Asynchronous Graph Generator

TL;DR

The Asynchronous Graph Generator (AGG) introduces a graph-based approach for imputation, classification, and prediction in multi-channel time series without assuming temporal regularity. By representing observations as nodes with learnable embeddings for measurements, timestamps, and channels, and by using encoder–generator blocks with a novel conditional attention generation mechanism, AGG learns expressive interdependencies through attention on a homogeneous, asynchronous graph. Empirical results across Beijing Air Quality, PhysioNet ICU 2012, and UCI Localization show state-of-the-art performance, robustness to sparsity, and competitive or superior results to recent attention-based methods, with additional validation on foreign exchange data. The work highlights data augmentation and transductive node generation as key mechanisms unlocking strong performance and broader applicability to downstream tasks.

Abstract

We introduce the asynchronous graph generator (AGG), a novel graph attention network for imputation and prediction of multi-channel time series. Free from recurrent components or assumptions about temporal/spatial regularity, AGG encodes measurements, timestamps and channel-specific features directly in the nodes via learnable embeddings. Through an attention mechanism, these embeddings allow for discovering expressive relationships among the variables of interest in the form of a homogeneous graph. Once trained, AGG performs imputation by \emph{conditional attention generation}, i.e., by creating a new node conditioned on given timestamps and channel specification. The proposed AGG is compared to related methods in the literature and its performance is analysed from a data augmentation perspective. Our experiments reveal that AGG achieved state-of-the-art results in time series imputation, classification and prediction for the benchmark datasets \emph{Beijing Air Quality}, \emph{PhysioNet ICU 2012} and \emph{UCI localisation}, outperforming other recent attention-based networks.

Figures (6)

• Figure 1: An illustration of the AGG self-supervised pipeline. a) Time series samples are collected (possibly) asynchronously, and comprise measurements, timestamps and channel features. b) Samples are ordered and a split into inputs and targets for self-supervised training. c) The input/target split is considered as instances of the asynchronous graph for training. d) The learnt graph encodes a rich representation of the underlying signal, where new samples to be generated through conditional attention generation; here, $c_g = blue$ and $\tau_g = t_N - t_*$.
• Figure 2: AGG architecture: The sections of the network are indicated at the top of the figure. Inputs and target are represented as circles and squares respectively, fixed operations are denoted by white blocks and learnable transformations in light grey blocks.
• Figure 3: AGG performance vs number of training samples from the same dataset via augmentation.
• Figure 4: A pictographic representation of how the time series data is converted into an input block and a imputation target after data is randomly removed. The stride, as depicted in the image is defined as the number of steps the block is moved before it is considered the next input to the AGG. In this image the stride has value of 2 and input block a size of 11.
• Figure 5: (a) Augmented training size vs stride as a percentage of block size. (b) Stride sensitivity.