The Book of Trees is now available!      See other retailers
Home     About     VC Book     Stats     Blog     Books     Links     Contact  
Search the VC database:
    Semantic Networks   < Prev | 231 of 1000 | Next >
The materials shown on this page are copyright protected by
their authors and/or respective institutions.
Semantic network of Flickr user tags
Ciro Cattuto
Project Description:
Ciro Cattuto wrote a simple piece of code that builds and visualizes a network of inferred semantic relations among the tags that Flickr users attach to their photos. He wrote a Python script that collects tagging info for the public photos of a given user. It does so by using FlickrClient, a Python interface to the Flickr API. The collected information is analyzed using this simple idea: if the presence of tag X and the presence of tag Y are statistically correlated, then X and Y must be somehow "semantically related".

The script focuses on the 50 most used tags of a given user. It loops over pairs of tags and uses a simple correlation heuristic to decide whether two tags are related or not. A graph is built by connecting related tags (nodes) with links (edges). Frequency information is used to measure the strength of a link and a simple conditional probability analysis is used to guess the link orientation, so that "less general" tags point to "more general" ones. The resulting partially directed graph is stored in DOT format and handed over to Graphviz (neato) for visualization.

Nodes correspond to tags, red nodes are the 10 most frequently used tags for the given user, label size increases with tag frequency, and thick edges represent statistically strong links. In this visualization model, strong links act as "elastic strings" pulling nodes close to each other: the closer two nodes are, the stronger is their correlation. Weak links are not shown, in an attempt to keep the graph readable.

Comments (0):
*Note* Before you submit your comment, bear in mind there's no guarantee it will be seen by this project's author. In case you want to contact the author directly, please follow the provided URL.
Leave a Comment:
(We're looking for the best solution to avoid unwanted SPAM)
Manuel Lima |