A hybrid model combining case-based reasoning and fuzzy decision tree for medical data classification

Presented By

Safi

Sumanth

Why?

  1. Data == Intelligence
  2. Binary / Multi-Class
  3. Accuracy > 90%

Why?

  1. Data == Intelligence
  2. Binary / Multi-Class
  3. Accuracy > 90%

Problems

  1. Need explanations
  2. Neural Nets == Black Box
  3. But......

Why?

  1. Data == Intelligence
  2. Binary / Multi-Class
  3. Accuracy > 90%

Problems

  1. Need explanations
  2. Neural Nets == Black Box
  3. But......

Solutions

  1. Neural Nets with fuzzy rules
  2. Cluster data + Fuzzy Decision Trees

Contents

  1. Literature Survey
    • ​abc
    • def
    • ghi
  2. CBFDT
  3. Experiments

Artificial Neural Nets

  1. Wu et. al used 133 training points with 43 features, to provide a binary output (ANN >= Human)
  2. Floyd et. al used 260 training points with 10 features to achieve Breast Cancer pred. accuracy of 50%
  3. Setiono et. al extracted rules from ANN and trained a model using the rules to get 96% accuracy
  4. All the above use Back Prop and needs extensive feature selection
  1. Fogel et. at proposed an evolutionary search algo. for solving this - but computationally expensive

Fuzzy Classifier

  1. Gadaras et. al proposed fuzzy classification to automatically extract rules
  2. Fernándas et. al used fuzzy rules to solve data imbalance problem

Fuzzy systems do not always use class labels but gain/loss func.

 

So we use them in combination with ANNs

Fuzzy Classifier

  1. Song et. al
  2. Gonçalves et. al
  3. Lee and Wang

Data Mining

  1. LDA and Time Series
    • ​But fail when relationships are complex / non-linear
  2. Alternatives
    • Bayesian
    • ANNs
    • Decision Trees
    • SVMs
    • Genetic Search Algorithms

Current paper used FDT and CBR

Fuzzy Decision Tree

  1. Similar to Decision Tree
  2. Recursive Binary Partition 
  3. Need to put some equations and add explanations

Current paper used FDT and CBR

Clustering Methods

  1. Unsupervised
  2. Exclusive / Fuzzy / Hierarchal 
  3. The paper used classic k-means
  4. Need to put some equations and add explanations

Current paper used FDT and CBR

Made with Slides.com