Recent advances in generative recommenders adopt a two-stage paradigm: items are first tokenized into semantic IDs using a pretrained tokenizer, and then large language models (LLMs) are trained to generate the next item via sequence-to-sequence modeling. However, these two stages are optimized for different objectives: semantic reconstruction during tokenizer pretraining versus user interaction modeling during recommender training. This objective misalignment leads to two key limitations: (i) suboptimal static tokenization, where fixed token assignments fail to reflect diverse usage contexts; and (ii) discarded pretrained semantics, where pretrained knowledge—typically from language model embeddings—is overwritten during recommender training on user interactions. To address these limitations, we propose to learn DEcomposed COntextual Token Representations (DECOR), a unified framework that preserves pretrained semantics while enhancing the adaptability of token embeddings. DECOR introduces contextualized token composition to refine token embeddings based on user interaction context, and decomposed embedding fusion that integrates pretrained codebook embeddings with newly learned collaborative embeddings. Experiments on three real-world datasets demonstrate that DECOR consistently outperforms state-of-the-art baselines in recommendation performance. Our code is available at https://github.com/yliuaa/DECOR.git.

The growing prevalence of false information originating from human–LLM mixed generation sources presents new challenges for veracity assessment, as different generation sources (e.g., human or LLM) exhibit distinct semantic patterns and retrieval needs. In particular, LLM-generated false information is often more fluent, persuasive, and difficult to detect than traditional human-written content. Existing retrieval-augmented generation (RAG) methods for veracity assessment often use a uniform retrieval strategy for all inputs, limiting their flexibility in adapting to varying content characteristics. Current adaptive RAG methods primarily focus on general input difficulty or model confidence, overlooking the distinct semantic characteristics and retrieval needs of different generation sources. We propose MoARE, a Mixture of Adaptive Retrieval Experts framework that dynamically determines whether to retrieve and how much evidence to retrieve for each post. MoARE leverages a Mixture-of-Experts (MoE) network trained via reinforcement learning to balance veracity assessment accuracy and time cost—without requiring knowledge of the input's generation source. Experiments on recent human–LLM mixed datasets demonstrate that MoARE outperforms state-of-the-art RAG baselines, achieving higher accuracy with lower time cost.

This paper studies a critical problem of explainable public health misinformation detection on social media, where clear explanations are essential for enhancing user understanding and trust, surpassing the limitations of black-box misinformation detection results. To tackle this problem, there is a growing trend of leveraging collective intelligence from diverse intelligence sources, such as deep neural networks (DNNs), human intelligence, and large language models (LLMs). However, integrating hybrid intelligence from different sources remains a challenge: DNNs excel in accurate and efficient classification, crowd workers provide contextual understanding and readable explanations, and LLMs offer extensive domain knowledge and advanced language generation. Moreover, current crowdsourcing and human-AI collaboration methods mainly focus on aggregating misinformation detection labels using traditional measures like consistency, often overlooking more complex and challenging inputs like textual explanations. We propose SynthX, a collective intelligence framework that incorporates a holistic prompting design to harness the language and reasoning capabilities of LLMs for synthesizing diverse detection and explanation results. It also integrates a novel estimation theory-LLM hybrid approach to assess the varying reliability of detection results from different intelligence sources. Our evaluation on a real-world social media misinformation dataset demonstrates that SynthX consistently outperforms a rich set of state-of-the-art baselines in both detection accuracy and explanation quality.

The proliferation of fake news on social media platforms disproportionately impacts vulnerable populations, eroding trust, exacerbating inequality, and amplifying harmful narratives. Detecting fake news in multimodal contexts-where deceptive content combines text and images-is particularly challenging due to the nuanced interplay between modalities. Existing multimodal fake news detection methods often emphasize cross-modal consistency but ignore the complex interactions between text and visual elements, which may complement, contradict, or independently influence the predicted veracity of a post. To address these challenges, we present Modality Interactive Mixture-of-Experts for Fake News Detection (MIMoE-FND), a novel hierarchical Mixture-of-Expert framework designed to enhance multimodal fake news detection by explicitly modeling modality interactions through an interaction gating mechanism. Our approach models modality interactions by evaluating two key aspects of modality interactions: unimodal prediction agreement and semantic alignment. The hierarchical structure of MIMoE-FND allows for distinct learning pathways tailored to different fusion scenarios, adapting to the unique characteristics of each modality interaction. By tailoring fusion strategies to diverse modality interaction scenarios, MIMoE-FND provides a more robust and nuanced approach to multimodal fake news detection. We evaluate our approach on three real-world benchmarks spanning two languages, demonstrating its superior performance compared to state-of-the-art methods. By enhancing the accuracy and interpretability of fake news detection, MIMoE-FND offers a promising tool to mitigate the spread of misinformation, with potential to better safeguard vulnerable communities against its harmful effects.

