Recent work has shown that large language models (LLMs) can enhance recommender systems by integrating collaborative filtering (CF) signals through hybrid prompting. However, most existing CF-LLM frameworks collapse explicit ratings into implicit or positive-only feedback, discarding the ordinal structure that conveys fine-grained preference strength. As a result, these models struggle to exploit graded semantics and nuanced preference distinctions. We propose Ordinal Semantic Anchoring (OSA), a hybrid CF-LLM framework that explicitly incorporates preference strength by modeling interaction-level user feedback. OSA represents ordinal preference levels as numeric textual tokens and uses their token embeddings as semantic anchors to align user-item interaction representations in the LLM latent space. Through strength-aware alignment across ordinal levels, OSA preserves preference semantics when integrating collaborative signals with LLMs. Experiments on multiple real-world datasets demonstrate that OSA consistently outperforms existing baselines, particularly in pairwise preference evaluation, highlighting its effectiveness in modeling fine-grained user preferences over prior CF-LLM methods.

With the increasing demand for histopathological specimen examination and diagnostic reporting, Multiple Instance Learning (MIL) has received heightened research focus as a viable solution for AI-centric diagnostic aid. Recently, to improve its performance and make it work more like a pathologist, several MIL approaches based on the use of multiple-resolution images have been proposed, delivering often higher performance than those that use single-resolution images. Despite impressive recent developments of multiple-resolution MIL, previous approaches only focus on improving performance, thereby lacking research on well-calibrated MIL that clinical experts can rely on for trustworthy diagnostic results. In this study, we propose Uncertainty-Focused Calibrated MIL (UFC-MIL), which more closely mimics the pathologists’ examination behaviors while providing calibrated diagnostic predictions, using multiple images with different resolutions. UFC-MIL includes a novel patch-wise loss that learns the latent patterns of instances and expresses their uncertainty for classification. Also, the attention-based architecture with a neighbor patch aggregation module collects features for the classifier. In addition, aggregated predictions are calibrated through patch-level uncertainty without requiring multiple iterative inferences, which is a key practical advantage. Against challenging public datasets, UFC-MIL shows superior performance in model calibration while achieving classification accuracy comparable to that of state-of-the-art methods.

Large-scale language models (LLMs) have shown exceptional performance in various tasks, particularly in zero-shot and few-shot settings. However, in sensitive domains like healthcare, detecting unanswerable questions remains a critical challenge. This task is challenging due to data imbalance, and existing methods are computationally expensive and inflexible to data distribution changes. To address these issues, we propose Retrieval-augmented Question Answerability Detection (RaQAD), a training-free method that uses LLMs to identify unanswerable questions by retrieving semantically similar examples as few-shot prompts. RaQAD ensures semantically similar sampling, adapts to schema changes, and eliminates the need for additional training. Extensive experiments on clinical datasets demonstrate its effectiveness in outperforming existing approaches while addressing data imbalance challenges.

Recent advancements in heterogeneous GNNs have enabled significant progress in embedding nodes and learning relationships across diverse tasks. However, traditional methods rely heavily on meta-paths grounded in node types, which often fail to encapsulate the full complexity of node interactions, leading to inconsistent performance and elevated computational demands. To address these challenges, we introduce MF2Vec, a novel framework that shifts focus from rigid node-type dependencies to dynamically exploring shared facets across nodes, regardless of type. MF2Vec constructs multi-faceted paths and forms homogeneous networks to learn node embeddings more effectively. Through extensive experiments, we demonstrate that MF2Vec achieves superior performance in node classification, link prediction, and node clustering tasks, surpassing existing baselines. Furthermore, it exhibits reduced performance variability due to meta-path dependencies and achieves faster training convergence. These results highlight its capability to analyze complex networks comprehensively. The implementation of MF2Vec is publicly available at https://github.com/kimjongwoo-cell/MF2Vec.

Recent advancements in large language models (LLMs) have boosted research on generating SQL queries from domain-specific questions, particularly in the medical domain. A key challenge is detecting and filtering unanswerable questions. Existing methods often relying on model uncertainty, but these require extra resources and lack interpretability. We propose a lightweight model that predicts relevant database schemas to detect unanswerable questions, enhancing interpretability and addressing the data imbalance in binary classification tasks. Furthermore, we found that LLM-generated schema descriptions can significantly enhance the prediction accuracy. Our method provides a resource-efficient solution for unanswerable question detection in domain-specific question answering systems.

