Multi-modal learning often struggles with incomplete alignment, modality-specific noise, and high-order inter-actions between heterogeneous inputs. This paper introduces HyperModal, a dynamic hypergraph-based model that learns instance-specific topologies and semantically coherent relation-ships across modalities. Each multi-modal input is represented as a hypergraph, where nodes are projected modality-specific feature units and hyperedges encode context-driven co-occurrence patterns. HyperModal integrates a dual-space contrastive objective that aligns within-modality consistency and cross-modality co-occurrence, enabling the model to learn modality-agnostic but semantically grounded representations. We evaluate the proposed framework on four diverse tasks drawn from the MultiBench benchmark suite, spanning classification, regression, and time-series forecasting. HyperModal outperforms strong baselines including Transformer-based and graph-based fusion models. The model demonstrates robust performance even under modality dropout and outperforms prior work in representation structure metrics. Extensive ablations confirm the value of dynamic topology construction and contrastive supervision. The proposed method offers a unified, generalizable solution to multi-modal fusion with promising implications for structured representation learning.

This paper proposes an adaptive hierarchical Growing Hierarchical Self-Organizing Map (GHSOM) integrated with Federated Learning (FL) for scalable and privacy-preserving anomaly detection in distributed Internet of Things (IoT) networks. In IoT environments, vast amounts of data are generated from heterogeneous sensors, requiring robust anomaly detection mechanisms to ensure system reliability. Traditional centralized approaches face significant challenges related to data privacy, communication overhead, and scalability. The proposed model addresses these issues by combining the local processing power of GHSOM with the privacy-preserving capabilities of FL, ensuring that only model updates, rather than raw data, are shared across the network. To further enhance efficiency, we introduce a novel adaptive hierarchy within the GHSOM structure. This hierarchy dynamically adjusts based on the context and characteristics of the incoming sensor data, such as its variance, frequency, and criticality. By embedding this context-awareness, the GHSOM model can tailor its clustering process to the specific requirements of each sensor node, improving anomaly detection accuracy. Additionally, FL is enhanced with a selective update mechanism, where only significant model updates—triggered by detected anomalies—are shared across the network, further reducing bandwidth usage and communication costs. We validate the proposed system on diverse IoT sensor datasets, demonstrating its superior ability to detect anomalies with reduced communication overhead compared to non-adaptive federated GHSOM approaches. This work highlights the potential of combining adaptive GHSOM hierarchies with federated learning to create scalable, privacy-preserving, and efficient anomaly detection systems suitable for real-time IoT applications.

This paper presents an innovative framework developed to identify, analyze, and generate memorable experiences in the hotel industry. People prefer memorable experiences over traditional services or products in today's ever-changing consumer world. As a result, the hospitality industry has shifted its focus toward creating unique and unforgettable experiences rather than just providing essential services. Despite the inherent subjectivity and difficulties in quantifying experiences, the quest to capture and understand these critical elements in the hospitality context has persisted. However, traditional methods have proven inadequate due to their reliance on objective surveys or limited social media data, resulting in a lack of diversity and potential bias. Our framework addresses these issues, offering a holistic solution that effectively identifies and extracts memorable experiences from online customer reviews, discerns trends on a monthly or yearly basis, and utilizes a local LLM to generate potential, unexplored experiences. As the first successfully deployed, fast, and accurate product of its kind in the industry, This framework significantly contributes to the hotel industry's efforts to enhance services and create compelling, personalized experiences for its customers.

This paper introduces HyperMAD, a novel Hypergraph Convolutional Network model designed for the multiclass classification of mental health advice in Arabic tweets. The model distinguishes between misleading and valid advice, further categorizing each tweet into specific classes of advice. HyperMAD leverages high-order relations between words in short texts, captured through the definition of four types of hyperedges that represent local and global contexts as well as semantic similarity. Extensive experiments demonstrate the effectiveness of HyperMAD, with results outperforming those from existing baselines. The study also includes an ablation study to investigate the significance and contribution of each hyperedge type. The paper presents a case study analyzing the accuracy and types of Arabic mental health advice on Twitter, revealing that about 9% of the advice in response to mental health expressions on Twitter was accurate in general. The paper concludes with the hope that the application of HyperMAD can be utilized in flagging misleading responses on social media, providing the correct resources for those who choose to share their mental health struggles online.

Edge prediction is a fundamental challenge in network science, with broad applications, notably in social networks. It plays a crucial role in unveiling complex system dynamics by forecasting connections between entities. Our paper introduces UniMHe (Unified Multi Hyperedge Prediction), a novel framework for predicting multiple hyperedges associated with each node using hypergraph representations. We present a case study focused on crime network analysis, where UniMHe reveals intricate patterns in criminal activities, including crime types, locations, and seasonal variations. Our research leverages extensive historical crime data encompassing geographical information, timestamps, points of interest, and crime categories. In an extensive evaluation, we benchmark UniMHe against state-of-the-art hypergraph deep learning techniques, highlighting its superior performance. These findings underscore the significance of UniMHe across various domains and problem-solving scenarios.

With the proliferation of bots on social media platforms, especially X, the need for effective and efficient bot detection mechanisms has never been more paramount. However, the inherent imbalance between the number of genuine users and bots presents a significant challenge, often leading to biased classifiers. In this paper, we introduce HyperSMOTE, a novel approach for imbalanced node classification leveraging the rich structure of hypergraphs. By representing X users as nodes and their interactions as hyperedges, we construct a hypergraph that captures the intricate relationships and interactions among users. This hypergraph-based representation allows for a more nuanced understanding of user behavior and interactions, providing a robust foundation for bot detection. HyperSMOTE addresses the class imbalance by generating synthetic bot accounts in the hypergraph, ensuring a balanced training dataset while preserving the hypergraph’s semantics. Our method significantly outperforms existing baselines across various metrics, demonstrating its efficacy. We further delve into the impact of different upsampling scales on classification performance, providing insights into the optimal configurations for HyperSMOTE.

