Simone Antonelli

Despite the widespread use of graph neural networks (GNNs) we lack methods to reliably quantify their uncertainty. We propose a conformal procedure to equip GNNs with prediction sets that come with distribution-free guarantees – the output set contains the true label with arbitrarily high probability. Our post-processing procedure can wrap around any (pretrained) GNN, and unlike existing methods, results in meaningful sets even when the model provides only the top class. The key idea is to diffuse the node-wise conformity scores to incorporate neighborhood information. By leveraging the network homophily we construct sets with comparable or better efficiency (average size) and significantly improved singleton hit ratio (correct sets of size one). In addition to an extensive empirical evaluation, we investigate the theoretical conditions under which smoothing provably improves efficiency.

Few-shot graph classification is a novel yet promising emerging research field that still lacks the soundness of well-established research domains. Existing works often consider different benchmarks and evaluation settings, hindering comparison and, therefore, scientific progress. In this work, we start by providing an extensive overview of the possible approaches to solving the task, comparing the current state-of-the-art and baselines via a unified evaluation framework. Our findings show that while graph-tailored approaches have a clear edge on some distributions, easily adapted few-shot learning methods generally perform better. In fact, we show that it is sufficient to equip a simple metric learning baseline with a state-of-the-art graph embedder to obtain the best overall results. We then show that straightforward additions at the latent level lead to substantial improvements by introducing i) a task-conditioned embedding space ii) a MixUp-based data augmentation technique. Finally, we release a highly reusable codebase to foster research in the field, offering modular and extensible implementations of all the relevant techniques.

Deep learning approaches have recently raised the bar in many fields, from Natural Language Processing to Computer Vision, by leveraging large amounts of data. However, they could fail when the retrieved information is not enough to fit the vast number of parameters, frequently resulting in overfitting and therefore in poor generalizability. Few-Shot Learning aims at designing models that can effectively operate in a scarce data regime, yielding learning strategies that only need few supervised examples to be trained. These procedures are of both practical and theoretical importance, as they are crucial for many real-life scenarios in which data is either costly or even impossible to retrieve. Moreover, they bridge the distance between current data-hungry models and human-like generalization capability. Computer vision offers various tasks that can be few-shot inherent, such as person re-identification. This survey, which to the best of our knowledge is the first tackling this problem, is focused on Few-Shot Object Detection, which has received far less attention compared to Few-Shot Classification due to the intrinsic challenge level. In this regard, this review presents an extensive description of the approaches that have been tested in the current literature, discussing their pros and cons, and classifying them according to a rigorous taxonomy.