Heterogeneous information networks (HINs) play an indispensable role in a wide range of domain-specific applications, from recommender systems to conversational platforms. Textual HINs are graphs with abundant textual information. Currently, most advanced approaches to mine textual features from HINs follow a “pre-training and fine-tuning” schema, which may cause a “negative transfer” problem since there is a gap between pre-training tasks and downstream tasks. We propose a prompt-learning framework P-HIN that provides a new angle to align textual information and graph information, while narrowing down the gap between the pre-trained models and various downstream tasks. To the best of our knowledge, we are among the first to introduce and exploit the idea of prompt learning to align HIN features and textual features. The proposed framework P-HIN is composed of a text encoder and a graph encoder and uses contrastive learning to align and fuse the graph–text pair. This pre-training operation naturally fits the few-shot learning setting. For the graph encoder, we introduce two graph pre-training tasks, masked node modeling and edge reconstruction, to exploit self-supervised information. During optimization, instead of handcrafted prompts, we use a learnable continuous text that enables more efficient and task-relevant transfer to downstream datasets. We consider a residual connection to use the context from the graph to prompt the text encoder. In experiments, P-HIN consistently and significantly outperforms state-of-the-art alternatives on all real-life datasets.
Large language models have facilitated the development of sophisticated dialogue planner policies. However, their reliance on LLM-based simulation and feedback for policy optimization may introduce preference bias. This article presents the first comprehensive analysis of preference bias in LLM-based dialogue planners. It systematically evaluates four state-of-the-art planning methods across three dialogue domains, utilizing multiple LLM families at varying scales. Our investigation reveals that all tested planners exhibit significant preference bias, favoring narrow strategy sets over maintaining balanced distributions. Notably, larger models paradoxically amplify bias. User simulation emerges as the primary bias driver, while diverse persona simulation proves ineffective as a mitigation strategy. Most concerning, preference bias leads planners toward ethically problematic strategies that achieve short-term success while compromising real-world effectiveness and ethical standards. These findings establish fundamental challenges for responsible deployment of LLM-based dialogue systems and provide crucial insights for developing more reliable and ethically aligned planning approaches.
Generative models for top- \(N\) recommendation have garnered significant attention, with Variational Autoencoder (VAE) emerging as a promising approach for modeling user preferences. Yet, traditional VAE-based models encounter two major challenges: simplistic priors may cause posterior collapse, resulting in ineffective latent structures; meanwhile, limited decoder capabilities hinder the processing of complex data and the extraction of task-relevant information, potentially affecting generative performance. To address these issues, we propose VAEinfox , short for Information-Bottlenecked Variational Autoencoder via Gating Mechanism, a simple yet effective framework that mitigates posterior collapse in VAE. VAEinfox introduces an Adaptive Compressed Representation (ACR) module based on the Information Bottleneck (IB) principle to efficiently and adaptively compress and select latent variables, mitigating posterior collapse. It leverages the Relaxed Bernoulli distribution for differentiable compressed representations and employs a dynamic compression probability selection mechanism with a trainable soft mask activation function to automatically adjust latent variable sparsity during training. Then, a joint training objective maximizes mutual information using the Jensen-Shannon estimator to minimize latent redundancy. In particular, the generative module in VAEinfox incorporates a Mixture-of-Experts (MoE) paradigm, enhancing representation learning and the decoder's ability to process complex data and extract task-relevant information. These dual enhancements improve the model's generative performance while also enhancing its universality and efficiency, particularly when handling implicit feedback. Extensive experiments on eight real-world datasets demonstrate that the proposed framework achieves better results than existing competitors and tackles the specified challenges in top- \(N\) recommendation tasks.
