Learning web-service task descriptions from traces

This paper considers the problem of learning task specific web-service descriptions from traces of users successfully completing a task. Unlike prior approaches, we take a traditional machine-learning perspective to the construction of web-service models from data. Our representation models both syntactic features of web-service schemas including lists and optional elements, as well as semantic relations between objects in the task. Together, these learned models form a full schematic model of the dataflow. Our theoretical results, which are the main novelty in the paper, show that this structure can be learned efficiently: the number of traces required for learning grows polynomially with the size of the task. We also present real-world task descriptions mined from tasks using online services from Amazon and Google.