While sequential recommendation achieves significant progress on capturing user-item transition patterns, transferring such large-scale recommender systems remains challenging due to the disjoint user and item groups across domains. In this paper, we propose a vector quantized meta learning for transferable sequential recommenders (MetaRec). Without requiring additional modalities or shared information across domains, our approach leverages user-item interactions from multiple source domains to improve the target domain performance. To solve the input heterogeneity issue, we adopt vector quantization that maps item embeddings from heterogeneous input spaces to a shared feature space. Moreover, our meta transfer paradigm exploits limited target data to guide the transfer of source domain knowledge to the target domain (i.e., learn to transfer). In addition, MetaRec adaptively transfers from multiple source tasks by rescaling meta gradients based on the source-target domain similarity, enabling selective learning to improve recommendation performance. To validate the effectiveness of our approach, we perform extensive experiments on benchmark datasets, where MetaRec consistently outperforms baseline methods by a considerable margin.

A fundamental issue in healthcare misinformation detection is the lack of timely resources (e.g., medical knowledge, annotated data), making it challenging to accurately detect emergent healthcare misinformation at an early stage. In this paper, we develop a crowdsourcing-based early healthcare misinformation detection framework that jointly exploits the medical expertise of expert crowd workers and adapts the medical knowledge from a source domain (e.g., COVID-19) to detect misleading posts in an emergent target domain (e.g., Mpox, Polio). Two important challenges exist in developing our solution: (i) How to leverage the complex and noisy knowledge from the source domain to facilitate the detection of misinformation in the target domain? (ii) How to effectively utilize the limited amount of expert workers to correct the inapplicable knowledge facts in the source domain and adapt the corrected facts to examine the truthfulness of the posts in the emergent target domain? To address these challenges, we develop CrowdAdapt, a crowdsourcing-based domain adaptive approach that effectively identifies and adapts relevant knowledge facts from the source domain to accurately detect misinformation in the target domain. Evaluation results from two real-world case studies demonstrate the superiority of CrowdAdapt over state-of-the-art baselines in accurately detecting emergent healthcare misinformation.

Drought poses significant challenges to sustainability across various sectors in our society, leading to substantial consequences on agriculture, environments, ecosystems, public health, and socioeconomic stability. While prior work has studied the impacts of drought using professionally measured data sources, the societal perspectives of drought impacts remain largely under-explored. In this work, we present SocialDrought, a novel and comprehensive dataset to facilitate research on the societal impacts of drought. In particular, SocialDrought consists of three major components: 1) over 1.5 million social media posts, 2) over 1,400 news articles collected and verified by domain experts, and 3) over 31,000 meteorological records from the U.S. Drought Monitor about drought severity. In addition, we also introduce an online analytical platform that enables interactive and real-time data exploration to gain timely insights into the societal impacts of drought. Our interdisciplinary dataset integrates both conventional meteorological data and unconventional social and news media data to provide a holistic understanding of drought impacts. SocialDrought opens new opportunities to study the societal impacts of drought through the lens of social and news media.

This paper studies a critical problem of emergent health misinformation detection, aiming to mitigate the spread of misinformation in emergent health domains to support well-informed healthcare decisions towards a Web for good health. Our work is motivated by the lack of timely resources (e.g., medical knowledge, annotated data) during the initial phases of an emergent health event or topic. In this paper, we develop a multi-source domain adaptive framework that jointly exploits medical knowledge and annotated data from different high-resource source domains (e.g., cancer, COVID-19) to detect misleading posts in an emergent target domain (e.g., mpox, polio). Two important challenges exist in developing our solution: 1) how to accurately detect the partially misleading and unverifiable content in an emergent target domain? 2) How to identify the conflicting knowledge facts from different source domains to accurately detect emergent misinformation in the target domain? To address these challenges, we develop MMAdapt, a multi-source multi-class domain adaptive misinformation detection framework that effectively explores diverse knowledge facts from different source domains to accurately detect not only the outright misleading but also the partially misleading or unverifiable posts on the Web. Extensive experimental results on four real-world misinformation datasets demonstrate that MMAdapt substantially outperforms state-of-the-art baselines in accurately detecting misinformation in an emergent health domain.

