Skip to main content
The Inference component runs trained models against feature frames and outputs a DataFrame of predictions. Both of its input ports accept as many wires as you like, and the node is the cross product of them: every model you connect, scored on every data set you connect.

​
Two axes on one node

Wire Data Splitting’s Train X, Val X and Test X all into the same Features port and one Inference node covers all three partitions. This is what lets a single downstream Evaluation report the train-versus-test gap — the number that says whether a score is a real result or a memorised one.

​
How a partition is recognised

Nothing has to be configured. The set a frame belongs to is read from the port it left its source node by: Train X means the training set, Val X the validation set, Test X the test set. A frame arriving from anywhere else is still a set — it is simply named after the node that produced it, so a hand-built holdout is a first-class partition.

​
Choosing what runs

Both rosters are picked on the node card, not in the inspector, because both are facts about the wiring and the options only exist once the wires do.
  • Models — tick a subset to compare three of five trained models without deleting the wires you will want back.
  • Sets — tick a subset to score only test while train and validation stay connected.
Un-ticking everything falls back to running all of them.

​
Input / Output

​
Output columns

A single-model, single-set run produces exactly the frame it always did: one prediction column and no __set.

​
Preview

Opening the preview lands on the Predictions page, which describes what a flat table of 0s and 1s cannot:
  • What this node scored — every partition and its row count.
  • Model agreement — the share of rows where every model predicted the same thing, per set. Two models at 0.85 accuracy that disagree on a fifth of the rows are a different situation from two that never disagree, and no accuracy score can tell them apart.
  • What each model predicted — the class split (or value spread) for each model on each set. A model that answered “survived” to every row is invisible in an accuracy number and obvious here.
The Data page carries a partition filter, so the raw table can be narrowed to one set.

​
Notes

  • The features must match what each model was trained on — same columns, same encoding, same scaling. This is why Inference belongs after the split, not before it.
  • A model trained on a different number of columns than it is being scored on is skipped by name, per set, rather than failing the whole run.
  • Send the predictions to an Evaluation node’s Predicted port, and the same split’s targets to its Actual port.