Using multi-armed bandit to solve cold-start problems in recommender systems at telco

Abstract

Recommending best-fit rate-plans for new users is a challenge for the Telco industry. Rate-plans differ from most traditional products in the way that a user normally only have one product at any given time. This, combined with no background knowledge on new users hinders traditional recommender systems. Many Telcos today use either trivial approaches, such as picking a random plan or the most common plan in use. The work presented here shows that these methods perform poorly. We propose a new approach based on the multi-armed bandit algorithms to automatically recommend rate-plans for new users. An experiment is conducted on two different real-world datasets from two brands of a major international Telco operator showing promising results.