Modelling a multi-channel messaging framework: a machine learning approach

Abstract

Multi-channel messaging (MCM) systems have been implemented to integrate heterogeneous messaging channels for message delivery to Financial Services Institutions (FSIs) customers worldwide. However, implementations utilizing machine learning techniques to determine channel availability, dynamic assignment, and monitoring of customer patterns are being explored with newer technological advancements. Such approaches are used to investigate how Integrated multi-channel messaging can be implemented using machine learning algorithms to enable effective and efficient dynamic channel selection and integration methods. This research explored and investigated the various machine learning algorithms for optimal channel selection. The study delved into applying these algorithms and their use to channel selection, including the end-user context, focusing on the multi-armed bandit (MAB) problem, Tug of War and Upper Confidence Bound algorithm in providing a novel approach to solving this problem. The study presented a framework that fully integrates different web channels (Facebook, WhatsApp, Instagram IM, SMS, and Email) with a decision-making module and machine learning for the model to learn customer patterns over time. This framework uses a minimal memory and computation capability which employs simple learning procedures from the machine learning algorithms while relying on the message acknowledgement feedback from each channel. This work also describes a software architecture to support this and evaluate its effectiveness in an MCM enabled customer alert system currently used by financial service institutions. The efficacy of the proposed solution was evaluated using a simulation-based performance analysis method. The designed framework took advantage of simple machine learning algorithms and the inherent flexibility of integrating heterogeneous channels for efficient resource utilization when messages are transmitted. Furthermore, an Enterprise Service Bus (ESB) assigns weights dynamically to channels that customer use frequently with the use of the Upper Confidence Bound algorithm. This method allows flexibility to realise requirements for geolocation sensing, scheduling schemes and user mobility. The framework can choose channels dynamically, incorporating a machine learning module for message delivery patterns and supporting an integration layer via a common data channel-agnostic to each channel integrated. Lessons learnt from this design should be further refined to motivate future work in this area