ML Glossary

How much of the variation in the target your model explains, compared with always guessing the average. 1.0 is perfect, 0 means no better than the average, and negative means worse than the average.

The average size of the model's mistakes, in the same units as the target. It treats a big miss and several small ones the same as long as they add up alike.

Like MAE, but each error is squared before averaging, so a few big misses raise it sharply. It is always at least as large as MAE; a wide gap between the two hints at occasional big errors.

The average of the squared errors — RMSE before taking the square root. It ranks models the same way as RMSE but is measured in squared units, which makes it harder to read directly.

The average error as a percentage of the true value — "off by 13% on a typical row". Easy to compare across different series, but it explodes when true values get close to zero.

A symmetric cousin of MAPE: each error is divided by the average of the actual and predicted values instead of the actual alone, which keeps the score bounded when true values approach zero.

The share of predictions that were right. Simple and intuitive, but misleading when one category dominates — always predicting the common category can score high while catching nothing.

Of the rows the model flagged as positive, how many really were. High precision means few false alarms.

Of the rows that really were positive, how many the model caught. High recall means few missed cases.

A single number that balances precision and recall (their harmonic mean). Useful when false alarms and missed cases both matter.

The average of each category's recall, so a rare category counts as much as a common one. Prefer it over plain accuracy when the data is imbalanced.

The probability that the model ranks a randomly chosen positive above a randomly chosen negative. 1.0 is a perfect ranking, 0.5 is a coin flip. It judges the model's scores, not just its final yes/no.

A formula the computer wrote by studying your data: it takes the feature columns in and returns a prediction for the target. In Bayanii, the winning model is saved and ready to use from the Models page.

The process of fitting a model to your data: make a guess, measure the error, adjust, and repeat until the error stops shrinking.

An input column the model is allowed to use when predicting — like size, city, or age. Bayanii can also build new features out of your columns, such as splitting dates into parts.

The column you want to predict — the answer the model learns to produce. Every training row must have it filled in; predictions fill it for new rows.

Your table of data — one row per example, one column per attribute, uploaded as a CSV file.

Predicting a number: a price, a duration, a quantity. Judged by how large the errors are (MAE, RMSE) and by R².

Predicting a category: yes/no, churn/stay, A/B/C. Judged with accuracy, precision, recall and F1, all read off a confusion matrix.

Predicting future values of something measured over time, where the order of rows matters — next month's sales, next week's demand. The model uses lags of past values and predicts a horizon ahead.

The model's answer for a row it has not seen before — a number for regression, a category for classification, future values for forecasting.

When a model memorizes the training data instead of learning the pattern — excellent scores on data it has seen, poor scores on new data. Bayanii guards against it with held-back test data and cross-validation.

When a model is too simple to capture the real pattern — it scores poorly on the training data and on new data alike.

A small table comparing predictions with reality: true positives, false positives, false negatives and true negatives. Accuracy, precision and recall are all read directly off it.

Instead of one single test, the data is split into several folds and each fold takes a turn as the test set; the scores are averaged. Bayanii uses it so the winning model is not just lucky on one split.

A setting of the learning process itself — like how deep a tree may grow — chosen before training rather than learned from the data. Bayanii tunes these automatically for the top models.

Past values of a time series used as inputs when forecasting. With 3 lags, the model predicts each month from the previous three months.

How far ahead the forecast goes, counted in time steps. Bayanii predicts one step at a time and feeds each answer back in — so the further out a step is, the more uncertainty it carries.

When information that will not exist at prediction time sneaks into training — like a column recorded after the outcome happened. It inflates scores during training and collapses in real use; Bayanii's data checks flag likely leaks.

When one category vastly outnumbers the others — say 95 customers stay for every 5 who churn. Plain accuracy becomes misleading; Bayanii applies balanced class weights and reports balanced accuracy.

A value far outside the usual range — a 10,000 m² "house", an age of 300. A few outliers can drag averages and models off course; the dataset report flags suspicious ones.