Coding

In this work, we introduce Directional Sheaf Hypergraph Networks (DSHN), a framework integrating sheaf theory with a principled treatment of asymmetric relations within a hypergraph. From it, we construct the Directed Sheaf Hypergraph Laplacian, a complex-valued operator by which we unify and generalize many existing Laplacian matrices proposed in the graphand hypergraph-learning literature.

Despite significant advances, Convolutional Neural Network (CNN)-based Sequential Recommender Systems are rapidly being overshadowed by the more performant attention-based models. In this paper, we present a novel modification of two widely used CNN-based SRSs, Caser and CosRec. We improve their training by adapting the convolution and pooling operations so that they can be trained simultaneously on the whole input sequence rather than just on the last element. Our experimental results show that these modified CNN-based models achieve up to +65% in NDCG@10 compared to their original versions.

In this work, we introduce GREEN (Guided Recommendations of Energy-Efficient Networks), a novel, inference-time approach for recommending Pareto-optimal AI model configurations that optimize validation performance and energy consumption across diverse AI domains and tasks. Our approach directly addresses the limitations of current eco-efficient model selection methods, which are often restricted to specific architectures or tasks. Central to this work is EcoTaskSet, a openly expandable dataset currently comprising training dynamics from over 1767 experiments across computer vision, natural language processing, and recommendation systems using both widely used and cutting-edge architectures.

In this work, we introduce a method to estimate label noise distribution using IAA statistics; we show how to use this estimate to train on noisy datasets;. We establish generalization bounds for our methods that rely on known quantities; we quantify the "distributional shift'" in the expected loss of a classifier between noisy and true-label distributions can be bounded by the spectral gap of the noise transition matrix and the class-prior matrix. Lastly, we provide experiments across different tasks to validate the proposed theory.

This work presents the conditions under which Majority Vote achieves the theoretically optimal lower bound on label estimation error. Our results capture the tolerable limits on annotation noise under which MV can optimally recover labels for a given class distribution. This certificate of optimality provides a more principled approach to model selection for label aggregation as an alternative to otherwise inefficient practices that sometimes include higher experts, gold labels, etc.,  that are all marred by the same human uncertainty despite huge time and monetary costs.

This work investigates the environmental impact of GNN-based recommender systems, an aspect that has been largely overlooked in the literature. Specifically, we conduct a comprehensive analysis of the carbon emissions associated with training and deploying GNN models for recommendation tasks. We evaluate the energy consumption and carbon footprint of different GNN architectures and configurations, considering factors such as model complexity, training duration, hardware specifica- tions and embedding size.

Our work tries to address the reproducibility problems in the domain of Sequential Recommendation Systems by standardising data pre-processing and model implementations, providing a comprehensive code resource, including a framework for developing SRSs and establishing a foundation for consistent and reproducible experimentation. We conduct extensive experiments on several benchmark datasets, comparing various SRSs implemented in our resource. We challenge prevailing performance benchmarks, offering new insights into the SR domain.
The proposed library can be used by research and practitioners to streamline the research in the field of Recommendation Systems. For details refer to the paper.

In this work, we propose a solution integrating a cutting-edge model inspired by category theory: Sheaf4Rec. Our approach takes advantage from sheaf theory and results in a more comprehensive representation that can be effectively exploited during inference. Our proposed model exhibits a noteworthy relative improvement of up to 8.53% on F1-Score@10 and an impressive increase of up to 11.29% on NDCG@10, outperforming existing state-of-the-art models such as NGCF, KGTORe and other recently developed GNN-based models. Sheaf4Rec shows remarkable improvements in terms of efficiency: we observe substantial runtime improvements ranging from 2.5% up to 37% when compared to other GNN-based competitor models.

All the code is written in Python and is based on Pytorch, Pytorch Geometric and the use of Wandb for logging purposes.

This is our approach to the task of identification of persuasion techniques in text, which is a subtask of the SemEval-2023 Task 3 on the multilingual detection of genre, framing, and persuasion techniques in online news. The subtask is multi-label at the paragraph level and the inventory considered by the organizers covers 23 persuasion techniques. 

Our solution is based on an ensemble of a variety of pre-trained language models fine-tuned on the propaganda dataset. 

The official evaluation shows our solution ranks 1st in English and attains high scores in all the other languages, i.e. French, German, Italian, Polish, and Russian. 

The aim of this project is to develop a safe navigation framework for the TIAGo robot moving in a human crowd. Our approach is based on the paper of Vulcano et al., where a sensor-based scheme is presented. This scheme consists of two modules, the Crowd Prediction and Motion Generation modules, which run sequentially during every sampling interval. Our setup is implemented in Python using ROS and to validate our implementation multiple experiments are performed on Gazebo in scenarios of different complexity.

Data augmentation techniques are used to increase the size and variability of training data for learning visual tasks. These techniques are well-known in computer vision and include rotation, cropping, scaling and other transformations to increase the size of a dataset.  However, no one has addressed modeling variations in the sensor domain. This paper proposes an automatic, physically-based, and straightforward augmentation pipeline to simulate, on real images, multiple effects which arise from non-ideal optics, such as spherical aberration, defocus, astigmatism, and coma. The introduction of these effects on a real dataset can improve the ability to perform multiple computer vision tasks on it. We validate this assumption on two popular computer vision tasks: object detection and semantic segmentation introducing sensor effects into the PASCAL VOC 2012 dataset. In the end, we show that these techniques can improve the performance of our models on the detection task while reaching very similar results on segmentation.