A New Trust Model using Hidden Markov Model Based Mixture of Experts Sarangthem Ibotombi Singh & Smriti Kumar Sinha Department of Computer Science & Engineering Tezpur University Napam, Tezpur – 284 028, Assam, India {sis,smriti}@tezu.ernet.in Abstract— In service-oriented computing, selection of an appropriate web service is a challenging problem. The more services are available, the more difficult is the service selection. Trust and reputation mechanisms have been used to filter good services from bad ones. Trust and reputation system of web services can often be modeled as a multi-agent system where agents are used to manage and reason about trust and reputation on behalf of their users providing or consuming services. In this paper, we propose a trust establishment framework for such a system based on direct experience and recommended trust. While making trust based decision of accessing a web service from a service provider, the value of the trust on which the decision is based is predicted from the direct trust values in the past. If the direct trust values in the past are not available, a recommended trust value is established by mixing the opinions obtained from a number of so-called “experts”. These experts are trained to learn regions of different volatilities in a time series constructed from the recommended trust values. The dynamics between the experts and the mixing weights are obtained using a coarse-grain Hidden Markov Model. Keywords-trust; reputation; Hidden Markov Model; expert; agent; time series. I. INTRODUCTION Service-oriented computing and Web services are becoming a popular technology for enabling organizations to use the Web as a market for selling their services and consuming existing services from others. Although the future of Web services looks bright and promising, to select the most appropriate service for a specific application from ever increasing bunch of services offering the same function remains a challenging issue. Current Web service security technologies based on Secure Socket Layer (SSL), Public Key Infrastructure (PKI), Web services Trust language [1] and Kerberos [2] etc. fail to account for those untrustworthy services with bad hidden motives. In such situations, trust and reputation mechanism can play an important role in service selection. For example, a client can decide the trustworthiness of a Web service based on his direct experience with the service provider or indirect trust based on collected recommendations of the service provider from other clients. Web services have functional and nonfunctional characteristics that can be difficult to present and control. According to [3], service behavior and quality of service (QoS) parameters can vary over time due to several sources listed below: • The service provider, depending on load being experienced at a particular time, may provide different quality of service. • A possible change in management or policy in the provisioning of the service by the service provider may lead to different quality of service of the provided service over time. • Possibility of slow deterioration in the quality of the provided service over time. Based of these observations, trust and reputation of a service provider can not be treated as static. They are dynamic, context-sensitive, transferable and history based[4] II. RELATED WORKS A number of trust and reputation models have been published in the literature [5][6][7]. Here we survey only those trust models that are based on Markov model. In [8] a trust model for autonomous agents in multi-agent environments based on Hidden Markov models(HMM) and reinforcement learning is proposed. Trusting agent in the system rates all other agents after an interaction and uses an HMM per agent to decide and predict whether or not the agent is malicious. The HMM is updated from observations which come in the form of ratings after direct experiences or recommendations requested from other intermediaries. The authors, in [9], provided an idea of a trust model based on HMM to cope with the inability of probabilistic trust models to capture the dynamic aspect of trust making over time. A HMM based approach to measuring an agent’s reputation as a recommender is proposed in [10]. They model the chained recommendation events as an HMM. The main features of the model are (1) no explicit requirement of chained recommended reputations, (2) a flexible recommendation network with presence of loops, and (3) integration of learning speed into trust evaluation reliability. In [11], an HMM and digital signatures based architecture for trust management in ubiquitous environments is proposed. Here the HMM is used to infer 502 978-1-4244-7818-7/10/$26.00 c 2010 IEEE