Source
Adapted from my own handwritten and hand-typed notes from Year 1 to Year 3 of my EEE degree
Most data science questions fall into five primary types. Each has a standard family of solutions.
Problems and Solutions
| Question | Problem type | Example | Solution | Model |
|---|---|---|---|---|
| How much? How many? | Prediction: numeric | expected sales of the next game in a franchise | Regression | |
| Is it type A or B? | Prediction: classes | will a student get into NTU? | Classification | |
| How is this organised? | Detection: structure | customer groups at FairPrice | Clustering | groups depend on "distance" |
| Is it weird behaviour? | Detection: anomaly | is a Boeing engine behaving unusually? | Anomaly detection | deviation from the regular pattern |
| What should be done next? | Decision: action | brake or accelerate at a yellow light? | Adaptive learning | optimise |
Typical Methods
Methods named in the lectures:
- regression: linear regression, tree models for regression, neural networks for regression
- classification: logistic regression, tree models for classification, neural networks for classification
- clustering: k-means, hierarchical clustering, using distances such as Euclidean or Jaccard
- anomaly detection: cluster analysis based detection, nearest neighbour models, support vector based detection
- adaptive learning: reinforcement learning, Monte Carlo, state action reward, Q-learning, deep reinforcement learning
Notes
Points worth remembering:
- classes prediction answers with a probability or a yes or no (0 or 1)
- structure detection is open ended, it asks whether the data is organised in groups
- for anomaly detection the deviations depend on the model of "normal" behaviour
- action decision models a profit or loss function at each state and tries to maximise or minimise it
The first four problems split into supervised and unsupervised learning, see Supervised and Unsupervised Learning.