...find key drivers
Jump to navigation
Jump to search
| AI Solver map |
|---|
|
Your current page is in bold
0 · Start
1 · Predict & analyze
tables, numbers, events
Predict values
Predict categories Forecast time series Cluster / segment Find outliers Make recommendations Find key drivers 2 · Understand unstructured data
text, documents, images, audio
3 · Generate & converse
4 · Decide & act
5 · Build it right
applies to every path
|
YouTube ... Quora ...Google search ...Google News ...Bing News
- AI Solver ... Dimensional Reduction ... Causation vs. Correlation ... Principal Component Analysis (PCA) ... Association Rule Learning ... ...cluster
Use this page when the question is why or what matters: which factors drive churn, which features a model relies on, whether a price change caused a drop in sales. Be clear which question you are asking, because "the model uses X" is not the same as "X causes Y".
Start here
- Which inputs matter most to my model?
- Permutation importance (how much accuracy drops when a feature is shuffled); SHAP values for global importance and for explaining single predictions; partial dependence plots to see the shape of an effect
- Too many features to see, store or model? → Dimensional Reduction
- Linear and fast → Principal Component Analysis (PCA)
- Visualize clusters in 2-D → UMAP or t-SNE (for pictures only; they distort distances)
- Nonlinear compression → Autoencoder (AE) / Encoder-Decoder
- Keep only useful original columns → feature selection (Lasso, mutual information, importance-based pruning)
- How do variables move together?
- Pairs of numbers → Correlation analysis; items that occur together → Association Rule Learning
- Will changing X cause a change in Y? (price → churn, campaign → sales)
- You can run an experiment → randomized A/B test (the gold standard)
- You cannot → causal inference on observational data: difference-in-differences, propensity-score matching, instrumental variables, double machine learning (DoWhy, EconML). To find who responds to an action → uplift modeling (CausalML)
- Want to ask questions of your data in plain English? → AI data-analysis assistants and text-to-SQL: ...answer questions from my data. Check the generated code or SQL.
Quick picks
| Question | Try first | Then compare |
|---|---|---|
| What drives my model's predictions? | SHAP values | Permutation importance, partial dependence |
| Too many columns | PCA | Feature selection, autoencoder |
| What goes together? | Correlation matrix, association rules | Clustering (...cluster) |
| Did my change cause the effect? | A/B test | Difference-in-differences, double ML |
| Who should get the offer? | Uplift model | Contextual bandit (...optimize decisions) |
Real-world examples
- Experiments over intuition — Booking.com, describing 150 customer-facing models, reported that better offline model metrics did not reliably translate into business gains, so it judged models by randomized controlled experiments. 150 successful machine learning models: 6 lessons learned at Booking.com | The Morning Paper
- Uplift modeling — Uber open-sourced CausalML to estimate which users respond to a treatment such as a promotion, so offers go where they change behavior. CausalML | Uber
Watch out
- Correlation is not causation: ice-cream sales and drownings rise together because of summer.
- SHAP explains the model, not the world. Correlated features share credit in unstable ways.
- Confounders and Simpson's paradox can reverse a trend when you split the data by group.
Measure success
For explanations: stability across retrains and agreement with domain experts. For causal effects: confidence intervals, robustness checks (placebo tests), and, where possible, confirmation by experiment. More: ...evaluate and trust AI
Learn more
- SHAP documentation
- DoWhy: causal inference library | PyWhy
- Permutation feature importance | scikit-learn
- Dimensional Reduction ... Causation vs. Correlation ... Time series analysis
Where next? ◀ Predict & analyze · ...predict values · ...cluster · ...optimize decisions · ▶ ...evaluate and trust AI