Please enable JavaScript.

Coggle requires JavaScript to display documents.

Data Foundations (Sources (Raw Data, Processed Data, Computations,…

- - - - Binning
        
        equal width
        
        n intervals, w = (max-min)/N
        
        outliers may dominate result
        
        equal depth
        
        handles skewed data well
        
        N intervals similar number of records
        
        Smooth by
        
        mean
        
        bin boundaries
      - errors
        
        random
        
        systematic
      - Regression
        
        linear
        
        non-linear
      - Clustering
    - - manual
        
        good for small datasets
        
        need to know value
        
        time consuming
      - constant
        
        allows other values to be kept
      - mean
      - ignore tuple/row
        
        no computational effort
        
        loss of information
        
        unnecessary if attribute not needed
      - most probable
        
        most accurate approximation
        
        most computational effort
  - - - rerun
        
        correct representation
        
        comp. expensive
        
        perception of previous is distorted
      - global const for too high/low
        
        same perception of previous
        
        comp. cheap
        
        high/low values inadequate repr.
        
        value range may be distorted
      - increase value range
        
        correct perception of results
        
        comp. cheap
        
        limits still unknown
        
        possible visual mapping space might be lost
  - - - feature extraction
      - Principal Component Ananlyis (PCA)
    - - sampling
        
        probabilistic
        
        simple random sampling
        
        least biased
        
        some characteristics not taken into account
        
        systematic random sampling
        
        good sample for odered attribute (on which we are sampling)
        
        bias in other attributes
        
        problem with periodicies
        
        stratified random sampling
        
        good sample of the data
        
        different sampling withing strata can be used
        
        more comp. intensive
        
        cluster random sampling
        
        cheap when geographically convenient (?)
        
        increasing sample size is easy
        
        least representative of population
        
        non-probabilistic
      - motivation
        
        speed up comp.
        
        represent larger dataset
  - - - k-Means
        
        Need to specify k
        
        Unable to handle noise
        
        Cannot detect non-convex shaped clusters
        
        O(tkn)
        
        Often terminates at local opt
        
        Only with mean
      - Hierarchical
        
        Single linkage
        
        Complete linkage
        
        Centroid linkage
      - Density based
        
        Maximality
        
        Connectivity
        
        (Directly) Density reachable
        
        Density Connected