Machine Learning Algorithms

Conventions:

• features = predictor variables = independent variables;
• target variable = dependent variable = response variable;
• training a model on the data = fitting a model to the data.

Supervised Learning

It uses labeled data as input, which are usually represented in a table structure. It divides in:

1. Classification: outputs a label, which is a value in a set C (categories).

   • Confusion matrix: table used to evaluate the performance of a classification model.

   1. k-nearest neighbors (KNN): predicts the label of a data point by looking at the k closest labeled data points. Code example: sklearn iris species example

   2. Logistic Regression: outputs probabilities.

      • One of the most commonly used ML algorithms for two-class classification;
      • Dependent variable follows Bernoulli Distribution;
      • Logistic regression threshold is 0.5, by default.
      • ROC curve (Receiver Operating Characteristic Curve): series of different false positive and true positive rates as we vary threshold;
      • Estimation: Maximum Likelihood Estimation (MLE);
      • Model fitness calculated through Concordance and KS-Statistics;
      • It divides in:
        1. Binary Logistic Regression: target variable has two possible outputs. Examples: spam detection, diabetes prediction, cancer detection, and if a user will click on an advertisement link or buy a product or not. Code example: sklearn breast cancer example
        2. Multinomial Logistic Regression: target variable has 3 or more nominal categories. Examples: types of iris flowers, and types of wine. Code example: sklearn iris species example
        3. Ordinal Logistic Regression: target variable has 3 or more ordinal categories. Example: restaurant or product rating (from 1 to 5), and classification of documents into categories.

   3. Support Vector Machines (SVM)

      • Code example: sklearn breast cancer example

   4. Kernel SVM

      • Code example: sklearn breast cancer example

   5. Naïve Bayes

      • Code example: sklearn breast cancer example

   6. Decision Trees

      • Code example: sklearn breast cancer example

   7. Random Forest Classification

      • Code example: sklearn breast cancer example

   8. Neural Networks models

2. Linear Regression: output is a value in R (continuous);

   • Estimation: OLS (Ordinary Least Squares), which is the sum of the squares of the residuals (same as minimizing the mean squared error of the predictions on the training set);
   • Code example: sklearn Boston housing prices example

Linear Regression X Logistic Regression:

• Linear regression gives a continuous output (example: house price), logistic regression gives a constant output (example: stock price);
• Linear regression estimated by Ordinary Least Squares (OLS), logistic regression estimated by Maximum Likelihood Estimation (MLE);

In Python we can use scikit-learn (sklearn), TensorFlow and Keras for Supervised Learning.

Unsupervised Learning

It tries to discover hidden patterns in unlabeled data. Main tasks:

1. Clustering: tries to discover the underlying groups (clusters) in a dataset;
   1. k-means clustering: finds clusters given a number of clusters (scikit-learn);
2. Data visualization: hierarchical clustering and t-SNE;
3. Dimension reduction techniques: Principal Component Analysis (PCA);
4. Dimension reduction techniques: Non-negative matrix factorization" ("NMF").

Reinforcement Learning

Interaction with environment to learn how to optimize behavior, using a system of rewards and punishments.

Useful links:

1. scikit-learn examples.
2. TensorFlow tutorials.
3. Keras.
4. Deep Learning Online Courses (free).