Solutions to Cold-start Problems for Latent Factor Models

In the data age, the "information overload" problem severely impacts the precise of people to choose what they prefer. However, recommendation systems are able to provide people related information from huge amounts of data, and effectively solve the "information overload" problem. Currently, Latent Factor Model(LFM) has become dominant in the recommendation field. For example, Matrix Factorization performs excellently on rating prediction problem. By optimizing a ranking criterion, LFM also has an outstanding performance on top-N recommendation problem, such as Bayesian Personalized Ranking. But LFM can't solve the cold-start problem. Aiming at solving the cold-start problem, we obtain the mapping concept to construct a hybrid model, in which we map new entities' (e.g. user or item) attributes to their latent features vector. Experiments on the cold-start problem show that the hybrid model provides much better recommendation precision.