Multi-Relational Topic Model-Based Approach for Web Services Clustering

Web service discovery is a significant and nontrivial task in the domain of Web service computing. With the rapid growth in the number of Web services on the Internet, e.g., an increasing number of enterprises tend to make public their software and other resources in the form of services within and outside the organizations, locating exactly the desired Web services is becoming increasingly hard for users. It has been shown that clustering Web services according to their functionalities is an efficient way to facilitate Web services discovery as well as services management. The clustering results can help us better understanding the more fine-granted categorically functional features of Web services, and meanwhile significantly reduce the searching space and retrieval time with respect to a given user query. Existing methods on this topic mainly focus on mining the semantic functional information of Web services, etc., adopting LDA to firstly elicit the functional semantics of Web services and then clustering Web services according to their topic distributions based on some clustering methods such as K-means. However, the natural description documents of Web services generally contain limit number of words. It is hard for most existing LDA-based methods to model short text documents, which may seriously degrade the Web service clustering accuracy. To narrow such negative effect, this paper aims to mitigate the data sparse issue by mining and leveraging some types of auxiliary information that is helpful to the service clustering problem. After a careful exploration of the Web service multi-relational network that is naturally established from users' frequent behaviors (e.g., invoking and annotating) of using Web services, we found that the composition relationships between Web services and the annotation relationships between Web services sharing identical tags could be used to improve the semantics extraction and service clustering processes, i.e., services with annotation relationships tend to share similar functional semantics, whereas services with composition relationships should follow dissimilar latent topic distributions. Based on these observations, we first propose a multi-relational probabilistic topic model, MR-LDA, to simultaneously model the composition relationships as well as the annotation relationships, where services with either composition or annotation relationship will exert an impact on each other during the topic sampling process for each word. Based on the topic model, we further propose an efficient Web service clustering algorithm, MR-LDA+, to firstly revise the obtained topic distribution probabilities of Web services such that above two kinds of relationships information can be explicitly encoded, and then based on it performs the Web services clustering. We extensively evaluate the proposed topic model and clustering approach on a real world dataset crawled from ProgrammableWeb. The experimental comparisons demonstrate that our approach significantly outperforms other state-of-the-art Web services clustering methods. In addition, we also design experiments to verify if the used auxiliary information can help to extract more accurate semantics by conducting service classification and vector visualization tasks based on the Support Vector Machine and t-SNE algorithms, respectively, and both the classification performance and vector visualization results demonstrate the positive impact of the introduced auxiliary relationships information. © 2019, Science Press. All right reserved.