面向高维稀疏时变数据的宏观趋势预测研究

Temporal data with high dimensionality and sparsity are frequently encountered in scientific research and industry applications. How to predict the macro tendency based on such sparse data is an essential issue related to many application scenes in the area of big data. For addressing it, this project models the temporal data with a temporal tensor, and then attempts to accumulate predictions for micro connections among individuals into predictions for macro tendency on the whole data space, from the aspects of mathematical analysis, network characteristics and social influence theory. The first topic focuses on each single slice corresponding to each single time point in the temporal tensor. The latent factor analysis and bipartite graph analysis will be respectively adopted for modeling its data and network characteristics; then the outputs will be fused for predicting unobserved static, micro connections in each single slice. The second topic considers the whole temporal tensor, stressing on time-aware predictions for its unknown dynamic connections among individuals. The tensor factorization and multiscale modeling will be adopted for extracting temporal patterns and knowledge from the target tensor for such purpose. With the reliable predictions for static and dynamic connections among individuals, the last topic adopts the important node analysis in complex networks for modeling and optimizing the individual influence in macro tendency, and then accumulates the predictions of micro connections into the predictions of macro tendency. The target applications of this project are social media popularity prediction and e-commerce sales rank prediction. The methods and algorithms will be designed and tuned to fit the target applications, and then applied to real industrial systems for validating their performance.

高维稀疏时变数据是科学研究和工程应用中普遍面临的研究对象。基于该数据的宏观趋势预测是大数据挖掘的一个核心问题，有着广泛的应用场景。本项目拟利用时变关系张量来对其进行刻画，从数值分析和网络性质入手，基于社会影响力理论，汇聚个体的微观预测为总体的宏观趋势。首先，针对时变关系张量在单一时间点的切片，分别利用隐特征分析方法和复杂网络分析方法获得数值特征和网络特征，并融合得到时间片静态数据的微观预测；其次，重点考虑引入时间信息的时变关系张量的预测问题，结合张量分析方法和多尺度建模方法，抽取能反映时间动态变化趋势的模式和知识，实现时序数据的微观预测；第三，在静态和动态的微观预测基础上，利用复杂网络重要节点分析来指导宏观趋势预测中的个体影响力学习和优化，通过汇聚微观预测结果来实现宏观趋势预测。项目将应用场景聚焦在社会媒体流行性预测和电商销售排行榜预测，以此考虑算法设计优化，以及应用和检验相关研究成果。

高维稀疏时变数据是科学研究和工程应用中普遍面临的研究对象。基于该数据的宏观趋势预测是大数据挖掘的一个核心问题。本项目利用时变关系张量来对其进行刻画，从数值分析和网络性质入手，通过汇聚个体的微观预测为总体的宏观趋势。首先，针对时变关系张量在单一时间点的切片，分别利用隐特征分析方法和复杂网络分析方法获得数值特征和网络特征，并融合得到时间片静态数据的微观预测；其次，重点考虑引入时间信息的时变关系张量的预测问题，结合张量分析方法和多尺度建模方法，抽取能反映时间动态变化趋势的模式和知识，实现时序数据的微观预测；第三，在静态和动态的微观预测基础上，利用复杂网络重要节点分析来指导宏观趋势预测中的个体影响力学习和优化，通过汇聚微观预测结果来实现宏观趋势预测。本项目已经超额完成研究任务，发表论文31篇（其中IEEE Trans. 11篇），培养人才13人，获得省部级一等奖2项。在2020年1月基金委组织的结题检查中获评优秀。