Ask for a recommendation framework that enables the provision of recommendations from multi-domains

User 363 | 6/13/2014, 9:58:16 AM

Hi, I am about to develop a recommendation framework that enables the provision of recommendations from multiple domains. The framework should contain the following elements. Do you know any implementation for such stuff?

1- A Semantic Logger to generate User profiles that should be implemented mainly using 4store (3store) , PHP and Java .

2- A Wikipedia crawler/disambiguator written in Java . This would find Wiki pages that correspond to user interests and download them.The pages need to then be translated to an adjacency matrix . This is a square matrix , with wikipedia pages indexing both columns and rows. When a link between 2 pages exists the entry is 1, otherwise 0.

3- The probabilistic algorithm using Markov Chain (Bayesian network) as the machine learning component of the framework. This should be implemented in MATLAB.

Comments

User 447 | 7/6/2014, 5:07:14 PM

  1. Establishing links between user profiles and sequences of topics is non-trivial. The closest system I know is by Olena Mendelev and Liu Anna Huang when they were at the University of Waikato. Olena is still in NLP and may have something.

  2. Wikipedia crawlers are strongly discouraged. You will need a data dump - which has license issues - but will have the whole of Wikipedia on disk afterwards. Included are a syntax for links which makes it rather easy to translate it into a graph.

  3. I wouldn't recommend markov chains for this (unless you have a strong time component to your user data.) Plenty of collaborative filtering algorithms exist that provide cheap alternatives for markov chains. SVD followed by nearest neighbor lookups also works particularly well.