The proliferation of harmful memes on online media poses significant risks to public health and stability. Existing detection methods heavily rely on large-scale labeled data for training, which necessitates substantial manual annotation efforts and limits their adaptability to the continually evolving nature of harmful content. To address these challenges, we present ALARM, the first lAbeL-free hARmful Meme detection framework powered by Large Multimodal Model (LMM) agent self-improvement. The core innovation of ALARM lies in exploiting the expressive information from "shallow" memes to iteratively enhance its ability to tackle more complex and subtle ones. ALARM consists of a novel Confidence-based Explicit Meme Identification mechanism that isolates the explicit memes from the original dataset and assigns them pseudo-labels. Besides, a new Pairwise Learning Guided Agent Self-Improvement paradigm is introduced, where the explicit memes are reorganized into contrastive pairs (positive vs. negative) to refine a learner LMM agent. This agent autonomously derives high-level detection cues from these pairs, which in turn empower the agent itself to handle complex and challenging memes effectively. Experiments on three diverse datasets demonstrate the superior performance and strong adaptability of ALARM to newly evolved memes. Notably, our method even outperforms label-driven methods. These results highlight the potential of label-free frameworks as a scalable and promising solution for adapting to novel forms and topics of harmful memes in dynamic online environments.

The inherent fluctuations in the stock market present significant challenges in understanding stock dynamics, especially for investment decisions based on stock ranking. Recent advancements in learning-based methods have led to promising results in exploring temporal dependencies to understand stock movements. However, they often assume stable, certain, and reliable environments, narrowing their insight into the complex and fluctuating nature of markets. This complexity is driven by two influential factors: the explicit consistency of dynamic yet stable trends across diverse temporal patterns, coupled with the implicit interplay of logic and possibility under uncertainty. Hence, we introduce aSensitivity-awareDependencyLearning solution (SDL) for stock ranking. With bridging the ideal and reality in mind, SDL captures short-term fluctuations under the guidance of long-term dependencies, associated with the augmentation of counterfactual knowledge. Specifically, SDL devises aShort-termCo-integrationDetector (SCD) that concentrates on capturing time-varying correlations and immediate market reactions, in addition to multi-period attention. Furthermore, aLong-termCo-movementsTracker (LCT) takes advantage of enduring industry relationships and incorporates counterfactual knowledge, allowing the model to generalize beyond observed patterns and identify diverse long-term trends. Comprehensive experiments on five real-world stock markets demonstrate that our proposed SDL outperforms several representative baselines.

High-quality embeddings for urban regions have enabled influential insights into urban structures and characteristics, facilitating the creation of more sustainable cities. However, the existing practices still face certain challenges, notably: (1) When multiple views contain distinct semantic information, ignoring the reliability and possibly inadequate collection differences (e.g., data missingness) among those views may degrade the representation robustness. (2) Consensus semantics extracted from different views are often fused in a simplistic manner, without considering the uniformity of embeddings (quality variations) and the complementarity between views. To address such challenges, we propose a novel Adaptive Risk-aware Consensus learning (ARC) solution for urban region embeddings. Specifically, we design both local- and region-level masking within the inter-view representation, following the paradigm of masked autoencoders, to better handle uncertainty risks. More importantly, we introduce a self-weighted contrastive mechanism in consensus learning to achieve maximum alignment and mitigate degradation. To enhance the uniformity of embeddings, we employ entropy, ensuring the diversity and complementarity of information. Ultimately, we apply the learned embeddings to down-stream tasks, demonstrating remarkable improvements compared to several representative baselines.

The automatic and accurate detection of online fake news is crucial to society, drawing significant attention from both industry and academia. With news content becoming increasingly multimodal, assessing its truthfulness has become more challenging. Existing efforts to combat multimodal fake news primarily follow a target-egocentric paradigm, which makes predictions based solely on features extracted from the target news and its associated social context. However, their performance is constrained by the inherent knowledge paucity within the target news. To address this challenge, we propose ReTIP, a novel retrieval-enhanced framework for multimodal fake news detection. ReTIP enriches the knowledge of target news by retrieving relevant news content, along with potential diffusion participants. Specifically, ReTIP retrieves relevant content from a local content pool using a key vector generated through the joint modeling of text and images, and employs a communitybased strategy to retrieve potential participants from a historical user interaction pool. Additionally, ReTIP employs a hypergraphbased information enhancement module to align knowledge across modalities and instances at a fine-grained level by capturing higher-order correlations. Finally, an attention-based fusion layer is employed to aggregate the multi-element knowledge from retrieved instances, which is then concatenated with the target news knowledge for the final prediction. Extensive experiments on three real-world multimodal fake news datasets not only demonstrate the superior performance of ReTIP compared to state-of-the-art baselines but also confirm the effectiveness of its individual components. Our code is made publicly available at https://github.com/xytitor/ReTIP.