Traditional multimodal video recommendation systems typically focus on semantic synergy across modalities (e.g., text, image, video) by employing attention mechanisms to weight and aggregate features, while overlooking the potential benefits of semantic divergence among modalities, which is often treated as noise. Moreover, in grid-based video recommendation platforms, existing methods fuse implicit video content into video encoders without explicit user visibility, leading to factually inconsistent representations. To address these limitations, we propose ESDvr (E xploring the S ynergetic and D ivergent Potentials of Multimodal Semantics for Feature Fusion-based V ideo R ecommendation ), a novel framework that jointly models cross-modal semantic divergence and synergy. Specifically, we introduce Mixture of Gaussian Blur Gating (MoGG) to capture semantic discrepancies between cover images and texts while integrating them with synergistic features. Additionally, based on cross-modal semantic synergy, we design a learnable threshold-based mechanism to dynamically simulate user hover behaviors, adaptively enhancing video encoding with content tailored to individual users. For user encoding, we leverage recently watched videos to model personalized interests. Furthermore, a contrastive loss is proposed to align video features derived from multimodal semantics with those from user-item interaction histories, ensuring semantic consistency. Rich experiments on MicroLens and MovieLens datasets demonstrate that ESDvr outperforms state-of-the-art (SOTA) baselines, achieving significant improvements of 10.15% (Recall@5) , 10.22% (NDCG@10) .
Conversational search addresses users’ information needs through multi-turn and context-aware interactions. Given that user queries are often ambiguous, the use of clarifying questions can effectively reduce uncertainty and enable a mixed-initiative conversational system. However, current datasets for clarifying questions remain limited in the following three aspects: (1) underrepresented multi-turn conversational data, (2) limited diversity, and (3) heavily reliance on crowdsourcing, thereby suffering from limitations such as high annotation cost. To address these issues, we propose a large language model (LLM)-based three-stage framework that relies on an existing community question answering dataset. It encompasses: (1) extracting essential information from the initial user query with the relevant contextual information, (2) generating clarifying questions paired with corresponding answers, and (3) refining conversations to ensure coherence and a natural conversational flow. We assess our multi-stage method against a baseline that directly prompts LLMs to generate conversations in a single-step process, evaluating on an answer retrieval task using recall, precision, normalized discounted cumulative gain and mean average precision. Results show that our three-stage generation approach consistently outperforms the baseline particularly in recall, while also achieving competitive results across other metrics. Human and automatic evaluations further indicate the high quality of generated conversations and fine-tuning on them improves retrieval performance, highlighting the pipeline’s potential.
Recommender systems are widely used to help users discover content of interest. However, due to their reliance on observational user–item interaction data, they often suffer from data bias. Such biases primarily stem from non-random exposure and users’ self-selection behavior, which distort the data distribution and lead to suboptimal performance of recommendation models. Existing debiasing methods, especially those based on loss reweighting strategies, have shown promising empirical results but still lack solid theoretical guarantees. In particular, they struggle to handle the complex, diverse, and often unidentifiable types of bias encountered in real-world scenarios. In this article, we revisit the problem of unbiased recommendation from the perspective of data bias and propose a unified debiasing framework that mitigates the effect of bias by aligning the distribution of training data with that of unbiased data collected under randomized exposure. We provide a thorough analysis of the theoretical limitations of existing reweighting methods, and we further propose a principled method, mEta Representation Alignment (ERA) , aiming to alleviate the inconsistency between user and item features under different distributions. Extensive experiments on real-world and semi-synthetic datasets demonstrate the effectiveness of ERA.
Website owner identification aims to link websites to their real-world owners, which is crucial for credibility assessment and information provenance in information retrieval and vital for applications in cybersecurity, Internet governance, and digital regulation. Existing approaches for website owner identification primarily rely on querying infrastructure registration records or analyzing webpage content. However, these methods often fail due to incomplete or outdated registration records and sparse webpage content. We observe that inter-website relationships, derived from shared infrastructure data such as primary domains, IP blocks, and geolocations, can provide valuable but underutilized ownership cues. To exploit this insight, we propose MetaRAG, a meta-path-guided dynamic graph retrieval-augmented generation framework that performs reasoning using large language models over ownership-relevant paths in a website-centric knowledge graph. MetaRAG consists of three components: (1) a knowledge graph construction module that integrates infrastructure data and crawled webpage content into a unified representation; (2) a meta-path-guided dynamic reasoning module that constrains retrieval to ownership-relevant meta-paths and adaptively decides whether to retrieve more information or perform inference based on evidence completeness; and (3) a multi-path evidence refinement module that aggregates and scores retrieved paths to suppress noise and distill high-confidence ownership signals. We evaluate MetaRAG on two constructed real-world datasets, achieving up to 6.82% improvement over strong baselines. The results demonstrate the effectiveness of our approach in combining structured web knowledge with large language model-based reasoning for more accurate website owner identification.
