Spatio-Temporal Bi-LSTM Based Variational Auto-Encoder for Multivariate IoT Data Imputation

Abstract

In the relam of the Internet of Things (IoT), prevalence of missing data due to continuous data collection by smart devices necessitates the essential preliminary step of data imputation before engaging in information mining activities. IoT data exhibit robust interconnections in both spatial and temporal dimensions, surpassing the limitations of Euclidean space. Yet, prevailing machine learning and deep learning approaches often focus solely on temporal attributes or capture spatial features exclusively within a Euclidean framework. To address these challenges, this paper introduces a novel network named ST-Bi-LSTM-VAE (Spatio-Temporal Bidirectional Long Short-Term Memory based Variational Auto-Encoder). The architecture of ST-Bi-LSTM-VAE is primarily grounded in the Variational Auto-Encoder (VAE) framework. This innovative approach incorporates two distinct types of VAEs. The first type is dedicated to computing the adjacent matrix of the device network, a crucial input for the Graph Convolutional Network (GCN) essential in capturing intricate spatial relationships among devices. The second type of VAE is specifically tailored for data imputation, leveraging both global spatial and temporal dependencies. Empirical experiments conducted on diverse publicly available datasets substantiate the efficacy of ST-Bi-LSTM-VAE. The results obtained consistently demonstrate that proposed method surpasses baseline techniques in maintaining pattern, structure, and trend across datasets even at 50% missing gap for imputation task with 4.91% performance improvement in case of Intel Berkley Research Laboratory (IBRL) dataset and 3.5% on PRSA dataset.