IJCAI 2024 Papers — Page 2
International Joint Conference on Artificial Intelligence · 790 papers
Bring Metric Functions into Diffusion Models
Jie An (University of Rochester), Jiebo Luo (University of Rochester)
GenerationDiffusion modelScore-based ModelImage
🎯 What it does: Propose a cascaded diffusion model called Cas-DM, incorporating metric functions such as LPIPS during training to enhance the quality of generated images.
By Fair Means or Foul: Quantifying Collusion in a Market Simulation with Deep Reinforcement Learning
Michael Schlechtinger (University of Mannheim), Heiko Paulheim (University of Mannheim)
Reinforcement LearningFinance Related
🎯 What it does: The study investigates collusive (hyper-competitive) pricing behavior in oligopolistic markets with repeated price competition, implementing and quantifying it using deep reinforcement learning agents (PPO, DQN), and validating its feasibility under different observation space constraints.
Bypassing the ASP Bottleneck: Hybrid Grounding by Splitting and Rewriting
Alexander Beiser (TU Wien), Stefan Woltran (TU Wien)
Computational EfficiencyBenchmark
🎯 What it does: This paper proposes a hybrid grounding technique that combines the traditional ground-and-solve paradigm with body-decoupled grounding (BDG), and designs new rewriting methods for handling aggregate expressions, significantly alleviating the ASP normalization bottleneck. The approach is evaluated in the implemented system NaGG.
C3L: Content Correlated Vision-Language Instruction Tuning Data Generation via Contrastive Learning
Ji Ma (Northwestern Polytechnical University), Yanning Zhang (Northwestern Polytechnical University)
Data SynthesisTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Generate visual-language instruction tuning (VLIT) data using open-source large vision-language models (LVLM), and enhance the content alignment between generated data and images through a content relevance module and a contrastive learning module.
CAP: A Context-Aware Neural Predictor for NAS
Han Ji (Sichuan University), Yanan Sun (Sichuan University)
Neural Architecture SearchGraph Neural NetworkSupervised Fine-TuningContrastive LearningGraphBenchmark
🎯 What it does: Proposes a context-aware neural predictor (CAP), which pretrains graph neural networks through self-supervised contrastive learning on unlabelled architectures, enabling accurate estimation of neural network performance with only a few labeled architectures;
Capturing Knowledge Graphs and Rules with Octagon Embeddings
Victor Charpenay (Mines Saint-Etienne), Steven Schockaert (Cardiff University)
Representation LearningGraph
🎯 What it does: This paper proposes an axis-aligned octagon knowledge graph embedding model (Octagon Embeddings), which uses closed geometric regions to achieve the intersection, union, and combination of relations, addressing the limitations of existing models in expressing closed-loop rules and combinations.
Causality-enhanced Discreted Physics-informed Neural Networks for Predicting Evolutionary Equations
Ye Li (Nanjing University of Aeronautics and Astronautics), Sheng-Jun Huang (Nanjing University of Aeronautics and Astronautics)
Computational EfficiencyBenchmarkPhysics Related
🎯 What it does: Propose a physics-informed neural network (TL-DPINN) based on implicit time discretization, which accelerates solving evolutionary partial differential equations through step-by-step training and transfer learning.
CausalNET: Unveiling Causal Structures on Event Sequences by Topology-Informed Causal Attention
Hua Zhu (Huazhong University of Science and Technology), Bang Liu (Universite de Montreal)
Explainability and InterpretabilityTransformerTime SeriesSequential
🎯 What it does: Constructed the CausalNET model, using Transformer and a trainable causal graph to perform causal structure learning on event sequences.
CausVSR: Causality Inspired Visual Sentiment Recognition
Xinyue Zhang (East China Normal University), Guitao Cao (East China Normal University)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: Proposes a visual emotion recognition framework called CausVSR based on emotional causality theory, which jointly learns the generation of pseudo-emotional regions and emotion classification through structural causal models and front-door adjustment.
