EAGER: FODAVA: Spectral Analysis for Fraud Detection in Large-scale Networks

摘要本课题采用统一的频谱变换方法，将数据频谱变换与网络拓扑可视化相结合，解决大规模动态网络中网络拓扑分析和欺诈模式识别的难题。大型社交和通信网络除了包含各种结构化、半结构化和非结构化数据外，还包含丰富的拓扑信息。迄今为止，致力于探索网络拓扑的工作很少，特别是从频谱分析的角度来看。如果所提出的基于图的简单邻接矩阵表示和图中基于k个社区的节点表示的方法可以被证明适用于非常大的数据集，那将是一个重大的进步。这项研究正在与信息可视化和可视化分析算法相结合，并有一个可用的银行数据测试平台，可以搜索欺诈行为。该研究描述了频谱投影空间中各种攻击的模式，并开发了基于频谱的方法来识别这些攻击。该方法利用了网络底层交互结构的频谱空间，与使用内容分析的传统方法是正交的。执行这种光谱分析的能力依赖于复杂数学技术的发展。正在探索的关键问题包括方法对非常大的数据集的可扩展性，以及谱空间中节点表示的维数的确定（这取决于图中集群的数量）。另一个问题是，每个节点的k维表示的每个组成部分都被解释为节点附着于k个社区的“可能性”。然而，必须保证表示每个节点的k维向量的分量都是非负的，否则必须开发出一种数学上一致的将负数解释为“可能性”的方法。这些和其他相关的数学问题正在探索中。

AbstractThis project takes a unified spectral transformation approach to address challenges of analyzing network topology and identifying fraud patterns in large-scale dynamic networks by using data spectral transformation with network topology visualization. Large-scale social and communication networks contain rich topological information embedded inside, in addition to various structured, semi-structured, and unstructured data. There has been little to date work dedicated to exploring network topology, especially from the spectral analysis point of view. If the proposed methods, which are based on the simple adjacency matrix representation of a graph and the node representation based on k communities in the graph, can be demonstrated to work for very large data sets, it will be a significant advance. The research is being integrated with information visualization and visual analytics algorithms and has a testbed of banking data available to allow for a search for fraud. The research is characterizing patterns of various attacks in the spectral projection space and developing spectrum based methods to identify these attacks. The approach, which exploits the spectral space of the underlying interaction structure of the network, is orthogonal to traditional approaches using content profiling. The ability to perform this spectral analysis is dependent upon the development of complex mathematical techniques. Critical issues that are being explored include the scalability of the methods to very large data sets and the determination of the dimensionality of the node representation in spectral space (which depends upon the number of clusters in the graph). Another issue is that each component of the k-dimensional representation of each node is interpreted as the 'likelihood' of a node's attachment to the k communities. However, it must be guaranteed that the components of the k-dimensional vector that represent each node will be all nonnegative or else an interpretation of the negative number as 'likelihood' must be developed that is mathematically consistent. These and other related mathematical issues are being explored.