Causal inference: XAI tab

Which variables drive one target in the current graph, and in which direction: SHAP values from the dataset's rows when they are available, otherwise a structural attribution from the graph's weights.

What the tab is for

XAI (explainable AI) shows which variables drive one target in the current graph and whether each pushes it up or down. Without a graph it says No graph to explain yet and offers Go to Discovery.

The card has the same shape as the Reasoning views: an intro (the question, what to enter, what you get, an example in the graph's own variable names), the Target variable input, Run attribution, the result, and How to read this.

Run it

Pick a Target variable and select Run attribution.

XAI for Erk on the Sachs graph

Reading it

• The method chip, first in the result: SHAP values (linear model) or Structural attribution, with a line saying what the figures were computed from. The engine's sentence under it names the largest contributors with their share of the total contribution.
• Contribution to the target: one row per variable, sorted by magnitude. The bar grows to the right of the centre line when raising the variable raises the target, to the left when it lowers it; the signed value is beside it. A variable with no path to the target reads 0.
• Variables that act together: pairs among the five largest contributors that the graph links within three steps, with the kind of link and their two |contributions| added. No interaction term is estimated.
• Each edge's role towards the target: every edge of the graph with its role (direct into the target, mediating on a directed route to it, lateral on no directed route to it, or undirected, drawn A — B, for a pair the run left without a direction), its stored weight (present for an algorithm that reports none, without a sign for one that reports only a strength), its measured confidence and an importance (the edge's |weight| times its confidence; not reported when no confidence was measured; not a SHAP value).

The method chip matters.

• SHAP values (linear model): the page sends the rows of the dataset selected on the Discovery tab (when it holds every variable of the graph); the engine fits a linear model of the target on the other variables and reports exact SHAP values, as each variable's mean absolute contribution in the target's units, signed by its coefficient. A linear summary of the data around the target, not a test of the graph.
• Structural attribution: without those rows (another dataset selected, or a saved graph whose dataset is not selected), the figures come from the graph's edge weights, not from data: for each variable, the products of the weights along each directed route to the target, summed. When the run's weights carry no sign (PC and the other structure-only algorithms, UCMTv4's strengths) only the size is shown. They summarise the graph's structure and inherit every error in it.

Attribution is not proof: a large contribution says the variable matters in this model or graph, not that changing it will change the target. That is what the Reasoning tab's counterfactual and the engine's checks on the Trust view are for. A value the run did not measure reads not reported and is never estimated in your browser.