arreason / CategoryLearning

Where Category Theory meets Machine Learning
1 stars 0 forks source link

customize n_hop sampler #41

Closed igorgarbuz closed 3 years ago

igorgarbuz commented 3 years ago

current n_hop implementation is not useful as the graph quickly becomes overpopulated edges. modify n_hop sampler to limit the branching factor (max degree) for each node. The idea is to eagerly rune extra edges when new node is added. It's an optimal heuristic, that will have a configuration parameter to limit the maximal subgraph degree (for more expansion), or just check the maximal degree of every node (faster, but would result lower expansion and higher branching factor).