Research On Relationship Prediction Technology Over Signed Heterogeneous Information Networks

We are living in an interconnected world. Most of data or information objects, groups or components are interconnected or interact with each other, forming numerous, large, interconnected networks. Without loss of generality, we call such interconnected networks as information networks. The analysis and mining of information network has gained extremely wide attentions nowadays from researchers in computer science, social science, physics, and so on.Information networks are divided into homogeneous information networks and heterogeneous information networks. In homogeneous information network, nodes are objects of the same entity type (e.g., person) and links are relationships from the same relation type (e.g., friendship). Heterogeneous information networks contain objects of multiple types and links of multiple types. For example, IMDB has objects of movie, director, actor/actress type and so on, and links of direct, star in type with different semantics and so on. With the development of Internet, people express their emotions more and more in social networks so the links in networks have sign, that is, the link can be positive(indicating trust, like, friend, etc) or negative (indicating distrust, unlike, disagrees, etc). The heterogeneous information networks with sign are called signed heterogeneous information networks.There has been a lot of analysis and mining methods research on information networks, among which relationship prediction is an important task. In signed heterogeneous information networks, relationship prediction contains link prediction and sign prediction, which predict the existence and sign of links respectively. Link prediction has significant value on network evolution, recommendation, cluster and so on. Sign prediction can be applied to many areas such as recommendation, decision-making, network evolution.Although relationship prediction has many researches, most of link prediction methods are based on non-signed information networks and the majority of sign prediction methods are based on homogeneous information networks. However, there are many signed heterogeneous information networks in real word, so the relationship prediction in such networks becomes a new challenge. This paper aims at signed heterogeneous information networks, exploring the relationship prediction on such networks. The innovative research works mainly include the following aspects.(1) Proposing the solution of link prediction problem on heterogeneous information network. In this paper, we propose a rule-based methodology called RulePredict to solve this problem. In RulePredict, we first extract all features systematically which contain positive features that promote the existence of a link and negative features that reduce the possibility reversely. The existence of link follows binomial distribution with probability p. p is a function of all features. Then, the best weights associated with different features will be learned by a supervised method which is based on generalized least squares (GLS). The learned weight will be used to predict the presence of links in test data.(2) Proposing the solution of sign prediction problem on heterogeneous information network. We propose a new method Hetesign to solve sign prediction. Firstly, we define the similar value between nodes under different relation, each value viewed as a feature with associated weight. The similarity between nodes is defined as mathematical expression of features and weighs. Calculate the score of sign, based on which to determine the sign of links. The score is a function of similarity of nodes and existed links. The importance of positive links and negative links are different, so they have corresponding coefficient. Then a supervised learning framework will be used to learn weights and coefficients by using maximum likelihood estimation algorithm.(3) We present experiments on a real network with positive and negative links, the IMDB and Epinions network, which demonstrate that our approach outperforms other approaches in terms of accuracy.