Optimizing stock selection through stock ranking is one of the critical but intricate tasks in quantitative trading areas because of the non-stationary dynamics and complicated interdependencies behind stock markets. Recent studies have made efforts to model historical market movements to enhance stock selection. However, they primarily borrowed the spirit of time series modeling and sought to build a deterministic paradigm without considering the uncertain fluctuations. In addition, some of these studies tailor to explore stock correlations from a predefined (e.g., binary) graph structure and use explicitly simple relations (such as first-order relations) to guide evolving interactions. Nevertheless, aggregating predefined but shallow relationships to collaborate with stock movements may affect selection generalizability and increase the risk of portfolio failure. This study introduces a novelRelational stock selection framework via probabilisticStateSpaceLearning (orRSSL) for stock selection. Specifically, RSSL first attempts to build a tree-based structure to explicitly expose higher-order relations in the stock market, primarily by discovering a hierarchical delineation of ties between stocks. Whereafter, it couples with time-varying movements via an attention mechanism to smoothly explore the interactive correlations among different stocks. Inspired by recent state space models (SSM) in probabilistic Bayesian learning, we devise a Probabilistic Kalman Network (PKNet) with uncertainty estimates to recursively simulate ever-changing stock volatility, enabling more promising return-risk trade-offs. The experimental results on several real-world stock market datasets demonstrate that RSSL outperforms several representative baseline methods by a significant margin.

Graph neural networks, as well as attention mechanisms, have gained widespread popularity for traffic flow forecasting due to their capacity to incorporate the complicated interactions behind flow dynamics. However, existing solutions either formulate a graph-based skeleton with narrow (e.g., static) interaction capture or build the spatiotemporal (e.g., dynamic) attention without proper comprehension of diverse risks, which inevitably burdens the generalization of high-accuracy traffic trends. In this study, we introduce Gboot (Graph bootstrap) enhancement framework for traffic flow forecasting. Gboot takes the traffic flow forecasting problem from a dependency dynamic learning perspective by treating each traffic sensor as the graph node while regarding the observed flows at each sensor as the node feature. In addition to exposing the explicit spatial connectivity behind traffic flows, we hierarchically devise temporal-aware and factual-aware graph learning blocks to consider temporal interactive dynamics and factual interactive dynamics. The former shows the trend dependencies behind flow signals and the latter uncovers different views of traffic situations (e.g., current observation vs. historical observation). More importantly, we present a Dual-view Bootstrap (DvBoot) mechanism in Gboot, which includes both risk-free and risk-aware stands. DvBoot attempts to flexibly align these two views in the latent space to enhance the generalization capability of capturing dynamic dependencies. Experiments on several real-world traffic datasets demonstrate the superiority of our Gboot over representative approaches.

Point-of-Interest (POI) recommendation involves predicting users' next preferred POI and is becoming increasingly significant in location-based social networks. However, users are often reluctant to trust recommended results due to the lack of transparency in these systems. While recent work on explaining recommender systems has gained attention, prevailing methods only provide post-hoc explanations based on results or rudimentary explanations according to attention scores. Such limitations hinder reliability and applicability in risk-sensitive scenarios. Inspired by the information theory, we propose a self-explainable framework with an ante-hoc view called \M~for next POI recommendation aimed at overcoming these limitations. Specifically, we endow self-explainability to POI recommender systems through compact representation learning using a variational information bottleneck approach. The learned representation further improves accuracy by reducing redundancy behind massive spatial-temporal trajectories, which, in turn, boosts the recommendation performance. Experiments on three real-world datasets show significant improvements in both model explainability and recommendation performance.

Exploring and modeling the spreading process of rumors have shown great potential in improving rumor detection performance. However, existing propagation‐based rumor detection models often overlook the uncertainty of the underlying propagation structure and typically require a large amount of labeled data for training. To address these challenges, we propose a novel rumor detection framework, namely, the Uncertainty‐Inference Contrastive Learning (UICL) model. Specifically, UICL innovatively incorporates an edge‐wise augmentation strategy into the general contrastive learning framework, including an edge‐inference augmentation component and an EdgeDrop augmentation component, which primarily aim to capture the edge uncertainty of the propagation structure and alleviate the sparsity problem of the original dataset. A new negative sampling strategy is also introduced to enhance contrastive learning on rumor propagation graphs. Furthermore, we use labeled data to fine‐tune the detection module. Our experiments, conducted on three real‐world datasets, demonstrate that UICL can not only significantly improve detection accuracy but also reduce the dependency on labeled data compared to state‐of‐the‐art baselines.

