动态社交网络表示学习理论与应用

在线社交网络的兴起彻底改变了人们沟通交流的渠道，研究与理解社交网络有着巨大的市场价值和社会意义。然而，在大数据时代，社交网络的动态性及其复杂的演化过程为相关研究工作带来了全新的挑战。具体来说，社交网络的动态性一方面表现在宏观层面其结构的不稳定性，另一方面则体现在微观层面个体行为的差异性，二者提升了社交网络的维度，进而使传统的数据结构或表示学习方法难以全面、高效地表征动态社交网络。针对此问题，项目拟研究动态社交网络表示学习理论与应用。研究内容包括：针对网络结构的不稳定性，研究结构感知的动态网络表示学习算法；针对用户行为差异性，研究角色感知的动态行为生成模型，将用户行为预测与表示学习融合在一个统一的框架之中；最后，将研究的理论成果应用于电信诈骗检测问题之中，体现研究成果的实用性。项目目标是突破动态社交网络表示学习的基础理论，解决其中的关键技术，构建工具包和应用系统，取得国内外领先的研究成果。

Online social networks have changed ways of communications between people. Understanding social networks has great scientific and marketing values. However, in the age of big data, the dynamics of social networks bring new challenge to relevant research work. In particular, at macro-level, the structure of a social network is unstable and evolves over time; at micro-level, users behave differently at different timestamps. The unstableness of network structure and the otherness of user behavior together increase the dimension of social networks. Thus, traditional data structures and representation learning methods can hardly capture sufficient information of a dynamic social network efficiently. To solve this problem, we propose a dynamic social network representation-learning framework. More specifically, at macro-level, we propose an embedding algorithm to preserve both structural information and temporal information of a given network; at micro-level, we propose the role-aware user behavior model, which bridges user behavior prediction and representation learning into a uniform model; at last, we apply our study on a telephone fraud detection problem. Overall, the goal of this work is to make a breakthrough at dynamic network representation learning theories, solve critical technical issues, build practical toolkits, and be in the lead of this filed.