Short-Term Load Forecasting with Missing Data using Dilated Recurrent Attention Networks

Abstract

Data without annotation are easy to obtain in the real-world, however, established supervised learning methods are not applicable to analyze them. Several learning approaches have been proposed in recent years to exploit the underlying structure of the data without requiring annotations. Semi-supervised learning aims to improve the predictive performance of these unsupervised approaches, by exploiting partially acquired annotations in the dataset. One recent promising line of work in this scheme makes use of graph neural networks (GNN). The data is expressed as a graph, where vertices are data samples and edges, given by an adjacency matrix A, represent pairwise relationships between data points. Although these approaches achieve promising performance, they have so far been limited to applications, where the graph, in form of the adjacency matrix, is available. This is a severe limitation, as most available datasets do not include a predefined graph structure. To address this shortcoming, we investigate if the adjacency matrix A can be replaced with affinity matrices obtained directly from the data. As a first step into this direction, and in order to analyze its potential, we provide ananalysis of how the current state-of-the art semi-supervised approach, Personalized Propagation of Neural Predictions (PPNP), is affected by changes in the affinity matrix.