This paper develops a symbiotic human-AI collective learning framework that explores the complementary strengths of both AI and crowdsourced human intelligence to address a novel Web-based healthcare-policy-adherence assessment (WebHA) problem. In particular, the objective of the WebHA problem is to automatically assess people's public health policy adherence during emergent global health crisis events (e.g., COVID-19, MonkeyPox) by exploring massive social media imagery data. Recent advances in human-AI systems exhibit a significant potential in addressing the intricate imagery-based classification problems like WebHA by leveraging the collective intelligence of both humans and AI. This paper aims to address the limitation of existing human-AI systems that often rely heavily on human intelligence to improve AI model performance while overlooking the fact that humans themselves can be fallible and prone to errors. To address the above limitation, this paper develops SymLearn, a symbiotic human-AI co-learning framework that leverages human intelligence to troubleshoot and fine-tune the AI model while using AI models to guide human crowd workers to reduce the inherent human errors in their labels. Extensive experiments on two real-world WebHA applications show that SymLearn clearly outperforms the state-of-the-art baselines by improving WebHA performance and reducing crowd response delay.

Robust out-domain uncertainty estimation has gained growing attention for its capacity of providing adversary-resistant uncertainty estimates on out-domain samples. However, existing work on robust uncertainty estimation mainly focuses on evasion attacks that happen during test time. The threat of poisoning attacks against uncertainty models is largely unexplored. Compared to evasion attacks, poisoning attacks do not necessarily modify test data, and therefore, would be more practical in real-world applications. In this work, we systematically investigate the robustness of state-of-the-art uncertainty estimation algorithms against data poisoning attacks, with the ultimate objective of developing robust uncertainty training methods. In particular, we focus on attacking the out-domain uncertainty estimation. Under the proposed attack, the training process of models is affected. A fake high-confidence region is established around the targeted out-domain sample, which originally would have been rejected by the model due to low confidence. More fatally, our attack is clean-label and targeted: it leaves the poisoned data with clean labels and attacks a specific targeted test sample without degrading the overall model performance. We evaluate the proposed attack on several image benchmark datasets and a real-world application of COVID-19 misinformation detection. The extensive experimental results on different tasks suggest that the state-of-the-art uncertainty estimation methods could be extremely vulnerable and easily corrupted by our proposed attack.

Web-based disaster response (WebDR) is emerging as a pervasive approach to acquire real-time situation awareness of disaster events by collecting timely observations from the Web (e.g., social media). This paper studies a class-wise inequality problem in WebDR applications where the objective is to address the limitation of current WebDR solutions that often have imbalanced classification performance across different classes. To address such a limitation, this paper explores the collaborative strengths of the diversified yet complementary biases of AI and crowdsourced human intelligence to ensure a more balanced and accurate performance for WebDR applications. However, two critical challenges exist: 1) it is difficult to identify the imbalanced AI results without knowing the ground-truth WebDR labels a priori; ii) it is non-trivial to address the class-wise inequality problem using potentially imperfect crowd labels. To address the above challenges, we develop CollabEquality, an inequality-aware crowd-AI collaborative learning framework that carefully models the inequality bias of both AI and human intelligence from crowdsourcing systems into a principled learning framework. Extensive experiments on two real-world WebDR applications demonstrate that CollabEquality consistently outperforms the state-of-the-art baselines by significantly reducing class-wise inequality while improving the WebDR classification accuracy.

