Glossary
168 terms
The report and its figures
- Accuracy
The share of rows classified correctly.
In prose, "accuracy" is also used loosely for how well any model predicts; as a figure it always means this one.
- AUC
The area under the ROC curve: how well the scores separate yes from no. 1 is perfect and 0.5 is chance.
- Autocorrelation
How strongly a value relates to the values before it in time.
Residuals that are autocorrelated show that the model has missed something that repeats.
ACF
- Balanced accuracy
Accuracy averaged over the classes, so a rare class counts as much as a common one.
- Baseline
A simple yardstick the model must beat.
Which yardsticks the report uses depends on the kind of target: the mean, the median and a linear regression over the numeric inputs for a continuous value; the most common class and a logistic regression for a yes/no value; the most common class for a category; the most common and the middle level for an ordered level; the average count and a Poisson regression for a count; and for a series over time also the last value, the value one season back and the recent trend extended. A yardstick is only a measuring stick: the report shows its scores and nothing else, and the analysis has one model. A model that barely beats its baseline is listed as a concern.
reference predictor, naive baseline
- Bootstrap
Drawing the held-out rows again many times and recomputing a figure each time, to see how far it moves.
The model itself is computed once and is not computed again.
- Brier score
The mean squared difference between the predicted probability and what happened. Lower is better.
- Calibration
A yes-or-no model is calibrated when a score of 70 % comes true about 70 % of the time.
The reliability diagram shows this as predicted probability against observed share.
- Cardinality
The number of distinct values in a column.
Symbol: |X|.
- Causal honesty
The chapter that states what a model does and does not say about cause.
A model describes how the columns move together; whether one causes the other is a question the data alone cannot answer.
- Class balance
How the rows spread over the classes of a target.
When one class is rare, accuracy alone flatters a model, which is why the report also gives balanced accuracy.
- Cluster
A group of similar rows, found without using the target.
- Complexity
How long the formula is, as a share of the complexity the search was allowed. Lower is simpler.
It describes the formula alone and is not measured on any rows.
model size
- Complexity limit
The largest formula the search may return.
It is set under Advanced in the Analyze Wizard.
complexity budget
- Concern
Something the analysis flagged about the data, the model or its use.
Each concern carries a level: critical, warning or info. The Concerns chapter lists them, the critical ones first.
- Confidence interval
The range a reported figure, such as R², could move in if a different set of rows had been held out.
It is about the figure, not about a single prediction. The report computes it by bootstrap.
- Confusion matrix
A table of actual class against predicted class: the diagonal holds the correct predictions.
- Constant
A fixed number inside the formula, tuned by the search.
- Contribution
What one term adds to the prediction for one row.
- Convergence
The search has converged when the curve of its progress flattens well before the last generation.
More time would not have found a better model.
- Correlation
How closely two columns move together, from −1 to 1.
Pearson's r measures a straight-line relation; Spearman's ρ and Kendall's τ measure whether the two rise and fall together in rank. A column of categories has no direction to move in: a number and a category are measured by the correlation ratio η, two categories by Cramér's V, both from 0 (unrelated) to 1. A correlation says nothing about cause.
- Covariance
How two columns vary together, in the columns' own units.
- Coverage
The share of held-out rows whose actual value fell inside the prediction interval.
A 95 % interval that covers about 95 % of the rows is well calibrated; far less means the interval is too narrow.
- Data leak
A column that gives the answer away because it was derived from the target after the fact.
An example is a tax amount computed from the price being predicted. The analysis flags columns that reproduce a target.
leakage
- Decile
The deciles D1 to D9 are every tenth percentile.
- Deployment readiness
A list of checks a model passes, passes with a warning or fails before it is deployed, with one verdict: ready, ready with caution, or not ready.
- ECE
The expected calibration error: the average gap between predicted probability and observed share. Lower is better.
- Effect size
How large a difference is, in units of spread.
The report uses Cohen's d: a difference in means divided by the standard deviation.
- Error
How far the predictions lie from the held-out values, as a share of the miss of always predicting the average. 0 % is exact, 100 % is no better than the average.
