Description
Kumo’s GNN-based predictions use node feature information to efficiently generate node embeddings for previously unseen data. Rather than train individual embeddings for each node, Kumo generates embeddings by sampling and aggregating features from a node's local neighborhood.
The num_neighbors
parameter determines the default number of neighbors to sample for each connection in each hop (i.e., how many neighbors are sampled for each node in each iteration).
- By default,
num_neighbors
will be determined using therun_mode
argument. - By default, two hops will be sampled. You can increase the depth of the sampled subgraph by increasing the length of the list.
- Max length of list:
5
- Min value:
1
(sampling one neighbor) - Max value:
128
(sampling 128 neighbors)
Supported Task Types
- All
Default Values
run_mode | Default Value |
---|---|
FAST | [12, 12] |
NORMAL | [16, 16] |
BEST | [24, 24] |
Updated 5 months ago