Please enable JavaScript.

Coggle requires JavaScript to display documents.

Machine Learning (Supervised Learning (Techniques (Tuning (Underfitting…

- - - - L1 Regularization
        
        Add the absolut value of the coefficients to the error
        
        Simple vs complex model
        
        Lambda Parameter that multiplies the L1 and L2 terms
        
        The highest the lambda, more will be penalized the complex models
        
        Computationally Inefficient
        
        Better for Sparse output
        
        Help in making feature selection
      - L2
        
        Add the squares of the coefficients to the error
        
        Better for non sparse outputs
        
        Not helping with feature selection
        
        Computationally Efficient
      - to improve models avoidind overfitting
      - add to the error the complexity of the model
    - - BoW - Bag of words
        
        counting word frequencies without taking in consideration order
        
        Count vectorizer scikit learn
        
        Tokenization : splitting in single words - split()
        
        Counting occurencies - Counter()
        
        LOWERCASE: convert to lowercase
        
        token_pattern: ignores all punctuation (string.punctuation)
        
        stop_words: remove most commonly used words
      - tfidf
    - - Underfitting
        
        not performing well in the training set, due to BIAS
      - Overfitting
        
        performing well in training but not in test, due to VARIANCE
      - CrossValidation
        
        we split the data in training, CrossValidation and Test set. !!!! the test will be used JUST with the final model
      - K-Fold CrossValidation
        
        Multi CrossValidation + Randomizing parameters
      - GridSearch on Hyper Parameters
    - - Reduce Overfitting
  - - - Decision Trees
        
        Hyperparameters
        
        Maximum Depth
        
        Small --> underfitting
        
        Large --> overfitting
        
        Minimum number of samples per leaf
        
        Small --> overfitting
        
        Large --> underfitting
        
        Minimum number of samples per split
        
        Maximum number of features
        
        tends to overfitting
        
        Random Forest is the solution
        
        Entropy
        
        Information Gain: Entropy of node father - Entropy of node children
      - C parameter
        
        Large C: focus on classyfing points
        
        Small C: focus on larger margin
      - Kernel
      - RBF Kernel
      - Lambda Parameter
        
        Large lambda: overfitting (left)
        
        Small lambda: underfitting (right)
      - Logistic Regression
      - Ensemble methods
        
        Adaboost
        
        Bagging
        
        Boosting
    - - Confusion Matrix
        
        Precision
        
        How many predicted positive on total of positive prediction are correct
        
        Recall
        
        How many sick on total of sick did we predict correctly
        
        F1 Score
        
        Harmonic mean between precision and recall and it is close to the smallest number of the 2
        
        F-Beta Score (Beta = 1)
        
        smaller Beta --> towards precision (Beta = 0 we have precision)
        
        Larger Beta --> towards Recall (Beta = inf we have recall
        
        Accuracy
      - ROC Curve
- - - - Center
        
        Mean
        
        Median
        
        strong to otuliers
        
        Mode
      - Spread
        
        Range
        
        Interquartile Range (IQR)
        
        Standard Deviation
        
        Variance
        
        5 Number Summary
        
        Minimum
        
        Q1 - 25%
        
        Q2 - 50% - Median
        
        Q3 - 75%
        
        Maximum
        
        Box Plot
      - Shape
        
        Right-skewed
        
        Left-skewed
        
        Symmetric
        
        Normal Distribution
        
        Mean = Mode = Median
      - Outliers
      - Histograms
        
        bins
      - Inferential Statistis
        
        population
        
        parameter
        
        Numeric summary about the population
        
        sample
        
        Statistic
        
        Numeric summary about a sample
        
        draw conclusion for larger population
  - - - One hot Encoding
- - - - Mean Squared Error
        
        Gradient or Derivative of the error function with respect to the weights
        
        Stochastic Gradient Descent
        
        Mini-Batch Gradient Descent
        
        Batch Gradient Descent
      - Mean Absolute Error
    - - Data has to be linear (or be transformed to linear)
      - The model is sensitive to the outliers
- - - - comparing the model to the simplest possible model
      - close to 1 if model is good
      - close to 0 if the model is bad