search

uhh-lt / path2vec

Learning to represent shortest paths and other graph-based measures of node similarities with graph embeddings
Apache License 2.0
33 stars 14 forks source link

Rank the nearest neighbours for the regularization #16

Closed alexanderpanchenko closed 6 years ago

alexanderpanchenko commented 6 years ago

1) Best NN strategy. Use the NLTK wordnet similarity package to rank the nearest neighbors used for the regularization (select the most semantically relevant).

2) All NN strategy Alternatively, just rotate the neighbours during different iterations. E.g. for k=1 nearest neighbour in the model, and n=5 neighbours, different iterations should use different neighbours.

alexanderpanchenko commented 6 years ago

Please compute the extra results as here https://docs.google.com/spreadsheets/d/1esXh-eNz76_86PikQe9-FwkeEpMz7Of6rzeiSS6lrD8/edit#gid=573783671 for the best configuration (NN=3 random) for all dimensionalities.