Discussion prompt
20 minutes. Find a claim that a model or algorithm “predicts” something. In 150 to 200 words, list the questions you would ask about its features, label, baseline and test data before trusting it.
Private: nothing you type here is sent or stored.
This course has no discussion board yet, so the prompt is a private reflection.
Recap
- A model maps features to an estimate learned from examples with known answers.
- Rows are examples; columns are features and one label; using the answer as a feature is leakage.
- Absolute error = |prediction − truth|; MAE is their mean; a model must beat a simple baseline.
- Fit on training data and score on unseen test data; a large gap signals over-fitting.
Module notes
Key terms. Feature: a known input. Label: the value to predict. Supervised learning: learning from labelled examples. Baseline: a simple model to beat. MAE: mean absolute error. Training set / test set: data to fit and data to score. Over-fitting: learning the training data too closely. Leakage: using information not known at prediction time.
Rules and formulas. Absolute error = |prediction − truth|. MAE = sum of absolute errors ÷ number of examples. Accuracy = correct ÷ total × 100.
Limits of this module. All tables are invented and tiny. Their numbers teach the ideas; they do not establish anything about real customers or homes.
Next
Modules 2 to 8 are in preparation. Return to the course overview or the University Track.
References and further reading
- Specialty Digest encyclopedia: Machine learning, Inference (machine learning), Deep learning, Regression Analysis and Occam’s Razor.
- All data in Module 1 are invented for teaching and are labelled as such.