Certified Policy Verification and Synthesis for MDPs under Distributional Reach-Avoidance Properties
S. Akshay (Indian Institute of Technology Bombay), Đorđe Žikelić (Singapore Management University)
Reinforcement LearningTabular
🎯 What it does: For Markov decision processes (MDP) under a distributional orientation, a method is proposed that simultaneously verifies and synthesizes strategies while providing formal proofs (certificates), specifically addressing distributed reach-avoid properties.
CF-Deformable DETR: An End-to-End Alignment-Free Model for Weakly Aligned Visible-Infrared Object Detection
Haolong Fu (Hunan University), Zhiyong Li (Hunan University)
Object DetectionTransformerContrastive LearningImageMultimodality
🎯 What it does: Designed an end-to-end alignment-agnostic visible-infrared object detection model called CF-Deformable DETR, achieving cross-modal point-level feature extraction and fusion directly on weakly aligned data.
ChatSpot: Bootstrapping Multimodal LLMs via Precise Referring Instruction Tuning
Liang Zhao (MEGVII Technology), Xiangyu Zhang (MEGVII Technology)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Propose ChatSpot, a unified end-to-end multimodal large language model that supports multiple precise pointing interactions (clicking, box selection, polygon drawing) and achieves fine-grained vision-language question answering and instruction following.
CIC: A Framework for Culturally-Aware Image Captioning
Youngsik Yun (Dongguk University), Jihie Kim (Dongguk University)
GenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Propose the CIC framework, which first generates cultural questions, uses VQA to extract cultural visual elements, and then employs an LLM to generate image captions containing cultural information.
Class-consistent Contrastive Learning Driven Cross-dimensional Transformer for 3D Medical Image Classification
Qikui Zhu (Case Western Reserve University), Shuo Li (Case Western Reserve University)
ClassificationTransformerContrastive LearningImageBiomedical Data
🎯 What it does: Proposed a cross-dimensional Transformer (CdTransformer) and class-consistent contrastive learning (CcCL) for 3D medical image classification, addressing the issues of traditional Transformers in 3D data, including loss of spatial structure, high computational complexity, and weak representation learning.
Class-Specific Semantic Generation and Reconstruction Learning for Open Set Recognition
Liu Haoyang (Hefei University Of Technology), Xuegang Hu (Hefei University Of Technology)
ClassificationRecognitionAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Propose an open-set recognition framework based on class-specific semantic generation and reconstruction learning (CSGRL), which generates boundary samples of known classes through a generator and uses an additional autoencoder to fit the joint boundary of the unknown space, thus considering both known and unknown classes in the reconstruction error to achieve more robust unknown sample rejection and known class classification.
CLIP-FSAC: Boosting CLIP for Few-Shot Anomaly Classification with Synthetic Anomalies
Zuo Zuo (Xi'an Jiaotong University), Zongze Wu (Xi'an Jiaotong University)
Data SynthesisAnomaly DetectionTransformerSupervised Fine-TuningContrastive LearningImageMultimodality
🎯 What it does: For the anomaly classification task in industrial scenarios with limited normal samples, the CLIP-FSAC framework is proposed, which utilizes the pre-trained CLIP model, two-stage trained image/text adapters, and combines cross-modal attention mechanisms from image to text for fine-tuning on few-shot anomaly classification.
ClothPPO: A Proximal Policy Optimization Enhancing Framework for Robotic Cloth Manipulation with Observation-Aligned Action Spaces
Libing Yang (East China Normal University), Long Chen (Hong Kong University of Science and Technology)
OptimizationRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: Through a two-stage framework combining pre-training and PPO, enabling robots to unfold fabric based on a pixel-level action space.
CLR-Face: Conditional Latent Refinement for Blind Face Restoration Using Score-Based Diffusion Models
Maitreya Suin (Johns Hopkins University), Rama Chellappa (Johns Hopkins University)
RestorationDiffusion modelScore-based ModelAuto EncoderImageStochastic Differential Equation
🎯 What it does: Proposes a latent space refinement method based on conditional diffusion for blind facial restoration of severely degraded face images under the VQGAN framework.
