Do you keep history of user & item actions? Even if an item is old, it has a correlation with users who are still active. User A,B,C buy X. X goes out of date. Users A,B,C buy Y. Y goes out of date. Users A,B buy Z. Should you recommend Z to C?
Algorithms are seductive to programmers. Start simple, understand your data and use a feedback loop to watch your responses. After this, you can pick&choose algorithms based on what makes sense in your context. On Sat, Aug 20, 2011 at 2:30 AM, Danny Bickson <[email protected]> wrote: > I advise taking a lot in some of the related papers: > A) Liang Xiong, Xi Chen, Tzu-Kuo Huang, Jeff Schneider, Jaime G. > Carbonell, Temporal Collaborative Filtering with Bayesian > Probabilistic Tensor Factorization. In Proceedings of SIAM Data > Mining, 2010. > B) Yehuda Koren. Collaborative Filtering with Temporal Dynamics. > http://research.yahoo.com/files/kdd-fp074-koren.pdf > C) Yahoo! Music Recommendations: Modeling Music Ratings with Temporal > Dynamics and Item Taxonomy. Gideon Dror, Noam Koenigstein and Yehuda > Koren > ACM Conference on Recommender Systems (RecSys), 2011 >) > All of the above papers bin ratings into time slots, and have the > flexibility to support temporal effects. In other words, the linear > model can learn availability of items per time bins > and not recommend items that do not exist in a certain time. (I assume > that item availability can be mapped to discrete time bins). > > Hope this helps, > > DB > >> >> Hi, >> My team is working on building a recommendation system to recommend items >> for the following use cases:1. Based on User similarity (using >> org.apache.mahout.cf.taste.hadoop.item.RecommenderJob as the Base)2. Based >> on item similarity >> The part where it gets tricky is that we have a temporal restriction on our >> items (they are valid only for x days). So in the ideal case, the >> recommender should/can use the rating information on all our historical >> items, but will never recommend any items that are not temporally available. >> Based on the historical rating information, we need the list of best matches >> from the temporally available items. >> Apart from ideas that involve any pre/post processing activities to filter >> temporally invalid item recommendations, we were reaching out to find if >> somebody out here has ever dealt with a similar requirement and has found an >> easier solution to deal with this edge case. >> Any piece of advice, word of caution or streak of brilliance is more than >> welcome. >> Thanks in advance. >> Bala > -- Lance Norskog [email protected]