Online item ranking systems are crucial for digital marketplaces, directly influencing user experience and vendor revenue. Traditional reputation-based ranking systems weight ratings according to user reputation scores. They have proven effective against manipulation but raise significant ethical concerns regarding user discrimination and privacy, and may raise concerns under emerging regulatory frameworks in certain application contexts. While a user-agnostic ranking system was recently introduced as an alternative approach that uses statistical filtering instead of user reputation scores, its theoretical foundations and resistance to bribing strategies remained unexplored. In this article, we provide the first comprehensive theoretical analysis of user-agnostic ranking system’s robustness properties. We establish formal bounds on bribing resistance by proving three key properties: strategy composition conditions, profitability constraints, and statistical validity requirements. Our theoretical framework demonstrates that profitable bribing strategies in this class of system must satisfy strict statistical conditions, making manipulation more difficult than in reputation-based systems. Experimental evaluation on three real-world datasets confirms our theoretical findings, showing that user-agnostic ranking systems can achieve superior bribing resistance while maintaining comparable effectiveness and efficiency.
For recommender systems in internet platforms, search activities provide additional insights into user interest through query-click interactions with items, and are thus widely used for enhancing personalized recommendation. However, these interacted items have not only transferable features that match users’ interests and are beneficial to the recommendation domain, but also have features related to users’ unique intents in the search domain. Such a domain gap of item features is neglected by most current search-enhanced recommendation methods. They directly incorporate these search behaviors into recommendation, and thus introduce partial negative transfer. Tackling this problem is challenging due to the lack of explicit supervision signals to disentangle features matching search-specific intent or general interest. To address this, we propose ClardRec, a c ounterfactual l e a rning-driven r epresentation d isentanglement framework for search-enhanced recommendation, based on the common belief that a user would click an item under a query not solely because of the item-query match but also due to the item’s query-independent general features (e.g., color or style) that interest the user. These general features exclude the reflection of search-specific intents contained in queries, ensuring a pure match to users’ underlying interests to complement recommendation. We perform the disentanglement based on a counterfactual thinking idea, how would user preferences and query match change for items if we removed their query-related features in search. Specifically, we leverage search queries to construct counterfactual signals to disentangle item representations, isolating only query-independent general features. These representations subsequently enable feature augmentation and data augmentation for the recommendation scenario. Comprehensive experiments on real datasets demonstrate that ClardRec is effective in both collaborative filtering and sequential recommendation scenarios. The source code is available at https://github.com/JJCui96/ClardRec .
Cognitive diagnosis (CD) is an important task in the field of intelligent education, aiming to discover the proficiency of students on knowledge concepts with response logs. In applications, different users of the tutoring system demand for a diagnosis of knowledge concepts at different granularities. However, recent methods assume that the concepts are of the same granularity and use explicit correlations between same-granularity concepts to improve the diagnosis performance. If required for diagnosing multi-granularity concepts, these methods will face diminished performance or partial invalidation. To this end, we make the first attempt for multiple-granularity cognitive diagnosis, i.e., diagnosis on coarse- and fine-grained concepts simultaneously. Specifically, in a skillful way, the same-granularity correlations are captured and embedded into concept representations in view of concept semantics and cross-granularity correlations to model the proficiency influence between concepts implicitly. Then, the specific loss for single-granularity diagnosis and the general loss for the consistency of multi-granularity are designed to train the model jointly, achieving multiple-granularity diagnosis. Extensive experiments demonstrate that our method can achieve state-of-the-art accuracy on both coarse- and fine-grained concepts.
The geographic location of social media users is crucial for understanding user behavior, optimizing advertising, and supporting location-based services such as emergency awareness and event monitoring services. However, existing Twitter user geolocation methods primarily focus on explicit social relationships between users while overlooking implicit location correlations, which affects the accuracy of user geolocation. To address this, this article proposes a Twitter user geolocation method (GeoICMF) based on implicit location correlations and multi-scale feature fusion. GeoICMF introduces a novel location association graph construction method to effectively capture implicit location correlations among users, an innovative multi-scale feature fusion model to dynamically fuse multi-scale features and generate richer user representations, and a pioneering geographic partitioning method to better adapt to user location distributions and enhance geolocation accuracy. Extensive experiments on three real-world datasets demonstrate that GeoICMF outperforms state-of-the-art baseline methods in Twitter user geolocation tasks, validating the effectiveness and superiority of the proposed method.
