【作者】 苏日启； 胡皓； 汪秉宏；

【Author】 Riqi SU1,Hao HU1,Binghong WANG12 1Department of Modern Physics University of Science and Technology of China,Hefei 230026,China 2Research Center for Complex System Science University of Shanghai for Science and Technology,Shanghai 200093,China

【机构】 中国科学技术大学近代物理系； 上海理工大学复杂系统研究中心；

【摘要】 随着互联网信息规模的急剧扩大以及用户搜索要求的多样性,设计一个高效的个性化推荐系统成为一个研究热点。其中,将物质扩散和热传导过程应用到推荐过程取得了很好的效果。这两种算法通过假设每个推荐对象具有一定的推荐资源,这些推荐资源按照行归一或者列归一的扩散规则在用户-对象二部份图上扩散,进而利用扩散结果来描述节点之间的相似性。在这个基础上,根据评价网络的拓扑信息,例如节点度,群落等对初始和末态评价资源分布,扩散过程等加以调节,可以得到一系列高效,个性化程度可调的推荐算法。然而,上述的算法都是在静态的环境下提出的,没有考虑评价关系的时效性;但是实际的信息评价网络是一个动态演化的网络。一方面,其拓扑结构随着评价关系的不断加入而演化,另一方面,评价关系的时效性变化也带来网络权重的变化。我们提出了基于网络的含时推荐算法,考虑评价关系的时间标签对基于物质扩散的推荐算法的影响。在这个算法中,我们对评价关系引入了"衰退-激活"过程。每一个时间步,已有的评价关系的时效性按照一定半衰期进行指数衰减;同时,赋予新引入的评价关系一个初始的时效性,并且激活周围的邻居节点。在推荐时,用时效因子来调节节点的推荐值。我们利用Netflix的数据对这个算法进行计算,并和原有的推荐算法进行比较。我们讨论了不同的激活方式,半衰期以及时效因子对算法的影响。更多还原

【Abstract】 It had become quite hot to design an effective and highly personal recommendation to meet the rapidly growing Information on Internet and the diversity of users’ demand.Among the varsity Algorithms,the Algorithms importing the heat/mass diffusion process seem surpassing for their simple expressions and accuracy,also the tunable diversity.These algorithm,which assuming some recommend power on each targets,diffuse these power on the opinion network by rows normalized or columns normalized,and make use of the power distribution to describe the similarity between nodes.They can be greatly improved by consider some topology information,such as degree,cluster,adjust the initial or final distribution on nodes.However,all these algorithm are run under the static network,while the real one is dynamic changing with their topology and weight of links.In this talk,we introduce the network-based recommendation algorithm with time dependence,which will consider the impact of links’ time tag on the mentioned mass-diffusion recommendation algorithm.We introduce the ’decay-active’ process for the critical links on the opinion networks.At each time step,the activity of each links will decay with exponent function,then set the newly added links an initial activity,and the neighbourhood of the links will also be actived.When making recommendation,the activity of each nodes is used to adjust the strength of recommend.Based on database of Netflix,we compare this algorithm with the origin one,and discuss the impact of half life decay,mode of active,and the initial activity on the accuracy and diversity.更多还原