0
$\begingroup$

Traditional library can be understood as a system, that archives the collective information from the mediums produced by our society, by indexing them to shelves. It is assumed that libraries have single ontology by which the books can be categorised to a single shelf. Many machine learning research papers exist for this problem.

Hyper-linked library could be understood in such way, that instead of a medium having one place in the ontology, there could be multiple. Imagine a book as a symbolic link to the physical book.

This aspect changes the machine learning problem slightly as instead of forming a clear hierarchy and simple tree structure, the hyper-linked library is a multi-label classification problem instead of hierarchical multi-class classification problem.

What are the important aspects that one has to consider when moving from traditional library problem towards a hyper-linked library problem? Links to research papers more than welcome!

$\endgroup$
0
$\begingroup$

To get discussion started, here are some of my intuitions.

Instead of hierarchy, it seems that the hyper-linked problem is more about optimization of information entropy in a tree structure: giving all labels to all documents would be a bad solution and also giving shortest path possible is far from optimal (due to being too vague). It seems that the multi-class classification problem of hyper-linked libraries would benefit from some cost factor in relation to how many labels a document is allowed to have and how (non-)generic they should be.

Scale-free networks might provide such cost factor as they can normalize network through-put even when the amount of documents grows unboundedly, e.g. the concept space can be frozen with a guarantee of keeping the navigation depth as constant, but still being able to distribute the documents "evenly" by preferential attachment based on some optimization goal that would give privilege for some documents over others (a scale-free network, where new links are also made between existing nodes).

This is actually the very same problem Google solved with PageRank. Internet is a scale-free network by topology and search engine provides us a symbolic link to the original based on keyword mapping.

However, while Google is entirely free from ontologies, hyper-linked libraries need to provide some baseline structure from existing ontologies. Also, if we would consider the reality of News Agencies classifying their own content, the life-cycle of some content is not as "eternal" as with content stored by libraries. Discovery of new media products would be important part of the problem. Again, seems like Google has solved these kind of problems.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.