Abstract:This paper introduces the Deep Functional Factor Model (DF2M), a Bayesian nonparametric model designed for analysis of high-dimensional functional time series. DF2M is built upon the Indian Buffet Process and the multi-task Gaussian Process, incorporating a deep kernel function that captures non-Markovian and nonlinear temporal dynamics. Unlike many black-box deep learning models, DF2M offers an explainable approach to utilizing neural networks by constructing a factor model and integrating deep neural networks within the kernel function. Additionally, we develop a computationally efficient variational inference algorithm to infer DF2M. Empirical results from four real-world datasets demonstrate that DF2M provides better explainability and superior predictive accuracy compared to conventional deep learning models for high-dimensional functional time series.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the analysis and prediction of high - dimensional function time series. Specifically, the paper proposes a method named Deep Functional Factor Model (DF2M), aiming to deal with the nonlinear and non - Markovian time - dependence in high - dimensional function time series through Bayesian non - parametric methods. DF2M combines the Indian Buffet Process (IBP), Multi - Task Gaussian Process (MTGP) and deep neural networks to improve the interpretability and prediction accuracy of the model. ### Main problems 1. **Processing of high - dimensional data**: With the progress of data collection techniques and computing power, high - dimensional data sets containing a large number of function time series are becoming more and more common. These data sets are characterized by each observation point being a function (for example, age - specific annual mortality curves, daily energy consumption curves of households, etc.), so it is necessary to handle the infinite - dimensional data characteristics. 2. **Capturing of time - dependence**: Traditional statistical methods often assume that time series have linear and Markov properties, which may not accurately capture complex nonlinear and non - Markovian time - dependence in practical applications. 3. **Interpretability of the model**: Although deep learning has achieved remarkable results in fields such as computer vision and natural language processing, its black - box nature makes it difficult to be applied in key fields such as finance, medical care and climate change, because decision - making in these fields requires the model to have high interpretability. ### Solutions DF2M solves the above problems in the following ways: 1. **Functional factor model**: Use the functional factor model for dimension reduction, represent the high - dimensional function time series as a combination of a small number of latent functional factors, thereby improving the interpretability of the model. 2. **Bayesian non - parametric method**: Adopt the Indian Buffet Process prior to encourage the column sparsity of the loading matrix, allowing the number of latent factors to change flexibly without prior specification. 3. **Multi - task Gaussian process**: Utilize the multi - task Gaussian process to capture the time - dependence of latent functional factors, and capture nonlinear and non - Markovian dynamics by designing appropriate kernel functions (such as deep kernel functions). 4. **Deep neural network**: Integrate the deep neural network in the kernel function to enhance the model's ability to model complex time - dependence. ### Experimental results The paper proves through experiments on four real - world data sets that DF2M performs better in terms of prediction accuracy of high - dimensional function time series compared with traditional deep - learning models, and also provides better interpretability. In conclusion, by proposing the DF2M model, this paper effectively solves multiple challenges in the analysis of high - dimensional function time series, especially in dealing with complex time - dependence and improving the interpretability of the model.

Deep Functional Factor Models: Forecasting High-Dimensional Functional Time Series via Bayesian Nonparametric Factorization

Strongly Hierarchical Factorization Machines and ANOVA Kernel Regression

Factor Models for High‐Dimensional Functional Time Series II: Estimation and Forecasting

Factor modelling for high-dimensional functional time series

Factor Models for High‐Dimensional Functional Time Series I: Representation Results

Covariance Function Estimation for High-Dimensional Functional Time Series with Dual Factor Structures

Deep Factors with Gaussian Processes for Forecasting

On the modelling and prediction of high-dimensional functional time series

Bayesian Dynamic Factor Models for High-dimensional Matrix-valued Time Series

Bayesian Temporal Factorization for Multidimensional Time Series Prediction

Deep Fundamental Factor Models

Deep Markov Spatio-Temporal Factorization

Time-Warped Sparse Non-negative Factorization for Functional Data Analysis

Generalized functional dynamic principal component analysis

Modeling High-Dimensional Time Series: A Factor Model With Dynamically Dependent Factors and Diverging Eigenvalues

High-dimensional functional time series forecasting: An application to age-specific mortality rates

Dynamic Matrix Factor Models for High Dimensional Time Series

Enhanced Bayesian Factorization with Variant Scale Partitioning for Multivariate Time Series Analysis

High-dimensional Factor Model and Its Applications to Statistical Machine Learning

Factor-guided estimation of large covariance matrix function with conditional functional sparsity