Knowledge Graphs (KGs) and heterogeneous graphs (HGs) offer a principled way to represent multi-entity, multi-relational systems, while also revealing a persistent tension between expressive modeling, scalable learning, and faithful reasoning.Two trends are rapidly reshaping the field: graph foundation models (GFMs), which seek transfer across graphs, tasks, and domains via large-scale pretraining, and the growing integration of large language models (LLMs) with graph-structured knowledge to improve grounding, interaction, and reasoning.Temporal settings add further challenges, as evolving facts and interactions demand time-consistent modeling and evaluation.This tutorial provides a structured survey of these directions: we introduce a unified background and notation for typed heterogeneous graphs, (temporal) KGs, and event-based temporal heterogeneous graphs; we then formalize the main task families (KG completion, query answering, node/graph prediction, and temporal variants), emphasizing evaluation protocols and leakage pitfalls.Finally, we review recent advances in GFMs and LLM-graph integration, and summarize the state of the art in learning over temporal heterogeneous graphs and temporal KGs.

The convergence of network science and artificial intelligence (AI) represents a rich area of research, where both fields can mutually enhance one another.Network science offers a comprehensive framework to analyze and model complex relationships, while machine learning (ML) and AI provide powerful tools for recognizing patterns and making predictions from large datasets.Combining these two disciplines can advance the study of complex systems and lead to new innovations in data-driven research.This tutorial paper reviews fundamental concepts of network science, describes the current and promising research direction for bridging network science and AI, and summarizes the contributions that have been accepted for publication in the ESANN 2025 special session on the topic.

Pre-trained vision-language models (VLMs), such as CLIP, demonstrate impressive zero-shot classification capabilities with free-form prompts and even show some generalization in specialized domains. However, their performance on satellite imagery is limited due to the underrepresentation of such data in their training sets, which predominantly consist of ground-level images. Existing prompting techniques for satellite imagery are often restricted to generic phrases like “a satellite image of …”, limiting their effectiveness for zero-shot land-use/land-cover (LULC) mapping. To address these challenges, we introduce SenCLIP, which transfers CLIP's representation to Sentinel-2 imagery by leveraging a large dataset of Sentinel-2 images paired with geotagged ground-level photos from across Europe. We evaluate SenCLIP alongside other state-of-the-art remote sensing VLMs on zero-shot LULC mapping tasks using the EuroSAT and BigEarthNet datasets with both aerial and ground-level prompting styles. Our approach, which aligns ground-level representations with satellite imagery, demonstrates significant improvements in classification accuracy across both prompt styles, opening new possibilities for applying free-form textual descriptions in zero-shot LULC mapping. Code, dataset and pretrained models are available at https://github.com/pallavijain-pj/SenCLIP

In many scientific fields, complex systems are characterized by a multitude of heterogeneous interactions/relationships that are challenging to model. Multilayer graphs constitute valuable tools that can represent such complex systems, thus making possible their analysis for downstream decision-making processes. Nevertheless, modeling such complex information still remains challenging in real-world scenarios. On the one hand, holistically including all relationships may lead to noisy or computationally intensive graphs. On the other hand, limiting the amount of information to model through the selection of a portion of the available relationships can introduce boundary specification biases. However, the current research studies are demonstrating that it is more beneficial to retain as much information as possible and at a later stage perform graph simplification i.e., removing uninformative or redundant parts of the graph to facilitate the final analysis. While simplification strategies, based on deep learning methods, have been already extensively explored in the context of single-layer graphs, only a limited amount of efforts have been devoted to simplification strategies for multilayer graphs. In this work, we propose the MultilAyer gRaph simplificAtion ( MARA ) framework, a GNN-based approach designed to simplify multilayer graphs based on the downstream task. MARA generates node embeddings for a specific task by training jointly two main components: (i) an edge simplification module and (ii) a (multilayer) graph neural network. We tested MARA on different real-world multilayer graphs for node classification tasks. Experimental results show the effectiveness of the proposed approach: MARA reduces the dimension of the input graph while keeping and even improving the performance of node classification tasks in different domains and across graphs characterized by different structures. Moreover, deep learning-based simplification allows MARA to preserve and enhance important graph properties for the downstream task. To our knowledge, MARA represents the first simplification framework especially tailored for multilayer graphs analysis.

Geographic bias in language models (LMs) is an underexplored dimension of model fairness, despite growing attention being given to other social biases. We investigate whether LMs provide equally accurate representations across all global regions and propose a benchmark of four indicators to detect undertrained and underperforming areas: (i) indirect assessment of geographic training data coverage via tokenizer analysis, (ii) evaluation of basic geographic knowledge, (iii) detection of geographic distortions, and (iv) visualization of performance disparities through maps. Applying this framework to ten widely used encoder- and decoder-based models, we find systematic overrepresentation of Western countries and consistent underrepresentation of several African, Eastern European, and Middle Eastern regions, leading to measurable performance gaps. We further analyse the impact of these biases on downstream tasks, particularly in crisis response, and show that regions most vulnerable to natural disasters are often those with poorer LM coverage. Our findings underscore the need for geographically balanced LMs to ensure equitable and effective global applications.

