With the increasing development and growth of Web services on the World Wide Web, the demand of appropriate Web service selection approaches are unprecedentedly strong, and Quality-of-Service (QoS) based service computing is becoming an important issue of service-oriented computing. In most of previous works, the QoS values of services to users are all conceived to be known, however, lots of them are unknown in practice application. Recently, lots of literatures aiming at predicting such missing QoS values are published, they all consider the unknown QoS values prediction as a fundamental step for the QoS-based service computing. Looking through existing works, we discover that the online cold-start scenario, in which some new coming Web services haven't been involved even once, hasn't been considered carefu lly. In this paper, we utilize a collaborative framework by integrating matrix factorization with probabilistic topic model to predict QoS values. Spec ifically, the basic idea of the proposed approach is collaborative filtering via matrix factorization, while the cold-start problem is handled by employing probabilistic topic model based on WSDL (Web Service Description Language) documents. The experiment are based on two real-world datasets (one contains 100 users and 150 Web services, and the other contains 339 users and 2344 Web services), and the results demonstrate the prediction accuracy of the proposed approach.