Running a prediction

This is how you train and run a model on one of your projects. From a project, choose Run prediction and work through the four-step wizard.

Step	What you do
1. Hyper Parameters	Set the model's training parameters (sensible defaults are pre-filled)
2. Context Parameters	Choose which attributes are conditions and which are labels
3. Run	Launch the run and watch live progress and logs
4. Explore Model	Inspect the trained model, then export it

Use Next and Back to move between steps.

1. Hyper Parameters

Parameters are split into two cards. Every field is pre-filled with a sensible default, so you can go straight to the next step if you're just getting started.

Rule Parameters

These shape the individual rules the model is allowed to learn.

Parameter	What it does	Range	Default
Minimum Precision	Minimum precision a rule must achieve to be retained	0–100	70
Minimum Coverage	Minimum proportion of examples a rule must cover	0–100	10
Maximum Support	Upper bound on rule support, to avoid trivial rules	0–100	0
Minimum Number of Conditions	Smallest conjunction size allowed per rule	1–20	1
Maximum Number of Conditions	Largest conjunction size allowed per rule	1–20	3

Precision vs coverage

Raising Minimum Precision gives you rules you can trust more, but fewer of them. Raising Minimum Coverage keeps only rules that apply broadly. Tune the two together.

Validation Parameters

These control how the model is evaluated and when it's accepted.

Parameter	What it does	Range	Default
Number of Runs	How many independent runs to execute	1–100	1
Number of Folds	K-fold cross-validation split count	2–20	2
F-measure Threshold	Minimum F-measure for accepting the model	0–100	50
Precision Threshold	Minimum global precision for accepting the model	0–100	50
Accuracy Threshold	Minimum overall accuracy for accepting the model	0–100	50
Model Size Limit	Maximum number of rules allowed in the final model	1–5000	100
Labels Provided	Whether the dataset already contains labeled examples	on / off	off

2. Context Parameters

Here you tell the model how to read your data by assigning attributes to:

• Conditions — the "if" side of a rule
• Labels — the outcome you want to predict
• Label Values — the specific values of interest within those labels

Use Add to Conditions and Add to Labels to move attributes across, and the filter box to find an attribute quickly in a large dataset.

3. Run

Launch the run and watch it work. The prediction executes on the server and streams progress in real time.

The Overall progress panel tracks three stages:

1. Loading Data — reading and encoding your dataset
2. Experiment — running the folds of the cross-validation
3. Model Building — generating and pruning the rules

Alongside it, the Logs panel streams the run's output; you can copy it, download it, or expand it to full screen. If you need to stop, use Cancel Prediction.

4. Explore Model

When the run finishes, move to Explore Model to inspect the outcome.

• Model Explorer — browse the trained model, switching between a JSON view and a Rules Explorer view.
• Model Quality — opens charts describing how the model performed.
  caution

  Model Quality needs Number of Folds set to 2 or greater. With fewer folds there's nothing to compare against, and the app will tell you so.

• Export to KBase — promote this model's rules into the shared Knowledge Base.
• Download the model — export as .pkl or .pmml for your downstream systems.
• New Prediction Run — start again with different parameters.

tip

Every run is recorded in Runs History, so you can always come back to review the parameters you used and the model you got.