Skip to content

Lesson 5 — Machine Learning and Mining by Problem Type

Classify methods by the problem and learning signal, not by how advanced the technique sounds.

Supervised / Unsupervised / Reinforcement

  • Supervised: known/labeled target. Example: emails labeled spam/not-spam train a classifier.
  • Unsupervised: no target labels; discover hidden structure. Example: natural customer groups.
  • Reinforcement: actions improve through reward/feedback toward a goal.

Fast discriminator: What tells the algorithm it is doing well?

Mining-method selector

Technique Best clue
Profiling understand typical structure/content/patterns
Data reduction make a large/complex set smaller while retaining important information
Association find items/events that occur or relate together
Clustering group similar observations/cases
Self-organizing map organize/visualize complex multidimensional similarity
Text mining analyze/classify content in text/documents
Sentiment analysis infer attitude/opinion/feeling from language in context

Association vs clustering

“Which products occur together?” → association.
“Which customers are naturally similar?” → clustering.

Mining vs prediction

Data mining can discover relationships you did not know to ask about. Predictive analytics estimates a specified likely future outcome. Mining can feed a predictive model, but they are not identical.

Sentiment trap

A positive word does not guarantee positive sentiment. Negation, sarcasm, and surrounding context make naive keyword counts unreliable.

Source: pp. 479–484.

← Lesson 4 · Next: Sources & Ingest →