Graph Node Embeddings for ontology-aware Sound Event Classification: an evaluation study

Multi-label Sound Event Classification (SEC) is a challenging task which requires to handle multiple co-occurring sound event classes. Recent works proposed an ontology-aware framework for SEC in which a Graph-Neural Network (GNN) approach is trained to exploit labels co-occurrence information and improve the performance of a standard audio-feature based classifier via late-fusion. This GNN is fed a graph-based representation of the training set labels. In this paper we adopt such framework and perform an in-depth study on how the labels embeddings used to construct the graph representation can affect the performance. We perform our experiment on the FSD50K dataset and compare different embeddings strategies: two from previous works and two which haven't been considered yet for SEC applications. Our results show that node2vec embeddings lead to substantial performance improvements with respect to other embeddings strategies used in previously ontology-aware SEC works. Our best node2vec model leads to an absolute improvement of 3.39% in mean average precision with respect to the best competing embedding strategy, with a lower number of trainable parameters.