CMACE: CMAES-based Counterfactual Explanations for Black-box Models
Xudong Yin (Ant Group), Yao Yang (Zhejiang Lab)
OptimizationExplainability and InterpretabilityTabular
🎯 What it does: This paper proposes a model-agnostic adversarial explanation method called CMACE based on CMA-ES for generating optimal counterfactual explanations for black-box models.
CMMU: A Benchmark for Chinese Multi-modal Multi-type Question Understanding and Reasoning
Zheqi He (Beijing Academy Of Artificial Intelligence), Hua Huang (Beijing Normal University)
Large Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Designed and released the CMMU Chinese multimodal and multi-type question-answering benchmark, which includes multiple disciplines and various difficulty levels of multiple-choice, multiple-select, and fill-in-the-blank questions.
CoAtFormer: Vision Transformer with Composite Attention
Zhiyong Chang (Peking University), Yan Wang (Zuoyebang)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposed the CoAtFormer visual Transformer and designed a composite attention module (dynamic channel attention + efficient spatial attention)
CoCoG: Controllable Visual Stimuli Generation Based on Human Concept Representations
Chen Wei (Southern University of Science and Technology), Quanying Liu (Southern University of Science and Technology)
GenerationData SynthesisExplainability and InterpretabilityVision Language ModelDiffusion modelImageBenchmark
🎯 What it does: Built a concept encoder to extract interpretable low-dimensional concept embeddings from visual stimuli, and trained a two-stage diffusion decoder using these embeddings to generate constrained images in the concept space, achieving controllable visual stimulus generation based on human concept representations.
CoFInAl: Enhancing Action Quality Assessment with Coarse-to-Fine Instruction Alignment
Kanglei Zhou (Beihang University), Xiaohui Liang (Beihang University)
ClassificationTransformerSupervised Fine-TuningVision-Language-Action ModelVideo
🎯 What it does: Treat action quality assessment (AQA) as a coarse-to-fine hierarchical classification task, proposing the CoFInAl framework that leverages coarse-level prototype learning and fixed ETF sub-grade prototypes to achieve coarse-to-fine instruction alignment, thereby improving evaluation performance.
Combinatorial Games with Incomplete Information
Junkang Li (NukkAI), Véronique Ventos (NukkAI)
Optimization
🎯 What it does: Propose a new combinatorial game with incomplete information (CGII) model and prove that its computational complexity in solving optimal strategies is comparable to that of general extensive-form games.
Combinatorial Routing for Neural Trees
Jiahao Li (Guangdong University of Technology), Yuguang Yan (Guangdong University of Technology)
Neural Architecture SearchImage
🎯 What it does: Proposes a neural tree method called CombRo based on multicast routing, which automatically searches and generates high-performance neural network architectures from a mother tree network.
Common-Individual Semantic Fusion for Multi-View Multi-Label Learning
Gengyu Lyu (Beijing University of Technology), Songhe Feng (Beijing Jiaotong University)
ClassificationOptimizationRepresentation LearningMultimodality
🎯 What it does: Propose a novel multi-view multi-label learning method called CISF, which leverages semantic-level fusion to separately capture shared semantics between views and individual semantics unique to each view, and constructs a multi-label classifier through low-rank and sparse constraints as well as dynamic label associations.
Comparing Ways of Obtaining Candidate Orderings from Approval Ballots
Théo Delemazure (Universite Paris Dauphine), Magdalena Tydrichova (Paris Saclay University)
Data SynthesisTabular
🎯 What it does: This paper proposes and compares five candidate axis sorting rules based on approval ballots, aiming to find the candidate order that best approximates an interval structure; it also conducts axiomatic analysis, complexity proofs, and empirical evaluations of these rules;
CompetEvo: Towards Morphological Evolution from Competition
Kangyao Huang (Tsinghua University), Huaping Liu (Tsinghua University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: Propose a competitive evolution framework called CompetEvo, enabling robots to co-evolve their body morphology and combat strategies through self-play.