The 15-minute city (15MC) is an urban planning concept that promotes sustainable and inclusive cities where residents can access essential services within a short amount of time (i.e., 15 minutes). However, evaluating 15MC compliance in hyper-dense cities remains challenging due to: (1) traditional manual assessments that are resource-intensive and difficult to scale, and (2) POI-based metrics that suffer from data unavailability and lack of spatial contexts. In this paper, we propose a novel evaluation framework that directly assesses 15MC compliance from geospatial imagery in three stages: image pre-processing, representation learning, and instance aggregation. To validate the framework, we have constructed a new dataset of 2,794 residential areas in Seoul, pairing high-resolution geospatial imagery with functional urban labels. Furthermore, we have developed a model, GeoTwin-MIL, on the basis of the proposed framework. The model includes two key components: (1) cross-modal contrastive learning that aligns satellite and map representations to capture both morphological (building density) and topological (road networks) features, enabling robust inference using only satellite images, and (2) multiple instance learning to efficiently aggregate geospatial details while detecting localized urban functions within high-resolution imagery. The experimental results obtained from various evaluation settings show that GeoTwin-MIL significantly outperforms single-modality approaches or vision baselines, validating the integrative effectiveness of the two key components and supporting the transferability of the model without POI dependencies. The code is available at https://github.com/20243439/geotwin_mil.git.

Vision-language models (VLMs) have recently been integrated into multiple instance learning (MIL) frameworks to address the challenge of few-shot, weakly supervised classification of whole slide images (WSIs). A key trend involves leveraging multi-scale information to better represent hierarchical tissue structures. However, existing methods often face two key limitations: (1) insufficient modeling of interactions within the same modalities across scales (e.g., 5x and 20x) and (2) inadequate alignment between visual and textual modalities on the same scale. To address these gaps, we propose HiVE-MIL, a hierarchical vision-language framework that constructs a unified graph consisting of (1) parent–child links between coarse (5x) and fine (20x) visual/textual nodes to capture hierarchical relationships, and (2) heterogeneous intra-scale edges linking visual and textual nodes on the same scale. To further enhance semantic consistency, HiVE-MIL incorporates a two-stage, text-guided dynamic filtering mechanism that removes weakly correlated patch–text pairs, and introduces a hierarchical contrastive loss to align textual semantics across scales. Extensive experiments on TCGA breast, lung, and kidney cancer datasets demonstrate that HiVE-MIL consistently outperforms both traditional MIL and recent VLM-based MIL approaches, achieving gains of up to 4.1% in macro F1 under 16-shot settings. Our results demonstrate the value of jointly modeling hierarchical structure and multimodal alignment for efficient and scalable learning from limited pathology data. The code is available at https://github.com/bryanwong17/HiVE-MIL.

Deep neural networks (DNNs) have transformed biomedical image analysis, particularly in histopathology with Whole Slide Images (WSIs) classification. However, training DNNs requires large annotated datasets, which is challenging due to the high heterogeneity and high resolution of WSIs. Multiple Instance Learning (MIL) has become a popular method for weakly supervised classification in this context, training with only slide-level labels. De-spite the advancements, ensuring the reliability of model performance is crucial in safety-critical domains including healthcare. Deep learning models in real-world decision-making systems must accurately predict probability estimates to reflect the true likelihood of correctness, known as confidence calibration. This study introduces a novel calibration framework, UDLS, which uses data-wise label smoothing based on predictive uncertainty to improve the calibration of MIL frameworks. This approach involves augmenting WSIs with PatchFeatureDropout, computing predictive uncertainty estimates for original data, and applying these estimates to each sample for label smoothing during model training. Experimental results on benchmark histopathology datasets show noticeable improvements in both calibration and classification performance, highlighting UDLS's potential for enhancing the reliability of pre-dictions from deep learning models in clinical settings.