For a quantity target it equals 1 − R² on the same rows. For a count or a category target the comparison is with the average count or the most common class. The word is also used in its everyday sense for a single miss; the entry Residual names that precisely.
- Extrapolation
Predicting outside the range of the training rows.
Predictions there are less reliable than inside the range, and the report shows how the formula behaves beyond it.
- F1
The harmonic mean of precision and recall.
Macro-F1 is F1 averaged over the classes. Support is the number of held-out rows in a class.
- Fairness
The chapter that checks whether predictions differ between the groups of a categorical column.
- Generalization
How the model's figures on the training rows compare with those on the held-out rows.
- Held-out rows
Rows set aside before the search, which the model never saw.
Quality and error are measured on them, which is why they say how the model does on new data. The words test and validation are not used for them.
unseen rows, rows the search never saw
- ICE curve
An individual conditional expectation curve: the trace of partial dependence for one single row.
- Input
A column the model may use to predict the target.
feature, variable, predictor, independent variable
- Input importance
How much the model's error rises when one input's values are shuffled across the rows.
An input whose shuffling changes nothing is not used by the model.
feature importance, permutation importance
- Input range
An automatically chosen slice of one input, for example a quartile of a numeric column, which the report uses to check where the model is weaker.
feature segment
- Interaction
Two inputs combined in one term, so that the effect of one depends on the value of the other.
- Known before the outcome
The wizard setting in which you confirm that a column is available before the target is known, so it is a legitimate input.
A confirmed column raises no concern.
- KS
The Kolmogorov-Smirnov statistic: the largest gap between two distributions, from 0 (identical) to 1.
- Lift
The model's margin over the best baseline, in the measure its kind of target is judged on. Positive means the model is ahead.
It is counted in points of variation explained (R²), of balanced accuracy or of accuracy, on the scale of weighted kappa, or for a count as the share of the baseline's Poisson deviance the model removes. The margin comes with a 95 % interval; when the interval includes zero, the lead is not settled on that many rows.
Baseline, R², Balanced accuracy, Quadratic-weighted kappa, Poisson deviance
- MAPE
The mean absolute percentage error: the average miss as a percentage of each actual value.
Shown on the Lab's model page.
- Missing value
A blank cell, or a code that stands for "no value".
The Missing Values chapter reports the share per column.
- Model constraints
Three pass or fail checks on a formula, run on the held-out rows.
Every prediction is a finite number, predictions stay within the range of the training target widened by two standard deviations, and no prediction runs away more than ten standard deviations from the training mean. They are guards and say nothing about quality.
- NMAE
The normalized mean absolute error: MAE divided by the range of the target on the training rows.
Shown on the Lab's model page. Because it is scaled by the range, it can be compared between targets of different scale.
- NMSE
The normalized mean squared error: the mean squared error divided by the square of the target's range on the training rows.
Shown on the Lab's model page. Because it is scaled by the range, it can be compared between targets of different scale.
- Out-of-distribution
A row that lies far from the rows the model was computed on.
Predictions there are less reliable than inside the range.
OOD
- Outlier
A value more than 1.5 times the IQR below Q1 or above Q3, by the Tukey rule.
An outlier is a value to look at, not necessarily a wrong one.
- Over-dispersion
A count target that varies more than a count model assumes.
The Count Distribution chapter reports it.
- Overfitting
A model that does much better on the training rows than on the held-out rows: its formula has absorbed noise from the training rows.
- p-value
How unlikely a correlation this large would be by chance alone.
A small one, for example below 0.05, means chance is an unlikely explanation.
- Parity ratio
The smallest group's mean prediction divided by the largest group's.
1.0 is perfect parity, and below 0.80 some group receives systematically different predictions.
- Partial dependence
How one term of the formula responds as one of its inputs moves across the training range while the others stay at their medians.
- Percentile
The value below which a given share of the rows lies.
P95 is the value below which 95 % lie.
- Poisson deviance
The error measure for a count target: how far the predicted counts lie from the actual ones, in a form suited to counts.
Lower is better and 0 is exact.
- Precision-recall curve
Precision plotted against recall over all thresholds.
It is the better picture when yes is rare.
- Prediction interval
The range a single prediction's true value is expected to fall in, at the stated level.