The drug abuse epidemic has been on the rise in the past few years, particularly after the start of COVID-19 pandemic. Our preliminary observations on Reddit alone show that discussions on drugs from 2018 to 2020 increased between a range of 45% to 200%, and so has the number of unique users participating in those discussions. Existing efforts focused on utilizing social media to distinguish potential drug abuse chats from unharmful chats regardless of what drug is being abused. Others focused on understanding the trends and causes of drug abuse from social media. To this end, we introduce PRISTINE (opioid crisis detection on reddit), our work dynamically detects-and extracts evolving misleading drug names from Reddit comments using reinforced Dynamic Query Expansion (DQE) and constructs a textual Graph Convolutional Network with the aid of powerful pre-trained embeddings to detect which type of drug class a Reddit comment corresponds to. Further, we perform extensive experiments to investigate the effectiveness of our model.

One of the strongest indicators of a mental health crisis is how people interact with each other or express them-selves. Hence, social media is an ideal source to extract user-level information about the language used to express personal feelings. In the wake of the ever-increasing mental health crisis in the United States, it is imperative to analyze the general well-being of a population and investigate how their public social media posts can be used to detect different underlying mental health conditions. For that purpose, we propose a study that collects posts from “reddits” related to different mental health topics to detect the type of the post and the nature of the mental health issues that correlate to the post. The task of detecting mental health related issues indicates the mental health conditions connected to the posts. To achieve this, we develop a multi-task learning model that leverages, for each post, both the latent embedding space of words and topics for prediction with a message passing mechanism enabling the sharing of information for related tasks. We train the model through an active learning approach in order to tackle the lack of standardized fine-grained label data for this specific task.

Drug overdose deaths are a dreadful crisis that leads to substantial societal impairments. Its harmful impact directly affects families and communities. To assist policymakers in mitigating this crisis, it is crucial to study the societal, economic, and criminal contributing factors linked to the crisis. Unfortunately, current data-driven works assume a singular factor, such as poverty being the cause and disregarding other realistic causes. Besides, recent works exhibited a lack of explainable models and spatial analysis of the crisis. Thus, DOD-Explainer links the gap by developing a realistic framework that predicts highly impacted counties of drug overdose deaths from crime and socioeconomic data. DOD-Explainer overcomes the challenge of data scarcity by proposing three data augmentation methods. Then, an algorithm is proposed to provide realistic explanations of the leading causes of the crisis. The results show that our application achieves the best predictive accuracy from several models, accurately identifies the most/least impacted counties by the crisis, and reveals the most contributing factors of drug overdoses.

Social media platforms have become an easy method of communication for many users. Content posted on social media can influence those who are exposed to it, and users who posted that content are referred to as influencers. Identifying influencers has many applications in marketing, politics, and even health awareness. While research identifying influential users across multiple fields has been studied extensively, users’ influence varies in different topics. Recent studies in topic-specific influence have shown that identifying influencers on the topic-level is more effective. However, most of the existing influencer detection approaches focus only on influential user identification and do not consider that some content can be influential regardless of who published it. This paper investigates the problem of detecting topic-specific influential users and tweets in Twitter datasets. We introduce HyperTwitter, a framework that uses a Twitter sub-graph consisting of users, tweets, and interactions as input. HyperTwitter generates a hypergraph with hyperedges of two types: networks and topic edges, then measures the topic distribution for both users and tweets. With this distribution and the constructed hypergraph, we create a local, topic-based influence ranking for each user and tweet. We conduct extensive experiments with two Twitter datasets and show that the proposed framework outperforms existing baselines significantly.

The vast presence of bots on Twitter requires reliable and accurate bot detection methods that differentiate legitimate bots from malicious ones. Despite the success of those methods, they fail to address the following challenges: (1) the huge size of datasets required to train a model to detect bots, (2) the constant evolution in bot accounts to evade automatic detection leads to scarcity in ground truth real-world datasets, and (3) the complexity in learning representations of a heterogeneous attributed network like Twitter. In this paper, we propose a novel framework, ADNET, to detect anomalies in Twitter-attributed networks using the least amount of labeled data. Specifically, we address the limitations of previous methods by proposing a topology-based active learning framework that uses a deep autoencoder to train the model and is able to handle large graphs better than previous methods. Our experimental results demonstrate that the proposed approach outperforms state-of-the-art methods in detecting anomalous bot accounts and reduces the annotation cost in Twitter attributed networks.

Existing COVID-19 prediction models focus on studying the dynamic nature of the virus spread by using pandemic-related temporal data. In this paper, we present a work that exclusively uses comprehensive socioeconomic factors to predict the high risk areas of COVID-19 infection based on fine-grained static spatial analysis. Moreover, the most and least influential socioeconomic factors on COVID-19 spread are identified. This paper uses a uniquely built dataset by combining local states’ cumulative COVID-19 statistics and their associated socioeconomic features on the zip code level. Further, the work solves the lack of data by augmentation. To evaluate the work, four case studies are conducted on Florida, Illinois, Minnesota, and Virginia. Experimental results show that the study provides accurate predictions with respect to ground truth data. By identifying high risk areas and socioeconomic factors, policymakers can use this study to take necessary measures to help disadvantaged communities.