While previous works on Multi-Behavior Sequential Recommendation (MBSR) have made great efforts to exploit the behavior types, they often fail to fully leverage the behavior and attribute information to learn different aspects of user preferences. Firstly, for predicting user preferences under target behaviors, there is considerable room for improvement by capturing both intra- and inter-behavior item transition relationships at the sequence level, as well as extracting the inter-behavior common and different aspects of a user’s preferences at each timestep based on the contextual sequential information. Secondly, in MBSR, the effect of behavior types on item attributes has not been revealed, which is very important in real-world applications since a behavior such as view representing a user’s weak preferences could not explicitly indicate the user’s interests on an attribute. To address the above two issues, we propose a novel solution called Behavior-aware Attribute-infused Sequential Encoders (BASE) . Specifically, a module named tri-path preference learning is designed to capture user multi-aspect interests via a sequence encoder, a behavior encoder and an attribute encoder. Besides, a behavior-aware dual-granularity contrastive learning module constructs sample pairs in item and attribute granularities to enhance the model’s ability to utilize potential item and attribute information. Extensive empirical studies on three public datasets show that our BASE significantly outperforms various state-of-the-art methods from five different categories. The datasets and our implementation codes are released at https://github.com/Erin-Gr/BASE .
Heterogeneous behavioral data provides comprehensive insights into user intentions and decision-making patterns. Contemporary multi-behavior recommendation models, which leverage such data to infer user preferences, typically capture high-order collaborative signals through graph neural networks on a multi-behavior heterogeneous graph or multiple behavior-specific subgraphs. However, auxiliary behaviors (e.g., view, cart) inherently contain noise that can mislead target behavior (e.g., purchase) prediction, and the incorporation of high-order collaborative signals further amplify such noise. Moreover, these approaches fail to adequately explore cross-behavior item dependencies, leading to inadequate modeling of dependencies across heterogeneous behaviors. To address these limitations, we propose Cross-behavior Item DEpendency modeling for multi-behavior Recommendation (CIDER) , a novel framework that explicitly models item dependencies across multiple types of behaviors for target behavior prediction (e.g., purchase). Specifically, our framework introduces the Hierarchical Behavior Sequence (HBS) , a data structure to systematically organize multi-behavior user–item interactions. Based on the HBS, we design a Cross-behavior Item Dependency Modeling (CIDM) module coupled with a multi-behavior cascading learning scheme to capture item-level dependencies. To enhance the robustness of the representations learned from the CIDM module, we develop an HBS-based denoising module that filters out noise inherent in auxiliary behaviors. Empirical evaluation on three benchmark datasets demonstrates the effectiveness of our model in harnessing multi-behavior data. The implementation is publicly available at https://github.com/SunJianier/CIDER .
Sequential recommendation systems require both temporal efficiency to handle long interaction histories and hierarchical representation to model complex user–item relationships. Existing approaches face a fundamental tension: Mamba-based methods offer linear-time efficiency ( \(\mathcal{O}(L)\) ) but operate in Euclidean space, which distorts hierarchical patterns; hyperbolic models capture taxonomies well but suffer quadratic complexity ( \(\mathcal{O}(L^{2})\) ). To solve this dual challenge , we propose Hyperbolic Mamba (HMamba), the first architecture that unifies curvature-aware state spaces with hyperbolic geometry. Our key insight is that hyperbolic curvature \(\kappa\) simultaneously governs: (1) state transition granularity through \(\mathbf{\bar{A}}=\exp(\Delta\mathbf{A}\odot\mathbf{K}(\kappa))\) and (2) hierarchical distance preservation via \(d_{\mathcal{L}}\propto\sqrt{\kappa}\log(\cdot)\) . This enables joint optimization of efficiency and hierarchy—addressing the previously unsolved problem of deep-long modeling . Experiments show HMamba achieves 3–11% accuracy gains while maintaining 3.2 \(\times\) faster training than attention-based models, establishing a new paradigm for hierarchy-aware sequential recommendation. The code and datasets accompanying our paper are publicly available at https://github.com/CoderPowerBeyond/HMamba .
