Machine Learning Fundamentals and Problem Framing
Know the problem type first
Every Alteryx Machine Learning project begins with framing: is the target categorical (classification) or continuous (regression)? Is there a label at all (supervised) or not (clustering, anomaly detection)? The tool asks you to select a target variable and infers the problem type from its data type, but you must confirm it. A binary target with rare positives is a classification problem with class-imbalance concerns, not a regression problem on 0/1 values.
Concepts to master
- Supervised vs unsupervised: prediction with a known target versus structure discovery.
- Train/validation/test split: the test set is held out until final evaluation; validation guides model selection and tuning.
- Overfitting vs underfitting: high training performance with poor holdout performance signals overfitting; poor performance on both signals underfitting.
- Leakage: any feature that encodes information unavailable at prediction time (or derived from the target) inflates scores.
- Baseline thinking: compare against a naive baseline (majority class, mean prediction) before celebrating a model.
Traps in questions
Questions often describe a scenario where the "obvious" target is actually a proxy that would not exist at scoring time, or where a target is a date or ID. Another common trap: confusing a regression metric expectation with a classification problem because the label is numeric-coded.
Rule of thumb
Ask three questions before building: What decision will this model support? What will be known at scoring time? What does a naive baseline score? If you cannot answer all three, the framing is not finished.