Collaborative Filtering on Familjeliv.se
Christian Wennerström · KTH Publication Database DiVA (KTH Royal Institute of Technology) · 2011
The purpose of this project was theimprovement of navigation on the forumsite Familjeliv.se. The forum containsover two million threads, and sitemanagement wanted a system to recommendthreads to users, alleviating the problemswith navigating all the data. The choicefell on a thread-to-thread similaritybased system, recommending threads similarto the one currently read. Two differentsimilarity measures were tried out, withthe first (cosine similarity) beingabandoned due to performance issues,replaced with a binary vector similaritysolution. The data still needed to bereduced to get execution times down toacceptable levels, and sampling andpartitioning were used. A live testshowed, after a week and about ten millionpage views, that the system provided abenefit a bit above 10% over a controlcondition in terms of number of clicks.