Learning representations for multivariate time series with missing data\n using Temporal Kernelized Autoencoders

2018/05/09 by Filippo Maria Bianchi, Lorenzo Livi, Bianchi, Filippo Maria +7 · 1 citation
Computer Science · Medicine · #Anomaly Detection Techniques and Applications #FOS: Computer and information sciences #Machine Learning in Healthcare #Neural and Evolutionary Computing (cs.NE) #Time Series Analysis and Forecasting #Traditional Chinese Medicine Studies

paper · pdf · doi:10.48550/arxiv.1805.03473

openalex publication_date 2018/05/09 · openalex created_date 2022/08/30 · openalex updated_date 2026/07/28

Abstract

Learning compressed representations of multivariate time series (MTS)\nfacilitates data analysis in the presence of noise and redundant information,\nand for a large number of variates and time steps. However, classical\ndimensionality reduction approaches are designed for vectorial data and cannot\ndeal explicitly with missing values. In this work, we propose a novel\nautoencoder architecture based on recurrent neural networks to generate\ncompressed representations of MTS. The proposed model can process inputs\ncharacterized by variable lengths and it is specifically designed to handle\nmissing data. Our autoencoder learns fixed-length vectorial representations,\nwhose pairwise similarities are aligned to a kernel function that operates in\ninput space and that handles missing values. This allows to learn good\nrepresentations, even in the presence of a significant amount of missing data.\nTo show the effectiveness of the proposed approach, we evaluate the quality of\nthe learned representations in several classification tasks, including those\ninvolving medical data, and we compare to other methods for dimensionality\nreduction. Successively, we design two frameworks based on the proposed\narchitecture: one for imputing missing data and another for one-class\nclassification. Finally, we analyze under what circumstances an autoencoder\nwith recurrent layers can learn better compressed representations of MTS than\nfeed-forward architectures.\n

Citations

• A Recurrent Latent Variable Model for Sequential Data
• Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks
• Adam: A Method for Stochastic Optimization
• Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling
• Long Short-Term Memory
• Inference and missing data
• Finding Structure in Time
• Distinguishing “Missing at Random” and “Missing Completely at Random”
• Sequence to Sequence Learning with Neural Networks
• Bidirectional recurrent neural networks
• Representation Learning: A Review and New Perspectives

Cited by

• A Causal-based Framework for Multimodal Multivariate Time Series Validation Enhanced by Unsupervised Deep Learning as an Enabler for Industry 4.0

Related

• Learning compressed representations of blood samples time series with missing data
• Recurrent Neural Networks for Multivariate Time Series with Missing Values
• Learning from Irregularly-Sampled Time Series: A Missing Data\n Perspective
• An End-to-End Model for Time Series Classification In the Presence of Missing Values
• Modeling Missing Data in Clinical Time Series with RNNs