A Web Service Recommendation Model based on Service Collaboration Network

Abstract

With the development of mobile internet, Web services tend to be fragmented and heterogeneous. The demands of users are expressed in multi-dimensionalities. Service recommendation is an effective way to help users discovering service resources that can make the Web service more convenient and intelligent. Mashup, which is a kind of light-weight method for integrating service functions, provides users with a way to integrate existing service resources to meet their needs. As a user-generated content that integrates service resources, mashups show the collaboration relationship between service resources. The goal of this study is to fill four gaps: (1) Extract the features of the resources from service collaboration network metrics and construct a service recommender; (2) Optimize the network projection weights to balance the influence of popular services; (3) Use MLP to extract user demands from textual information; (4) Match the users' demands with service resources via service collaboration network. The main work of this study is constructing a service recommendation framework, including two significant modules: Recommendation computation based on the service collaboration network extraction and context extraction. The recommendation computation uses network matrix multiplication and resource-allocation process to transform the network into service features. Then the module uses the cosine method to compute the similarity between application functions and services. The context extraction module transforms the text into structured features by using TF-IDF and TSVD. The features are then used to train a multilayers perceptron, which can learn application context from input features. This module also applies a cross-validation a model selector to train the supervised model. In the empirical part, this study uses the data from ProgrammableWeb to implement this service recommendation system. In the evaluation part, the performance of MLP in the context-aware module is better than another two supervised models: KNN and SVC. For the whole recommender's results, the F1-score can arrive at 0.55 when the length of the recommendation list is set to 8.