ML-Elections

Machine Learning 101

Steps of work

1. Data Cleansing

   1. Missing Data
      1. fill missing columns with relevant values
         • median for numeric fields
         • most common for label for categorial fields
      2. (optional) look for linear correlation with other feature.
      3. (optional) create boolean feature for missing values.
      4. (optional) closest fit (WHAT??)
   2. Noisy Data
      1. Outlier Detection
         1. Nearest Neighbours - remove by choosing outlier.
   3. Data transformation.
      1. Change categories to boolean (0/1) columns.
      2. Change boolean columns to binary.
      3. (Optional) Create categories by grouping. (linear - age, Logarithmic - num of employees )
      4. Scaling
         1. linear ( Xi / max(X) )
      5. (Optional) balance data
         1. if we have one label with way more occurrences than other we should scale it.

2. Feature selection

   1. variance filter - remove features with low variance
   2. filter methods
      1. select top 25% features with f_classif
      2. select top 25% features with mutual_information
   3. wrapper method
      1. run RFECV with fold = 3
      2. run RFECV with StratifiedKFold(2)
   4. Embedded methods
      1. decisions tree - pick 25% of the highets weight features
   5. Sum up all the features together.

3. Evaluating model

   1. Trials – trying out different models.
   2. Examination – after the trials we zoomed in on best models.
   3. Training – Training the chosen models
   4. Prediction - predicting based on the chosen models.