Parallel construction of module networks
Parallel construction of module networks
复制标题
模块网络的并行构建
DOI：
10.1145/3458817.3476207
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Aluru, Srinivas
中科院分区：
文献类型：
--
作者：
Srivastava, Ankit;Chockalingam, Sriram P.;Aluru, Maneesha;Aluru, Srinivas
模块网络（monet）是贝叶斯网络的参数共享专门化，用于对多维实体进行推理，并在变量组之间进行协调交互。monet的构建是计算密集型的，顺序方法需要几个月的时间来学习包含几千个变量的网络。在本文中，我们提出了第一个可扩展的分布式内存并行解决方案，通过并行化柠檬树，一个广泛使用的顺序软件来构造monet。我们在monet的一个关键应用上展示了我们的并行方法的可扩展性-基因组规模基因调控网络的构建。使用4096个内核，我们的并行实现分别在24分钟和4.2小时内构建了两个模式生物的5,716和18,373个基因的调控网络，而使用柠檬树分别需要49天和1561天来生成完全相同的网络。我们的方法是应用不可知的，并且广泛适用于高维莫奈的任何广泛应用的学习。
Module networks (MoNets) are a parameter-sharing specialization of Bayesian networks that are used for reasoning about multi-dimensional entities with concerted interactions between groups of variables. Construction of MoNets is compute-intensive, with sequential methods requiring months for learning networks with a few thousand variables. In this paper, we present the first scalable distributed-memory parallel solution for constructing MoNets by parallelizingLemon-Tree, a widely used sequential software. We demonstrate the scalability of our parallel method on a key application of MoNets - the construction of genome-scale gene regulatory networks. Using 4096 cores, our parallel implementation constructs regulatory networks for 5, 716 and 18, 373 genes of two model organisms in 24 minutes and 4.2 hours, compared to an estimated 49 and 1561 days usingLemon-Treefor generating exactly the same networks, respectively. Our method is application-agnostic and broadly applicable to the learning of high-dimensional MoNets for any of its wide array of applications.