A major difference between ground-level and satellite imagery of landscapes lies in their semantic granularity: ground-level images tend to offer details on objects and human activities, while satellite images provide broader geographic context but, typically, with coarser semantics. This study aims to leverage this complementary information by integrating fine-grained insights from a ground-level view into the analysis of satellite image data. To achieve this integration, we propose to align a satellite image representation with co-located geo-tagged ground-level image CLIP representations. This method focuses on enriching satellite image visual features by leveraging the inherent visual characteristics found in ground-level images as a reference in a contrastive manner, without relying on additional textual information to guide the learning process. We evaluate the quality of the learned representations on the EuroSAT benchmark in various few-shot settings.

With the advent of the Sentinel-2 mission and its high revisit frequency, high-resolution time series of optical images, the use of satellite image time series for automatic land cover mapping has fostered. However, one of the main limitations related to this kind of imagery is the presence of clouds, which often hinders its descriptive potential by reducing the actual temporal resolution. Although some common practices exist to enable their use in land cover processing chains, the majority of them aims at reconstructing the time series upstream to the classification task, hence introducing a heavy, error-prone pre-processing step. With the aim of exploiting the capacity of deep learning networks to adaptively combine tasks, in this preliminary study we propose an end-to-end framework that simultaneously perform cloud removal and classification of a Sentinel-2 image time series for the downstream task of land cover mapping. The proposed framework is evaluated over an agricultural area in Northern Benin. Our first results show comparable performances with respect to using state-of-the-art gap filling pre-processing on Sentinel-2 time series, hence motivating further exploration.

Event based disease surveillance (EBS) systems are biosurveillance systems that have the ability to detect and alert on (re)-emerging infectious diseases by monitoring acute public or animal health event patterns from sources such as blogs, online news reports and curated expert accounts. These information rich sources, however, are largely unstructured text data requiring novel text mining techniques to achieve EBS goals such as epidemiological text classification. The main objective of this research was to improve epidemiological text classification by proposing a novel technique of enriching thematic features using a weak supervision approach. In our approach, we train and test a mixed domain language model named EpidBioELECTRA to first enrich thematic features which are then used to improve epidemiological text classification. We train EpidBioELECTRA on a large dataset which we create consisting of 70,700 annotated documents that includes 70,400 labelled thematic features. We empirically compare EpidBioELECTRA with both general purpose language models and domain specific language models in the task of epidemiological corpus classification. Our findings shows that epidemiological classification systems work best with language models pre-trained using both epidemiological and biomedical corpora with a continual pre-training strategy. EpidBioELECTRA improves epidemiological document classification by 19.2 F1 score points as compared to its vanilla implementation BioELECTRA. We observe this by the comparison of BioELECTRA verses EpidBioELECTRA on our most challenging dataset PADI-WebXL where our approach records 92.33 precision score, 94.62 recall score and 93.46 F1 score. We also experiment the impact of increasing context length of train documents in epidemiological document classification and found out that this improves the classification task by 7.79 F1 score points as recorded by EpidBioELECTRA’s performance. We also compute Almost Stochastic Order (ASO) scores to track EpidBioELECTRA’s statistical dominance. In addition, we carry out ablation studies on our proposed thematic feature enrichment approach using explainable AI techniques. We present explanations for the most critical thematic features and how they influence epidemiological classification task We found out that biomedical features (such as mentions of names of diseases and symptoms) are the most influential while spatio-temporal features (such as the mention of date of a given disease outbreak) are the least influential in epidemiological document classification. Our model can easily be extended to fit other domains.

Huge amount of data are nowadays produced by a large and disparate family of sensors, which typically measure multiple variables over time. Such rich information can be profitably organized as multivariate time-series. Collect enough labelled samples to set up supervised analysis for such kind of data is challenging while a reasonable assumption is to dispose of a limited background knowledge that can be injected in the analysis process. In this context, semi-supervised clustering methods represent a well suited tool to get the most out of such reduced amount of knowledge. With the aim to deal with multivariate time-series analysis under a limited background knowledge setting, we propose a semi-supervised (constrained) deep embedding time-series clustering framework that exploits knowledge supervision modeled as Must- and Cannot-link constraints. More in detail, our proposal, named conDetSEC (constrained Deep embedding time SEries Clustering), is based on Gated Recurrent Units (GRUs) with the aim to explicitly manage the temporal dimension associated to multi-variate time series data. conDetSEC implements a procedure in which an embedding generation step is combined with a clustering refinement step. Both steps exploit the small amount of available knowledge provided by Must- and Cannot-link constraints. More specifically, during the data embedding generation the constraints are used by jointly optimizing the network parameters via both unsupervised and semi-supervised tasks, while at the refinement step they are used in conjunction with the goal to stretch the embedding manifold towards the clustering centroids to recover a more clear cluster structure. Experimental evaluation on real-world benchmarks coming from diverse domains has highlighted the effectiveness of our proposal in comparison with state-of-the-art unsupervised and semi-supervised time-series clustering methods.

