Introduction

  • Usage of algorithms to derive predictive insights from data and make repeated decisions
  • Scaling up business intelligence and decision making
  • Forwards looking—compare with analytics (backwards looking)
    • Predictions/inferences
  • A computer writes a program for another computer
    • Computer figures out best program by looking at examples

Vs AI

  • AI
    • Discipline (c.f. Physics)
    • Abilities of computers to mimic human behaviours
  • ML
    • Methodology
    • Generally refers to supervised and unsupervised learning

What Problems Can ML Solve?

  • Things currently done via ‘rules’
  • Scale, automate, personalise
  • Not for problems were one set of rules fits all—generic
    • Personalisation—lots of data required

Best Practices

Concepts and Definitions

TermDefinition
Example/instance/observationRow in data set
LabelValue to predict. Can be numeric (requires regression model) or categorical (requires logistic model)
FeaturesColumns which aren’t labels.
Feature engineeringCombine/transform features.
One-hot encodingTransform categorical data to numeric. 1 or 0 to indicate whether category applies to the observation
HyperparametersUser-defined parameter used to control the learning process.

Model Types

  • Supervised learning—task driven, identifies a goal, labelled data
    • Classification—predict a category
    • Regression—predict a continuous number
  • Unsupervised learning—data-driven, identifies a pattern, unlabelled data
    • Clustering
    • Association—identify underlying relationships
    • Dimensionality reduction—transform data from a high-dimensional space to a low-dimensional space, combine features

Model Architectures

  • Convolutional Neural Networks (CNNs)—image classification, object detection, recommendation systems
  • Recurrent Neural Networks (RNNs)—sequence modelling, next word prediction, translation, speech-to-text
  • Sorting/clustering—anomaly detection, pattern recognition
  • Generational Adversarial Networks (GANs)—anomaly detection, cyber security, self-driving cars, reinforced learning

Strategy

  • Follow The Path to Machine Learning
  • Starts as data strategy
  • Start with manual data analysis
    • Fail fast, try new ideas in an agile way
    • Then move to ML
  • Data collection is the longest/hardest part—most likely to fail
  • If you can’t do analytics, you can’t do ML

ML Effort Allocation

  • Expectation:
    • Optimizing ML algorithm is most significant effort—practitioners focus on latest ‘shiny’ thing
  • Reality:
    • Data collection and building infrastructure often overlooked
    • By spending time collecting data, KPIs often become apparent—rather than relying on intuition

Graph View