 , 75%) B is not noise: Method 1 >Method 3> Method 2 B: Method 1 : (1 , 50%) Method 2 : (2 , 60%) Method 3 : (1 , 0%) 2 Method 1 & Method 2 is too extreme. So we choose the Method 3.
ACIN –Analysis #3 • retrain local classifier Initial ( ICA ) D A B 1 2 + E C retrain ( 3 , 70%) ( 1 , 80%) D 3 A ( 2 , 70%) 1 B 2 1 1 + 2 E C ( 2, 60%) ( 1 , 90%)
Experiments • Data sets: • Cora • CiteSeer • WebKB