Research On Web Service Classification To Address Semantic-sparsity And Sample-sparsity Issues

With the continuous development of service computing and cloud computing,the number of Web services is increasing year by year,and service discovery has become an increasingly important issue.The success of service discovery largely depends on whether the categories of Web services are labeled accurately,but manual annotation is time-consuming,labor-intensive and prone to errors.Therefore,automatic classification of Web services becomes very important.The categories and corresponding quantities of service contained in the existing Web service repositories usually follow the long-tail distribution,suggesting that the vast majority of Web service samples only belong to a small part of the common categories(head categories),while most categories(tail categories)contain a very limited number of Web services.The classification experiments of each category show that the classification performance of the tail category is far worse than that of the head category due to the sparseness of labeled samples.At present,a large number of Web service automatic classification methods often ignore the tail category directly,and only conduct research on the head category.In addition,these researches also have not proposed a good solution for the semantic-sparsity issue of Web service descriptions.Web service multi-classification and multi-label classification are two typical scenarios of service classification.Service multi-classification refers to assigning a unique category label to a service,while service multi-label classification refers to assigning a corresponding category set to a service.This thesis gives corresponding solutions for these two tasks on the tail category:(1)In order to alleviate the semantic-sparsity and sample-sparsity issues faced by fewshot Web service multi-classification problem,a multi-information fusion based few-shot Web service multi-classification method called MIF-FWSMC is proposed based on meta-learning.MIF-FWSMC can fuse the classification information in the head categories and the service category name to overcome the sample-sparsity.It can also fuse the unsupervised word distribution information on the training set and the supervised word distribution information in the few-shot episode to overcome semantic-sparsity.By transforming the multi-classification problem into multiple regression problems,MIF-FWSMC greatly improves the efficiency of classification,so that the model can quickly find the most relevant category of the service in multiple categories.(2)For few-shot Web service multi-classification problem,a multi-information fusion based few-shot Web service multi-label classification method called MIF-FWSMLC is proposed based on MIF-FWSMC.MIF-FWSMLC modifies the word vector generator to fuse the classification information in multiple service category names.It also modifies the classification component by changing the regression object from probability to logarithmic probability to improve the accuracy and stability of the model.MIF-FWSMLC can also automatically learn the classification threshold of positive and negative to find the most relevant category set of the service quickly and accurately.For these two types of problems,the experimental part of the thesis constructs the corresponding task scenarios by controlling the number of categories contained in each few-shot episode on two real Web Service datasets,and carries out relevant experiments from multiple dimensions to verify the effectiveness of the proposed method.