RI: Small: GraphLab 2: An Abstraction and System for Large-Scale Parallel Machine Learning on Natural Graphs

随着Web的发展和科学数据收集技术的改进，数据集在大小和复杂性上都在迅速增加，这就需要对机器学习算法进行相应的扩展。然而，设计和实现高效的并行机器学习算法是具有挑战性和耗时的。为了应对这一挑战，我们最近发布了GraphLab，这是一个框架，提供了一种富有表现力和高效的高级抽象，满足了广泛的机器学习算法的需求。我们系统的性能已经引起了人们的极大关注，收到了来自许多大学和公司的数千次下载。目前，GraphLab只解决多核环境下的批处理问题。在这个项目中，我们正在开发GraphLab 2：解决更具挑战性的在线和分布式环境，解决以下问题：1)基于云的分布式机器学习。2)自然图，具有不适用于图划分方法的极高度顶点。3)在线任务，其中数据和查询随着时间的推移而流动。4)核外计算，因为巨大的问题可能无法存储在内存中，即使是在云中也是如此。该项目的关键贡献之一是不断传播和转移我们的技术。我们的开源软件版本将继续支持大规模机器学习在科学和工程领域的应用。我们雄心勃勃的更广泛的影响目标，超越理论和系统，包括开发一门新课程，重点是为学生在该领域的工业和科学需求做好准备。我们建议的课程包括“网络上的机器学习”和“面向大型机学习和数据挖掘的云计算”。

With the growth of the Web and improvements in data collection technology in Science, datasets have been rapidly increasing in size and complexity, necessitating comparable scaling of machine learning algorithms. However, designing and implementing efficient parallel machine learning algorithms is challenging and time consuming. To address this challenge, we recently released GraphLab, a framework providing an expressive and efficient high-level abstraction satisfying the needs of a broad range of machine learning algorithms. The performance of our system has attracted significant attention, receiving thousands of downloads from many universities and companies.Currently, GraphLab only addresses batch processing in multicore settings. In this project, we are developing GraphLab 2: addressing the much more challenging online and distributed settings, tackling: 1) Cloud-based distributed machine learning. 2) Natural graphs, with very high-degree vertices that are not amenable to graph partitioning methods. 3) Online tasks, where data and queries are streaming over time. 4) Off-core computation, since huge problems may not fit into memory, even across the cloud.One of the key contributions of the project is the continual dissemination and transfer of our technology. Our open-source software releases will continue to enable large-scale machine learning applications in science and engineering.Our ambitious broader impact goals, beyond theory and systems, include the development of a new curriculum focused on preparing students for the industrial and scientific needs in this field. Our proposed courses include "Machine Learning on the Web" and "Cloud Computing for Big Machine Learning and Data Mining."