High-Efficiency P-Rank Computation Through Asynchronous Accumulative Updates in Big Data Environment

DOI：10.13328/j.cnki.jos.004637

 作者 单位 E-mail 王旭丛 中国人民大学 信息学院 计算机系, 北京 100872 wangpu9003@126.com 李翠平 中国人民大学 信息学院 数据仓库与商务智能实验室, 北京 100872 陈红 中国人民大学 信息学院 数据仓库与商务智能实验室, 北京 100872

P-Rank是SimRank的扩展形式,也是一种相似度度量方法,被用来计算网络中任意两个结点的相似性.不同于SimRank只考虑结点的入度信息,P-Rank还加入了结点的出度信息,从而更加客观准确地评价结点间的相似程度.随着大数据时代的到来,P-Rank需要处理的数据日益增大.使用MapReduce等分布式模型实现大规模P-Rank迭代计算的方法,本质上是一种同步迭代方法,不可避免地具有同步迭代方法的缺点:迭代时间(尤其是迭代过程中处理器等待的时间)长,计算速度慢,因此效率低下.为了解决这一问题,采用了一种迭代计算方法——异步累积更新算法.这个算法实现了异步计算,减少了计算过程处理器结点的等待时间,提高了计算速度,节省了时间开销.从异步的角度实现了P-Rank算法,将异步累积更新算法应用在了P-Rank上,并进行了对比实验.实验结果表明该算法有效地提高了计算收敛速度.

P-Rank enriches the traditional similarity measure, SimRank. It is also a method to measure the similarity between two objects in graph model. Different from SimRank which only considers the in-link information, P-Rank also takes the out-link information into consideration. Consequently, P-Rank could effectively and comprehensively measure “how similar two nodes are”. P-Rank is applied widely in graph mining. With the arrival of big-data era, the data scale which P-Rank processes is increasing. The existing methods which implement P-Rank, such as the MapReduce model, are essentially synchronous iterative methods. These methods have some shortcomings in common: the iterative time, especially the waiting time of processors during iterative computing, is long, thus leading to very low efficiency. To solve this problem, this paper uses a new iterative method—the Asynchronous Accumulative Update method. Different from the traditional synchronous methods, this method successfully implementes asynchronous computations and as a result reduces the waiting time of processors during computing. This paper implements P-Rank using the asynchronous accumulative update method, and the experiment results indicate that this method can effectively improve the computation speed.
HTML  下载PDF全文  查看/发表评论  下载PDF阅读器