【作者】 张瑜；

【导师】 彭玉青；

【作者基本信息】 河北工业大学 ， 计算机应用技术， 2003， 硕士

【摘要】 数据挖掘(Data Mining)就是从大量的、不完全的、有噪声的、模糊的、随机的数据中，提取隐含在其中的、人们事先不知道的、但又是潜在有用的信息和知识的过程。被信息产业界认为是数据库系统最重要的前沿之一，是信息产业最有前途的交叉学科。 数据挖掘算法的好坏将直接影响到所发现知识的好坏。目前大多数的研究都集中在数据挖掘算法和应用上。本文根据数据挖掘领域的需求和现状，研究了复杂数据库中的时序数据库挖掘技术。主要从事的是，在时序数据中发现时序模式。 本文概述了时序模式发现所涉及的领域和学科的基本理论知识、目前的研究现状，并分析了它的研究意义。同时给出了本文的具体的工作，主要是：对在时序数据序列中发现模式问题进行了描述，并介绍了一种新的趋势逻辑表示方法，给出了其算法及算法的实验结果；对时序数据进行处理，提出了利用线段的斜率反正切值作为模式识别的样本，从而在分类时忽略模式的畸变；另外，还提出了一个新的基于高阶神经网络的时序模式发现算法。在这一算法中，作者运用了单层的高阶神经网络和分层方法。单层高阶神经网络没有隐含层节点的困扰，训练速度更快；模式分类能力更为强大；不存在局部最小的问题，模式分类精度更高；模式的不变性构建于网络结构之中等优点。分层方法能加快各节点神经网络的学习速度，也能提高模式划分精度。这两种技术使得算法的分类效果更好。给出了算法思想，并实现了此算法。 实验结果表明，作者提出的算法在分类的精度上有了一定的改善。 本文提出的算法，可以应用到具有时序数据的领域。同其他的时序模式发现算法相比较，具有几点创新：1)从分类精度的角度来改善时序模式发现算法；2)通过高阶神经网络，直接利用网络的特性忽略了时序模式的某些畸变；3)利用了分层方法，进一步改善分类精度。更多还原

【Abstract】 Data mining is the process of abstracting unaware, potential and useful information and knowledge from plentiful, incomplete, noisy, fuzzy and stochastic data, which is deemed to one of a foreland of data mining system and a promising cross-subject. Association rule is one of more important part in data mining, and finding frequent item sets plays an important role in abstracting association rules. Most of existing algorithms cost much time and space because they need plenty of repetition.Data mining algorithm will directly influence the finding knowledge. At present, most of researches focus on the algorithm and application of data mining. According to requirement and actuality of data mining field, mining technology of time-series database in complex database is studied in this paper. I study on mostly finding time-series patterns in time-series data.The field and subject basic theory knowledge of time-series patterns finding and the actuality of research are summarized, and the meaning of its research is analyzed. The detailed works are as follows: the finding patterns problems in the time-series data sequence are described, and a new trend logic expression method is introduced, and its algorithm and experiment result of algorithm are given; time-scries data are disposed, and using the arctg. slope of line as the sample of pattern recognition, so ignoring the aberrance of pattern in the classified. In addition, a new time-series pattern finding algorithm based on higher-order neural network is put forward. In this algorithm, author used two technology - the higher-order neural network of single_ layer and hierarchy. Higher-order neural network of single_layer has no trouble of hiding layer node, so training speed is fast; Ability of pattern classified increases; The classification accuracy of pattern is improved without local optimality problem; Hierarchy can make training speed of each node neural network, and improve the classified accuracy of pattern. The two technologies make the classified effect of the algorithm better. The thinking of the algorithm is provided and performed.The result of experiment indicates that the algorithm which author put toward improve the classified accuracy.The algorithm that the paper provides may be applied in the field of time-series data. Comparing with the other time-series pattern finding algorithm, this one has a few of innovations. Firstly, improving time-series pattern finding algorithm from the classified accuracy. Secondly, ignoring some aberrance of time-series pattern by the high-order neural network. At last, improving the classified accuracy by the classified way.更多还原