Reliable Uncertainty Quantification (UQ) and failure prediction remain open challenges for Vision-Language Models (VLMs). We introduce ViLU, a new Vision-Language Uncertainty quantification framework that contextualizes uncertainty estimates by leveraging all task-relevant textual representations. ViLU constructs an uncertainty-aware multi-modal representation by integrating the visual embedding, the predicted textual embedding, and an image-conditioned textual representation via cross-attention. Unlike traditional UQ methods based on loss prediction, ViLU trains an uncertainty predictor as a binary classifier to distinguish correct from incorrect predictions using a weighted binary cross-entropy loss, making it loss-agnostic. In particular, our proposed approach is well-suited for post-hoc settings, where only vision and text embeddings are available without direct access to the model itself. Extensive experiments on diverse datasets show the significant gains of our method compared to state-of-the-art failure prediction methods. We apply our method to standard classification datasets, such as ImageNet-1k, as well as large-scale image-caption datasets like CC12M and LAION-400M. Ablation studies highlight the critical role of our architecture and training in achieving effective uncertainty quantification. Our code is publicly available and can be found here: https://github.com/ykrmm/ViLU.

Fully connected Graph Transformers (GT) have rapidly become prominent in the static graph community as an alternative to Message-Passing models, which suffer from a lack of expressivity, oversquashing, and under-reaching. However, in a dynamic context, by interconnecting all nodes at multiple snapshots with self-attention,GT loose both structural and temporal information. In this work, we introduce Supra-LAplacian encoding for spatio-temporal TransformErs (SLATE), a new spatio-temporal encoding to leverage the GT architecture while keeping spatio-temporal information. Specifically, we transform Discrete Time Dynamic Graphs into multi-layer graphs and take advantage of the spectral properties of their associated supra-Laplacian matrix. Our second contribution explicitly model nodes' pairwise relationships with a cross-attention mechanism, providing an accurate edge representation for dynamic link prediction. SLATE outperforms numerous state-of-the-art methods based on Message-Passing Graph Neural Networks combined with recurrent models (e.g, LSTM), and Dynamic Graph Transformers, on~9 datasets. Code is open-source and available at this link https://github.com/ykrmm/SLATE.

Reasoning on inconsistent and uncertain data is challenging, especially for Knowledge-Graphs (KG) to abide temporal consistency. Our goal is to enhance inference with more general time interval semantics that specify their validity, as regularly found in historical sciences. We propose a new Temporal Markov Logic Networks (TMLN) model which extends the Markov Logic Networks (MLN) model with uncertain temporal facts and rules. Total and partial temporal (in)consistency relations between sets of temporal formulae are examined. We then propose a new Temporal Parametric Semantics (TPS) which allows combining several sub-functions leading to different assessment strategies. Finally, we present the new NeoMaPy tool, to compute the MAP inference on MLNs and TMLNs with several TPS. We compare our performances with state-of-the-art inference tools and exhibit faster and higher quality results.

Markov Logic Networks (MLN) are used for reasoning on uncertain and inconsistent temporal data. We proposed the TMLN (Temporal Markov Logic Network) which extends them with sorts/types, weights on rules and facts, and various temporal consistencies. The NeoMaPy framework integrates it as a knowledge graph based on conflict graphs which offers flexibility for reasoning with parametric Maximum A Posteriori (MAP) inferences, efficiency with an optimistic heuristic and interactive graph visualization for results explanation.

The Heterogeneous Information Network (HIN) formalism is very flexible and enables complex recommendations models. We evaluate the effect of different parts of a HIN on the accuracy and the diversity of recommendations, then investigate if these effects are only due to the semantic content encoded in the network. We use recently-proposed diversity measures which are based on the network structure and better suited to the HIN formalism. Finally, we randomly shuffle the edges of some parts of the HIN, to empty the network from its semantic content, while leaving its structure relatively unaffected. We show that the semantic content encoded in the network data has a limited importance for the performance of a recommender system and that structure is crucial.

We consider the emerging field of music score libraries, where documents rely on a music notation markup language such as MusicXML or MEI. We propose to model as synchronized time-series the music structure that can be extracted from such documents, together with an algebra that operates in closed form and allows manipulations, restructurings, and combinations of music scores stored in a database. We formally present the model, its algebraic operators, and finally show how our approach can serve as a building block for a query and analytic language on large collections of XML-encoded music scores.

Peer-to-peer (P2P) networks are popular to exchange large volumes of data through the Internet. Pedophile activity is a very important topic for our society and some works have recently attempted to gauge the extent of pedophile exchanges on P2P networks. A key issue is to obtain an efficient detection tool, which may decide if a sequence of keywords is related to the topic or not. We propose to use social network analysis in a large dataset from a P2P network to improve a state-of-the-art filter for pedophile queries. We obtain queries and thus combinations of words which are not tagged by the filter but should be. We also perform some experiments to explore if the original four categories of paedophile queries were to be found by topological measures only.

Increasing knowledge of paedophile activity in P2P systems is a crucial societal concern, with important consequences on child protection, policy making, and internet regulation. Because of a lack of traces of P2P exchanges and rigorous analysis methodology, however, current knowledge of this activity remains very limited. We consider here a widely used P2P system, eDonkey, and focus on two key statistics: the fraction of paedophile queries entered in the system and the fraction of users who entered such queries. We collect hundreds of millions of keyword-based queries; we design a paedophile query detection tool for which we establish false positive and false negative rates using assessment by experts; with this tool and these rates, we then estimate the fraction of paedophile queries in our data. We conclude that approximately 0.25% of queries are paedophile. Our statistics1are by far the most precise and reliable ever obtained in this domain.