Retweeting prediction is of great importance to the event detection and influence evaluation, which has attracted wide attention from both academic and industrial fields.Existing prediction methods are mostly concentrated in the study of the microblogs properties and diffusion network characteristics.They have not fully considered the dynamics of retweeting behavior and regularity of user’s historical behavior.This paper investigated the microblog retweeting prediction problem from the view of microblog visibility and user behavior features,and (1 )proposed a method to recognize retweeting behavior,ignoring behavior and un-received behavior based on user’s activity and dynamic time window,which provided more accurate dataset for model training and effectiveness analysis;(2)presented user interest model based on dynamic user’s interest and time attenuation, which is proved to be an effective measurement of user’s interest and its change characteristics;(3)proposed several user behavior features of the user’s retweeting rate and interaction frequency, which can effectively measure the impact of user’s historical behavior patterns and user’s influence transfer effect.Finally,this paper proposed a classification model based on retweeting behavior prediction method which is blend with upstream user’s characteristics,microblog’s characteristics, forwarding user’s interest and user’s historical behavior characteristics.Experimental results on real data show that the proposed method can improve prediction accuracy effectively,and achieve good results in the smaller size of the training set.%微博转发预测对微博话题检测和微博影响力评估具有重要意义，引起了学界和产业界的广泛关注。现有方法主要集中在微博属性及微博传播网络特征的研究，没有充分考虑转发行为的动态性和用户历史行为的规律性。文中从微博能见度和用户行为特征角度研究微博转发预测问题，（1）提出了基于用户活跃期和时间窗的转发行为、忽略行为、未接收行为识别方法，为模型训练和效果分析提供了更为准确的数据基础；（2）提出了基于时间衰减的用户兴趣计算模型，有效度量用户兴趣及其变化特性对用户转发行为的影响程度；（3）提出了用户转发率、交互频率等用户行为特征，有效度量了用户历史行为模式和用户影响力传递效应的差异性对用户转发行为的影响，最后融合上游用户特征、微博特征、转发用户兴趣和历史行为特征，提出了基于分类模型的转发行为预测方法。在真实数据上的实验结果表明，该方法能够有效提升预测准确性，并且能够在较小规模的训练集上取得好的预测效果。
展开▼
