On Reducing Classifier Granularity in Mining ConceptDrifting Data

  • Slides: 20
Download presentation
On Reducing Classifier Granularity in Mining Concept-Drifting Data Streams Peng Wang, H. Wang, X.

On Reducing Classifier Granularity in Mining Concept-Drifting Data Streams Peng Wang, H. Wang, X. Wu, W. Wang, and B. Shi Proc. of the Fifth IEEE International Conference on Data Mining (ICDM’ 05) Speaker: Yu Jiun Liu Date : 2006/9/26

Introduction o State of the art n n o The incrementally updated classifiers. The

Introduction o State of the art n n o The incrementally updated classifiers. The ensemble classifiers. Model Granularity n n Traditional : monolithic This paper : semantic decomposition

Motivation o o The model is decomposable into smaller components. The decomposition is semantic-aware

Motivation o o The model is decomposable into smaller components. The decomposition is semantic-aware in the sense.

Monolithic Models o o o Stream : Attributes : Class Label : Window :

Monolithic Models o o o Stream : Attributes : Class Label : Window : Model (Classifier) : Ci

Rule-based Models o A rule form : minsup = 0. 3 and minconf =

Rule-based Models o A rule form : minsup = 0. 3 and minconf = 0. 8 Valid rules of W 1 are: o Valid rules of W 3 are: o o

Algorithm o Phase 1 : Initialization n n o Use the first w records

Algorithm o Phase 1 : Initialization n n o Use the first w records to train all valid rules for window W 1. Construct the RS-tree and REC-tree. Phase 2 : Update n n When record arrives, insert it into the REC-tree and update the sup. and conf. of the rules matched by it. Delete oldest record and update the value matched by it.

Data Structure

Data Structure

RS-Tree o o o A prefix tree with attribute order Each node N represents

RS-Tree o o o A prefix tree with attribute order Each node N represents a unique rule R : P Ci N’ (P’ Cj) is a child node of N, iff:

REC-Tree o o Each record r as a sequence Node N points to rule

REC-Tree o o Each record r as a sequence Node N points to rule in the RS-tree if :

Detecting Concept Drifts o percentage V. S. the distribution of the misclassified records. The

Detecting Concept Drifts o percentage V. S. the distribution of the misclassified records. The percentage approach cannot tell us which part of the classifier gives rise to the inaccuracy.

Definition

Definition

Finding Rule Algorithm

Finding Rule Algorithm

Update Algorithm

Update Algorithm

Experiments o o o CPU : 1. 7 GHz Memory : 256 MB Datasets

Experiments o o o CPU : 1. 7 GHz Memory : 256 MB Datasets : synthetic and real life dataset. n Synthetic : o n Real life dataset : o 10, 344 recodes and 8 dimensions.

Effect of model updating o o Synthetic 10 dimensions Window size 5000 4 dimensions

Effect of model updating o o Synthetic 10 dimensions Window size 5000 4 dimensions changing

The relation of concept drifts and

The relation of concept drifts and

Effect of rule composition

Effect of rule composition

Accuracy and Time o o o Window size : 10, 000 EC : 10

Accuracy and Time o o o Window size : 10, 000 EC : 10 classifiers, each trained on 1000 records. Synthetic data.

Real life data

Real life data

Conclusion o o Overcome the effects of concept drifts. By reducing granularity, change detection

Conclusion o o Overcome the effects of concept drifts. By reducing granularity, change detection and model update can be more efficient without compromising classification accuracy.

Data Mining FrequentItemset Mining Data Mining Some miningData Mining FrequentItemset Mining Data Mining Some mining
Data Mining What Is Data Mining Data miningData Mining What Is Data Mining Data mining
Data Mining Web Data Mining Web Data MiningData Mining Web Data Mining Web Data Mining
Data Mining Data Mining Simply stated data miningData Mining Data Mining Simply stated data mining
Lecture outline Classification Nave Bayes classifier Nearestneighbor classifierLecture outline Classification Nave Bayes classifier Nearestneighbor classifier
Classifier Systems Anil Shankar Classifier Systems Overview IntroductionClassifier Systems Anil Shankar Classifier Systems Overview Introduction
Lecture outline Classification Nave Bayes classifier Nearestneighbor classifierLecture outline Classification Nave Bayes classifier Nearestneighbor classifier
Lecture outline Classification Nave Bayes classifier Nearestneighbor classifierLecture outline Classification Nave Bayes classifier Nearestneighbor classifier
Data Mining Association Rule Mining 2262021 Data MiningData Mining Association Rule Mining 2262021 Data Mining
Data Mining FrequentItemset Mining Data Mining Discovery ofData Mining FrequentItemset Mining Data Mining Discovery of
Data Mining FrequentItemset Mining Data Mining Discovery ofData Mining FrequentItemset Mining Data Mining Discovery of
DATA MINING RULE CLASSIFIERS The first alternative classifierDATA MINING RULE CLASSIFIERS The first alternative classifier
SLIQ A Fast Scalable Classifier for Data MiningSLIQ A Fast Scalable Classifier for Data Mining
SPRINT A Scalable Parallel Classifier for Data MiningSPRINT A Scalable Parallel Classifier for Data Mining
DATA MINING LECTURE 10 Classification knearest neighbor classifierDATA MINING LECTURE 10 Classification knearest neighbor classifier
SPRINT A Scalable Parallel Classifier for Data MiningSPRINT A Scalable Parallel Classifier for Data Mining
SLIQ A Fast Scalable Classifier for Data MiningSLIQ A Fast Scalable Classifier for Data Mining
Data Mining Algorithms Web Mining Web Mining OutlineData Mining Algorithms Web Mining Web Mining Outline
Data Mining for Data Streams 1122020 Data MiningData Mining for Data Streams 1122020 Data Mining