arXiv subjects

Djellel Difallah

Publications and source records attributed to Djellel Difallah.

5 recordsLinked to original sources

arXiv2026

MLaGA: Multimodal Large Language and Graph Assistant

Large Language Models (LLMs) have demonstrated substantial efficacy in advancing graph-structured data analysis. Prevailing LLM-based graph methods excel in adapting LLMs to text-rich graphs, wherein node attributes are text descriptions. However, their applications to multimodal graphs--where nodes are associated with diverse attribute types, such as texts and images--remain underexplored, despite their ubiquity in real-world scenarios. To bridge the gap, we introduce the Multimodal Large Language and Graph Assistant (MLaGA), an innovative model that adeptly extends LLM capabilities to facilitate reasoning over complex graph structures and multimodal attributes. We first design a structure-aware multimodal encoder to align textual and visual attributes within a unified space through a joint graph pre-training objective. Subsequently, we implement a multimodal instruction-tuning approach to seamlessly integrate multimodal features and graph structures into the LLM through lightweight projectors. Extensive experiments across multiple datasets demonstrate the effectiveness of MLaGA compared to leading baseline methods, achieving superior performance in diverse graph learning tasks under both supervised and transfer learning scenarios.

cs.AI↗

arXiv2026

UniST-Pred: A Robust Unified Framework for Spatio-Temporal Traffic Forecasting in Transportation Networks Under Disruptions

Spatio-temporal traffic forecasting is a core component of intelligent transportation systems, supporting various downstream tasks such as signal control and network-level traffic management. In real-world deployments, forecasting models must operate under structural and observational uncertainties, conditions that are rarely considered in model design. Recent approaches achieve strong short-term predictive performance by tightly coupling spatial and temporal modeling, often at the cost of increased complexity and limited modularity. In contrast, efficient time-series models capture long-range temporal dependencies without relying on explicit network structure. We propose UniST-Pred, a unified spatio-temporal forecasting framework that first decouples temporal modeling from spatial representation learning, then integrates both through adaptive representation-level fusion. To assess robustness of the proposed approach, we construct a dataset based on an agent-based, microscopic traffic simulator (MATSim) and evaluate UniST-Pred under severe network disconnection scenarios. Additionally, we benchmark UniST-Pred on standard traffic prediction datasets, demonstrating its competitive performance against existing well-established models despite a lightweight design. The results illustrate that UniST-Pred maintains strong predictive performance across both real-world and simulated datasets, while also yielding interpretable spatio-temporal representations under infrastructure disruptions. The source code and the generated dataset are available at https://anonymous.4open.science/r/UniST-Pred-EF27

cs.LG↗

arXiv2025

RAW-Explainer: Post-hoc Explanations of Graph Neural Networks on Knowledge Graphs

Graph neural networks have demonstrated state-of-the-art performance on knowledge graph tasks such as link prediction. However, interpreting GNN predictions remains a challenging open problem. While many GNN explainability methods have been proposed for node or graph-level tasks, approaches for generating explanations for link predictions in heterogeneous settings are limited. In this paper, we propose RAW-Explainer, a novel framework designed to generate connected, concise, and thus interpretable subgraph explanations for link prediction. Our method leverages the heterogeneous information in knowledge graphs to identify connected subgraphs that serve as patterns of factual explanation via a random walk objective. Unlike existing methods tailored to knowledge graphs, our approach employs a neural network to parameterize the explanation generation process, which significantly speeds up the production of collective explanations. Furthermore, RAW-Explainer is designed to overcome the distribution shift issue when evaluating the quality of an explanatory subgraph which is orders of magnitude smaller than the full graph, by proposing a robust evaluator that generalizes to the subgraph distribution. Extensive quantitative results on real-world knowledge graph datasets demonstrate that our approach strikes a balance between explanation quality and computational efficiency.

cs.LG↗

arXiv2021

A Multilingual Entity Linking System for Wikipedia with a Machine-in-the-Loop Approach

Hyperlinks constitute the backbone of the Web; they enable user navigation, information discovery, content ranking, and many other crucial services on the Internet. In particular, hyperlinks found within Wikipedia allow the readers to navigate from one page to another to expand their knowledge on a given subject of interest or to discover a new one. However, despite Wikipedia editors' efforts to add and maintain its content, the distribution of links remains sparse in many language editions. This paper introduces a machine-in-the-loop entity linking system that can comply with community guidelines for adding a link and aims at increasing link coverage in new pages and wiki-projects with low-resources. To tackle these challenges, we build a context and language agnostic entity linking model that combines data collected from millions of anchors found across wiki-projects, as well as billions of users' reading sessions. We develop an interactive recommendation interface that proposes candidate links to editors who can confirm, reject, or adapt the recommendation with the overall aim of providing a more accessible editing experience for newcomers through structured tasks. Our system's design choices were made in collaboration with members of several language communities. When the system is implemented as part of Wikipedia, its usage by volunteer editors will help us build a continuous evaluation dataset with active feedback. Our experimental results show that our link recommender can achieve a precision above 80% while ensuring a recall of at least 50% across 6 languages covering different sizes, continents, and families.

cs.CY↗

arXiv2019

Non-Parametric Class Completeness Estimators for Collaborative Knowledge Graphs -- The Case of Wikidata

Collaborative Knowledge Graph platforms allow humans and automated scripts to collaborate in creating, updating and interlinking entities and facts. To ensure both the completeness of the data as well as a uniform coverage of the different topics, it is crucial to identify underrepresented classes in the Knowledge Graph. In this paper, we tackle this problem by developing statistical techniques for class cardinality estimation in collaborative Knowledge Graph platforms. Our method is able to estimate the completeness of a class - as defined by a schema or ontology - hence can be used to answer questions such as "Does the knowledge base have a complete list of all {Beer Brands|Volcanos|Video Game Consoles}?" As a use-case, we focus on Wikidata, which poses unique challenges in terms of the size of its ontology, the number of users actively populating its graph, and its extremely dynamic nature. Our techniques are derived from species estimation and data-management methodologies, and are applied to the case of graphs and collaborative editing. In our empirical evaluation, we observe that i) the number and frequency of unique class instances drastically influence the performance of an estimator, ii) bursts of inserts cause some estimators to overestimate the true size of the class if they are not properly handled, and iii) one can effectively measure the convergence of a class towards its true size by considering the stability of an estimator against the number of available instances.

cs.DB↗