​
Overview of Dimensions

There are four orthogonal dimensions that influence model architecture selection in Kumo:
  1. Temporal vs. Static Graph
  2. Use of ID-GNN Features
  3. RHS Model Type: GNN or Shallow
  4. Scoring Mechanism: Softmax or Contrastive
Not all combinations are supported. Kumo offers internally defined modules to handle key scenarios.

​
Architecture Matrix

RHS Model TypeScoringTemporal (Add ID-GNN)Temporal (No ID-GNN)Static (Add ID-GNN)Static (No ID-GNN)
GNNSoftmaxN/AN/AN/AN/A
GNNContrastN/AN/AN/ATwo-tower module (embedding)
ShallowSoftmaxContextGNN module (ranking)RHSShallow module (embedding)N/ARHSShallow module (ranking)
ShallowContrastN/AN/AN/AN/A

​
Supported Modules

  • ContextGNN: Used for temporal models with shallow RHS and softmax scoring. Optimized for ranking tasks. Read more about ContextGNN here.
  • RHSShallow: Applicable to static, shallow RHS models with or without ID-GNN. Can be used for both ranking and embedding.
  • Two-tower: Used for contrastive learning with static graphs and GNN on the RHS. Learns embeddings independently for LHS and RHS nodes.

​
Notes

  • The feature, lookup, and fusion modes are specific to RHSShallow architectures and is passed to the target_embedding_mode model parameter.
  • The term two-tower is used internally to denote that the GNN architecture exists for both the LHS and RHS entities. This may differ from broader industry definitions.
  • GNN-based RHS architectures are currently not supported in Kumo for temporal LP.
  • ContextGNN is not suitable for generating inductive embedding representations.
  • LHS refers to the entity declared by your predictive query. This is the table.entity that follows after the FOR EACH statement.
  • RHS refers to the targety declared by your predictive query. This is the table.target that follows after the PREDICT statement.

​
Recommendations

  • Use ContextGNN if you require temporal modeling and ranking behavior with a shallow RHS.
  • Use RHSShallow if you’re working with static graphs and want lightweight embedding or ranking models.
  • Use Two-tower if your goal is contrastive embedding with a static graph and GNN on RHS.
This architecture guide helps ensure optimal model selection when setting up LP models on the Kumo platform. Future updates may unify terminology and expand support.