Causal inference: Discovery tab

Choose an algorithm and a dataset, run causal discovery, and read the result: summary tiles, the graph, the relationships, the edge table and the metrics.

What the tab is for

Causal discovery looks for the cause-and-effect structure in a table: which variables drive which. On the left you choose and run; on the right the result appears. The header line under the page title repeats the algorithm, whether a dataset is selected and, after a run, the number of nodes and edges.

Choose an algorithm

The algorithm list grouped as Constraint-Based, Score-Based, Optimization, Non-Gaussian and UCMT, with UCMT v2 Global selected — Select Causal Discovery Algorithm

Group	Algorithms	In short
Constraint-Based	PC, FCI	Independence tests between variables. FCI allows for hidden common causes.
Score-Based	GES	Searches for the structure with the best BIC score.
Optimization	NOTEARS	Fits the whole graph at once under an acyclicity constraint.
Non-Gaussian	LiNGAM, DirectLiNGAM	Linear relationships with non-Gaussian noise.
UCMT	v0, v1, v1 Sparse, v2 Hybrid, v2 Global, V3	UCMT Labs' own models.

UCMT v2 Global is selected when the page opens, in both modes; every algorithm of the catalogue can be chosen in Executive mode too. It orients links with dedicated direction tests (TRO, ANM, IGCI) and reports its gates after the run.
UCMT V3 (experimental): its description notes that its internal benchmarks have not yet been independently reproduced. Treat its results like any other algorithm's: compare, and check against what you know.
About this algorithm, under the list, explains the selected algorithm in a sentence and, under Scientific details, its assumptions, what it returns, how it scales and its published reference (the UCMT models have none).
The descriptions in the list name internal stages of the UCMT models (phases, blocks, gates, REF-001). They describe how the model works; the result reports those stages for your run in the Discovery quality panel.

Choose the dataset

Select Dataset lists the built-in examples and your uploads with their domain. To add one, use the Data tab.

Advanced options

Estimated Run Time: a rough figure from the algorithm and the dataset size.
Compare Multiple Algorithms: tick two or more algorithms and select Compare … Algorithms to run each on the dataset and see them side by side under Algorithm Comparison Results.
Each algorithm runs with the engine's own settings: the run request carries the dataset, the domain and the algorithm, so the page offers no hyperparameter, bootstrap or preprocessing switches. These options work on every plan.

Consensus discovery

Consensus runs several algorithms on the same data and keeps the edges enough of them agree on. Tick at least two algorithms (three or more is more useful), set Agreement Threshold (40–100 %; the line under it says how many algorithms must find an edge) and select Run Consensus Discovery. Consensus Results groups the edges by agreement (high, medium, low, single algorithm). Agreement between algorithms is not proof of causation.

Consensus needs a plan with advanced analytics (Professional or Enterprise as created by this release). On another plan the card says so above the button, which stays disabled, with a link to the plans.

Run

Select Run Causal Inference. How the run proceeds, and how to cancel a queued one: How a causal run works.

Reading the result

Summary tiles.

Three tiles: Confidence 80.0%, DAG Valid with a check mark, Nodes/Edges 11/10 — The summary tiles of a UCMT v2 Global run on the Sachs data

Confidence: the average of the confidences the run measured. Only some algorithms measure one: UCMT v2 Global and V3 when their resampling step measured an edge's stability, UCMT V4 when its subsample vote is turned on. PC, GES and the other classical algorithms measure none, and the tile then reads not reported; no number is put in its place. Compare confidence between runs of the same algorithm only.
DAG Valid: the graph has no cycles (no variable causes itself through a loop).
Nodes/Edges: variables and discovered relationships.

The graph. Boxes are variables, arrows go from cause to effect: green for a positive effect, red for a negative one, thicker for a larger weight. A dashed line without arrowheads is an edge whose direction the algorithm did not decide; two heads mean both directions were reported. Drag a box to move it, drag empty space to pan, scroll or pinch to zoom; Fit shows the whole graph, Reset restores the automatic layout. Edge labels (weight and confidence) are hidden until you turn on Edge labels; a confidence the run did not measure is not shown. Hover a box to highlight its connections; long names are shortened and shown in full on hover. Expand opens the graph full screen and Export downloads it as PNG, SVG or PDF on a sheet with the UCMT Labs header, the algorithm, the date, the counts and a legend. The line under the card's title says the graph was saved.

The discovered causal graph of the Sachs data: protein boxes such as Raf, Mek and Erk joined by weighted arrows — Discovered Causal Graph

Causal Relationships (Executive: What Causes What) lists each edge as cause → effect with its weight, confidence, effect strength and a sentence of interpretation, strongest first.

The Causal Relationships card: totals at the top, then numbered cause-effect pairs such as Akt to Erk with weight +0.858 — Causal Relationships

Key Discovered Edges is the same set as a table with confidence bands (Very High ≥ 90 %, High 70–90 %, Medium 50–70 %, Low < 50 %), the strongest edge and the most confident one. Export Edges (CSV) downloads it.

The Key Discovered Edges table with ten rows and the Export Edges (CSV) button — Key Discovered Edges

Node Statistics lists each variable's type, mean and standard deviation; its Distribution column reads "normal" for every variable in this release.

The evaluation metrics

The Causal Discovery Evaluation Metrics card with graph structure, causal effects, edge weights, confidence, DAG quality and performance figures — Causal Discovery Evaluation Metrics

What each group is calculated from:

Group	Calculated from
Graph Structure, Causal Effects, Edge Weights, Confidence Analysis	The discovered edges: counts, degrees, sums and averages of weights and confidences.
DAG Quality	Valid DAG is the run's acyclicity flag, checked on the returned edges.
Performance	The time the engine reported, and the counts of nodes and edges.

Ground truth and advanced analytics

For a built-in example with known relationships, Ground Truth Validation compares the graph with them (precision, recall, F1, structural Hamming distance). Your own uploads have no ground truth, so the card does not appear.
Advanced Analytics (switch on the card) adds the distribution of edge weights and confidences, an Adjacency Matrix, Node Centrality, and ROC/PR Curves, which need ground truth.

The Advanced Analytics card switched on, showing the weight summary and the edge weight and confidence distributions — Advanced Analytics

Executive mode

The Analyze step shows one card: the algorithm (UCMT v2 Global is pre-selected; any algorithm of the catalogue can be chosen, and About this algorithm describes the selected one), the dataset, Run Analysis, and after the run View Insights. The result is the same, worded for non-specialists: Result Reliability is the confidence, Analysis Valid the DAG check, Factors / Connections the nodes and edges.

Under Result Reliability and on the Validation tile, the verdict rests on the engine's own checks, never on the confidence alone: Engine checks passed only when every gate passed and REF-001 reported the data ready; Engine checks not passed says which did not; Not checked by the engine for an algorithm that reports no gates. None of them is a proof of causation: review the key relationships before acting on them.