基于时序数据的结构学习与模式预测联合优化算法研究

发布时间：2018-03-07 06:34

本文选题：时序数据分析　切入点：结构学习　出处：《浙江大学》2016年博士论文　论文类型：学位论文

【摘要】：时序数据分析是一项重要的数据挖掘技术。在现实中,该科学技术不仅可以用于发现时序数据中潜藏的发展模式或演化规律,而且也可以用于预测时序数据的未来发展趋势。时序数据分析通常包含两个方面的科学研究,即结构学习和模式预测。其中,结构学习的目标是利用时序数据样本之间的内在关系来提炼出新型的样本特征或者样本标签；模式预测的目标则是在样本特征和样本标签之间建立起有一种有效的映射关系。相比于传统的时序数据分析方法通常将上述两者视为相互独立的模块,本文认为结构学习和模式预测存在着内在的互动联系,并且从理论分析与模型设计上阐明结构学习和模式预测是相互支持与相互促进的。本文深入分析了实际应用中可能出现的各种问题和挑战,在研究大量前人先进工作的基础上,提出了多种联合结构学习模型和模式预测模型的新型时序数据分析算法。论文工作的主要内容和创新之处在于：一,提出了一种基于消费者社会影响力分析的商品行为研究算法,并实际应用于女装商品的销量预测。该算法首先利用电子商务平台中的商品数据和消费者数据来建立商品特征和消费者特征,并以此研究商品之间的竞争合作关系和消费者之间的社会影响力。其次,考虑到实际中的商品销量变化同时存在平缓性以及波动性,该算法将商品销量分解为销量主体部分和销量噪声部分,并分别进行回归建模。最后,该算法建立一种有效的优化模型来联合实现消费者社会影响力的挖掘与商品销量的预测。二,提出了一种基于慢特征分析的人脸年龄估计算法。该算法首先联合人脸样本的时序结构和上下文结构来建立结构化人脸序列,并以此表征人脸样本的内在结构。其次,该算法利用上述的结构化人脸序列来提取人脸样本的慢特征,并将其作为一种新型的人脸特征。最后,该算法建立一种有效的优化模型来联合实现结构化人脸序列的构造和人脸样本年龄标签的预测。三,提出了一种基于自适应年龄分布学习的人脸年龄估计算法。该算法首先利用子空间技术来发现人脸样本的上下文结构。其次,该算法基于样本上下文结构来建立人脸样本的年龄分布,并将其作为一种新型的预测目标；并且,样本年龄分布的建立可以有利于研究者理解上下文相关样本中的时序年龄关联性以及分析年龄标签模糊性问题。最后,该算法建立一种有效的优化模型来联合实现人脸样本上下文结构的发现与人脸样本年龄分布的预测。
[Abstract]:Time series data analysis is an important data mining technology. In reality, this technology can not only be used to discover the hidden development patterns or evolution laws in time series data. It can also be used to predict the future trend of time series data. Time series data analysis usually involves two aspects of scientific research, namely, structure learning and pattern prediction. The goal of structure learning is to extract a new type of sample feature or sample label by using the inherent relationship between the samples of time series data. The goal of pattern prediction is to establish an effective mapping relationship between sample features and sample labels. This paper argues that there is an internal interaction between structural learning and model prediction. From the theoretical analysis and model design, it is explained that structural learning and model prediction support each other and promote each other. In this paper, various problems and challenges that may arise in practical application are deeply analyzed. In this paper, a variety of new time series data analysis algorithms for joint structural learning models and pattern prediction models are proposed. The main contents and innovations of this paper are as follows: 1. In this paper, a research algorithm of commodity behavior based on consumer social impact analysis is proposed. The algorithm first uses the commodity data and consumer data in the e-commerce platform to establish the commodity characteristics and consumer characteristics. Secondly, considering the smoothness and volatility of the real changes in commodity sales, we study the relationship of competition and cooperation between commodities and the social influence between consumers. The algorithm decomposes the sales volume into the main part of the sales volume and the noise part of the sales volume, and respectively carries on the regression modeling. Finally, The algorithm establishes an effective optimization model to jointly realize the mining of consumer social influence and the prediction of sales volume. In this paper, a face age estimation algorithm based on slow feature analysis is proposed, which combines the temporal structure and context structure of the face sample to construct the structured face sequence, and represents the inherent structure of the face sample. The algorithm uses the structured face sequence to extract the slow feature of the face sample and regards it as a new type of face feature. Finally, This algorithm establishes an effective optimization model to combine the construction of structured face sequences and the prediction of face sample age tags. In this paper, a face age estimation algorithm based on adaptive age distribution learning is proposed, which firstly uses subspace technology to discover the context structure of face samples. The algorithm is based on the sample context structure to establish the age distribution of face samples and take it as a new prediction target. The establishment of sample age distribution can help researchers understand the temporal correlation of contextual samples and analyze the fuzziness of age labels. This algorithm establishes an effective optimization model to combine the discovery of face sample context structure with the prediction of face sample age distribution.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：O211.61;TP301.6

【相似文献】