Machine Learning

Gives computers the ability to learn without being explicitly programmed.

Machine Learning v.s. Data Mining v.s. Artificial Intelligence

• Overlap significantly
• ML: learning properties and adapt to new data
• DM: discovering unknown properties in data
• AI: machines performing tasks that are characteristic of human intelligence
  • ML is a way of achieving AI
  • In the 1980s, AI all about expert systems
    • Expert system = knowledge base + inference engine
    • Problems
      • Knowledge created by hand
      • Things in real world not always true/false
    • ML improvement
      • Knowledge learned from data
      • Probability to represent the real world

Types

Human supervision or not

• Supervised learning
  • k-NN
  • Linear regression
  • Logistic regression
  • SVM
  • Decision trees & random forests
  • Neural networks
• Unsupervised learning
  • Clustering
    • k-means
    • Hierarchical cluster analysis (HCA)
    • Expectation maximization
  • Visualization & dimensionality reduction
    • Principal component analysis (PCA)
    • Kernel PCA
    • Locally-linear embedding (LLE)
    • t-distributed stochastic neighbor embedding (t-SNE)
  • Anomaly detection
  • Association rule learning
    • Apriori
    • Eclat
• Semi-supervised learning
• Reinforcement learning
  • Observe environment, select & perform action, get reward or penalty, update policy

Learning incrementally on the fly or not

• Online
  • "Data in motion"
  • Model updated as data arrive
  • Advantage
    • Suitable for streaming data
    • Suitable for resource-constrained systems
    • Suitable for out-of-core learning (dataset cannot fit memory)
  • Disadvantage
    • Bad data needs to be monitored
• Offline (batch)
  • "Data at rest"
  • Model estimated all data at once
  • Advantage
    • Faster convergence
  • Disadvantage
    • Not efficient if new data frequently come in
    • Not applicable to reinforcement learning

Try to find patterns to build a model or not

• Instance-based
  • Memorize every training instances
  • Generalize to new instances using similarity measure
• Model-based
  • Build a model of the given training data, and use it to make predictions
  • Less vulnerable to bad data

Data

• More data
  • Generalize better
  • May have sampling bias
• Less data
  • Sample noise (non-representative data)

Data v.s. Algorithm

Very different ML algorithms performed almost identically well on a complex problem given enough data.

But small- and medium-sized datasets are still very common.

How to better detect patterns?

• Data cleaning
  • Remove outliers
  • Ignore/fill in missing values
• Feature engineering
  • Select the most useful features
  • Combine features
  • Create new features
• Prevent overfitting (regularization)
  • Simplify model
    • Fewer parameters
    • Reduce features
  • Gather more data
  • Reduce noise in data
  • Add in hyperparameters to tune the pattern
• Prevent underfitting
  • Build more powerful model
    • More parameters
  • Feed better features

Testing & Validating

To see how well the derived model performs, try it on new data instances.

• Split dataset
  • Training set
  • Test set
    • Generalization error (out-of-sample error)

No Free Lunch Theorem

No one model works best for every problem. -> It is common in ML to try multiple models.