Content recommender systems (RSs), which recommend content (e.g., news, videos) to users, can inadvertently facilitate the spread of misinformation (e.g., fake news, inaccurate content) due to their ignorance of content credibility. The widespread misinformation can lead to serious consequences, including public confusion and social unrest. Despite the pressing need to mitigate misinformation in content RSs, only a few studies have attempted to tackle this. Moreover, they generally require retraining the entire model from scratch, which is time-consuming and impractical for real-world applications. To this end, we propose MisEraser , a novel misinformation unlearning framework that effectively mitigates misinformation in content RSs without retraining. Specifically, MisEraser consists of three key components: (1) A misinformation disentanglement network , which effectively separates misinformation-specific information from item content while preserving misinformation-free information to maintain recommendation accuracy; (2) A misinformation fusion network , which fuses misinformation-specific signals from both users’ interaction sequences and the entire misinformation item set to enhance the comprehensiveness of misinformation unlearning; and (3) A misinformation erasing network , which effectively removes the fused misinformation-specific signals from trained recommendation models, enabling them to generate misinformation-suppressed and more responsible recommendations. Extensive experiments demonstrate the effectiveness of MisEraser in mitigating misinformation while maintaining high recommendation accuracy and efficiency.
Clarifying questions enable Conversational Search (CS) systems to resolve underspecified queries by eliciting missing information from users. However, how prompting strategies shape the quality of clarifying questions and how such questions should be evaluated at scale remains understudied. We present Automatic GENeration and evaluaTion of Clarifying Questions (AGENT-CQ) , a framework for systematically generating and evaluating clarifying questions and simulated user responses using Large Language Models (LLMs) . To support scalable and multi-perspective evaluation, we introduce CrowdLLM , an LLM-based evaluation paradigm that simulates diverse annotator judgments through distinct evaluator personas. Our experiments span both open-domain CS and a regulatory question-answering setting, allowing us to examine the extent to which clarification strategies generalize across domains with different interaction constraints. Across settings, temperature-variation prompting leads to higher quality clarifying questions than baseline prompting and human-authored questions on several dimensions of the task. In addition, LLM-generated clarifying questions lead to improved downstream retrieval performance than human-authored questions in open-domain search. Together, AGENT-CQ and CrowdLLM provide a practical framework for studying and improving clarification strategies in conversational IR systems.
The increasing specialization and segmentation of modern medical practice, while improving expertise, pose significant challenges in efficiently connecting patients with the right healthcare professionals. The vast array of medical specializations, coupled with sparse data on doctor profiles, overwhelms traditional recommendation algorithms. This study introduces CLEAR-Med: A Contrastive Learning-Enhanced knowledge grAph Recommender designed to match patients with healthcare providers in specific Medical subfields. CLEAR-Med leverages a domain-specific Knowledge Graph (KG) and advanced Contrastive Learning (CL) techniques to capture the nuanced expertise and preferences of doctors, effectively addressing data sparsity and information overload in Online Healthcare Communities (OHCs). The system constructs a comprehensive KG enriched with diverse information, including doctors’ social relationships, professional networks, and specialized attributes derived from OHC data. By embedding key entities and attributes through CL, CLEAR-Med generates robust representations, supported by a flexible attribute encoding module that integrates both efficient LSTMs and powerful Transformer-based models. Its modular prediction layer, featuring options from a stable Multilayer Perceptron (MLP) to an advanced generative diffusion model, then produces highly accurate and personalized recommendation sequences. CLEAR-Med demonstrates superior recommendation performance in baseline comparison experiments, excelling in adaptability and accuracy within OHC settings. Ablation studies confirm the effectiveness of individual components, while further experiments exploring advanced architectures like Transformers and diffusion models highlight the strong balance our framework strikes between performance and computational efficiency. Beyond addressing data sparsity challenges, CLEAR-Med establishes a strong foundation for future advancements in specialized medical matching systems, filling critical research gaps in the domain.
