基于众包标注的多元时间序列流早期预警及其应用

An unconventional emergency has multiple signs and causes serious damages. Therefore, it is an urgent need to research the evolutionary mechanisms and early warning of unconventional emergency based on group intelligence of experts and data mining technology. According to the characteristics of high dimension, strong noise and structural complexity of multivariate time series (MTS), here we will establish an interpretable and confident early warning model on MTS stream. Its aim is to identify unconventional emergencies that people focus on as early as possible so as to prevent or dispose of them early. The research contents include: (1) To handle the scarcity of the labeled samples and the imbalance of data, we will study a cost-sensitive PU learning method based on crowdsourcing mode, which could increase the diversity and scale of training samples at least cost. (2) To discover the local characteristics of all classes, we will analyze the intrinsic variables of different classes and the sequential relations among features of different intrinsic variables, and will propose an approach to effectively mine interpretable multivariate sequential association rules (MSAR). (3) We will build a MSAR-based hierarchical ensemble for early classification, and study evaluation criteria and balance strategy of early degree and confidence. (4) To realize incremental learning of the model, we will study parallel evolution and update strategies of the model in terms of multi-layer and multi-granularity, which could effectively improve the performance of early classification. This module could be applied to early warning of unconventional emergencies, such as sudden cardiac death. Therefore, the research has great theoretical significance and wide application prospects.

非常规突发事件预兆形式多样且危害严重，迫切需要基于专家们的群体智慧和数据挖掘技术分析其演化机理并早期预警。本项目针对多元时间序列高维度、高噪声和结构复杂等特点，构建一个具有可解释性和可信度的早期预警模型，尽早预测非常规突发事件，以便及早防范或处置。研究内容:(1)针对已标注样本稀少且数据不平衡，研究基于众包模式的代价敏感性PU学习方法，以最小成本增大训练样本集的多样性和规模；(2)为了表征各类别的局部特性，分析各类别的本征维变量及变量特征间复杂的时序关系，挖掘具有可解释性的多元时序关联规则；(3)建立一个基于多元时序关联规则的层次式集成早期分类器模型，研究分类早期度和可信度的评估准则及其平衡策略；(4)提出分类器模型多层、多粒度的并行演化和更新策略，实现模型的增量学习，提高早期分类器模型的性能。该模型可应用于心脏性猝死等非常规突发事件的早期预警，具有重大理论意义和广泛应用前景。

时间序列流早期预警是时间序列挖掘领域的重要课题，其目的是尽早预测非常规突发事件，以便及早防范或处置。本项目基于多元时间序列高维度、高噪声且结构复杂的特点和训练数据的不完备性，对多元时间序列早期预警问题进行深入研究，主要研究成果包括：样本选择策略和基于众包模式的可信标注，以最小成本增大训练样本集的多样性和规模；协同聚类方法和多元时序关联规则挖掘，挖掘具有可解释性的各类局部特征，提升早期分类性能和可解释性；设计早期度和可信度的评估准则，平衡分类早期度与可信度的控制策略，确保早期分类结果满足指定的可信度；提出规则和分类器的更新策略和评估机制等，使得模型不仅具有增量学习能力，提高早期分类器模型的性能。已发表SCI期刊论文7篇，其中IEEE Trans论文5篇，培养硕士9人和国内访问学者1名，实现了项目预期研究目标。本项目研究成果可用于非常规突发事件早期预警，如心脏异常事件、智能设备维护和网络舆情等。