It is derived from the residuals of the held-out rows. The report uses a 95 % interval, from the 2.5th to the 97.5th percentile of those residuals. A deployment returns a 90 % interval, from the 5th to the 95th percentile, so nine in ten held-out rows had their actual value inside it.
- PSI
The Population Stability Index: it compares how a column's values spread over ranges in two sets of rows.
The larger the figure, the more the column has moved.
- Q-Q plot
Residuals plotted against a normal distribution.
Points on a straight line mean the residuals are normally spread.
- Quadratic-weighted kappa
Agreement between predicted and actual ratings that counts a near miss as better than a far one.
1 is perfect agreement and 0 is chance.
- Quality
How well the model predicts on held-out rows, marked down for the size of its formula. 0 % is no better than always predicting the average, 100 % is exact.
Because it is measured on rows the search never saw, a model cannot score well by memorizing the table. For a quantity target, quality sits at or just below the R² of the same rows, and the gap between the two is what the formula's size costs. Quality is not R² and is not a share of variance. During the search, the same measure on the training rows is what candidates are ranked by.
- Quartile
The quartiles Q1 and Q3 are the 25th and 75th percentiles.
- Question
A question typed before launch, in your own words.
The report answers it in the Questions & Answers chapter.
- R²
The share of the target's variation the model accounts for on a set of rows. 1 is exact, 0 is no better than always predicting the average, and a value below 0 is worse than the average.
The report states R² on held-out rows unless it says otherwise.
R-squared, coefficient of determination
- Residual
Actual value minus predicted value, for one row.
A positive residual means the model predicted too low.
- RMSE
The root mean squared error: a typical miss in the target's own units that counts large misses more heavily.
It is never smaller than MAE. The report gives RMSE for count targets.
- ROC curve
The true positive rate plotted against the false positive rate over all thresholds.
- Search progress
The chapter that shows quality on the training rows over the generations of the search.
- Segment
One value of the column chosen under "Segment by" in the wizard; the report shows results for each segment.
segment key, segment column
- Sensitivity
How strongly the prediction changes as one input moves across its training range, with the other inputs left as they are.
- Skewness
Whether a distribution leans to one side: positive means a long tail towards high values.
- Standard deviation
How widely values spread around the mean, in the column's own units: the square root of the variance.
Symbol: σ.
- Target
The column the model predicts.
label, dependent variable, output
- Target kind
How the target column is read: a quantity (regression), a count, a rating (ordinal), yes or no (binary), or a category (multiclass).
The kind decides which chapters the report contains.
- Term
One additive part of a formula: the formula is the sum of its terms.
- Term ablation
How much the held-out R² drops when a term is replaced by its average contribution.
It says how much predictive work the term does.
- Threshold
The score above which a yes-or-no model answers yes.
- Training rows
The rows the search used to find the model.
- Variance
The average squared distance of the values from the mean.
Symbol: σ².
- What-if
The prediction for input values you set yourself.
- Worst predictions
The held-out rows with the largest residuals.
The platform
- Activity log
The record of what happened in a team.
Part of the Team plan.
- Admin console
Where a team's owner and admins manage members, seats and billing.
Part of the Team plan.
- Analysis trigger
A schedule attached to an analysis that starts a new run when it fires.
It is a different thing from a function's trigger.
schedule
- Assistant
The conversational interface to the platform.
It answers questions about a report, the product and your project.
Chat
- Credit
The unit that measures compute, storage, transfer and AI usage.
One credit is about one vCPU-minute of compute.
- Credit budget
The most credits a run may use. Only the credits used are charged.
budget
- Data source
A lasting connection between remote data and a table, refreshed on request or on a schedule.
- Deep merge
A merge that also copies the entities the merged entity links to.
- Deployment
One target of a finished analysis as an authenticated prediction endpoint.
- Depth
How long the search may run and what it may spend.
There are five: Express (1 hour, 500 credits), Quick (2 hours, 2,000), Standard (4 hours, 10,000), Thorough (8 hours, 40,000) and Deep (24 hours, 150,000).
- Entity
The central unit of information in a project: a table, a function, a notebook, a model and so on.
