I finally joined netflix. The laborious process of introducing myself to its recommender system combined with some recent reading on collaborative filtering inspired me to look at these systems again.
So basic approaches to web personalization for recommender systems are
- manual decision rule systems usually collected via site registration
- collaborative filtering systems usually base their predictions on user set preferences and ratings run through a correlation engine
- content based filtering agents rely on finding similarity between personal profiles created by the users themselves and documents
Problems with these traditional approaches
- user entered data about themselves is prone to biases
- static profiles cause the system to degrade over time
- collaborative filtering doesn’t scale well. It seems that a number of studies show that applying this technique to a large number of items cause prediction performance and accuracy to suffer since many items aren’t rated, etc.
Last time I poked around wondering how these work was around the same time I was looking at different ways to squeeze information out of web log data to inform design. Clustering together frequent paths taken through a site is particularly informative. In real time, this information can be used in various ways for personalization including next link recommendation. I discovered Bamshad Mobasher’s work on web usage mining and web personalization. In general he discusses some of the problems with web usage analysis and then some ways to use this data (sometimes in conjunction with other techniques) for web personalization. Really interesting work.