Please enable JavaScript.

Coggle requires JavaScript to display documents.

Machine Learning (Methods (Decision Trees (Creation of a tree (Start at…

- - - - One Versus All (OvA), one binary classifier for each class
      - One versus One (OvO), one classifier to distinguish between two classes.
      - Random Forest (?)
    - - Output of 1, 0, 1, ... to assign multiple classes
    - - Same as Multilabel, but each label can be multiclass
    - - Requires numeric features
        
        Can be transformed from nominal features via labeling
        
        LabelEncoder
      - Perzeptron
        
        Graphic display of linear function
        
        Training
        
        Iterative
        
        Random start weights
        
        Gradual improvement of error, gradient descent
  - - - Number of categories predefined
      - Initialisation
        
        Initialisation of centers
        
        Farthes away points
        
        Risk for choosing outliers
        
        Random
        
        Risk of two points lying next to each other
        
        k-means++
        
        Like farthest away, choose far but randomly
      - Initialisation of k
        
        Using total squared error
        
        Total Error = SQ error + complexity (penalty)
  - - - Find best weight vector that makes width of street a maximum, see slide 17 in presetation
        
        Use slack variables to allow for outliers. Problem then comes to: Maximize distance (minimize inverted w), minimize slack penalty overall
        
        Optimal value of slack variable via Grid Search
        
        class_weight to weigh penalty for classes
    - - Adapt dimensions to each other
  - - - Start at root node
        
        use measure for homogeneity to find optimal split of node
  - - - Hard Voting
      - Soft Voting
    - - AdaBoosting
- - - - Feature Selection
      - Feature Transformation (combine multiple features)
- - - - Predict discrete function y(x)
      - Immediate transition in classes
      - 3D with color
      - Measuring amounts of errors
    - - Scatter plots, suggest / predict relationship between x and y
        
        explicit function
        
        Nearest neighbor regression (step function)
        
        Linear regression, linear function r = th0 + th1 * x1
        
        Measuring error, error residual