A Fast Learning Recommender Estimating Preferred Ranges of Features

We propose a recommender system which is based on a semisupervised classification algorithm designed for estimating users’ preferred ranges of features. The system is targeted for new users, and it infers likings incrementally by presenting two alternatives to users in each step. To create the learning model, multidimensional scaling is employed to reduce the original feature space to a low-dimensional space. Then, the proposed classification algorithm, called geometrical exclusion, effectively finds a region which is not preferable for users and is to be excluded from the model in the reduced space. The algorithm consists of simple geometrical operations that are based on preference information obtained from users. An experiment by simulation is conducted to measure the performance of the system, and the result indicates that it can produce good recommendations with a practical number of user interaction steps. We also report statistics collected from our system deployed in a commercial web service.