%0 Journal Article %T 融合时间特征的社交媒介用户影响力分析<br>User influence analysis of social media with temporal characteristics %A 廖祥文 %A 张凌鹰 %A 魏晶晶 %A 桂林 %A 程学旗 %A 陈国龙< %A br> %A LIAO Xiang-wen %A ZHANG Ling-ying %A WEI Jing-jing %A GUI Lin %A CHENG Xue-qi %A CHEN Guo-long %J 山东大学学报（理学版） %D 2018 %R 10.6040/j.issn.1671-9352.0.2017.371 %X 摘要： 针对现有张量影响力模型未能充分考虑用户的时间特征以及在线学习等问题,提出了一种融合时间特征的社交媒介用户影响力分析方法。该方法首先将用户观点、活跃度、网络中心度等特征加入张量模型中,并将张量分解过程中的用户潜在特征矩阵加入时间特征约束;其次,采用随机梯度下降的方法进行张量的分解;最后,通过融合不同张量片的影响力信息得到用户影响力得分。该方法的优点是能够快速分解张量并准确预测特定话题领域下的用户社会影响力,同时能够在已有模型参数的基础上进行新数据的在线训练。实验结果表明,与现有TwitterRank、OOLAM、受限非负张量分解模型等相比, 该方法在平均预测准确率上提升了2%~6%。同时,该方法的时间消耗仅为受限非负张量分解模型的30%~50%。<br>Abstract: Since both the temporal characteristics and online learning are not fully considered in exsiting tensor influence models, a novel method with temporal characteristics is proposed in this paper. This method constructs tensor with users opinion, activity and network centrality information. Then, a factorizes tensor with stochastic gradient descent algorithm which is constrained by temporal characteristics matrix is deployed in our model. Base on these two steps above, this method calculates user influence by combining different slices of tensor in the end. The advantages of this method are that it can decompose tensor efficiently and satisfy the need of online learning. Experimental results show that the average accuracy of the proposed method is 2% to 6% better than the baseline method such as TwitterRank, OOLAM and constrained nonnegative tensor factorization method. Besides, the running time of the proposed method is only 30% to 50% of constrained nonnegative tensor factorization method %K 时间特征 %K 张量 %K 社会影响力 %K < %K br> %K Temporal Characteristics %K Social Influence %K Tensor %U http://lxbwk.njournal.sdu.edu.cn/CN/10.6040/j.issn.1671-9352.0.2017.371