Software Journal

Predictive analytics is frequently used to build decision support systems or to present the items that are likely to be of interest to the users. Pandora, Netflix, Last.fm, etc are some prominent examples of predictive analytics in the consumer place. There are primarily two mechanisms to do this – (i) content based filtering, wherein we look at the internals of the item, or (ii) collaborative filtering, wherein we look at the users’ behaviors and recommend what other “similar” users are finding to be of interest. Of course we can also use a hybrid model that uses two underlying models and use a combination technique to combine the results of the models.

Recent research focus has been on building better information filtering systems (better content based filtering system or a better collaborative filtering system or a better combination model). If collaborative filtering is of interest, you can read my Computing Reviews review for More reputable recommenders give more accurate recommendations?

However, an equally important question is if we can characterize which particular model is more likely to be effective in a given system? Further, can we characterize the systems in which collaborative filtering can be simplified to use just a set of “power users”?