Generating images directly from tabular data while ensuring an accurate representation of ground truth is a useful application in manufacturing. Simply embedding tabular data to use it as a condition in image generation models often fails to learn the correspondence between tabular features and their impact on the generated image. To overcome this limitation, we propose Contrastive Tabular-Image Pre-training (CTIP), inspired by the CLIP framework. These pre-train methods help improve the quality of the embedding of the tabular encoder on the tabular data, which then helps improve the performance of the image generation model. CTIP uses contrastive learning on multiple tabular and image data pairs, allowing the model to learn how changes in certain tabular features affect images. This approach is particularly crucial in manufacturing, where accurate capture of product outcomes under varying conditions is essential. We demonstrate that applying CTIP enhances image generation performance, yielding images that closely match ground truth images, even in Feature Few-shot or Feature Zero-shot scenarios where specific features are sparse or novel. We further show the application of CTIP in tire development, where tire footprint images are generated based on tire specifications and test conditions. CTIP produces high-quality embeddings that align well with ground truth images and effectively handle the scarcity or sparseness of specific features, addressing common challenges in new product development. Our code is available in https://github.com/Noverse0/CTIP.git.

Chatbots possess great potential benefits, yet concerns persist regarding users adopting inappropriate, offensive language. This research delved into the influence of user characteristics on verbally aggressive behaviours towards social chatbots. Employing a mixed-method study, we examined individual characteristics such as personal dispositions, offensive language patterns, academic majors, and prior experiences with conversational agents. Findings from a ten-day field experiment involving 33 participants using a real-world Telegram-based chatbot app unveiled that users' anthropomorphism, computer-related major, and gender significantly impact their moral emotions and evaluations of the chatbot's capabilities. Moreover, employing offensive language towards the chatbot detrimentally impacted users' perceptions of its abilities, helpfulness, and likability. The research findings advocate for ongoing monitoring and effective resolution of users' behaviours regarding the use of offensive language in their interactions with a chatbot. Additionally, the results underscore the importance of incorporating diverse perspectives into chatbot design to address biases and offensive utterances.

Abstract The application of language models in essay scoring has gained significant attention in recent years, typically on the basis of evaluating a single model across multiple prompts. However, in a multi-prompt setup, it is crucial to understand the varying aspects of different prompts. In such settings, there exist notable variations even in a trait with the same name across prompts. This semantic variation on the same traits underscores the need to treat them differently at a fine-grained level according to each prompt. In this study, we propose a multi-level disentanglement framework for multi-prompt essay scoring, designed to achieve fine-grained disentanglement of semantic differences across such traits. Our method not only improves the quality of the essay scoring, but also reduces memory usage and latency. Experimental results highlight that our framework surpasses seven state-of-the-art essay scoring methods and large language model(LLM)-based zero-shot and few-shot approaches, achieving the highest agreement with human essay ratings.

Hypergraphs provide a superior modeling frame-work for representing complex multidimensional relationships in the context of real-world interactions that often occur in groups, overcoming the limitations of traditional homogeneous graphs. However, there have been few studies on hypergraph-based contrastive learning, and existing graph-based contrastive learning methods have not been able to fully exploit the high-order correlation information in hypergraphs. Here, we propose a Hypergraph Fine-grained contrastive learning (HyFi) method designed to exploit the complex high-dimensional information inherent in hypergraphs. While avoiding traditional graph augmentation methods that corrupt the hypergraph topology, the proposed method provides a simple and efficient learning augmentation function by adding noise to node features. Furthermore, we expands beyond the traditional dichotomous relationship between positive and negative samples in contrastive learning by introducing a new relationship of weak positives. It demonstrates the importance of fine-graining positive samples in contrastive learning. Therefore, HyFi is able to produce high-quality embeddings, and outperforms both supervised and unsupervised baselines in average rank on node classification across 10 datasets. Our approach effectively exploits high-dimensional hypergraph information, shows significant improvement over existing graph-based contrastive learning methods, and is efficient in terms of training speed and GPU memory cost. The source code is available at https://github.com/Noverse0/HyFi.git.

Going beyond the user–item rating information, recent studies have utilized additional information to improve the performance of recommender systems. Graph neural network (GNN) based approaches are among the most common. However, existing models that utilize text data require a lot of computing resources and have a complex structure that makes them difficult to utilize in real-world applications. In this research, we propose a new method, keyword-enhanced graph matrix completion (KGMC), which utilizes keyword sharing relationships in user–item graphs. Our model has a simpler structure and requires less computing resources than existing models that utilize text data, but it has the advantage of cross-domain transferability while providing an intuitive understanding of the inference results. KGMC consists of three steps: (1) keyword extraction from the review text, (2) subgraph extraction and keyword-enhanced subgraph construction, and (3) GNN-based rating prediction. We have conducted extensive experiments over eight benchmark datasets to examine the relative superiority of the proposed KGMC method, compared to state-of-the-art baselines. Additional experiments and case studies have been also conducted to demonstrate the transferability as well as keyword-based explainability of KGMC. Our findings highlight the practical advantages of our model for recommender systems and support its effectiveness in inductive graph-based link prediction.

