The prevalent surge in the development of web services and application programming interfaces (API) has given birth to vast data stores that need to be efficiently and automatically classified for the purposes of service discovery and integration. Traditional supervised learning methods are highly sensitive to the availability of large amounts of labeled training data which, in practice, are potentially costly and hard to acquire in real-world repositories of services. In fact, most available web service data is unlabeled and this restricts the power of purely supervised classification methods. Current semi-supervised methods are often based on a single method that may have certain weaknesses, such as error propagation and limited scalability. This paper proposes a hybrid semi-supervised learning framework to use both labeled and unlabeled data in order to improve the classification accuracy of web services. The approach that is proposed combines the semantic feature extraction from service descriptions and Quality of Service (QoS) characteristics, such as response time, availability and reliability. The framework consists of a combination of self-training and graph-based label propagation, which deploys with unlabeled data to iteratively train them to improve the training process. Experimental evaluation on Web service data sets indicates that the proposed framework has improved the classification accuracy around 5-7% compared to the traditional supervised classifiers such as the Support vector machine and Random Forest. The results have shown that semi-supervised learning algorithm offers a scalable and cost-effective solution for large-scale classification in web services and supports the development of intelligent service discovery systems in distributed computing environments.
Landage et al. (Thu,) studied this question.