Recommender Systems (RS) have been widely adopted to provide personalized suggestions based on historical user behaviors. However, some users are hesitant to allow RS to learn their preferences at the expense of their privacy information. Therefore, these users prefer to hide their preferences from RS. In this article, we consider this practical yet important question: can privacy-concerned users make RS unavailable to learn their preferences? The challenge lies in achieving this goal while complying with real-world constraints. Normal users still expect accurate recommendations; the scope should target privacy-concerned users. Also, as most companies do not allow users to delete their implicit feedback, the solution cannot rely on data deletion. To this end, we propose a novel ULRec from the perspective of fake interaction generation, a general method for making preferences of privacy-concerned users U n L earnable to personalized Rec ommendation algorithms. First, we formulate the constraints of the data attack based on practical considerations. Then, we define a bi-level optimization process, where the outer loop updates data addition, and the inner loop dynamically updates RS parameters. After that, we propose a loss function that simultaneously considers the requests of both privacy-concerned users and normal users. To ensure the feasible range and model efficiency, we adopt projected gradient descent and automatic differentiation. Finally, extensive experiments on three real-world datasets have demonstrated the effectiveness of our proposed ULRec .
Knowledge tracing aims to model learners’ cognitive state dynamically from interaction sequences to support personalized instructional decisions. While existing methods achieve good prediction accuracy, they often overlook the transfer effects between knowledge concepts (KCs) and their propagation, limiting fine-grained, structured modeling of mastery and overall performance. Although some studies incorporate knowledge transfer using predefined KC similarity graphs, they assume static transfer structures for all learners, neglecting the continuous evolution of transfer abilities due to interventions and self-regulation, and focus solely on KC-hierarchy relations. To overcome these limitations, we propose MHAKT. MHAKT operates across multiple hierarchies of knowledge components through three modules: (1) Transfer perception module utilizes a masked attention mechanism to identify the contribution of the Top- \( K \) most relevant historical interactions for the target knowledge component, dynamically updating the learner-specific transfer structure; (2) Knowledge transfer module employs hypergraph neural networks to comprehensively model many-to-many transfer processes among knowledge components; (3) Cognition update module consolidates new knowledge while applying forgetting mechanisms to update the learner’s cognitive state. Extensive experiments on benchmark datasets demonstrate that MHAKT significantly outperforms thirteen baseline models. In particular, under data sparsity and generalization settings designed to simulate cold-start knowledge components, MHAKT shows strong robustness and maintains superior predictive accuracy. Ablation studies and exploratory experiments further validate the essential contribution of each module, and visualization analyses further reveal MHAKT’s potential for explainable modeling.
Retrieval-Augmented Generation (RAG) improves the performance of Large Language Models (LLMs) by retrieving and integrating relevant information from external knowledge bases, which helps generate more accurate responses. However, RAG is vulnerable to retrieval poisoning attacks , where attackers can induce LLM to produce inaccurate responses by injecting malicious documents into the retrieval process. In this article, we propose ShieldRAG , a novel defense framework designed to counteract retrieval poisoning attacks by reshaping the retrieval embedding space. ShieldRAG leverages a dual-strategy effect realized via a majority-consensus mechanism: ① Push : Implicitly forces the embedding of a user query away from malicious documents by filtering out their minority signals, reducing their influence. ② Pull : Aligns the embedding of a user query closer to that of benign documents, reinforcing accurate retrieval. These strategies work synergistically to preserve retrieval integrity and enhance the quality of LLM-generated responses. Specifically, ShieldRAG operates through three key steps: Sliding Retrieval Explanation Generation , Keyword Aggregation , and Query Targeting Optimization . These three steps collectively ensure the effective integration of information from benign sources while filtering out malicious interference, thereby significantly enhancing the robustness of RAG systems against retrieval poisoning attacks. We evaluate ShieldRAG on four open-domain Question Answering (QA) datasets: Natural Questions, MS-MARCO, HotpotQA, and 2WikiMultiHopQA, using seven representative LLMs. Extensive experiments demonstrate that ShieldRAG significantly improves response accuracy while mitigating adversarial effects, showcasing strong generalization across multiple datasets and LLM architectures.
