【作者】 曹锋；

【导师】 周傲英；

【作者基本信息】 复旦大学 ， 计算机软件与理论， 2006， 博士

【摘要】 近年来，许多应用中的数据是以流的形式产生的，例如网络流，传感器数据，以及网页点击流等。分析和挖掘这类数据日益成为一个热点问题。作为一种基础的数据挖掘手段，聚类分析在数据流环境下得到了学术界和工业界的广泛关注。与传统数据库不同，数据流具有如下特点：(1)数据总量的无限性；(2)数据到达的快速性；(3)数据到达次序的无约束性；(4)除非可以保存，每个元素均只能被处理一次。数据流的上述特点对数据流上的聚类挖掘提出了如下要求：首先，算法必须能够进行实时在线挖掘，快速处理每一个元组，并实时输出挖掘处理结果。其次，相对于无限规模的数据流内存通常是有限的，算法的空间复杂度要低，往往需要在数据量的对数范围内。再次，由于算法实时在线挖掘以及对空间复杂度的限制，算法往往只能得到近似解，且需要具有一定的精确度保证。最后，算法要具有较强的适应性，包括对数据流不断进化的底层模型的适应性，处理离群点的能力，以及挖掘任意形状簇的能力等。学术界已经对数据流上的聚类分析问题进行了不少研究工作，但仍存在许多问题尚待研究和解决。本文研究了滑动窗口内的数据流聚类分析问题，数据流中具有任意形状簇的挖掘问题，利用图形处理器加速数据流聚类问题以及分布式数据流的数据聚类问题，旨在为现有的数据流系统提供更为多样的聚类分析功能。本文的主要贡献有如下四个方面：1．本文提出了一种新算法CluWin来解决滑动窗口内数据流聚类分析问题。我们设计了一种新的概要结构—聚类特征指数直方图—来保持滑动窗口中簇的统计信息。CluWin算法仪需要维护O(k/∈log(∈[N/k]))个时间聚类特征结构，就能够估算长度为N的滑动窗口中所有记录的聚类结果，且窗口最大相对差不超过∈。此外，它还被扩展用于解决N-n窗口(滑动窗口扩展模型)数据聚类问题。2．本文提出了一种新算法DenStream用于挖掘进化数据流中具有任意形状的簇。我们引入一种“密”微簇称为核心微簇(core-micro-cluster)用于描述数据流中任意形状的簇，并提出潜在核心微簇(potential core-microcluster)和离群微簇(outlier micro-cluster)结构分别用于维护并区分数据流中潜在的簇和离群点。DenStream基于这些概念包含了一种新颖的淘汰策略，该策略可利用次线性空间的内存维护并保证各微簇权值的精度。3．本文利用性能强大、日趋廉价且在数据流领域尚未引起足够重视的图形处理器(GPU)处理数据流聚类挖掘问题。我们提出一类基于GPU的快速聚类方法，包括基于k-means的基本聚类方法，基于GPU的数据流聚类以及数据流簇进化分析方法。这些方法的共同特点就是充分利用GPU强大的处理能力和流水线特性。与以往具有独立框架的数据流聚类算法不同，基于GPU的聚类算法具有同一框架和多种聚类分析功能，为数据流聚类分析提供了统一平台。4．本文提出了一个分布式聚类处理框架CluDistream。该框架可高效地实时处理分布式数据流中海量数据，有噪声、有损或不完整数据记录，以及有交叠的数据集。在CluDistream基于期望最大化(Expectation Maximization)的算法中，每个数据记录可以以不同的隶属度属于不同的簇。这种软聚类方式能较好地反映簇的交叠性。对有噪声、损坏的或不完整的数据记录，算法可通过最大化数据簇的似然度来学习数据流的底层分布。此外，CluDistream算法中测试后聚类的策略可有效地减少算法的平均处理代价，这对分布式数据流的在线实时聚类挖掘非常有效。总之，本文研究了数据流聚类分析的四个基本问题并分别提出了新的解决方案。滑动窗口是处理数据流的基本模型之一，如何在滑动窗口内对数据流进行聚类分析是一个基本问题；具有任意形状簇相对于球形簇是更为一般的数据簇模型，如何挖掘任意形状的簇也是一个基本问题；如何提高数据流聚类算法的处理速度是一个基本问题，这是由数据流聚类算法实时在线挖掘的特点所决定的；分布式数据流的数据聚类问题，其基础性在于现实应用中数据流往往是在分布式环境中产生的。本文算法是对现有数据流上的聚类分析技术的有益补充和改进。理论分析和实验结果表明本文算法能够高效地解决相应问题，与现有数据流聚类方法相比，本文算法在存储空间开销、挖掘处理速度以及结果准确性上具有优势。更多还原