Understanding human mobility has become an important aspect of location-based services in tasks such as personalized recommendation and individual moving pattern recognition, enabled by the large volumes of data from geo-tagged social media (GTSM). Prior studies mainly focus on analyzing human historical footprints collected by GTSM and assuming the veracity of the data, which need not hold when some users are not willing to share their real footprints due to privacy concerns—thereby affecting reliability/authenticity. In this study, we address the problem of Inferring Real Mobility (IRMo) of users, from their unreliable historical traces. Tackling IRMo is a non-trivial task due to the: (1) sparsity of check-in data; (2) suspicious counterfeit check-in behaviors; and (3) unobserved dependencies in human trajectories. To address these issues, we develop a novel Graph-enhanced Attention model called IRMoGA , which attempts to capture underlying mobility patterns and check-in correlations by exploiting the unreliable spatio-temporal data. Specifically, we incorporate the attention mechanism (rather than solely relying on traditional recursive models) to understand the regularity of human mobility, while employing a graph neural network to understand the mutual interactions from human historical check-ins and leveraging prior knowledge to alleviate the inferring bias. Our experiments conducted on four real-world datasets demonstrate the superior performance of IRMoGA over several state-of-the-art baselines, e.g., up to 39.16% improvement regarding the Recall score on Foursquare.

Deep neural networks have recently achieved considerable improvements in learning human behavioral patterns and individual preferences from massive spatial-temporal trajectory data. However, most of the existing research concentrates on fusing different semantics underlying sequential trajectories for mobility pattern learning which, in turn, yields a narrow perspective on comprehending human intrinsic motions. In addition, the inherent sparsity and under-explored heterogeneous collaborative items pertaining to human check-ins hinder the potential exploitation of human diverse periodic regularities as well as common interests. Motivated by recent advances in disentanglement learning, we propose a novel disentangled solution called SSDL for tackling the next POI prediction problem. SSDL primarily seeks to disentangle the potential time-invariant and time-varying factors into different latent spaces from massive trajectories, providing an interpretable view to understand the intricate semantics underlying human diverse mobility representations. To address the data sparsity issue, we present two realistic trajectory augmentation approaches to enhance the understanding of both the human intrinsic periodicity/habits and constantly-changing intents. In addition, we devise a POI-centric graph structure to explore heterogeneous collaborative signals underlying historical check-ins. Extensive experiments conducted on four real-world datasets demonstrate that SSDL significantly outperforms the state-of-the-art approaches–for example, it yields up to 8.57% averaged improvement on ACC@1.

Fraud detection poses a fundamental yet challenging problem to mitigate various risks associated with fraudulent activities. However, existing methods are limited by their reliance on static data within single geographical regions, thereby restricting the trained model’s adaptability across different regions. Practically, when enterprises expand their business into new cities or countries, training a new model from scratch can incur high computational costs and lead to catastrophic forgetting (CF). To address these limitations, we propose cross-regional fraud detection as an incremental learning problem, enabling the development of a unified model capable of adapting across diverse regions without suffering from CF. Subsequently, we introduce Cross-Regional Continual Learning (CCL), a novel paradigm that facilitates knowledge transfer and maintains performance when incrementally training models from previously learned regions to new ones. Specifically, CCL utilizes prototype-based knowledge replay for effective knowledge transfer while implementing a parameter smoothing mechanism to alleviate forgetting. Furthermore, we construct heterogeneous trade graphs (HTGs) and leverage graph-based backbones to enhance knowledge representation and facilitate knowledge transfer by uncovering intricate semantics inherent in cross-regional datasets. Extensive experiments demonstrate the superiority of our proposed method over baseline approaches and its substantial improvement in cross-regional fraud detection performance.

Learning classification models from real-world data often requires substantial human effort devoted to instance annotation. As the instance-based annotating process can be very time-consuming and costly, we propose a novel active learning framework that builds classification models from human-annotatedregions. A region is defined by a set of conjunctive patterns that are formed by value ranges over the input features. A region label is a human assessment of the classproportionin the data population covered by the region. By leveraginglearning from label proportionsalgorithms, regions and their class proportions can be used to train instance-based classification models. However, the key challenge is that in practice, very few regions are defined already. Therefore, to identify regions important for model learning, we design ahierarchical active learning(HAL) framework, which actively builds a hierarchy of regions. Similar to the decision-tree learning process, our approach progressively divides the input data space into smaller sub-regions, solicits labels for the new regions, and retrains the base classification model with all the leaf regions. And we further develop amulti-hierarchy(forest) solution, which builds multiple shallower hierarchies that have more informative, diverse, and simpler regions. We evaluate our HAL framework on numerous impactful classification datasets as well as on a real user study - on the survival analysis of colorectal cancer patients. The results demonstrate that region-based active learning methods can learn high-quality classifiers from very few labeled regions. Hence, our framework is shown very effective in reducing the human annotation effort needed for building classification models.