We present EpidBioBERT, a biosurveillance epidemiological document tagger for disease surveillance over PADI-Web system. Our model is trained on PADI-Web corpus which contains news articles on Animal Diseases Outbreak extracted from the web. We train a classifier to discriminate between relevant and irrelevant documents based on their epidemiological thematic feature content in preparation for further epidemiology information extraction. Our approach proposes a new way to perform epidemiological document classification by enriching epidemiological thematic features namely disease, host, location and date, which are used as inputs to our epidemiological document classifier. We adopt a pre-trained biomedical language model with a novel fine tuning approach that enriches these epidemiological thematic features. We find these thematic features rich enough to improve epidemiological document classification over a smaller data set than initially used in PADI-Web classifier. This improves the classifiers ability to avoid false positive alerts on disease surveillance systems. To further understand information encoded in EpidBioBERT, we experiment the impact of each epidemiology thematic feature on the classifier under ablation studies. We compare our biomedical pre-trained approach with a general language model based model finding that thematic feature embeddings pre-trained on general English documents are not rich enough for epidemiology classification task. Our model achieves an F1-score of 95.5% over an unseen test set, with an improvement of +5.5 points on F1-Score on the PADI-Web classifier with nearly half the training data set.

Radar and Optical Satellite Image Time Series (SITS) are sources of information that are commonly employed to monitor earth surfaces for tasks related to ecology, agriculture, mobility, land management planning and land cover monitoring. Many studies have been conducted using one of the two sources, but how to smartly combine the complementary information provided by radar and optical SITS is still an open challenge. In this context, we propose a new neural architecture for the combination of Sentinel-1 (S1) and Sentinel-2 (S2) imagery at object level, applied to a real-world land cover classification task. Experiments carried out on the Reunion Island, a overseas department of France in the Indian Ocean, demonstrate the significance of our proposal.

While microblogging-based Online Social Networks have become an attractive data source in emergency situations, overcoming information overload is still not trivial. We propose a framework which integrates natural language processing and clustering techniques in order to produce a ranking of relevant tweets based on their informativeness. Experiments on four Twitter collections in two languages (English and French) proved the significance of our approach.

With the growing complexity of scenarios relating to online social networks (OSNs), there is an emergence of effective models and methods for understanding the characteristics and dynamics of multiple interconnected types of user relations. Profiles on different OSNs belonging to the same user can be linked using the multilayer structure, opening to unprecedented opportunities for user behavior analysis in a complex system. In this paper, we leverage the importance of studying the dichotomy between information-producers (contributors) and information-consumers (lurkers), and their interplay over a multilayer network, in order to effectively analyze such different roles a user may take on different OSNs. In this respect, we address the novel problem of identification and characterization of opposite behaviors that users may alternately exhibit over multiple layers of a complex network. We propose the first ranking method for alternate lurker-contributor behaviors on a multilayer OSN, dubbed mlALCR. Performance of mlALCR has been assessed quantitatively as well as qualitatively, and comparatively against methods designed for ranking either contributors or lurkers, on four real-world multilayer networks. Empirical evidence shows the significance and uniqueness of mlALCR in being able to mine alternate lurker-contributor behaviors over different layer networks.

Research on influence maximization ofter has to cope with marketing needs relating to the propagation of information towards specific users. However, little attention has been paid to the fact that the success of an information diffusion campaign might depend not only on the number of the initial influencers to be detected but also on theirdiversityw.r.t. the target of the campaign. Our main hypothesis is that if we learn seeds that are not only capable of influencing but also are linked to more diverse (groups of) users, then the influence triggers will be diversified as well, and hence the target users will get higher chance of being engaged. Upon this intuition, we define a novel problem, namedDiversity-sensitive Targeted Influence Maximization (DTIM), which assumes to model user diversity by exploiting only topological information within a social graph. To the best of our knowledge, we are the first to bring the concept of topology-driven diversity into targeted IM problems, for which we define two alternative definitions. Accordingly, we propose approximate solutions of DTIM, which detect a size-$k$set of users that maximizes the diversity-sensitive capital objective function, for a given selection of target users. We evaluate our DTIM methods on a special case of user engagement in online social networks, which concerns users who are not actively involved in the community life. Experimental evaluation on real networks has demonstrated the meaningfulness of our approach, also highlighting the opportunity of further development of solutions for DTIM applications.

