​
Overview

The Explanations tab on Training detail page provides model-level explainability and insights into individual entity-level predictions for a selected subset of entities.

​
Entity-Level Explainability

This includes:

  • Entity-level Explainability to explore model predictions and compare them against ground-truth data.

  • Subgraph Visualization to understand the model’s input and debug unexpected results.

  • Column Importance Scores to identify key features influencing predictions.

  • Pre-selected Entities categorized by prediction confidence and correctness for easy exploration.

​
Entity Selection

Entities are grouped into categories based on model confidence and correctness (e.g., true positives, false negatives, uncertain predictions, cold-start entities). Each category allows exploration of up to 50 entities, along with their anchor time.

​
Model Prediction Analysis

For a selected entity, Kumo displays its model prediction and ground-truth value. In link prediction tasks, predictions are shown in three searchable panes:

  • Historical Items

  • Ground Truth

  • Predictions

The interface supports image comparisons, with correctly predicted items.

​
Subgraph

The Subgraph Explorer visualizes the portion of the graph that influenced a prediction. This helps debug whether the model is using expected inputs. The traversal depth is determined by the number of layers set in the model planner. If a table is missing, it may be unreachable based on the configured depth.

Importance scores highlight influential nodes and can detect target leakage or unexpected data relationships.

Example predictive query for customer churn:

PQL
PREDICT COUNT(transactions.*, 0, 90, days) = 0
FOR EACH customers.customer_id
WHERE COUNT(transactions.*, -90, 0, days) > 0

Analyzing false positives may reveal missing status indicators, while uncertain predictions often indicate mixed signals within the data.

​
Global Explainability

​
Contribution Score

Kumo scores the importance of entire tables and individual columns, showing their impact on final predictions.

The Variation of Predictions column quantifies how each column influences predictions based on variance in the training data.

​
Detecting Data Leakage

High contribution scores may indicate data leakage, where the model unintentionally relies on future information unavailable at prediction time.

Example: A column indicating subscription cancellation within 30 days could lead to target leakage in a churn prediction task.

​
Column Analysis

Clicking a column provides a distribution plot comparing model predictions and actual labels in the holdout data split. This helps identify:

  • Bias in model predictions across different entity populations.

  • Feature Importance and its effect on prediction outcomes.

  • Potential Data Gaps requiring additional feature engineering or dataset improvements.

By understanding column-level contributions, data scientists can refine predictive queries to improve accuracy and interpretability.

EvaluationBatch Predictions