Explaining Model Predictions by Hand: Contributions, Shapley Values, Permutation Importance, Partial Dependence & Counterfactuals

What you'll build

What you will be able to do

Each lab ends with a concrete result that you build yourself and the checker confirms, cell by cell.

• Predict the rent of eight flats with a toy linear model, compute the baseline (the average prediction), split every prediction into one contribution per feature and check that baseline plus contributions rebuilds the prediction.
• Build all eight coalitions of three features for one machine, let absent features fall back to the fleet baseline, score every mix with a toy model that has an interaction term and measure the total effect that the explanation has to share out.
• Let three features join in all six possible orders, record each feature's marginal contribution, average them into exact Shapley values, recompute one with the coalition-size weights and confirm that the three values add up to the total to explain.
• Score the four coalitions of a two-feature sales model with an interaction term, compute both Shapley values, isolate the interaction effect and find out how much of it each feature receives.
• Measure the toy rent model's error on ten test flats, replace one feature at a time with a shuffled copy, measure the error again and rank the features by how much the model relies on them.
• Set vibration to a grid of values for four machines, draw one individual (ICE) curve per machine, average them into the partial-dependence curve, measure its slope and chart both to see where the average hides an interaction.
• Score M-17 on a grid of smaller vibration values and a grid of fewer hours since service, find the smallest change in each feature that brings the toy score below the service threshold with MINIFS, and compare the two counterfactuals.
• Turn local contributions into a global importance ranking, check how stable permutation importance is across three shuffles, chart the ranking and write down what the explanation can and cannot say.