【Abstract】 Recently, various applications generate a large number of streaming data, such as network flow, sensor data, and web page clicks. Mining and analyzing such kind of data has been becoming a hot topic. As a basic method of data mining, clustering in data stream settings has been wildly concerned from academia and industry. Different from traditional data bases, data streams have the following distinguished characteristics: (1)unbounded volume of data; (2)rapid arriving rate of data; (3)uncontrollability of tuples’ arriving order; (4) being processed only once for each tuple, except being reserved.Above characteristics propose the following requirements on clustering over data streams: Firstly, algorithms should be online mining, be able to fast process each tuple, and output the mining result on time; Secondly, because of the limited memory compared to unlimited data streams, algorithms should have low space complexity, the space complexity should be in the range of log bound of data volume. Thirdly, by the constraint of online mining and low space complexity, algorithms can only get approximate results, however, that results should be guaranteed on precision. At last, algorithms should be adaptive, including the applicability on data streams with evolving underlying model, the abilities on handling outliers and arbitrary-shaped clusters.A lot of methods on stream clustering have been proposed, however there are many problems need to be researched and resolved. In this paper, we study the the problem of clustering data streams over sliding windows, the problem of discovery arbitrary-shaped clusters in data streams, the problem of clustering data streams via graphic processors and the clustering problem in distributed data stream settings, in order to provide the existing data stream systems more functions on clustering analysis. The main contributions of this thesis include the following aspects:1. We propose a novel method, CluWin, to cluster data streams over sliding windows, and a new synopsis named Exponential Histogram of Cluster Featuresto maintain the statistic information of clusters in sliding windows. CluWin algorithm can provide the clustering results in sliding windows ( window size = N) with relative error no more than e by maintaining O(k/ε log(ε[N/K]) Time Clustering Feature synopses. In addition, it has been extended to deal with clustering problems over the n-of-N sliding window (an extended model of sliding window).2. We propose a new approach, DenStream, to discover clusters with arbitrary shapes in evolving data streams, and introduce a dense micro-cluster(called core-micro-cluster) to describe the arbitrary-shaped clusters in data streams. In addition, we propose potential core-micro-cluster and outlier micro-cluster synopses to maintain and distinguish the potential clusters and outliers in data streams. Based on these concepts, DenStream includes a novel pruning strategy, which can maintain and ensure the precision of the weight of micro-clusters by memory space with log complexity.3. We utilize the powerful, cheaper and cheaper and not enough considered graphic processing units (GPUs) to cluster data streams. We propose a group of fast clustering methods via GPUs, including the basic k-means method, the stream clustering algorithm, and the evolving analysis method over data streams. The common characteristics of these methods are making use of the strong computational and pipeline power of GPUs. Different from pervious clustering methods with individual framework, our methods share the same framework with multi-function, which provides a uniform platform for stream clustering.4. We propose a new framework, CluDistream, for clustering distributed data streams, which can effectively handle a huge volume of data in real-time, the noisy, corrupted or incomplete data records, the overlapping properties of data set in distributed data streams. In CluDistream, the EM-based (Expectation Maximization) algorithms, each data record is assigned to a cluster with certain degree of membership. This soft clustering reflects more naturally the overlapping property of the clusters. In the presence of noisy, corrupted or incomplete data records, our algorithms learn the parameters governing the distribution of underlying data streams by maximizing the likelihood of the data clusters. Moreover, a test-and-cluster strategy is proposed to reduce the average processing cost, which is especially effective for online clustering over large data streams.All in all, we study four principal problems of clustering data streams and propose novel solutions for them in this thesis. Since sliding window is a basic model for processing data streams, how to cluster data streams over sliding windows becomes a basic problem; arbitrary-shaped clusters model is more general than spherical clusters model, therefore discovery arbitrary-shaped clusters in data streams is a basic problem; the characteristic of online mining makes how to improve the processing speed of algorithm a basic issue in clustering data streams; clustering distributed data streams is a basic problem, because in real life applications, the data streams are often generated in a distributed environment. Our methods are great complementarity and improvement to existing stream clustering methods. Theoretic analysis and experimental results show that our methods can solve their corresponding problems efficiently and outperform previous clustering methods in space complexity, processing rate and result quality.更多还原