Please enable JavaScript.

Coggle requires JavaScript to display documents.

Knowledge discovery - Coggle Diagram

- - - - IRS-RS-Day4-Slide 20
    - - IRS-RS-Day4-Slide 21
  - - - IRS-RS-Day4-Slide27
  - - - Too many variables lead to clusters that are not homogenous/heterogenous from each other
    - - Intra class similarity is high, interclass similarity is low
      - Cluster cohesion: Sum of square differences between objects in the cluster
        Cluster separation: Distance between cluster centroid and overall centroid multiply by number of objects in the cluster
      - Popular stats IRS-RS-Day4-Slide33
    - - Prediction based on how close a point is to a cluster?
- - - - Be careful to avoid facade of strong rules by calculating lift (P(LHS and RHS)/P(LHS)P(RHS)
        If Lift is < 1, negatively correlated. If Lift is > 1 then positively correlated
    - - Solution - Apriori Algorithm
        
        Find large item sets that occur tog often
        
        Generate rules for these
        
        Find all single item data occurs e.g., apple and how many times it appear in each data set
        
        If the number of time > threshold, keep it
        
        When all single items are found e.g., apple pear milk, find all combinations that have the all kept data e.g., {apple, pear}, {milk, apple} etc.
        
        Generate association rules
      - Solution - Rule Pruning
        Only keep rules with min support and confidence
    - - Equal width binning (same interval size), equal frequency binning(same no. of objects), cluster-based binning/