高维空间下大数据多模态聚类与预测及精准教育服务研究

It is significant to improve the precision level of educational services through the multimodal clustering and prediction approaches for educational big data by exploiting the multi-dimensional correlative analysis under high-dimensional space. The project aims at providing the personalized educational services, and focuses on the following issues: high-dimensional fusion approach for the multi-source and heterogeneous data, multi-modal clustering method, multi-modal prediction, and precise educational service model. First, a set of tensor-based representation and cross-domain fusion approaches are presented for the multi-source and heterogeneous educational data, which lay the foundation for the subsequent correlative analysis. Second, based on the extracted key characteristics by implementing tensor train decomposition for the clustering objects, the tensor distances between any two clustering objects are calculated when selecting different situation spaces, and then the multi-modal clustering algorithm is further proposed by exploiting the tensor-distance based measurement. Third, based on the learners’ clustering results, we explore the interaction effects among the similar learners and determine the multimodal influence factors which influence learners to choose their learning resources, and put forward a prediction algorithm based on the prominent joint eigen tensor to realize the multimodal sequence prediction. Fourth, to tackle the computational challenges caused by the curse of dimensionality, we study the parallel computation method for the tensor calculations which are directly implemented based on the tensor train decomposition results. Finally, from the integrative perspective of educational theories and computing technologies, a precise educational service model based on the pervasive learning process is explored to provide the real-time personalized educational services. This study will provide a new method for the multi-modal clustering and prediction with their efficient calculations under the high-dimensional space, and it is conducive to resolving the problem of precise educational services and promoting the adaptive learning for everyone.

在高维空间对教育大数据进行整体关联分析及多模态聚类和预测，有利于提升教育服务的精准化水平。项目以个性化教育服务为目的，致力于多源异构数据高维融合、多模态聚类、多模态预测及精准教育服务模式研究。首先利用张量模型对多源异构数据进行高维表示和跨域融合，奠定后续关联分析基础；基于张量链分解提取的关键特征，计算不同情境空间选择系数下聚类对象间的张量距离，以此为度量方式提出多模态聚类方法；基于学习者聚类结果，探寻同类学习者之间的相互影响，分析学习资源选择的多模态影响因素，提出基于主特征联合张量的多模态序列预测算法；针对维度灾难计算难题，探求直接基于张量链分解结果进行张量操作的并行计算方法；并从教育与技术融合的视角，探索嵌入泛在学习过程的精准教育服务模式，提供应时应景的个性化教育服务。本研究将为高维空间下大数据多模态聚类与预测及高效计算提供新方法，并破解教育服务中的精准化难题，促使人人能享自适应学习。

为提升教育服务的精准化水平，本项目重点研究在高维空间下对教育大数据进行整体关联分析及多模态聚类和预测。以个性化教育服务为目的，通过多源异构数据高维融合、多模态聚类、多模态预测实现精准教育服务。首先利用张量模型对多源异构的教育大数据进行高维表示和跨域融合，得到融合张量，利用张量链分解技术提取融合张量的关键特征。然后，通过计算不同情境空间选择系数下聚类对象间的张量距离，并以此为度量方法进行聚类，实现了在选择不同情境系数时求取聚类结果。随后，基于学习者聚类结果，结合学习过程中的局部强关联和认知延续性规律，提出多元多阶马尔科夫预测模型的构建方法，研究基于主特征联合概率张量进行多模态预测的机制，实现在不同影响因素下的精准预测。同时，针对维度灾难计算难题，提出了直接基于张量链分解后的低阶张量直接进行多元状态转移的方法，并根据其运算规则及其分布存储特点，研究包含核间和核内的两层并行计算策略，解决计算效率低下难题。并从教育与技术融合的视角，从认知心理学、学习科学、知识服务、数字教育资源配置等相关理论角度，研究嵌入泛在学习过程的精准教育服务模式，提出了自适应的个性化教育服务理念，并在实践中验证了其效果。最后，基于以上理论框架搭建了一个基于张量对教育大数据进行高维表示、跨域融合、多模态聚类和预测的示范平台。本研究为解决教育大数据的多源异构难题，以及大数据多模态聚类与预测及高效计算提供了新方法，可在泛在自适应学习中实现精准的教育资源推荐，有利于破解规模化教育的个性化培养难题，促使人人能享自适应学习。