The problem of local community detection refers to the identification of a community starting from a query node and using limited information about the network structure. Existing methods for solving this problem however are not designed to deal with multilayer network models, which are becoming pervasive in many fields of science. In this work, we present the first method for local community detection in multilayer networks. Our method exploits both internal and external connectivity of the nodes in the community being constructed for a given seed, while accounting for different layer-specific topological information. Evaluation of the proposed method has been conducted on real-world multilayer networks.

The participation inequality phenomenon in online social networks between the niche of super contributors and the crowd of silent users, a.k.a. lurkers, has been witnessed in many domains. Within this view, understanding the role that lurkers take in the network is essential to develop innovative strategies to delurk them, i.e., to engage such users into a more active participation in the social network life. In this work, we leverage the boundary spanning theory to enhance our understanding of lurking behaviors, with the goal of improving the task of delurking in social networks. Assuming the availability of a global community structure, we first analyze how lurkers are related to users that take the role of bridges between different communities, unveiling insights into the bridging nature of lurkers and their tendency to acquire information from outside their own community. Moreover, based on a targeted influence maximization method designed for delurking, we also analyze how the learning of users that can best engage lurkers is related to the community structure. We found that the best users to engage lurkers belonging to any particular community, are more often found outside that community, and more specifically they are located in the adjacent communities.

Lurkers are silent members of a social network (SN) who gain benefit from others' information without significantly giving back to the community. The study of lurking behaviors in SNs is nonetheless important, since these users acquire knowledge from the community, and as such they are social capital holders. Within this view, a major goal is to delurk such users, i.e., to encourage them to more actively be involved in the SN. Despite delurking strategies have been conceptualized in social science and human-computer interaction research, no computational approach has been so far defined to turn lurkers into active participants in the SN. In this work we fill this gap by presenting a delurking-oriented targeted influence maximization problem under the linear threshold (LT) model. We define a novel objective function, in terms of the lurking scores associated with the nodes in the final active set, and we show it is monotone and submodular. We provide an approximate solution by developing a greedy algorithm, named DEvOTION, which computes a k- node set that maximizes the value of the delurking-capital-based objective function, for a given minimum lurking score threshold. Results on SN datasets of different sizes have demonstrated the significance of our delurking approach via LT-based targeted influence maximization.

Mining the silent members, also called lurkers, of an online community has been recognized as an important problem that accompanies the extensive use of social networks. Existing solutions to the ranking of lurkers can aid understanding the lurking behaviors in social networks, however they ignore any information concerning the time dimension. In this work we push forward research in lurker mining by providing an analysis of temporal aspects that aims to unveil the behavior of lurkers and their interrelations with other users. Our analysis builds upon four research questions, which encompass relations between lurkers and inactive users, relations between lurkers and active users, the responsiveness behavior of lurkers, and the evolution of lurking trends across time. Evaluation has been conducted on Flickr, FriendFeed and Instagram networks.

The massive presence of silent members in online communities, the so-called lurkers, has long attracted the attention of researchers in social science, cognitive psychology, and computer-human interaction. However, the study of lurking phenomena represents an unexplored opportunity of research in data mining, information retrieval and related fields. In this paper, we take a first step towards the formal specification and analysis of lurking in social networks. Particularly, focusing on the network topology, we address the new problem of lurker ranking and propose the first centrality methods specifically conceived for ranking lurkers in social networks. Using Twitter and FriendFeed as cases in point, our methods' performance was evaluated against data-driven rankings as well as existing centrality methods, including the classic PageRank and alpha-centrality. Empirical evidence has shown the significance of our lurker ranking approach, which substantially differs from other methods in effectively identifying and ranking lurkers.

An emerging trend in research on recommender systems is the design of methods capable of recommending packages instead of single items. The problem is challenging due to a variety of critical aspects, including context-based and user-provided constraints for the items constituting a package, but also the high sparsity and limited accessibility of the primary data used to solve the problem. Most existing works on the topic have focused on a specific application domain (e.g., travel package recommendation), thus often providing ad-hoc solutions that cannot be adapted to other domains. By contrast, in this paper we propose a versatile package recommendation approach that is substantially independent of the peculiarities of a particular application domain. A key aspect in our framework is the exploitation of prior knowledge on the content type models of the packages being generated that express what the users expect from the recommendation task. Packages are learned for each package model, while the recommendation stage is accomplished by performing a PageRank-style method personalized w.r.t. the target user's preferences, possibly including a limited budget. Our developed method has been tested on a TripAdvisor dataset and compared with a recently proposed method for learning composite recommendations.