The widespread misinformation on the Web has raised many concerns with serious societal consequences. In this paper, we study a critical type of online misinformation, namely fauxtography, where the image and associated text of a social media post jointly convey a questionable or false sense. In particular, we focus on a sparse semi-supervised fauxtography detection problem, which aims to accurately identify fauxtography by only using the sparsely annotated ground truth labels of social media posts. Our problem is motivated by the key limitation of current fauxtography detection approaches that often require a large amount of expensive and inefficient manual annotations to train an effective fauxtography detection model. We identify two key technical challenges in solving the problem: 1) it is non-trivial to train an accurate detection model given the sparse fauxtography annotations, and 2) it is difficult to extract the heterogeneous and complicated fauxtography features from the multi-modal social media posts for accurate fauxtography detection. To address the above challenges, we propose ContrastFaux, a multi-view contrastive learning framework that jointly explores the sparse fauxtography annotations and the cross-modal fauxtography feature similarity between the image and text in multi-modal posts to accurately detect fauxtography on social media. Evaluation results on two social media datasets demonstrate that ContrastFaux consistently outperforms state-of-the-art deep learning and semi-supervised learning fauxtography detection baselines by achieving the highest fauxtography detection accuracy.

This paper focuses on an important problem of early misinformation detection in an emergent health domain on social media. Current misinformation detection solutions often suffer from the lack of resources (e.g., labeled datasets, sufficient medical knowledge) in the emerging health domain to accurately identify online misinformation at an early stage. To address such a limitation, we develop a knowledge-driven domain adaptive approach that explores a good set of annotated data and reliable knowledge facts in a source domain (e.g., COVID-19) to learn the domain-invariant features that can be adapted to detect misinformation in the emergent target domain with little ground truth labels (e.g., Monkeypox). Two critical challenges exist in developing our solution: i) how to leverage the noisy knowledge facts in the source domain to obtain the medical knowledge related to the target domain? ii) How to adapt the domain discrepancy between the source and target domains to accurately assess the truthfulness of the social media posts in the target domain? To address the above challenges, we develop KAdapt, a knowledge-driven domain adaptive early misinformation detection framework that explicitly extracts rel-evant knowledge facts from the source domain and jointly learns the domain-invariant representation of the social media posts and their relevant knowledge facts to accurately identify misleading posts in the target domain. Evaluation results on five real-world datasets demonstrate that KAdapt significantly outperforms state-of-the-art baselines in terms of accurately detecting misleading Monkeypox posts on social media.

In the real-world application of COVID-19 misinformation detection, a fundamental challenge is the lack of the labeled COVID data to enable supervised end-to-end training of the models, especially at the early stage of the pandemic. To address this challenge, we propose an unsupervised domain adaptation framework using contrastive learning and adversarial domain mixup to transfer the knowledge from an existing source data domain to the target COVID-19 data domain. In particular, to bridge the gap between the source domain and the target domain, our method reduces a radial basis function (RBF) based discrepancy between these two domains. Moreover, we leverage the power of domain adversarial examples to establish an intermediate domain mixup, where the latent representations of the input text from both domains could be mixed during the training process. Extensive experiments on multiple real-world datasets suggest that our method can effectively adapt misinformation detection systems to the unseen COVID-19 target domain with significant improvements compared to the state-of-the-art baselines.

Modern machine learning (ML) is one of the prevailing tools for big data applications of face attribute recognition. However, due to the commonly observed imbalanced distribution of the training data, well-trained models could suffer severely from undesired performance bias across different demographic groups. Motivated by the fact that neural networks could be extremely sensitive to adversarial examples, we argue that there exists the possibility of properly leveraging adversarial examples to address the imbalanced data distribution, and guiding the training convergence towards the direction of improved fairness. That is, we propose to use adversarial examples to alleviate the performance bias issue from the origin: the data source. In this paper, we present a novel adversarial training framework that generates adversarial features in the latent space to automatically balance the distribution of training features and adjust the deep classification layers of the face attribute classifiers to be more fair. Extensive experimental results on the CelebA face dataset show that our method is able to boost the model fairness more effectively compared to the state-of-the-art adversarial debiasing algorithms.

Human face images represent a rich set of visual information for online social media platforms to optimize the machine learning (ML)/AI models in their data-driven facial applications (e.g., face detection, face recognition). However, there exists a growing privacy concern from social media users to share their online face images that will be annotated by unknown crowd workers and analyzed by ML/AI researchers in the model training and optimization process. In this paper, we focus on a privacy-aware face recognition problem where the goal is to empower the facial applications to train their face recognition models with images shared by social media users while protecting the identity of the users. Our problem is motivated by the limitation of current privacy-aware face recognition approaches that mainly prevent algorithmic attacks by manipulating face images but largely ignore the potential privacy leakage related to human activities (e.g., crowdsourcing annotation). To address such limitations, we develop FaceCrowd, a web crowdsourcing based face partition approach to improve the performance of current face recognition models by designing a novel crowdsourced partial face graph generated from privacy-preserved social media face images. We evaluate the performance of FaceCrowd using two real-world human face datasets that consist of large-scale human face images. The results show that FaceCrowd not only improves the accuracy of the face recognition models but also effectively protects the identity information of the social media users who share their face images.

