Research On Link Prediction Methods In Personalized Recommendations Of Military Information

Information is the key resource of information warfare. With the rapid development of the Global Information Grid, the quantity of information has been unprecedentedly huge. Although the mass, dynamic and heterogeneous information can satisfy all the demands of all the users’, it is quite time-consuming and almost impossible for the commanders to get what they want with high accuracy and fast speed. How to discover the right information from the limitless data ocean and send them to the right commanders? Personalized Recommendation of Military Information(PRMI) offers a method. At present, the researches on PRMI are still very scarce and most achievements are gained in civil fields.Owing to the big differences between PRMI and civil recommender systems, this thesis firstly proposes the description method of PRMI, defining the relationship of users’ preferences, the relationship of information’s value and the utilization relationship between users and information. Recommendations are produced by predictions of the three types of relationships. The Graph of User and Information is constructed to describe the key elements and corresponding relationships of PRMI. It is different from ordinary graphs, including two types of nodes and three kinds of links.Secondly, this thesis proposes two link prediction algorithms, namely the Sub-graph Similarity-based Link Prediction Algorithm(SS) and the Graph Kernel-based Link Prediction Algorithm(GKLP). The SS algorithm can be divided into USS algorithm and ISS algorithm, according to the sub-graphs, which includes the users’ sub-graphs and information’s sub-graphs. The aim of PRMI is to discover the utilization relationships between users and information, the important foundations of which are the relationship of users’ preferences and the relationship of information’s value. The USS(ISS) algorithm computes the similarity of two users’(information’s) sub-graphs and the user(information) pair with higher sub-graph similarity has a higher probability of preference relationship(value relationship). The USS(ISS) algorithm with weighed links is called WUSS(WISS), which has higher prediction accuracy. However, the basic premise of the SS algorithm is the utilization relationships between users and information, which do not exist in new ones. The SS algorithm has a bad cold-start problem, which means it cannot applied to new users or new information. The GKLP algorithm maps the User(Information) Graph to their corresponding space of adjacent matrix’s eigenvalues and finds a proper link prediction function by applying the polynomial curve fitting to the eigenvalues. The GKLP can predict the preference relationships(value relationships) through the relationships of the User(Information) Graph except the utilization relationships, so it can solve the cold-start problem.Finally, this thesis discusses the application of link prediction in Personalized Recommendation of Military Information and proposes two recommender strategies, namely recommendation based on the similarity of users’ preferences and recommendation based on the relationships of information’s value. The applications of link prediction algorithms proposed in this thesis can not only work for the personalized recommendation of military information, but also can solve the cold-start problem, which is a big challenge for civil recommender systems.