Continual text classification is an important research direction in Web mining. Existing works are limited to supervised approaches relying on abundant labeled data, but in the open and dynamic environment of Internet, involving constant semantic change of known topics and the appearance of unknown topics, text annotations are hard to access in time for each period. That calls for the technique of weakly supervised text classification (WSTC), which requires just seed words for each category and has succeed in static text classification tasks. However, there are still no studies of applying WSTC methods in a continual learning paradigm to actually accommodate the open and evolving Internet. In this paper, we tackle this problem for the first time and propose a framework, named Continual Learning for Weakly Supervised Text Classification (CL-WSTC), which can take any WSTC method as base model. It consists of two modules, classification decision with delay and seed word updating. In the former, the probability threshold for each category in each period is adaptively learned to determine the acceptance/rejection of texts. In the latter, with candidate words output by the base model, seed words are added and deleted via reinforcement learning with immediate rewards, according to an empirically certified unsupervised measure. Extensive experiments show that our approach has strong universality and can achieve a better trade-off between classification accuracy and decision timeliness compared to non-continual counterparts, with intuitively interpretable updating of seed words.

In this work we propose DeepTrip -- an end-to-end method for better understanding of the underlying human mobility and improved modeling of the POIs' transitional distribution in human moving patterns. DeepTrip consists of: a Trip Encoder to embed a given route into a latent variable with a recurrent neural network (RNN); and a Trip Decoder to reconstruct this route conditioned on an optimized latent space. Simultaneously, we define an Adversarial Net composed of a generator and critic, which generates a representation for a given query and uses a critic to distinguish the trip representation generated from Trip Encoder and query representation obtained from Adversarial Net. DeepTrip enables regularizing the latent space and generalizing users' complex check-in preference. We demonstrate the effectiveness and efficiency of the proposed model, and the experimental evaluations show that DeepTrip outperforms the state-of-the-art baselines on various evaluation metrics.

An important task in Location based Social Network applications is to predict mobility - specifically, user's next point-of-interest (POI) - challenging due to the implicit feedback of footprints, sparsity of generated check-ins, and the joint impact of historical periodicity and recent check-ins. Motivated by recent success of deep variational inference, we propose VANext (Variational Attention based Next) POI prediction: a latent variable model for inferring user's next footprint, with historical mobility attention. The variational encoding captures latent features of recent mobility, followed by searching the similar historical trajectories for periodical patterns. A trajectory convolutional network is then used to learn historical mobility, significantly improving the efficiency over often used recurrent networks. A novel variational attention mechanism is proposed to exploit the periodicity of historical mobility patterns, combined with recent check-in preference to predict next POIs. We also implement a semi-supervised variant - VANext-S, which relies on variational encoding for pre-training all current trajectories in an unsupervised manner, and uses the latent variables to initialize the current trajectory learning. Experiments conducted on real-world datasets demonstrate that VANext and VANext-S outperform the state-of-the-art human mobility prediction models.

Learning explicit and implicit patterns in human trajectories plays an important role in many Location-Based Social Networks (LBSNs) applications, such as trajectory classification (e.g., walking, driving, etc.), trajectory-user linking, friend recommendation, etc. A particular problem that has attracted much attention recently - and is the focus of our work - is the Trajectory-based Social Circle Inference (TSCI), aiming at inferring user social circles (mainly social friendship) based on motion trajectories and without any explicit social networked information. Existing approaches addressing TSCI lack satisfactory results due to the challenges related to data sparsity, accessibility and model efficiency. Motivated by the recent success of machine learning in trajectory mining, in this paper we formulate TSCI as a novel multi-label classification problem and develop a Recurrent Neural Network (RNN)-based framework called DeepTSCI to use human mobility patterns for inferring corresponding social circles. We propose three methods to learn the latent representations of trajectories, based on: (1) bidirectional Long Short-Term Memory (LSTM); (2) Autoencoder; and (3) Variational autoencoder. Experiments conducted on real-world datasets demonstrate that our proposed methods perform well and achieve significant improvement in terms of macro-R, macro-F1 and accuracy when compared to baselines.