This paper focuses on a critical problem of explainable multimodal COVID-19 misinformation detection where the goal is to accurately detect misleading information in multimodal COVID-19 news articles and provide the reason or evidence that can explain the detection results. Our work is motivated by the lack of judicious study of the association between different modalities (e.g., text and image) of the COVID-19 news content in current solutions. In this paper, we present a generative approach to detect multimodal COVID-19 misinformation by investigating the cross-modal association between the visual and textual content that is deeply embedded in the multimodal news content. Two critical challenges exist in developing our solution: 1) how to accurately assess the consistency between the visual and textual content of a multimodal COVID-19 news article? 2) How to effectively retrieve useful information from the unreliable user comments to explain the misinformation detection results? To address the above challenges, we develop a duo-generative explainable misinformation detection (DGExplain) framework that explicitly explores the cross-modal association between the news content in different modalities and effectively exploits user comments to detect and explain misinformation in multimodal COVID-19 news articles. We evaluate DGExplain on two real-world multimodal COVID-19 news datasets. Evaluation results demonstrate that DGExplain significantly outperforms state-of-the-art baselines in terms of the accuracy of multimodal COVID-19 misinformation detection and the explainability of detection explanations.

Social sensing has emerged as a pervasive and scalable sensing paradigm to obtain timely information of the physical world from "human sensors". In this paper, we study a new extremely-sparse disaster damage assessment (DBA) problem in social sensing. The objective is to automatically assess the damage severity of affected areas in a disaster event by leveraging the imagery data reported on online social media with extremely sparse training data (e.g., only 1% of the data samples have labels). Our problem is motivated by the limitation of current DDA solutions that often require a significant amount of high-quality training data to learn an effective DDA model. We identify two critical challenges in solving our problem: i) it remains to be a fundamental challenge on how to effectively train a reliable DDA model given the lack of sufficient damage severity labels; ii) it is a difficult task to capture the excessive and fine-grained damage-related features in each image for accurate damage assessment. In this paper, we propose ContrastDDA, a deep contrastive learning approach to address the extremely-sparse DDA problem by designing an integrated contrastive and augmentative neural network architecture for accurate disaster damage assessment using the extremely sparse training samples. The evaluation results on two real-world DDA applications demonstrate that ContrastDDA clearly outperforms state-of-the-art deep learning and semi-supervised learning baselines with the highest DDA accuracy under different application scenarios.

Human face images represent a rich set of visual data information that is utilized by various big data driven human facial applications. However, the performance of these applications is usually biased towards the majority demographic group due to the data imbalance issue. In this paper, we focus on a fair human face data exchange problem where the goal is to exchange visual features of human face images between different human face datasets and obtain a set of augmented datasets that improve the fairness and performance of human facial applications. Our problem is motivated by the limitations of current fairness approaches that only focus on a single human face dataset from a particular application and require a large amount of pre-annotated demographic attribute labels to develop fair human facial models. To address these limitations, we develop ExgFair, a crowdsourcing-based fair data exchange framework to generate a set of augmented fair face image datasets by leveraging the crowdsourced demographic attribute labels of human face images. We evaluate ExgFair using a set of real-world human face image datasets with different demographic distributions. The results show that ExgFair not only reduces demographic biases of the datasets but also improves the accuracy of human facial applications trained on the augmented fair datasets.

