AdaFlow: Learning and Utilizing Workflows for Enhanced Service Recommendation in Dynamic Environments

Abstract

Service provisioning represents a nuanced form of recommendation, offering a bundle of services (APIs) tailored to the specifics needs of an application (mashup) as defined by the developer, significantly easing development efforts. Unlike standard product recommendations, service recommendations face unique challenges, including cold-start, long-tail phenomena, constraints, dynamic environments, and workflows. While the first four issues have seen some resolution in the literature, the workflow mining and integration among services remains underexplored. In this article, we focus on this gap by introducing AdaFlow, a model designed to understand and leverage service workflows within mashups, identifying viable service patterns for recommendations. AdaFlow employs a Graph Neural Network (GNN)-based framework, AdaptiveNN, to capture and learn service interactions. This learned workflow knowledge feeds into a dynamic GNN, enhancing service evolution representations that inform our recommendation process. Moreover, AdaFlow exhibits superior performance in managing dynamic and imbalanced scenarios.