Multi-armed bandits

: a simulation gym for customer retention in a TelCo

Abstract

The highly competitive telecommunications industry poses a significant challenge for TelCos in retaining customers. To achieve customer retention, TelCos often utilize Machine Learning (ML) algorithms to align their product offerings with client preferences. However, these algorithms have limitations in exploiting past client interactions that may contain biases from various sources. This thesis investigates the potential of multi-armed bandits (MABs) to address this challenge. MABs are a type of reinforcement learning algorithm that maximizes long-term rewards and has been used in various industries to optimize resource allocation. This study uses historical data from proactive customer retention to build a gym, simulating the daily operations of a TelCo. The gym allows for the evaluation of multiple policies and scenarios in an offline environment. The study’s findings show that MABs can balance exploration and exploitation and outperform classical algorithms in certain cases. However, when the number of possible arms increases dramatically, simpler MAB algorithms may struggle. The study also shows how we can strategically limit these arms to increase performance without changing the algorithm to one more complex. Overall, this research enhances the understanding of the potential and limitations of MABs for customer retention in TelCos and provides insights for their successful implementation and deployment.

Date of Award	3 May 2023
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Miguel Godinho de Matos (Supervisor)