Abstract
Hyperdimensional Computing (HDC) developed by Kanerva is a computationalmodel for machine learning inspired by neuroscience. HDC exploitscharacteristics of biological neural systems such as high-dimensionality,randomness and a holographic representation of information to achieve a goodbalance between accuracy, efficiency and robustness. HDC models have alreadybeen proven to be useful in different learning applications, especially inresource-limited settings such as the increasingly popular Internet of Things(IoT). One class of learning tasks that is missing from the current body ofwork on HDC is graph classification. Graphs are among the most important formsof information representation, yet, to this day, HDC algorithms have not beenapplied to the graph learning problem in a general sense. Moreover, graphlearning in IoT and sensor networks, with limited compute capabilities,introduce challenges to the overall design methodology. In this paper, wepresent GraphHD$-$a baseline approach for graph classification with HDC. Weevaluate GraphHD on real-world graph classification problems. Our results showthat when compared to the state-of-the-art Graph Neural Networks (GNNs) theproposed model achieves comparable accuracy, while training and inference timesare on average 14.6$\times$ and 2.0$\times$ faster, respectively.