High-dimensional time series often exhibit hierarchical structures represented by tensors, while statistical methodologies that can effectively exploit the structural information remain limited. We propose a supervised factor modeling framework that accommodates general hierarchical structures by extracting low-dimensional features sequentially in the mode orders that respect the hierarchical structure. Our method can select a small collection of such orders to allow for impurities in the hierarchical structures, yielding interpretable loading matrices that preserve the hierarchical relationships. A practical estimation procedure is proposed, with a hyperparameter selection scheme that identifies a parsimonious set of action orders and interim ranks, thereby revealing the possibly latent hierarchical structures. Theoretically, non-asymptotic error bounds are derived for the proposed estimators in both regression and autoregressive settings. An application to the IPIP-NEO-120 personality panel illustrates superior forecasting performance and clearer structural interpretation compared with existing methods based on tensor decompositions and hierarchical factor analysis.
翻译:高维时间序列常呈现以张量表示的层次结构,而能有效利用该结构信息的统计方法仍较为有限。本文提出一种监督因子建模框架,通过沿符合层次结构的模态顺序依次提取低维特征,以适配一般化的层次结构。该方法可选择少量此类顺序以容忍层次结构中的不纯性,从而生成保留层次关系的可解释载荷矩阵。我们提出了一种实用的估计流程,并配备超参数选择方案,以识别简约的动作顺序与中间秩,进而揭示可能的潜在层次结构。理论上,本文推导了所提估计量在回归与自回归设定下的非渐近误差界。应用于IPIP-NEO-120人格面板的实例表明,相较于基于张量分解与层次因子分析的现有方法,本方法具有更优的预测性能与更清晰的结构可解释性。