Menu
Home
Log in / Register
 
Home arrow Computer Science arrow Social Computing, Behavioral-Cultural Modeling and Prediction
< Prev   CONTENTS   Next >

Who Will Follow a New Topic Tomorrow?

Abstract. When a novel research topic emerges, we are interested in discovering how the topic will propagate over the bibliography network, i.e., which author will research and publish about this topic. Inferring the underlying influence network among authors is the basis of predicting such topic adoption. Existing works infer the influence network based on past adoption cascades, which is limited by the amount and relevance of cascades collected. This work hypothesizes that the influence network structure and probabilities are the results of many factors including the social relationships and topic popularity. These heterogeneous information shall be optimized to learn the parameters that define the homogeneous influence network that can be used to predict future cascade. Experiments using DBLP data demonstrate that the proposed method outperforms the algorithm based on traditional cascade network inference in predicting novel topic adoption.

1 Introduction

Information cascade has been well studied to explain how individuals adopted information, but not as much to predict future cascades especially when the information is new and possibly not quite relevant to past cascades. An individual adopts the information when she receives enough influence from her infected neighbors [4], one who already adopted the information, or randomly if the underlying decision making process is unclear [4]. A set of work was developed to infer the inherent influence network [2] [5] [1] based on a number of cascades; the influence network inferred is the one which best fits all cascades. In general, the more cascades the more accurate influence network one can recover.

The inferred influence network can recover the most likely influence flows based on the distribution of collected past topic adoption cascades. However, it is unclear whether the future topic cascade can be explained by the distribution of past ones, due to two reasons. First, past cascades can be limited to cover the relationships among all actors. Second, the new topic propagation could be irrelevant to how past topics propagated. Though the topic cascade can be volatile and in many cases there are not sufficient cascades available, the way an author being influenced to work on a new topic is relative stable; an author is likely to be influenced by her colleagues or other researchers with social connections. These social connections contain rich information describing different relationships between authors and can be used to infer the inherent influence flows.

Existing research on heterogeneous information networks mostly focus on ranking and clustering [3], similar objects searching [8] and link prediction [6]. Differently, this work tries to predict how a new topic is being adopted by authors as a cascade without knowing the influence network structure, but not to predict an additional new link over the existing network. This work proposes to leverage the rich heterogeneous bibliography network information to complement the past topic cascades for determining the inherent influence network. Besides the social connections, the popularity of the topic itself also affects the adoption process. In general, authors are more likely to follow popular topics than less widely accepted ones. To this end, this paper aims at developing an algorithm that finds an influence network by optimizing over the social connections and topic popularity subject to past cascades. The influence networks will then be used to predict new topic adoption. DBLP data is used to demonstrate the performance improvements of the proposed method in Mean Average Precision (MAP) and Area under ROC Curve (AUC).

 
Found a mistake? Please highlight the word and press Shift + Enter  
< Prev   CONTENTS   Next >
 
Subjects
Accounting
Business & Finance
Communication
Computer Science
Economics
Education
Engineering
Environment
Geography
Health
History
Language & Literature
Law
Management
Marketing
Philosophy
Political science
Psychology
Religion
Sociology
Travel