Hello,

actually Louvaine is a clustering method for graphs, which maximizes the intra-cluster similarity and minimizes the inter-cluster similarity. Even though it is usually used for social networks, you can actually apply it to any kind of graph structure.

E.g. I have used it to cluster the similarity of words,as follows: the words are the nodes in the graph, weighted edges between nodes are established if the similarity of the two corresponding words is above a certain threshold.

It would be important to know how you set up data fed into Louvaine, especially which role tf-idf plays in your setting.

Using your problem setting, I would do it with Louvaine as follows. If Google gives you per image several labels (actually I don't know whether this is true for the Google API ), then you can consider a label obtained from two different images as an unweighted edge, connecting both images. Thus, the set of all images forms a network, where labels represent edges connecting the nodes. Depending on these connections Louvaine can cluster this network and as a result images annotated with similar labels, will probably be located in the same cluster, even if the label's content are not used.

I think this would make sense. Imagine two images one annotated by "cow, mountain, snow, gras", the other with ""cow, mountain, snow, gras, church", the four labels will impose a strong connection for Louvaine.

I would be interested to know a bit more about your setting and the final results you will obtain.

Hope this helps a bit.

Thomas Hoppe

Hello Thomas, this sounds very promising and feasible. Thank you very much! I wonder if you have published the research that you clustered the similarity of words? Or if you recommended some literature using Louvaine that interprets image as a network? Ive been searching for it but still couldn't really find one.

Much appreciate!!!!

Hello Chung-En,

I haven't published this work yet. And I actually do not know of any articles of using Louvaine this way. I have used this idea befor for building a network of postcards through the annotations they shared over the social media tags attached to them. The idea is not very different.

Maybe I'll propose this also as a project for my students in the next semester ...

Thank you for inspiring me.