Glossary

Use this glossary to understand terms specific to Kumo Relational and the NVIDIA Kumo Relational Client. Follow the links for procedures, examples, and complete reference information.

A

Anchor timestamp. The reference time associated with a prediction or context example. Kumo Relational uses this boundary to determine which historical events are available for the prediction. See Configuration.

C

Context example. A historical example that pairs an entity with a known outcome. Context examples show Kumo Relational how the prediction target relates to the surrounding data. Kumo Relational can derive them from a Predictive Query Language (PQL) query or accept them directly through predict_task().

Custom context. User-selected context examples and prediction entities supplied as pandas DataFrames through predict_task(), instead of being derived from a PQL query. See Predict with Custom Context.

D

Data type. The physical representation of a column, such as integer, floating point, string, Boolean, timestamp, or list. See Data Types and Semantic Types.

E

Entity table. A table representing objects, such as customers or products, for which Kumo Relational can make predictions. An entity table typically has a stable, unique primary key.

End-time column. Table metadata identifying when a row stops being valid. For the prediction entity table, the end time is exclusive, so the row is no longer valid when its end time equals the prediction’s anchor timestamp. See Time and End-Time Columns.

Event table. A table containing time-dependent activities or transactions related to one or more entity tables. Its time column identifies when each event became available to the prediction workflow.

Explanation. The object returned when you call predict() with explain=True. It contains the prediction and machine-readable details about the associated data patterns and related records. When available, it also contains grouped patterns from context examples, local subgraphs with attribution values, and a natural-language summary. See Prediction Explainability.

F

Fan-out. The maximum number of neighboring rows sampled during one step, or hop, through graph relationships. Each value in num_neighbors sets the fan-out for the corresponding hop. See Neighborhood sampling.

Foreign key. A column whose values refer to the primary key of another graph table. Kumo Relational relationships connect foreign keys to their destination entity tables.

H

Hop. One traversal from a table to a related table in the neighborhood around a prediction entity. The number of entries in num_neighbors determines the number of sampled hops.

I

Indices. Entity primary-key values passed through the indices argument to predict(). They select the entities to score and override entity IDs embedded in the PQL query. See Make Predictions.

Inference configuration. Task-specific model settings passed through inference_config. These settings control how Kumo Relational combines multiple component predictions, called estimators, and what it includes in the output. See Inference configuration.

N

Neighborhood. The related rows that Kumo Relational samples around a prediction entity to provide context for the prediction. The values in num_neighbors control the neighborhood’s depth and the maximum number of rows sampled at each hop.

NVIDIA NIM. A deployed service that hosts Kumo Relational and exposes an endpoint for prediction requests. The SDK connects to the NIM and sends requests; it does not deploy or run the model itself. See Deploy, Install, and Connect.

P

Prediction entity. The object, such as a customer or product, for which Kumo Relational returns a prediction. In PQL, the table key in the FOR or FOR EACH clause identifies this entity.

Prediction horizon. The future interval covered by a time-dependent prediction target. The start, end, and time unit in a PQL aggregation define this interval. See Query Kumo Relational.

Predictive Query Language (PQL). The language used to describe a predictive task over a relational graph. A runnable Kumo Relational query identifies a target and the entity or entities to score. See Write Predictive Queries.

Primary key. A non-null, unique column that identifies each row in a graph table. Relationships from other tables can refer to this key.

R

Relational graph. The tables, table metadata, and relationships that Kumo Relational uses to gather context for a prediction. See Create a Graph.

RelationalClient. The public SDK client that manages the HTTP connection to one NIM endpoint. Bind a graph with client.relational(graph) before making Kumo Relational predictions. See Python API Reference.

Relationship. A graph connection from a foreign key in an event or junction table to the primary key of an entity table. See Define Relationships.

Run mode. A prediction preset that controls the maximum number of context examples and the default neighborhood fan-out. Run modes let you choose how much context to include, which affects request size and latency. See Run modes and defaults.

S

Semantic type. Metadata describing how Kumo Relational should interpret a column’s values, regardless of how the values are physically stored. Examples include ID, numerical, categorical, text, timestamp, and sequence. See Data Types and Semantic Types.

T

Target. The value or event to predict. In PQL, the target can be a column, a future aggregation, or a condition applied to an aggregation.

Task type. The prediction category, such as classification, regression, forecasting, ranking, or link prediction. See Prediction Types.

Time column. Table metadata identifying when an event became available to the prediction workflow. A datetime column does not automatically become the table’s time column. See Time and End-Time Columns.