Please enable JavaScript.

Coggle requires JavaScript to display documents.

Provost Chapter 6 ("Nearest Neighbors" (analyzing which single…

- - - - use "distance" measure to compare whiskey similarities
        
        order scotches by similarity
        
        powerful conceptual tool
  - - - use collection of data to predict target value based on nearest neighbors
        
        credit card example, predicting response to ad based on what similar (and non-similar) customers did
        
        score needs to be estimated - more accurate than binary "yes/no" response - need probability estimation
        
        score can then be used for regression
  - - - the > the k, the more smooth the estimation
    - - "contribution" (nn/100) shows weight of each neighbor
- - - - k is a complexity parameter
    - - conduct cross-validation to find best k value
- - - - Netflix "recs" justifiable for NN
      - Mortgage default predictability - probably not justifiable
  - - - ex: when models need fast prediction speed, NN is an issue
- - - - can lead to useful/profitable discoveries
    - - choose strategy for broad and eclectic collection
        
        choose one popular and one less-popular scotch from each flavor cluster
  - - - rings dont overlap - represents hierarchy of clusterings
        
        dendogram is graphic representation of this
      - clusters are merged based on the distance function chosen
        
        called a linkage function
    - - allows analyst to see the landscape of data similarity
    - - represent each cluster by its cluster center
        
        geometric center of a group of instances
      - k-means clustering
        
        arithmetic means of the values along each dimension for instances in the cluster
        
        k = how many clusters you want in the data
        
        "nearest neighbor" clustering
        
        k-means algorithm is efficient and runs fast
  - - - each data point itself conveys meaning just by being in a particular cluster
    - - where supervised machine learning comes in
        
        run a ML algorithm to identify classifiers of each cluster
- - - - distance between 2-dimensional points
        
        "Euclidian Distance"
        
        for numeric objects - can be on 3 dimensional plane as well
- - - - variable-scaling front ends
      - apportion a fixed # of bins
  - - - N/S, E/W
    - - treats objects as a set of characteristics
        
        unions/intersections