Compilation and Fast Model Counting beyond CNF
Alexis de Colnet (TU Wien), Tianwei Zhang (TU Wien)
Computational Efficiency
🎯 What it does: Proposes a fixed-parameter tractable (FPT) knowledge compilation method that compiles constraint systems satisfying specific 'Slim' properties into d-DNNF circuits, enabling linear-time model counting.
Computational Aspects of Progression for Temporal Equilibrium Logic
Thomas Eiter (Technische Universitaet Wien), Davide Soldà (Technische Universitaet Wien)
🎯 What it does: This paper investigates the monitoring complexity of temporal equilibrium logic (TEL) and temporal here-there logic (THT), proposes a three-valued monitoring method based on progression, namely P(THT) and PTEL, and further introduces an incremental progression method PincTEL. It proves their computational complexity and polynomial-time feasibility in acceptable subclasses (e.g., ordinary programs and head-cycle-free programs).
Computational Complexity of Verifying the Group No-show Paradox
Farhad Mohsin (College of the Holy Cross), Lirong Xia (Rensselaer Polytechnic Institute)
Computational EfficiencyTabularBenchmark
🎯 What it does: Studied the computational complexity of verifying the group non-attendance paradox (GNSP) under various common voting rules (Copeland, Maximin, STV, and all Condorcet-based integer scoring rules), and proposed two classes of algorithms based on integer linear programming (ILP) and breadth-first search (BFS) to achieve this verification;
Computing Optimal Equilibria in Repeated Games with Restarts
Ratip Emin Berker (Carnegie Mellon University), Vincent Conitzer (Carnegie Mellon University)
OptimizationTabular
🎯 What it does: In anonymous multi-player repeated games, a strategy sequence with a 'restart' mechanism is proposed to ensure stability, and the study focuses on how to compute its optimal form.
CONC: Complex-noise-resistant Open-set Node Classification with Adaptive Noise Detection
Qin Zhang (Shenzhen University), Junyang Chen (Shenzhen University)
ClassificationAnomaly DetectionGraph Neural NetworkAuto EncoderContrastive LearningGraph
🎯 What it does: Propose a framework named CONC that performs robust open-set node classification on graph data containing out-of-distribution (OOD) noise and in-distribution (IND) label noise.
Concentration Tail-Bound Analysis of Coevolutionary and Bandit Learning Algorithms
Per Kristian Lehre (University of Birmingham), Shishen Lin (University of Birmingham)
OptimizationGraph
🎯 What it does: This paper proposes a new drift theorem that provides exponential tail bounds for scenarios including positive drift, weak drift, zero drift, and even negative drift, thereby analyzing the first hitting time of random algorithms;
Concept-Level Causal Explanation Method for Brain Function Network Classification
Jinduo Liu (Beijing University of Technology), Junzhong Ji (Beijing University of Technology)
ClassificationConvolutional Neural NetworkGraphBiomedical Data
🎯 What it does: Propose a causal concept-based explanation method named CLCEM, which automatically extracts interpretable ROI concepts from brain functional networks and drives classification decisions through concept contributions;
Conflict-Alleviated Gradient Descent for Adaptive Object Detection
Wenxu Shi (Chongqing University), Bochuan Zheng (China West Normal University)
Object DetectionDomain AdaptationAutonomous DrivingConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a Conflict-Aware Gradient Descent (CAGrad) method to address gradient conflicts that arise during optimization in unsupervised domain adaptation for object detection, and designs an equivalent model named CAGrad+DAOD to enhance detection performance.
Constrained Intrinsic Motivation for Reinforcement Learning
Xiang Zheng (City University of Hong Kong), Cong Wang (City University of Hong Kong)
Reinforcement LearningBenchmark
🎯 What it does: This paper proposes Constrained Intrinsic Motivation (CIM), designing new intrinsic goals and adaptive coefficients for Reward-Free Pre-Training (RFPT) and Exploration with Intrinsic Motivation (EIM) tasks, achieving more efficient skill discovery and exploration.