This paper studies an emerging and important problem of identifying misleading COVID-19 short videos where the misleading content is jointly expressed in the visual, audio, and textual content of videos. Existing solutions for misleading video detection mainly focus on the authenticity of videos or audios against AI algorithms (e.g., deepfake) or video manipulation, and are insufficient to address our problem where most videos are user-generated and intentionally edited. Two critical challenges exist in solving our problem: i) how to effectively extract information from the distractive and manipulated visual content in TikTok videos? ii) How to efficiently aggregate heterogeneous information across different modalities in short videos? To address the above challenges, we develop TikTec, a multimodal misinformation detection framework that explicitly exploits the captions to accurately capture the key information from the distractive video content, and effectively learns the composed misinformation that is jointly conveyed by the visual and audio content. We evaluate TikTec on a real-world COVID- 19 video dataset collected from TikTok. Evaluation results show that TikTec achieves significant performance gains compared to state-of-the-art baselines in accurately detecting misleading COVID-19 short videos.

Social sensing has emerged as a pervasive and scalable sensing paradigm to collect observations of the physical world from human sensors. A key advantage of social sensing is its infrastructure-free nature. In this paper, we focus on a streaming urban infrastructure monitoring (Streaming UIM) problem in social sensing. The goal is to automatically detect the urban infrastructure damages from the streaming imagery data posted on social media by exploring the collective power of both AI and human intelligence from crowdsourcing systems. Our work is motivated by the limitation of current AI and crowdsourcing solutions that either fail in many critical time-sensitive UIM application scenarios or are not easily generalizable to monitor the damage of different types of urban infrastructures. We identify two critical challenges in solving our problem: i) it is difficult to dynamically integrate AI and crowd intelligence to effectively identify and fix the failure cases of AI solutions; ii) it is non-trivial to obtain accurate human intelligence from unreliable crowd workers in streaming UIM applications. In this paper, we propose StreamCollab, a streaming crowd-AI collaborative system that explores the collaborative intelligence from AI and crowd to solve the streaming UIM problem. The evaluation results on a real-world urban infrastructure imagery dataset collected from social media demonstrate that StreamCollab consistently outperforms both state-of-the-art AI and crowd-AI baselines in UIM accuracy while maintaining the lowest computational cost.

Social media sensing has emerged as a powerful sensing paradigm to collect the observations of the physical world by exploring the “wisdom of crowd”. In this paper, we focus on a migratable disaster damage assessment problem in social media sensing applications. Our goal is to accurately identify the damage severity of affected areas in an unfolding disaster event using unlabeled social media data feeds (e.g., image posts on social media). Two fundamental challenges exist in solving our problem: i) different disaster events often have distinct characteristics (e.g., damage types, affected areas) that cannot be easily migrated; ii) it is non-trivial to modify a damage assessment model from a previous event to adapt to a new event without using the labeled data from the new event. To address the above challenges, we develop SocialTrans, a hybrid deep transfer learning framework, to enable effective model migration for accurate damage assessment without using any training data from the studied disaster event. The evaluation results on four real-world disaster events show that SocialTrans consistently outperforms the state-of-the-art baselines in accurately assessing the damage level of disasters.

Social sensing has emerged as a new sensing paradigm to observe the physical world by exploring the "wisdom of crowd" on social media. This paper focuses on the abnormal traffic event localization problem using social media sensing. Two critical challenges exist in the state-of-the-arts: i) "content-only inference": the limited and unstructured content of a social media post provides little clue to accurately infer the locations of the reported traffic events; ii) "informal and scarce data": the language of the social media post (e.g., tweet) is informal and the number of the posts that report the abnormal traffic events is often quite small. To address the above challenges, we develop SyntaxLoc, a syntax-based probabilistic learning framework to accurately identify the location entities by exploring the syntax of social media content. We perform extensive experiments to evaluate the SyntaxLoc framework through real world case studies in both New York City and Los Angeles. Evaluation results demonstrate significant performance gains of the SyntaxLoc framework over state-of-the-art baselines in terms of accurately identifying the location entities that can be directly used to locate the abnormal traffic events.

With the increasing popularity of portable devices with cameras (e.g., smartphones and tablets) and ubiquitous Internet connectivity, travelers can share their instant experience during the travel by posting photos they took to social media platforms. In this paper, we present a new image-driven poetry recommender system that takes a traveler's photo as input and recommends classical poems that can enrich the photo with aesthetically pleasing quotes from the poems. Three critical challenges exist to solve this new problem: i) how to extract the implicit artistic conception embedded in both poems and images? ii) How to identify the salient objects in the image without knowing the creator's intent? iii) How to accommodate the diverse user perceptions of the image and make a diversified poetry recommendation? The proposed iPoemRec system jointly addresses the above challenges by developing heterogeneous information network and neural embedding techniques. Evaluation results from real-world datasets and a user study demonstrate that our system can recommend highly relevant classical poems for a given photo and receive significantly higher user ratings compared to the state-of-the-art baselines.

