Abstract:Missing value imputation is a fundamental problem in spatiotemporal modeling, from motion tracking to the dynamics of physical systems. Deep autoregressive models suffer from error propagation which becomes catastrophic for imputing long-range sequences. In this paper, we take a non-autoregressive approach and propose a novel deep generative model: Non-AutOregressive Multiresolution Imputation (NAOMI) to impute long-range sequences given arbitrary missing patterns. NAOMI exploits the multiresolution structure of spatiotemporal data and decodes recursively from coarse to fine-grained resolutions using a divide-and-conquer strategy. We further enhance our model with adversarial training. When evaluated extensively on benchmark datasets from systems of both deterministic and stochastic dynamics. NAOMI demonstrates significant improvement in imputation accuracy (reducing average prediction error by 60% compared to autoregressive counterparts) and generalization for long range sequences.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is **the missing value imputation problem in long - time series**, especially in spatio - temporal modeling, such as in fields like motion tracking and physical system dynamics. Specifically, existing deep autoregressive models are prone to error accumulation when dealing with long - time series, resulting in poor imputation effects. Therefore, this paper proposes a non - autoregressive multi - resolution sequence imputation model (NAOMI) to address this challenge. ### Problem Background 1. **Importance of the Missing Value Problem**: - Missing values often occur in actual sequential data. For example, in motion tracking, trajectories may contain missing data due to object occlusion, trajectory crossing, or unstable camera movement. - Missing values introduce observation bias, affecting the quality of training data and making the learning process unstable. Therefore, imputing missing values is crucial for downstream sequence learning tasks. 2. **Limitations of Existing Methods**: - Statistical methods usually rely on strong assumptions about missing patterns, such as missing at random (MAR), and perform poorly when dealing with high - proportion missing data and long sequences. - Although deep generative models provide a flexible framework for dealing with missing data, all existing methods are autoregressive, that is, they predict the value of the current time step through the value of the previous time step. This sequential decoding method is prone to error accumulation, especially in long - time series. ### Core Contributions of the Paper 1. **Proposing the Non - autoregressive Multi - resolution Imputation Model (NAOMI)**: - NAOMI adopts a non - autoregressive method and uses the multi - resolution structure of spatio - temporal data to impute missing values recursively from coarse - grained to fine - grained. - This model not only considers historical information but also combines (predicted) future information, thus avoiding the error accumulation problem. 2. **Introducing Adversarial Training**: - Use generative adversarial networks (GAN) for adversarial training to reduce variance and improve imputation accuracy. 3. **Experimental Verification**: - Extensive experiments were carried out on multiple benchmark datasets, including traffic time series, billiard ball trajectories, and basketball game trajectories. The experimental results show that NAOMI is significantly superior to autoregressive models in imputation accuracy, with an average error reduction of 60%. ### Formula Summary - **Conditional Distribution Modeling**: \[ q(H_f|I)=\prod_{t = 1}^Tq(h_f^t|h_f^{<t},I_{\leq t}) \] \[ q(H_b|I)=\prod_{t = 1}^Tq(h_b^t|h_b^{>t},I_{\geq t}) \] - **Decoding Function**: \[ p(x_t^*|H)=g^{(r)}(h_f^i,h_b^j) \] - **Adversarial Training Objective Function**: \[ \min_{\theta}\max_{\omega}\mathbb{E}_{X^*\sim C}\left[\sum_{t = 1}^T\log D_{\omega}(\hat{x}_t,x_t)\right]+\mathbb{E}_{X^*\sim C,M\sim p(M),\hat{X}\sim G_{\theta}}\left[\sum_{t = 1}^T\log(1 - D_{\omega}(\hat{x}_t,x_t))\right] \] Through these innovations, NAOMI can provide more accurate and robust imputation results when dealing with missing values in long - time series.

NAOMI: Non-Autoregressive Multiresolution Sequence Imputation

Multiple Imputation with Denoising Autoencoder using Metamorphic Truth and Imputation Feedback

Imitative Non-Autoregressive Modeling for Trajectory Forecasting and Imputation

Uncertainty-Aware Deep Attention Recurrent Neural Network for Heterogeneous Time Series Imputation

MIDIA: exploring denoising autoencoders for missing data imputation

MuSDRI: Multi-Seasonal Decomposition Based Recurrent Imputation for Time Series

Missing Value Imputation on Multidimensional Time Series

Sequence-to-Sequence Imputation of Missing Sensor Data

Smoothed LSTM-AE: A spatio-temporal deep model for multiple time-series missing imputation

Missingness-Pattern-Adaptive Learning With Incomplete Data

Time Series Imputation with Multivariate Radial Basis Function Neural Network

Missing Value Imputation for Radar-Derived Time-Series Tracks of Aerial Targets Based on Improved Self-Attention-Based Network

Long-Term Missing Value Imputation for Time Series Data Using Deep Neural Networks

MBGAN: An improved generative adversarial network with multi-head self-attention and bidirectional RNN for time series imputation

SSIM—A Deep Learning Approach for Recovering Missing Time Series Sensor Data

Deep Generative Imputation Model for Missing Not At Random Data

Parallel Generative Adversarial Imputation Network for Multivariate Missing Time-Series Reconstruction and Its Application to Aeroengines

Online Missing Value Imputation for High-Dimensional Mixed-Type Data via Generalized Factor Models

ContrAttNet: Contribution and attention approach to multivariate time-series data imputation

Missing Features Reconstruction Using a Wasserstein Generative Adversarial Imputation Network

Neural ODEs for Informative Missingness in Multivariate Time Series