Recommendation based on user experiences
Recommender systems follow 2 main strategies: contentbased filtering and collaborative filtering. Collaborative is often the preffered approach as it requires no domain knowledge and no feature gathering effort. The 2 primary methods for collaborative filtering are latent factor models and neighborhood methods. In user-user neighbourhood methods, similarity between users is measured by transforming them into the item space. Similar logic applies to item-item similarity. In latent factor methods, both user and items are transfomed into a latent featuee space. An item is recommended to a user if thu are similar, their vector representation in the latent feature spase is relatively high. We select latent factor model because it allows us to identify the hidden feature of the users. These features are time indepedent. We first discuss standard recommeder system that discover time-depedent features in order to make suggestion to users. We then discuss new model (introduced in ) to deal with time-depedent feature that we named “user experience”. We experiment with a movie data set, evaluate the improvment over standard recommender system, and discuss other benefits of discovering user experiences.
Research Paper Link: Download Paper