- Entity branch
A parallel version of one entity inside a project branch.
- Evaluation
One execution of a function.
- Extension Pack
Additional monthly credits on the Ultimate or the Team plan.
- Fork
A new project branch created as an independent copy of an existing one.
- Formula
The written form of a model.
The developer documentation calls it an expression.
expression, equation
- Function
Code that runs on the platform and can read and change the project.
It is written in Java, Kotlin, Scala, JavaScript or Python against the FDK.
intelligent function
- Locked project
A project that cannot start new work until credits are available again.
Its data stays safe and visible.
- MCP
The Model Context Protocol: how an AI agent connects to Diafunc.
The server is at diafunc.com/mcp.
- Merge
Copying an entity from one project branch into another.
- Model
What an analysis returns for a target: a short piece of readable code.
The report prints it as a formula, which is the written form of the model. "Model" also names the entity that holds it.
- Model export
The model as standalone code in another language.
- Notebook
An entity that holds a report, or your own text, tables, charts and code.
- Project
The top-level unit of work in Diafunc: it holds your data, analyses and models and is private unless you share it.
- Project branch
A parallel copy of a whole project; every project has at least one, called main.
- Project graph
Everything in a project, as entities connected by links.
- Report
What an analysis writes: a notebook in the browser, and the same content as PDF, EPUB or Excel.
- Role
What a person may do in a team or in a project.
Team roles are owner, admin and member; project roles are guest, contributor, maintainer and owner.
- Run
One execution of an analysis; an analysis that is run again has several runs.
- SCIM
A standard through which the organization's identity provider adds and removes team members automatically.
Part of the Team plan.
- Selection
In the FDK, a table-like view on data that functions filter, join and analyze.
- Settings
What an analysis is launched with: targets, questions, segment and depth.
recipe
- Skeleton
A formula template you give the search to try, with placeholders the search fills in.
- SSO
Single sign-on: members sign in through the organization's identity provider.
Part of the Team plan.
- Table
An entity that holds rows and columns.
An analysis reads one.
- Team
A shared subscription with pooled credits.
- Time limit
The longest the whole run may take.
time budget, computation time
- Transaction
The unit of change to a project: all of its changes apply, or none do.
- Trigger
A rule that evaluates a function automatically.
The Lab offers four types: Cron fires on a calendar pattern, Interval fires at a regular distance in time, Continuous keeps the function running, and Event fires when an event is raised on an entity.
- Version
A stored earlier state of an entity, kept when versioning is enabled for that entity.
The Lab switches versioning on and off per entity.
- Welcome credits
The credits a new account starts with. They do not expire.
welcome grant
Ufinq
- Candidate
One formula the search is considering.
individual, solution, program
- Constant tuning
Adjusting the numbers in a formula to the data.
It happens to every candidate during the search.
- DEBOHAC
The name of Ufinq's method for distributing a search over a cluster.
It stands for Distributed Evolution through Budget-Driven Orchestration over Hierarchically Aggregated Capacity.
- Decomposition
Splitting a hard problem into smaller searches whose results are combined.
On the Lab's model page the same word names the split of a formula into its terms.
- Evolutionary branch
An independent group of candidates; Ufinq searches in several at once.
An evolutionary branch is not a project branch.
- Generation
One round of the search, in which candidates are varied, measured and selected.
- One model
An analysis returns one model for each target, not a list to choose from.
Candidates are ranked by a single quality value that already weighs size against accuracy, and the one that holds up best is returned with its evidence.
- Population
The candidates of one evolutionary branch at one moment.
- Readable model
A short piece of explicit code that says how the target depends on the inputs.
You can read every line and check it against what you know.
white-box model
- Simplification
Rewriting a formula into a shorter one that computes the same.
It happens to every candidate during the search.
- SRBench
The public Symbolic Regression benchmark Ufinq is evaluated on.
- Symbolic AI
AI whose result is explicit structure a person can read: code and formulas, not trained numbers.
- Symbolic Regression
A search for the form of a model and its constants together, directly from the data.
It is the method underneath Symbolic AI.
- Ufinq
Diafunc's Symbolic AI technology, built on Symbolic Regression.
It discovers the model in an analysis.