Skip to main content
The terms you will meet in the console, the API and these docs.

​
Platform terms

Feature refinement — A Feature Refinery run. Takes a labelled train and test pair, removes the columns that do not help, and returns a reduced pair plus a per-column decision. See Feature Refinery. Column status — The per-column outcome of a refinement: KEEP or REMOVE, one per original column in header order. Also downloadable as a CSV so the same columns can be dropped from a later file. Dataset — The labelled data a training run reads: an 80% train file and a 20% test file that keeps its labels. Built either from uploads or from a finished refinement. Buckets — How finely the data is split for internal cross-validation. More buckets means more thorough validation and a longer run. 4, 10 or 20 in a refinement; 4 to 1000 on a dataset. Target (performanceThreshold) — The accuracy each training attempt aims for, from 0 to 1. A dataset that came from a refinement carries the target that refinement reached, as recommendedThreshold. Use it. See Tuning. Scaling (scalingFactor) — An internal training parameter, 8 to 499. Default 19. Training run — One Model Crucible job. Many attempts across three algorithm families, all trained at once and scored on the held-back test data. Attempt — One training inside a run. A run has dozens. Each has a number and reports its own score. Candidate — One of the models a finished run offers, ranked best first. Champion — The candidate predictions currently run against. Starts as the best-ranked, and you can move it. Moving it never changes what an earlier prediction meant. Model version — v1, v2, v3 of one model: the best three candidates a run kept. The version number is the ranking. Prediction — Scoring unlabelled rows against a champion. Batch: a CSV in, a CSV out. Also called inference. Blind / unseen data — Rows with no target column, the input to a prediction.

​
How performance is reported

False decisions — How many rows the model got wrong on the held-back test set. Fewer is better, and this is what candidates are ranked on. It is the number that maps most directly to what a mistake costs. Accuracy — The share of all rows predicted correctly. On imbalanced data it flatters: a model that predicts “no fraud” on a 3%-fraud dataset scores 97% and is useless. Precision — Of the rows the model flagged, how many were really positive. Low precision means false alarms.
Recall — Of the rows that were really positive, how many the model caught. Low recall means misses.
F1 — The harmonic mean of precision and recall, in one number. 1 is perfect. On imbalanced problems this is usually a better summary than accuracy.
A metric returned as null is undefined, not zero. A test slice with no positive cases leaves recall undefined, which is not the same as a model that recalled none of them. Show it as absent.

​
Reading them together

Precision and recall trade against each other, and which one you want depends on what a mistake costs. This is exactly the choice the candidates table puts in front of you.

​
General terms

Feature — A column used as input to the prediction. Everything between the ID and the target. Label / target — The outcome column: what you are predicting. Binary classification — Two possible outcomes: fraud or not, good or bad, pass or fail. Multi-class classification — Three or more outcomes. Handled by Model Crucible. Feature Refinery takes binary targets today, with multi-class planned. Class imbalance — When one outcome is much rarer than the other. The normal state of affairs for the problems MathFi.ai is built for, and the reason accuracy alone misleads. Overfitting — A model that has learned the training data including its noise, and so performs well there and poorly on anything new. Scoring against held-back data is how you catch it. Deterministic — The same input produces the same output, every time. Every MathFi.ai model is, which is what makes a decision replayable and auditable. Predictive AI vs generative AI — Predictive AI estimates an outcome from historical data. Generative AI produces new content. This platform is the first kind, and it is the kind you want when a decision has to be right, repeatable and defensible.