Road traffic accidents are a major challenge in urban transportation systems. An effective countermeasure to address this problem is to accurately forecast the traffic risks in a city before accidents actually happen. Current traffic accident prediction solutions largely rely on accurate data collected from infrastructure-based sensors, which is not always available due to various resource constraints or privacy and legal concerns. In this paper, we address this limitation by exploring social sensing, a new sensing paradigm that uses humans as sensors to report the states of the physical world. In particular, we consider two types of publicly available social sensing data sources: social media data (e.g., traffic posts on Twitter) and open city data (e.g., traffic data from the city web portal). In this paper, we develop the RiskCast, an inductive multi-view learning approach to accurately forecast the traffic risk by exploiting the social sensing data under a principled co-regularization framework. The evaluation results on a real world dataset from New York City show that RiskCast significantly outperforms the state-of-the-art baselines in forecasting the traffic risks in a city.

Urban land usage classification is a critical task in big data based smart city applications that aim to understand the social-economic land functions and physical land attributes in urban environments. This paper focuses on a migratable urban land usage classification problem using remote sensing data (i.e., satellite images). Our goal is to accurately classify the land usage of locations in a target city where the ground truth land usage data is not available by leveraging a classification model from a source city where such data is available. This problem is motivated by the limitation of current solutions that primarily rely on a rich set of ground-truth data for accurate model training, which encounters high annotation costs. Two important challenges exist in solving our problem: i) the target and source cities often have different urban characteristics that prevent the direct application of a model learned from the source city to the target city; ii) the complex visual features in satellite images make it non-trivial to “translate” the images from the target city to the source city for an accurate classification. To address the above challenges, we develop TransLand, an adversarial transfer learning framework to translate the satellite images from the target city to the source city for accurate land usage classification. We evaluate our scheme on the real-world satellite imagery and land usage datasets collected from live different cities in Europe. The results show that TransLand significantly outperforms the state-of-the-art land usage classification baselines in classifying the land usage of locations in a city.

Social media sensing has emerged as a new application paradigm to collect observations from online social media users about the physical environment. A fundamental problem in social media sensing applications lies in estimating the evolving truth of the measured variables and the reliability of data sources without knowing either of them a priori. This problem is referred to as dynamic truth discovery. Two major limitations exist in current truth discovery solutions: i) existing solutions cannot effectively address the missing truth problem where the measured variables do not have any reported measurements from the data sources; ii) the latent correlations among the measured variables were not fully captured and utilized in current solutions. In this paper, we proposed a Reliable Missing Truth Finder (RMTF) to address the above limitations in social media sensing applications. In particular, we develop a novel data-driven technique to identify the lagged and latent correlations among measured variables, and incorporate such correlation information into a holistic spatiotemporal inference model to infer the missing truth. We evaluated the RMTF using the real-world Twitter data feeds. The results show that the RMTF scheme significantly outperforms the state-of-the-art truth discovery solutions by correctly inferring the missing truth of the measured variables.

With the increasing popularity of online video sharing platforms (such as YouTube and Twitch), the detection of content that infringes copyright has emerged as a new critical problem in online social media. In contrast to the traditional copyright detection problem that studies the static content (e.g., music, films, digital documents), this paper focuses on a much more challenging problem: one in which the content of interest is from live videos. We found that the state-of-the-art commercial copyright infringement detection systems, such as the ContentID from YouTube, did not solve this problem well: large amounts of copyright-infringing videos bypass the detector while many legal videos are taken down by mistake. In addressing the copyright infringement detection problem for live videos, we identify several critical challenges: i) live streams are generated in real-time and the original copyright content from the owner may not be accessible; ii) streamers are getting more and more sophisticated in bypassing the copyright detection system (e.g., by modifying the title, tweaking the presentation of the video); iii) similar video descriptions and visual contents make it difficult to distinguish between legal streams and copyright-infringing ones. In this paper, we develop a crowdsourcing-based copyright infringement detection (CCID) scheme to address the above challenges by exploring a rich set of valuable clues from live chat messages. We evaluate CCID on two real world live video datasets collected from YouTube. The results show our scheme is significantly more effective and efficient than ContentID in detecting copyright-infringing live videos on YouTube.