With the advancement of artificial intelligence, conversational agents are now capable of displaying intelligent and emotionally empathetic responses, which are essential for the continued use of AI-based agents. However, there is a scarcity of formal measures that can comprehensively evaluate how well they react to the user’s emotional needs. The objective of this research is to develop and validate a set of new measures, collectively called the agent empathic reactivity index (AERI), an adaptation of the interpersonal reactivity index (IRI) developed for the human-human relationship evaluation to the human-agent interaction context. By rigorously following the measure development procedures suggested by prior research, four dimensions of AERI measures of empathic concern, perspective-taking, fantasy, and personal distress were developed. Multiple pilot tests and surveys involving various conversational agents were conducted to validate the four AERI measures. The study results show that the new measures have strong psychometric properties and nomological validity.

Existing graph neural network (GNN) based recommendation models depend highly on initial node features and graph structures. Most prior studies do not support the use of additional information on user-item interactions and adopt a transductive method, which has limitations for actual web service. To overcome these problems, we propose Context-aware Inductive Graph Matrix Completion (CGMC), which utilizes context vectors that represent user-item interactions and user-item bipartite graph structures in an inductive manner. Our method constructs context vectors from the review, timestamp, and rating and uses them as edge features of the user-item bipartite graph. Relations between homogeneous nodes are also constructed based on common neighbors. Our model combines context vectors and relational information using Context-aware Graph Attention Networks and Edge Fusion Graph Convolutional Networks. We conducted extensive experiments using six real-world datasets, and the results show that the proposed model achieves superior performances over other competing models. Furthermore, we analyzed the inductive characteristics of CGMC through the cross-domain transferability. The source code is available in the repository at [https://github.com/venzino-han/CGMC_SBERT].

Conversational agents (CAs) offer new functionality and convenience. While their sales have been soaring, they have also rapidly become victims of verbal abuse by their users. Without proper handling of abusive usage, abusers’ actions can be reinforced and transferred to real life. This study investigates whether alternative response styles of empathy orientation and emotional expressivity of voice-activated virtual assistants influence users’ moral emotions found to reduce verbal aggression as well as whether they affect user perceptions of the agent’s capability. Ninety-eight participants were assigned to one of the three emotional expressivity conditions (no-facial expression, fixed-facial expression, varied-facial expression) and interacted with two alternative empathy orientation conditions (other-oriented, self-oriented) of agents. The experimental results show that, regardless of the emotional expressivity types, the agent’s empathy orientation has a significant effect on the moral emotions and agent capability perceptions. Overall, an agent that employed other-oriented empathy style elicited most positive responses from the users. However, the preference was not across the board, as about one-third of the participants showed preference to the self-oriented CA. Users valued agents’ verbal contents and vocal characteristics above their facial expressions. Based on the study findings, we draw several design guidelines and suggest avenues for future research.

Fueled by the power of AI, chatbots are becoming more personal. Prior research showed that a chatbot has great potential to elicit its user’s self-disclosure because it does not judge the user. However, the chatbot’s features beyond the conversational characteristics in eliciting a user’s self-disclosure are not as well researched. In this study, we have developed a chatbot and implemented two non-conversation features: (1) co-activity (COA), conducting an activity together, and (2) conversation atmosphere visualization (CAV), visually displaying the emotional feelings conveyed in the conversation, to examine their effects on self-disclosure and user experience. We conducted a field study involving 87 participants who were randomly assigned to one of the four experimental conditions (control, COA only, CAV only, CAV + COA) and asked to use the assigned chatbot for 10 days in their natural life setting. Our results from this field study show that both the COA and CAV features have significant effects on a user’s self-disclosure. In addition, interaction effects between COA and CAV have been found to affect a user’s intention to use. Based on the findings, we provide design implications for a user’s self-disclosure and trusting relationship development with a chatbot.

With the popularity of AI-infused systems, conversational agents (CAs) are becoming essential in diverse areas, offering new functionality and convenience, but simultaneously, suffering misuse and verbal abuse. We examine whether conversational agents' response styles under varying abuse types influence those emotions found to mitigate peoples' aggressive behaviors, involving three verbal abuse types (Insult, Threat, Swearing) and three response styles (Avoidance, Empathy, Counterattacking). Ninety-eight participants were assigned to one of the abuse type conditions, interacted with the three spoken (voice-based) CAs in turn, and reported their feelings about guiltiness, anger, and shame after each session. The results show that the agent's response style has a significant effect on user emotions. Participants were less angry and more guilty with the empathy agent than the other two agents. Furthermore, we investigated the current status of commercial CAs' responses to verbal abuse. Our study findings have direct implications for the design of conversational agents.

