On social networks, investigating how the influence is propagated is crucial in understanding the network evolution and the social impact of different topics. In previous study, the influence propagation is either modeled based on the static network structure, or the infection between two connected users is recovered from some given event cascades. Unfortunately, existing solutions are incapable of identifying the user susceptibility delivered by user generated content. In this paper, we propose RegInfoIbp, a general regularized learning framework for modeling topic-aware influence propagation in dynamic network structures. Specifically, the observed time-sequential user topic preference and user adjacency information are factorized by the prior information reflected by a user-influential bipartite relation graph. The influence propagation is approximated with a nonparametric regularized Bayesian matrix factorization model with tractable polynomial complexity. and the influential users are identified by several sampling algorithms with slightly different approximation qualities. To further model dynamic temporal evolution, we construct Markov conditional probabilistic model on the compact latent feature representation. By integrating both topic and structure information into the regularized non-parametric probabilistic learning process, RegInfoIbp is more efficient and accurate in discovering the key factors in the content and influential users in dynamic network structure. Extensive experiments demonstrate that RegInfoIbp better adapts to real data, and achieves better approximation in influence propagation over existing approaches.

中文翻译：

---

动态关系网络中正则化的主题感知潜在影响的传播

在社交网络上，调查影响的传播方式对于理解网络的演变和不同主题的社会影响至关重要。在以前的研究中，影响传播是基于静态网络结构建模的，或者是从某些给定的事件级联中恢复两个连接的用户之间的感染。不幸的是，现有的解决方案无法识别用户生成的内容所传递的用户敏感性。在本文中，我们提出了RegInfoIbp，这是一个通用的正规学习框架，用于对动态网络结构中主题感知的影响传播进行建模。具体地，观察到的时间序列用户主题偏好和用户邻接信息由用户影响的二元关系图反映的先验信息分解。用具有可处理多项式复杂度的非参数正则贝叶斯矩阵分解模型来近似影响传播。并通过几种近似质量略有不同的采样算法来确定有影响力的用户。为了进一步模拟动态时间演化，我们在紧凑的潜在特征表示上构造了马尔可夫条件概率模型。通过将主题和结构信息整合到正规化的非参数概率学习过程中，RegInfoIbp在发现内容的关键因素以及动态网络结构中有影响力的用户方面更有效，更准确。大量实验表明，RegInfoIbp可以更好地适应实际数据，并且可以在现有方法上更好地近似影响传播。