With the ever-increasing number of road traffic accidents worldwide, the road traffic safety has become a critical problem in intelligent transportation systems. A key step towards improving the road traffic safety is to identify the locations where severe traffic accidents happen with a high probability so the precautions can be applied effectively. We refer to this problem as risky traffic location identification. While previous efforts have been made to address similar problems, two important limitations exist: i) data availability: many cities (especially in developing countries) do not maintain a publicly accessible database for the traffic accident records in a city, which makes it difficult to accurately estimate the accidents in the city; ii) location accuracy: many self-reported traffic accidents (e.g., social media posts from common citizens) are not associated with the exact GPS locations due to the privacy concerns. To address these limitations, this paper develops the RiskSens, a multi-view learning approach to identifying the risky traffic locations in a city by jointly exploring the social and remote sensing data. We evaluate RiskSens using a real world dataset from New York. The evaluation results show that RiskSens significantly outperforms the state-of-the- art baselines in identifying risky traffic locations in a city.

Copyright infringement detection is a critical problem in large-scale online video sharing systems: the copyright-infringing videos must be correctly identified and removed from the system to protect the copyright of the content owners. This paper focuses on a challenging problem of detecting copyright infringement in live video streams. The problem is particularly difficult because i) streamers can be sophisticated and modify the title or tweak the presentation of the video to bypass the detection system; ii) legal videos and copyright-infringing ones may have very similar visual content and descriptions. We found current commercial copyright detection systems did not address this problem well: a large amount of copyrighted content bypasses the detection system while legal streams are taken down by mistake. In this paper, we develop the StreamGuard, an unsupervised Bayesian network based copyright infringement detection system that addresses the above challenges by leveraging live chat messages from the audience. We evaluate StreamGuard on real-world live video streams collected from YouTube. The results show that StreamGuard is effective and efficient in identifying the copyright-infringing videos.

Social media sensing has emerged as a new big data application paradigm to collect observations and claims about the measured variables in physical environment from common citizens. A fundamental problem in social media sensing applications lies in estimating the evolving truth of claims and the reliability of data sources without knowing either of them a priori, which is referred to as dynamic truth discovery. We identified two critical challenges that are not fully addressed by solutions from current literature. The first challenge is “physical constraint-awareness” where the transition of truth is constrained by some physical rules that must be followed to ensure correct estimation of the evolving truth. The second one is “noisy and incomplete data” where the social media sensing data is sparse in nature and contains a lot of rumors and misinformation, making it difficult to capture the constantly evolving truth of measured variables. In this paper, we developed a new Constraint-Aware Dynamic Truth Discovery (CA-DTD) scheme to address the above challenges. To address the physical constraint-awareness challenge, CA-DTD develops a new constraint-aware Hidden Markov Model to effectively infer the evolving truth of measured variables by incorporating physical constraints. To address the noisy and incomplete data challenge, CA-DTD fuses sensing observations from online social media with information from traditional news media using a principled approach. We evaluate CA-DTD scheme using two real-world social media sensing data traces and the results show that CA-DTD significantly outperforms the state-of-the-art baselines.

Point-of-Interest (POI) recommendation is an important application in Location-based Social Networks (LBSN). The category prediction problem is to predict the next POI category that users may visit. The predicted category information is critical in large-scale POI recommendation because it can significantly reduce the prediction space and improve the recommendation accuracy. While efforts have been made to address the POI category prediction problem, several important challenges still exist. First, existing solutions did not fully explore the temporal dependency (e.g., “long range dependency”) of users' check-in traces. Second, the hidden contextual information associated with each check-in point has been underutilized. In this work, we propose a Context-Aware POI Category Prediction (CAP-CP) scheme using Natural Language Processing (NLP) models. In particular, to address temporal dependency challenge, we develop a novel Temporal Adaptive Ngram (TA-Ngram) model to capture the dynamic dependency between check-in points. To address the challenge of hidden context incorporation, CAP-CP leverages the Probabilistic Latent Semantic Analysis (PLSA) model to infer the semantic implications of the context variables in the prediction model. Empirical results on a real world dataset show that our scheme can effectively improve the performance of the state-of-the-art POI recommendation solutions.