Can we apply community detection algorithms for word vector space?


As I understand we can apply community detection algorithms such as Louvain to detect communities in a social network (i.e. involves people).

But I am quite interested in knowing if we can use the same community detection algorithms such as Louvain to identify communities in word vector space (e.g., word2vec), instead of clustering?


Posted 2017-10-17T00:41:52.770

Reputation: 641

How are you defining the edges, by thresholding a distance? I suppose you could, as long as your embeddings are not too high dimensional; just try it. – Emre – 2017-10-17T01:26:28.323

@Emre is it essential to have edges in between nodes (i.e, my word2vec word vectors) to use community detection algorithms? – Volka – 2017-10-17T01:30:20.587

You can't have a network or graph without edges, but your algorithm might be defining them implicitly, like I suggested. – Emre – 2017-10-17T01:40:36.567

Thanks a lot. I never thought about it. Btw do you have any suggestions to implicitly define these threshold values? – Volka – 2017-10-17T02:04:53.600

1No, but I'd take a step back to ask what your ultimate goal is and whether this is the best approach. – Emre – 2017-10-17T02:27:09.450

@Emre Thanks. I actually want to cluster my vectors. I am currently using k-means. But it does not seem to work well. That is why I am looking for other possibilities of clustering such as community detection algorithms. – Volka – 2017-10-17T04:41:42.503



"Improving Community Detection in in Wikipedia Articles using Semantic Features" This paper talks about various methods of community detection and might be helpful.


Posted 2017-10-17T00:41:52.770

Reputation: 967