Release Notes

Highlights

Entity-level XAI

Entity-level XAI has been enabled for all task types. Users are now able to see an “Explorer” tab on the PQ evaluation page, which shows insights into individual entity-level subgraphs for a selected subset of entities. In particular:

Entity-level XAI shows the full subgraph that is used inside the model to derive a prediction. This can be tremendously useful for debugging purposes (e.g, to see whether the model sees the correct input).
Entity-level XAI lets you observe column importance scores for individual nodes in the surrounding of an entity (e.g., importance scores of transaction dates, article categories, etc). Column importance scores are generated through backpropagation of gradients w.r.t input columns. The higher the score, the more important this particular row/column pair is for your prediction.
Entities to inspect are pre-selected based on a set of different criterias (e.g., true positives, true negatives, false positives, false negatives, cold start entities, etc.) This lets you easily explore a diverse and rich set of entities across a wide range of different model predictions.

Baselines

Baselines have been enabled for temporal binary classification, regression and link prediction tasks. This lets you easily compare and verify Kumo’s model performance across a range of baselines.

Baselines are currently computed on a per-entity level (i.e., by using the past label as a forecast for regression, or by recommending past user behavior for link prediction). Global baselines will follow in a later release.

Monitoring of Fact Table Gaps

Batch prediction has been robustified by letting users know about unintentional data changes over time. In particular, Kumo will now notify users about fact table gaps during batch prediction.

Other Updates

Improved link prediction performance on inductive tasks (via handle_new_target_entities: True). Users can expect up to 2x performance improvement on inductive link prediction tasks.
Predictive query endpoint speed has been significantly increased.
MULTILABEL_RANKING tasks in predictive query have been temporarily disabled. Users are currently blocked from writing LIST_DISTINCT(category_id, 0, *) RANK TOP K queries.

Bugfixes

Columns with all N/A values are now correctly handled
Table and column names are now evaluated early in ModelPlan.encoder_overrides to fail gracefully