Please enable JavaScript.

Coggle requires JavaScript to display documents.

Applied Machine Learning (WEKA (.ARFF, % - Single comment, relation <…

- - - - Remove missing instances completely from the data set
      - Impute mean values for the missing values in the data set
      - Reasons:
        
        Is the data big enough to consider removing instances
        
        What is the variance on the attribute and will imputing the mean value effect this in any way
        
        Does this results in under or over fitting?
        
        Ethically is this acceptable?
        
        Realistically is this acceptable?
  - - - Do not know the distribution of the data
      - No Gaussian
    - - Gaussian Distribution
  - - - Info Gain measures the amount of info in bits about the class prediction, if only info available is the presence of a feature class prediction, if the only information available is the presence of a feature and the corresponding class distribution. Concretely, it measures the expected reduction in entropy
    - - This works for both classification and regression algoritms
      - This runs Pearson correlation analysis on each attribute and compares it to the class
- - - - The predictor is too complex
      - The training set is not representative of the population
      - The training methods may not transfer to unseen accuracy
    - - The predictor is too simplistic
      - The predictor is simply not powerful enough to recognize patterns in the data
      - Accuracy reported is lower than future unseen accuracy
  - - - Training Dataset
        
        Use all of the data as the training set as well as the test set
      - Supplied Test Set
        
        Use a separate supplied test set to examine your training set
      - Percentage Split (Hold Out) & Stratification
        
        Use some of the training set for training and some of it for testing
        
        Stratification is the organizing of the data, so the training set and the test set has an equal representation of the instances
        
        Possible issues:
        
        Data in training set is biased to test set, resulting in over-fitted model
        
        Data in training set is significantly different/novel to the test set, resulting in a very under-fitted
      - Cross Validation
        
        (Rotation Estimates) train's and tests, using data subsets from the training data, where each instance is trained and tested, but is never a member of each set at the same time
        
        Uses seed value in testing to get responsibility
    - - Sensitivity = TP / (TP+ FN)
      - Specificity = TN/ (TN + FP)
      - Accuracy = TP + TN / (TP + TN + FN + FP)
    - - Used to measure performance on classification problems only
    - - TP - the correct identification of what you are looking to identify
      - FP - the incorrect identification of what you are looking to identify
      - TN - the correct identification of the opposite of what you are trying to identify
      - FN - the incorrect identification of the opposite of what you are trying to identify