Electrophoresis (EP) test separates protein components based on their density. Patterns exhibited by this test mostly show very close approximation, making it difficult to examine test results within a short amount of time as it has many variations of patterns and requires a significant amount of knowledge to discern them accurately. To help clinical examiners save time and produce consistent results, a new deeplearning model optimized for EP graphic images was developed. Extending recent work on capsule network, which is a stateof- the-art deep learning model, this study was carried out to develop a best-performing model in classifying abnormal and normal electrophoresis patterns. Instead of extracting features from the image, we used the whole slide image as an input to the classifier. This study used 39,484 electrophoresis 2D graph images and utilized capsule network as the foundation of the deep learning architecture to learn the images without data augmentation. The formulated models were compared for a multitude of performance metrics including accuracy, sensitivity, and specificity. Overall, the study results show that our proposed architecture EP-CapsNet, which combines capsule network with Google’s inception module, is the best performing model, outperforming the baseline and alternative models in almost all comparisons.

Stickers, though similar in appearance to emoji, have distinct characteristics because they often contain animation, diverse gestures, and multiple characters and objects. Stickers can convey richer meaning than emoji, but their complexity and placement constraint may result in miscommunication. In this paper, we aim to understand how people perceive emotion in stickers, as well as how miscommunication related to sticker occurs in actual chat contexts. Toward this goal, we conducted an online survey (n = 87) and in-depth interviews (n = 28) in South Korea. We found emotional and contextual aspects of sticker misinterpretation. In particular, emotion misinterpretation mostly happened due to stickers' ambiguous (multiple) facial/bodily expressions and different perception of dynamism in gestures. In real chat settings, there were also contextual misinterpretations where senders and receivers differently interpret stickers' visual representation/reference, or/and corresponding textual messages. Based on these findings, we provide several practical design implications such as context awareness support in sticker interaction.

What critical factors contribute to knowledge workers’ effective information management and consequent job performance? This paper begins to address this important question by developing a conceptual definition of a new construct called personal information management effectiveness (PIME) and its constituent dimensions. Specifically, we theorize that PIME consists of two underlying dimensions: personal information management motivation (PIMM) and personal information management capability (PIMC). Synthesizing the extant literature on information management and information orientation, we further conceptualize PIMM as having four sub-dimensions of proactiveness, sharing, transparency, and formality, and PIMC as possessing five sub-dimensions of sensing, collecting, organizing, processing, and maintaining. Moreover, we develop a theoretical model that positions PIME as a mediator between two selected individual characteristics (IT self-efficacy and need-for-cognition) and job performance. New measures for PIME dimensions were developed and shown to have strong psychometric properties. The proposed model was empirically tested using data collected from 352 knowledge workers. As theorized, PIME was found to have significant effects on job performance (41%) and fully mediate the effects that the selected individual characteristics have on job performance. Responding to recent calls for advanced research on personal information management, the measures of PIMM and PIMC developed in this study have practical value as research and diagnostic tools and the findings provide useful insights to help organizations improve knowledge workers’ information management practices.

Current movie title retrieval models, such as IMDB, mainly focus on utilizing structured or semi-structured data. However, user queries for searching a movie title are often based on the movie plot, rather than its metadata. As a solution to this problem, our movie title retrieval model proposes a new way of elaborately utilizing associative relations between multiple key terms that exist in the movie plot, in order to improve search performance when users enter more than one keyword. More specifically, the proposed model exploits associative networks of key terms, called knowledge structures, derived from the movie plots. Using the search query terms entered by Amazon Mechanical Turk users as the golden standard, experiments were conducted to compare the proposed retrieval model with the extant state-of-the-art retrieval models. The experiment results show that the proposed retrieval model consistently outperforms the baseline models. The findings have practical implications for semantic search of movie titles particularly, and of online entertainment contents in general.

Collaborative filtering relies on numerical ratings for recommendations. While users consider various aspects of content as a basis of their evaluation, a numeric rating provides only an aggregated report of final assessment. The performance of a collaborative recommender system could be enhanced if