| With the development of informational technologies, the computing model is transferring from traditional desktop computing to the pervasive computing. In the pervasive environment, the informational space is merged with the physical space, users will enjoy various kinds of services provided by the environment anytime, anywhere. However, with the dramatic increase of information available on the Internet, users may be encountered with serious problems of information overloading, and thus there is obviously a trend to provide users with active and personalized recommendation. Personalized recommendation is becoming a research hotspot in recent years, which has been widely used in many service areas. Almost all Electronic Commerce Systems have used the personalized recommendation technology, such as Amazon, CDNOW and Netflix. But currently recommendation systems have many areas to improve, such as prediction accuracy and the quality of recommendation. It had better to combine users'interests, but users'interests are strongly deponded on users'current behavior. Thus, the key issue of this paper is to research a personalized recommending system based on users'current behavior.This paper summarized and analyzed context-recommending techonology that based on users'current behavior in three related areas: context-aware computing, ontology and personalized recommendation. This paper's primary content and related research achievement are as follows:1. Propose a context-awareness computing sub model based on users'behavior. By using the client and the server side agent, this model separately perceives the users'service contents and software contexts, and simplify the perceived contents, and constructs a model of the representative contents and software types by OWL method. Through probability-statistics method, the model can get three tables that are users'state-time, software type-users'state and service type-users'state by learning the historical context of users'behavior. Then, we can get the users'current state by calculating the current context of users'behavior. The final experiment result proves that the model can precisely recognize the current user-state (the average accuracy rate reaches 87%),and it's a effectively filter and guidance for the following personalized recommendation(the average acceptance rate increased 14.3% compared with traditional methods).2. Propose a method which transfers the various ontology OWL and store to the relation database through share inline technology and method part path. It can actualize the context's effective storage, retrieval and calculation. Simulation experiment indicates that this method has good effects at producing datasheet's amount, data redundancy, difficulty of querying the datasheet and change or delete the data, and it has good applying-prospect at the filed of store context.3. Propose a kind of recommendation sub model based on hybrid filtering (HR) algorithm on the basis of combination of Content Filtering and Item-based Collaborative Filtering. First, it selects m subclass services by using Content Filtering method according to server content's classify characteristics and current user-state predicted by users'behavior context-awareness computing sub model. Next, users'specific ratings are predicted by K-nearest neighbor method (KNN). Finally, the Top-N services in the m subclass will be recommended. In Collaborative Filtering algorithm, we use the Baseline Estimate to solve the user-item rating matrix sparse problem, which makes all ratings based on the convergence rating, and use improved similarity measure to make similarity measure calculation based on the support of a large number of users. The experiment results show that HR provides better recommendation results than traditional collaborative filtering algorithms at the part of prediction accuracy (RMSE dropped to 0.947, compared with traditional methods decreased 7%).4. At last, we integrate two sub models mentioned above into a system. This system is able to predict users'current state by using hybrid recommending model. Then, determine which service subclass should be recommended by using the current user-state computed by context-aware model. Finally, recommend services to users according to certain recommendation rules. The final experiment results indicates that the model has great service acceptance rate (the highest user Service acceptance rate reach 71% in the Top-10 recommendation) and good scalability (in correctly predict the users'state situation of the user Service acceptance rate reach 79%)... |
PDF Full Text Request