Online and active learning for recommender systems

Description

Recommender systems require the ability to deal with high velocity of data and strong concept drift, since the users express their preference sequentially, and their preference changes rapidly and often.

Jorge Silva explains how SAS uses a Bayesian variant of the highly successful factorization machine model for recommender systems, in which the company employs a probabilistic model and performs variational Bayesian inference, to tackle both problems by making the learning process online and active. SAS leverages the model to actively select the best subset of observations at each step of the online learning process. This active learning procedure is motivated by the fact that each measurement is often obtained sequentially and requires user interaction, which may be expensive and time consuming. Prioritizing the most informative users and items potentially helps guide the rating acquisition process in a more efficient way.

In general, the problem of finding the exact best ratings to acquire is NP-complete. SAS has avoided this difficulty by developing a greedy algorithm based on the estimated variance of the user and item factors returned by the Bayesian inference. Experiments on benchmark datasets show highly promising results.