Figuring out the User in a Few Steps: Bayesian Multifidelity Active Search with Cokriging

Nikita Klyuchnikov, Davide Mottin, Georgia Koutrika, Emmanuel Müller, Panagiotis Karras

KDD 2019

Can the system help the user in the decision?

Can the system help the user in the decision?


Efficient interactive method to learn user preferences in an active way:

  • Fast – Only 10 steps to build a model of the user
  • Multiple information – incorporate shallow preferences from the system to help the user
  • Flexible – recommendations for multi-dimensional and graph data

In this paper:

  • A novel active search formulation that fuses continuous user scores with correlated computationally derived scores
  • Experiments on synthetic data, in which MF-ASC outperforms multifidelity methods for function optimization
  • Two real case-studies on tabular consumer ratings and information graphs.
  • Real-user experiments in which MF-ASC outperforms state-of-the-art single-fidelity active search methods
  • A general and powerful method that adapts to different data models and situations
  • Few interactions steps with the user capture their preference starting from scratch