Release Notes - v1.34

Highlights

Multi-label Ranking

Multi-label ranking is now available in PQLv2. This allows you to rank the target categories in the output. A multi-label ranking query looks like the following:

LIST_DISTINCT(category_id, 0, *) RANK TOP K

where K is the number of target entities included in the rank.

Faster Training Table Generation on Temporal Queries in Snowflake

Cross joins with user-defined table functions (UDTF) in Snowflake have been optimized, resulting in faster training table generation for temporal queries (over 20x speedup over certain queries).

Past Sequence Encoder

You can now specify which encoders to use on autoregressive labels in regression and forecasting tasks. By specifying past_encoder in the model plan, you can now encode the autoregressive labels using multi-layer perceptron (MLP) or transformer architecture. By adding this layer of nonlinearity, Kumo can now achieve better performance on those tasks.

training_table_generation:
  lag_timesteps: 16
  ...
model_architecture:
  ...
  past_encoder:
  - mlp
  - transformer

Kumo Quick Start Guide

Kumo's Quick Start Guide provides guidance on how to use Kumo to solve your machine learning task with optimal performance in the shortest amount of time possible.

Other Updates

• Training is now more robust against GPU OOMs and requires less RAM
• Improved speed in UI
• Better validations on Link Prediction Timeframes

Bugfixes

• Encoder override bug is fixed in the Model Planner.