search

JudyYe / zero-shot-gcn

Zero-Shot Learning with GCN (CVPR 2018)
MIT License
917 stars 178 forks source link

How to build the adjacent matrix? #10

Closed iwtw closed 4 years ago

iwtw commented 6 years ago

As the paper says that "As for knowledge graph, we use the sub-graph of the WordNet [32], which includes around 30K object nodes." As I know that Wordnet provides the shortest path distance between two words, but how to build the adjacent matrix for Imagenet labels from Wordnet? The building process is mentioned for NELL and NEIL in the paper in details, could you please explain the process for Imagenet and Wordnet as well?

iamkissg commented 6 years ago

Hi @iwtw ,

In my opinion, you could simply choose a threshold for the task. If the shortest path distance between two words is shorter than the chosen threshold, make a connection between the corresponded nodes. (There are different kinds of similarities besides the shortest path, which you could consider to use)

fungtion commented 5 years ago

@iwtw the same problem, I obtain some adjacent matrixs from wordnet, by setting the threshold of the shortest path distance to some values, but the performance of gcn is not better than fully-connected layer on my custom dataset.

JudyYe commented 4 years ago

We use http://www.image-net.org/archive/wordnet.is_a.txt provided by ImageNet instead of directly building tree from wordnet. These folks provided some insight into how the graph structure affects the performance