Alteryx Machine Learning key terms

A modeling approach where an algorithm learns from labeled examples, mapping input features to a known target outcome so it can predict that outcome on new, unseen records.

A modeling approach that finds structure in data without a labeled outcome, such as grouping similar records or reducing many variables into fewer components.

The outcome column a supervised model is trained to predict, such as a churn flag, a sales amount, or a risk category.

A supervised task whose target is a discrete category, such as yes/no, fraud/not fraud, or one of several customer segments.

A supervised task whose target is a continuous numeric value, such as revenue, demand, or time to failure.

A classification problem with exactly two possible outcomes, typically a positive class and a negative class, evaluated with metrics like precision and recall.

A classification problem with three or more mutually exclusive outcome categories, where each record belongs to exactly one class.

An input column or derived attribute used by a model to make predictions, such as age, transaction count, or days since last purchase.

Data organized into rows and columns with defined fields, such as tables from databases, spreadsheets, or CSV files, which is the typical input for tabular modeling.

The declared category of a field, such as numeric, string, date, or boolean, which determines how a tool reads, transforms, and models the values.

An absent or null entry in a field. Models usually require handling it through removal, imputation, or an indicator so the algorithm can process the record.

Filling in missing values with a substitute such as the mean, median, mode, or a model-based estimate, so records remain usable for training or scoring.

A value far outside the typical range of a field, which can distort statistics and model behavior if not examined, capped, or otherwise handled.

The work of cleaning, reshaping, joining, and validating raw data so it is consistent, complete, and suitable for model training and scoring.

Creating, transforming, or selecting input variables, such as ratios, date parts, or aggregations, to help a model capture meaningful patterns.

Converting categorical values into a numeric form a model can use, for example one-hot encoding categories into separate indicator columns.

Rescaling numeric features to a comparable range or distribution so that no variable dominates purely because of its units or magnitude.

A field whose values are labels or groups rather than measured quantities, such as region, product type, or customer tier.

A modeling mistake where information unavailable at prediction time, or drawn from the test set, influences training and inflates measured performance unrealistically.

The portion of labeled data used to fit a model's parameters, from which the algorithm learns the relationship between features and the target.

A held-out portion of data used during model development to compare candidates and tune settings without touching the final test data.

A held-out portion of data used once at the end to estimate how a finished model is likely to perform on new, unseen records.

A technique that splits data into multiple folds, training and validating repeatedly so performance estimates are more stable and less dependent on one split.

When a model learns noise and idiosyncrasies of the training data so closely that it performs well in training but poorly on new data.

When a model is too simple or insufficiently trained to capture the underlying pattern, producing weak results on both training and new data.

A model's ability to make accurate predictions on data it was not trained on, which is the real measure of practical usefulness.

A setting chosen by the modeler before training, such as tree depth or learning rate, that controls how the learning algorithm behaves.

The process of comparing candidate algorithms and configurations using validation results and choosing the one best suited to the problem and data.

A penalty applied during training that discourages overly complex models, helping reduce overfitting and improve performance on new data.

A model that splits data through a series of if-then rules on feature values, producing an interpretable tree structure for prediction.

An ensemble of many decision trees trained on varied samples and feature subsets, whose combined predictions are usually more accurate and stable than a single tree.

A modeling strategy that combines predictions from multiple models, such as bagging or boosting, to improve accuracy and robustness over any single model.

An ensemble technique that builds models sequentially, each new one correcting errors of the previous ones, often yielding strong accuracy on tabular data.

A table comparing predicted versus actual classes, showing true positives, true negatives, false positives, and false negatives to reveal error patterns.

The share of all predictions that are correct. It can be misleading when classes are imbalanced, since a majority-class guess may score high.

Of the records a model flagged as positive, the proportion that truly are positive, reflecting how often a positive prediction is correct.

Of the records that truly are positive, the proportion the model successfully identified, reflecting how many real cases were caught.

A situation where one outcome category is far more common than another, which can bias models toward the majority class and distort accuracy.

A measure of how much each input variable contributes to a model's predictions, useful for interpretation, trust, and simplifying the model.

The degree to which a person can understand why a model produced a given prediction, which matters for trust, review, and regulatory obligations.

Applying a trained model to new records to generate predictions, probabilities, or scores that can be stored and acted upon.

Moving a validated model into production so it can score new data on a schedule or on demand and feed results into downstream processes.

The gradual decline in a deployed model's performance as real-world data and relationships change away from what was seen during training.

Ongoing tracking of a deployed model's inputs, outputs, and performance so drift, data issues, or degradation are detected and addressed.

The practice of building and operating models that are fair, transparent, accountable, and respectful of privacy and applicable rules.

Systematic error that unfairly favors or disadvantages groups or outcomes, often arising from unrepresentative data or flawed problem framing.

The policies, roles, documentation, and controls that ensure models are developed, reviewed, deployed, and retired consistently and accountably.