Scholarlink

<p>Missing data often undermines machine learning and effective data analysis in industries like healthcare, industrial sensor networks, and finance. Despite the existence of various imputation methods, the majority of modern models focus on random, unstructured missingness and fail to capitalize on the temporal dependencies present, including recurrent sparsity patterns. Furthermore, when handling consecutive missing values, conventional iterative methods frequently experience error propagation. The recurrent imputation network with patterned output (RINPO), a deep learning framework designed to handle both structured and unstructured sparse data, is presented in this study. RINPO combines a novel dual-output layer with the temporal modeling capabilities of recurrent neural networks (RNNs), particularly LSTMs and GRUs. RINPO’s design enables the simultaneous prediction of multiple subsequent missing values at each time step, directly capturing the structure of the sparsity pattern, in contrast to standard architectures that predict a single step ahead. The model was thoroughly tested on four different datasets: PAMAP2, SYNTH (MIMO radar), smart home, and NASA C-MAPSS, as well as cutting-edge deep learning techniques (GAIN, VAE, and standard LSTM) and traditional baselines (Mean, k-NN). Results from experiments show that RINPO consistently attains higher accuracy. With an MSE of 0.0180, RINPO-GRU had the lowest error on the PAMAP2 dataset. With an MSE of 0.1882, RINPO-LSTM greatly outperformed rivals in complex industrial scenarios like the NASA C-MAPSS dataset (FD001). Furthermore, the model demonstrated high robustness when handling random missingness, demonstrating its adaptability outside of structured patterns. According to the study’s findings, RINPO offers a reliable, highly accurate, and computationally efficient method for data imputation. This is a significant development for practical applications involving high-dimensional, sparse time-series data, as it enables efficient interpolation of sensor readings at irregular intervals. The code for this work is available at the following link: <a href="https://github.com/saood321/RINPO.git">https://github.com/saood321/RINPO.git</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Recurrent imputation network with patterned output (RINPO) for sparse data

  • 期刊论文

  • 发表日期：

    2026年4月6日

查看全文
  • Shahzaib Ur Rehman,
  • Muhammad Saood Sarwar

摘要

Missing data often undermines machine learning and effective data analysis in industries like healthcare, industrial sensor networks, and finance. Despite the existence of various imputation methods, the majority of modern models focus on random, unstructured missingness and fail to capitalize on the temporal dependencies present, including recurrent sparsity patterns. Furthermore, when handling consecutive missing values, conventional iterative methods frequently experience error propagation. The recurrent imputation network with patterned output (RINPO), a deep learning framework designed to handle both structured and unstructured sparse data, is presented in this study. RINPO combines a novel dual-output layer with the temporal modeling capabilities of recurrent neural networks (RNNs), particularly LSTMs and GRUs. RINPO’s design enables the simultaneous prediction of multiple subsequent missing values at each time step, directly capturing the structure of the sparsity pattern, in contrast to standard architectures that predict a single step ahead. The model was thoroughly tested on four different datasets: PAMAP2, SYNTH (MIMO radar), smart home, and NASA C-MAPSS, as well as cutting-edge deep learning techniques (GAIN, VAE, and standard LSTM) and traditional baselines (Mean, k-NN). Results from experiments show that RINPO consistently attains higher accuracy. With an MSE of 0.0180, RINPO-GRU had the lowest error on the PAMAP2 dataset. With an MSE of 0.1882, RINPO-LSTM greatly outperformed rivals in complex industrial scenarios like the NASA C-MAPSS dataset (FD001). Furthermore, the model demonstrated high robustness when handling random missingness, demonstrating its adaptability outside of structured patterns. According to the study’s findings, RINPO offers a reliable, highly accurate, and computationally efficient method for data imputation. This is a significant development for practical applications involving high-dimensional, sparse time-series data, as it enables efficient interpolation of sensor readings at irregular intervals. The code for this work is available at the following link: https://github.com/saood321/RINPO.git.