Negative sign prediction for signed social networks

Highlights

• Existing works predict signs by analyzing features according to Structural Balance Theory, Status Theory, or both.

• However, they only involve partial information related to negative signs, which leads to the limited prediction performances.

• We propose a novel negative sign prediction method involving negative sign related features from different categories.

• We contribute to generalize three main categories of negative sign related features: nodes features, triad features and user similarity features.

• By merging features belonging to these categories, our method shows superior performances to existing works.

Abstract

Sign prediction reveals the underlying relationships between users of signed social networks. Though negative signs usually dominate the final user decision in most real applications, since negative signs cannot be directly propagated between users like positive signs, the research of negative sign prediction is still at its beginning stage. Existing works predict negative signs by analyzing features according to Structural Balance Theory, Social Status Theory, or both. However, these works only involve partial information related to negative signs, which leads to the limited negative sign prediction performances. We therefore propose a novel negative sign prediction method involving negative sign related features comprehensively from different categories. The proposed method contributes to generalize three main categories of negative sign related features: nodes features, triad features and user similarity features. By merging features belonging to these categories via Logistic Regression Model, the proposed method shows superior performances to existing works: the negative sign prediction accuracy can be improved around $5\text{\%}$ and F1 score can be improved up to $32.69\text{\%}$. The generalization performance and embeddedness can also be significantly improved by using the proposed method.

Introduction

Social networks in many real world applications are signed social networks [1], i.e., links between nodes of these social networks have positive or negative signs. Positive sign means two end nodes of the link are positively related [2], e.g., they trust each other, they are friends or they agree with each other. Negative sign means two end nodes of the link are negatively related [3], e.g., they distrust each other, they are foes, or they disagree with other each other.

Sign prediction predicts signs for links of signed social network. It reveals the underlying relationships between users. It can be widely used in many applications such as recommendation systems and abnormal user detections [4]. Most works of sign prediction focus on positive sign prediction [5], [6]. The research of negative sign prediction is still at the beginning stage. This is because: (a) It is easier to predict positive signs than predict negative signs [7]. Positive signs can be propagated between users of social networks. E.g., if A trusts B and B trusts C, A will trust C to some extent. While negative signs cannot be propagated between users of social networks, which makes negative sign prediction more complex. E.g., if A distrusts B and B distrusts C, it is hard to judge the relationships between A and C directly. (b) The number of publicly available datasets for negative sign prediction research is limited [8]. Compared with positive signs, the information of negative signs between users of social networks is more sensitive. So there are only several available real world datasets containing negative signs.

Though the research of negative sign prediction is far less than the research of positive sign prediction, it is extremely important in the real world applications. The effects of negative signs and the effects of positive signs are unbalanced in signed social networks [9], [10]. In real world applications, the number of negative signs is far less than the number of positive signs. Users rarely express their antipathy to others due to their manners or fear of being retaliated [11]. However, it is the negative signs which dominate the final decision made by users in most cases. E.g., in the recommendation system, even if 100 users express positive opinions on a target user, a user may not choose the items recommended by this user if several users express negative opinions on this target user. Existing works [11], [12], [13], [14], [15] of negative sign prediction are based on Structural Balance Theory [16], Social Status Theory [17], or both. Structural Balance Theory and Social Status Theory analyze the triad relationship between nodes of networks, which include positive relationships and negative relationships. Structural Balance Theory regards the enemy of enemy as friend, and regards the enemy of friend as enemy [16]. It considers four possible triad relationships with positive and/or negative relationships in signed networks. Two of these triad relationships are regarded as balanced. Social Status Theory regards the sign between users as the gap between their statuses [17]. A positive sign for the link pointing from A to B means A thinks the status of B is higher than A, while a negative sign means A thinks the status of B is lower than A. Social Status Theory therefore defines eight balanced triad relationships in signed networks. Existing works use some attributes based on Structural Balance Theory and/or Social Status Theory and predict the negative signs. However, the attributes evolved in existing works only reflect partial information related to negative signs, which leads to the limited negative sign prediction accuracy. Attributes reflect different aspects of signs shall be considered comprehensively to get better prediction performances in the negative sign predictions.

To solve the problems of existing works, we propose a novel negative sign prediction method which involves negative sign related features of different categories. The proposed method contributes to generalize three main categories of negative sign related features: nodes features which reflect a node’s tendency of giving/receiving negative signs, triad features which reflect the properties of nodes according to Structural Balance Theory and Social Status Theory, and user similarity features which reflect the similarities between users giving/receiving negative signs to/from the target user. The proposed model totally selects 7 features from these three categories: negative indegree ratio, negative outdegree ratio, negative triad ratio, source node positive similarity, source node negative similarity, destination node positive similarity and destination node negative similarity. Logistic Regression Model is used to merge these 7 features to predict negative signs for signed social networks. Experiments held on three publicly available datasets verified the effectiveness of the proposed method: compared with existing works, the negative sign prediction accuracy can be improved around $5\text{\%}$, while the negative sign prediction F1 score can be significantly improved, up to $32.69\text{\%}$. The generalization performance of the proposed method is also superior to existing works, up to $21.37\text{\%}$ improvement comparing with existing works. In addition, the proposed method also has better prediction accuracy and F1 score for signed social networks with different embeddedness, especially for signed social networks whose embeddedness equals to 0.

The rest of the paper is organized as follows: Section 2 introduces the related works, Section 3 presents the proposed negative sign prediction method in details, Section 4 gives the experimental results of the proposed method, and Section 5 concludes this paper and points out the future directions.