Real-time Mining of Dynamic, Fast Evolving Data

社会面临的最大挑战之一是理解当前产生的大量流数据。由此产生的数据爆炸挑战了计算机科学界开发新的算法来描述，探索和建模这些快速发展的存储库。迫切需要新颖的解决方案来检测流量波动，分析智能电话使用模式，跟踪商品在线销售的趋势，用于移动的人群感测，或者跟踪社交网络中思想、观点和运动的传播。数据流挖掘是一个活跃的研究领域，开发技术解决方案来应对这些挑战。虽然现有的算法已经取得了一些成功，但仍然存在大量的研究挑战。在这个领域工作的大多数研究人员将注意力集中在挖掘同步流上，其中存在恒定的信息流并且数据属性集保持固定，而不是解决当今数据流的异步和动态特性。由于这些数据可能是异步到达的，特别是在分布式环境中，因此需要新的解决方案来选择数据和构建模型。此外，这些大规模、快速发展的存储库需要近乎即时的适应性解决方案，提供解释当前数据的模型，并可用于预测不久的将来。在这种情况下，好的近似值往往比成熟但过时的模型更好。本文提出的研究计划致力于开发新的数据挖掘解决方案，以帮助我们了解在这种动态演变的异步数据流中正在发生的事情。在我们提出的工作中，我们创建了一个能够产生即时模型的各种技术的库。不断地选择最佳模型并提供给用户。我们提出的解决方案提供了对数据当前状态的近乎即时、准确的解释，从而帮助决策者预测在不久的将来会出现什么趋势。

One of the greatest challenges facing society is to make sense of the vast amount of streaming data that are currently being produced. The resulting explosion of data has challenged the computer science community to develop new algorithms to describe, explore and model these fast evolving repositories. There is an urgent need for novel solutions to detect traffic congestions, to analyze smart phone usage patterns, to track the trends in the online sales of merchandise, for mobile crowd sensing, or to trace the spread of ideas, opinions and movements in social networks. Data stream mining is an active research area, developing technical solutions to address these challenges. Although existing algorithms have had some success, a significant number of research challenges remain. Most researchers working in this domain focus their attention on mining synchronous streams, where there is a constant flow of information and the set of data attributes remain fixed, rather than addressing the asynchronous and dynamic nature of today's data streams. Because such data may arrive asynchronously, especially in distributed environments, new solutions are needed to select the data, as well as to build the models. Further, these massive, fast evolving repositories require near-instant adaptive solutions, providing models that explain the current data and may be used to predict the immediate future. In this case, good approximations are often better than full-fledged, but outdated models. The research program proposed herein is devoted to the development of novel data mining solutions to aid our understanding of what is happening right now in such dynamically evolving, asynchronous data stream. In our proposed work, we create a reservoir of diverse techniques that are able to produce just-in-time models. Continuously, the best models are selected and provided to the users. Our proposed solution provides a near-instant, accurate explanation of the current state of the data, thus aiding decision makers to predict what trends to expect in the near future.