- GEANN: A Scalable Graph Extension for Multi-Horizon Time Series Forecasting (arXiv)
Authors: Sitan Yang, Malcolm Wolfe, Shankar Ramasbramanian, Vincent Quenneville-Berea, Ronak Mehta, Michael W. Mahoney
Abstract: Encoder/decoder deep neural networks are being increasingly studied for multi-horizon time series forecasting, especially in real-world applications. However, to make accurate forecasts, these advanced models typically rely on a large number of time series examples with substantial history. A topic of rapidly growing interest is forecasting time series that lack sufficient historical data. This is often referred to as the “cold start” problem. In this paper, we present a novel and straightforward way to address this problem by leveraging graph neural networks (GNNs) as data augmentations to enhance the encoders used by such predictors. These GNN-based features can capture relationships between complex series and their generation process can be optimized end-to-end with forecasting tasks. We show that our architecture can use either data-driven graphs or domain knowledge-defined graphs and can be extended to incorporate information from multiple very large graphs with millions of nodes. For our target application of demand forecasting for a large e-commerce retailer, both on a small dataset of 100,000 products and a large dataset of over 2 million products, we demonstrated that the method improves overall performance over competing baseline models. More importantly, it shows significantly greater returns for ‘cold-start’ products, such as newly launched or recently out-of-stock products.
2.ENGAGE: Explanation-Guided Data Augmentation for Graph Representation Learning (arXiv)
Author : Yucheng Shi, Kaixiong Zhou, Ninghao Liu
Abstract: Recent contrastive learning methods are widely applied to modeling graph data due to their effectiveness in representation learning. Random perturbations are widely used to construct contrasting views of graph data, but they can inadvertently break the graph structure and cause suboptimal performance. Moreover, graph data is usually highly abstracted, making it difficult to extract intuitive meaning and design more informed enhancement schemes. An effective representation should preserve the main characteristics of the data and discard superfluous information. In this paper, we propose ENGAGE (ExplaNation Guided data AuGmEntation). In ENGAGE (ExplaNation Guided data AuGmEntation), explanations guide the contrast augmentation process, preserving important parts in the graph and considering removing superfluous information. Specifically, we design an efficient unsupervised explanation method called smoothed activation map as an index of node importance in representation learning. Then on the graph he designs two data augmentation schemes to perturb the structural and feature information respectively. We also show the justification of the proposed method in the framework of information theory. To demonstrate the effectiveness and flexibility of ENGAGE, experiments are conducted for both graph-level and node-level tasks on different model architectures and different real-world graphs. The ENGAGE code can be found at https://github.com/sycny/ENGAGE.
