The loss function for Graph Neural Networks (GNNs) in graph classification tasks is defined as:

$\mathcal{L} = \sum_{\mathcal{G}_i \in \mathcal{T}} \| \text{MLP}(z_{\mathcal{G}_i}) - y_{\mathcal{G}_i} \|_2^2$

Where:

• $z_{\mathcal{G}_i}$ represents the graph-level embeddings.
• mathcal{T} = {mathcal{G}_1, dots, mathcal{G}_n} is the set of labeled training graphs.
• $\text{MLP}$ is a densely connected feed-forward neural network.
• $y_{\mathcal{G}_i} \in \mathbb{R}$ is the target value for the training graph $\mathcal{G}_i$.