The British Magazine Economist has published an interesting article about Collaborative Filtering. This is a well written article giving a light introduction of the status quo of research in CF.
Although there are fairly many systems using CF nowadays to recommend products, there is still much room left for further research. Put it in the simplest form, a recommender system is able to “tell me what I want”. This can be done either through a user-user framwork or item-item framework:
In theory, the item-item framework works better because it does not require constant updating of user preferences. In reality, the sparseness of the data can also help us to avoid doing tedious calculations on a million-row by million-column matrix.
As mentioned in the article Prof. John Canny‘s research helps to protect the privacy and at the same time use the aggregate data for providing recommendations. Although interesting, it is only of a second-order importance. In my opinion, as everyone in the system would only be referred to as an ID, it does not make sense to anonymize it further. (In the end, in order to get the creators create better products, the users need to give feedbacks about the existing products). A first order important problem is one about strategic gaming of the system, it is an important question because as people rely more on the recommendations, it is crucial to provide unbiased recommendations.
Another problem is that “There is no pint in making suggestions any more finely tuned than the variations in an individual’s own opinons”. That means if I rated a movie as 4.5/10, and I could have rated it 3.6/10, because internal to my mind, these scores are not that different, both these scores show that I dislike the movie, but I do not hate it. This randomness of numerical ratings resulted from psychological variations would be as severe a problem as mis-reporting.