Machine learning (an overview)

Machine learning refers to the gathering, processing and analyzing of data by some human-made device, based on statistics and patterns:

Machine learning types

Machine learning could fall under:

Supervised learning

• training a model to recognize inputs, based on previously labelled datasets, to classify new, unseen data or predict an outcome or an output
• further divided into:
  • classification
    • assigning a label based on inputs
  • regression
    • predicting a numeric quantity based on inputs
    • quantity based on other records with similar characteristics

Unsupervised learning

• training a model to classify or group data without the benefit of previously-labelled data
• further divided into:
  • clustering
    • finding groups of similar or similarly-positioned inputs
  • dimensionality reduction
    • reducing columns (but not rows) of data
      • e.g. for real estate prices, remove duplicate features such as "area in square metres" if another column has "area in square feet"
  • association rule learning
    • finding tendencies of one variable as a good predictor of another
      • e.g. "customers who bought X also bought Y"

Reinforcement learning

• a device will interact with an environment and adjust based on feedback (either via "rewards" [positive] or "punishment" [negative])

Semi-supervised learning

• a mix between supervised and unsupervised learning (in which some data would have labels, while others won't)

Deep learning

• convolutional neural networks
  • recognizes spatial patterns
  • great for image-based ("spatially-oriented") data
• recurrent neural networks
  • recognizes sequential data
  • great for text and speech ("temporally-oriented") data

Machine learning process

Machine learning happens in a process of steps also known as machine learning pipeline; many versions of this pipeline exist, steps and order of steps may vary, but most versions include some form of the "input + processing + output" shape:

Data collection

• gather data from reliable sources

Data processing

• preparation and transformation before th emodel
• check for missing or incorrect values
• check for correct data types
• feature scaling
  • transform (or "normalize") the values to a more interpretable range
    • e.g. transform from a range of "-337.28 to 5828.91", to "0 to 1"
  • this helps both humans and computers more easily distinguish between low and high values
  • an optional step if the data values already make sense to everyone using the data

Model training

• split data into "training data" and "testing data"
  • ~80% for training and ~20% for testing

Model building

• formula derivation with data visualization
  • use the training data to create a model (such as a formula)
    • e.g. an equation in the form of y = mx + b or y = m1x1 + m1x2 + ... + mnxn + b (or whatever linear or non-linear equation)
  • graph the model if possible and/or desired
  • model typologies include:
    • k-nearest neighbours
    • decision trees
    • random forest
    • boosting
    • support vector machines
    • neutral networks

Model testing

• formula validation
  • use the testing data to test the model against the model
  • plug the independent variables (x) of the testing data into the model to compare its prediction (y_pred) with the actual dependent (y_test) variable of the data
  • use Python libraries and metrics such as:
    • mean absolute error (MAE)
    • mean square error (MSE)
    • root mean square error (RMSE)
  • see which x (or combination of x's) has the error metrics closest to 0

Backward elimination

• noisy data removal
  • for multiple regression models, remove any variable(s) that would cause the model to classify or predict poorly

Model deployment

• taking that model
  • to the internet
  • to an intranet
  • to some private users for their own use cases