Classical ML
On this page6
- Linear models
- Trees and ensembles
- Distance and probability based
- Unsupervised
- Applied problem shapes
- The two comparisons worth memorising
Classical ML
Still what wins on tabular data in 2026. Gradient boosting beats deep learning on most structured problems, and interviews for AI-facing backend roles routinely check that you know the classics rather than only the LLM layer.
Linear models
| # |
File |
Why it’s asked |
| 01 |
Linear regression |
assumptions, and which ones matter for prediction vs inference |
| 02 |
Logistic regression |
the default baseline; why cross-entropy not MSE; threshold != model |
Trees and ensembles
Distance and probability based
Unsupervised
| # |
File |
Why it’s asked |
| 08 |
Clustering |
k-means assumptions, DBSCAN/HDBSCAN, evaluating without labels |
| 09 |
Dimensionality reduction |
PCA vs UMAP, and compressing embeddings to cut RAG cost |
Applied problem shapes
The two comparisons worth memorising
Bagging vs boosting. Bagging trains independent deep trees in parallel and averages them to cut variance; more trees never overfit. Boosting trains shallow trees sequentially, each fitting the last one’s errors, to cut bias; more trees eventually do overfit, so early stopping is mandatory. Same ensemble family, opposite mechanism, opposite base learner.
Why trees still beat neural networks on tabular data. Heterogeneous feature types, piecewise-constant relationships, robustness to uninformative features, and far lower data and tuning cost. Deep learning wins when you need to embed high-cardinality relations or fuse tabular data with text or images.