ConstrainedZero: Chance-Constrained POMDP Planning Using Learned Probabilistic Failure Surrogates and Adaptive Safety Constraints
Robert J. Moss (Stanford University), Mykel J. Kochenderfer (Stanford University)
Optimization
🎯 What it does: This paper proposes ConstrainedZero, an extension of the BetaZero POMDP planning algorithm, designed to address chance-constrained POMDP (CC-POMDP) problems. It can simultaneously learn value functions, action policies, and failure probabilities in the Bayesian space, and employs an adaptive safety threshold in online Monte Carlo Tree Search (MCTS) for safe decision-making.
Constructive Interpolation and Concept-Based Beth Definability for Description Logics via Sequents
Timothy S. Lyon (Technische Universität Dresden), Jonas Karge (Technische Universität Dresden)
🎯 What it does: Propose a constructive method based on inference trees, using a sequent system to prove and compute the concept base Beth definability (CBP) and its interpolation for description logic RIQ;
Contextualized Speech Recognition: Rethinking Second-Pass Rescoring with Generative Large Language Models
Yixuan Tang (National University of Singapore), Anthony K. H. Tung
RecognitionTransformerLarge Language ModelPrompt EngineeringTextAudio
🎯 What it does: Propose a secondary generation framework based on large language models (LLMs) for contextual semantic ASR tasks, replacing traditional secondary re-ranking methods;
Continual Compositional Zero-Shot Learning
Yang Zhang (Beijing Jiaotong University), Jiazheng Yuan (Beijing Open University)
ClassificationRecognitionKnowledge DistillationConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Propose and address the task of Continuous Compositional Zero-Shot Learning (CCZSL), enabling models to recognize unseen attribute-object combinations while continuously expanding the primitive set.
Continual Multi-Objective Reinforcement Learning via Reward Model Rehearsal
Lihe Li (Nanjing University), Lei Yuan (Nanjing University)
Recurrent Neural NetworkReinforcement LearningBenchmark
🎯 What it does: Proposed a continuous multi-objective reinforcement learning (CMORL) framework and designed the CORE3 algorithm to address the problem of objectives evolving with learning stages.
Continual Multi-View Clustering with Consistent Anchor Guidance
Chao Zhang (Nanjing University), Huaxiong Li (Nanjing University)
OptimizationComputational EfficiencyRepresentation LearningMultimodality
🎯 What it does: Proposed a consistent Anchor-guided continuous multi-view clustering method (ACMVC), which employs a two-phase (initialization and continuous learning) alternating optimization of a low-rank Anchor graph model to address clustering problems for streaming multi-view data.
Continual Multimodal Knowledge Graph Construction
Xiang Chen (Zhejiang University), Huajun Chen (Zhejiang University)
Representation LearningTransformerMultimodalityBenchmark
🎯 What it does: Proposed the 'Continual Multimodal Knowledge Graph Construction (MKGC)' framework, and constructed two incremental benchmarks (IMNER and IMRE) for this task, as well as the Multimodal Stable-Plastic Transformer (MSPT) model for continual learning;
Contract Scheduling with Distributional and Multiple Advice
Spyros Angelopoulos (Sorbonne University), Bertrand Simon (CNRS)
Optimization
🎯 What it does: This paper proposes and analyzes a learning-enhanced framework using distributed prediction and multiple prediction in contract scheduling problems, designing a scheduling algorithm that achieves optimal consistency while maintaining 4-robustness.
Contrastive and View-Interaction Structure Learning for Multi-view Clustering
Jing Wang (Beijing Jiaotong University), Songhe Feng (Beijing Jiaotong University)
Representation LearningGraph Neural NetworkAuto EncoderContrastive LearningGraph
🎯 What it does: Propose a contrastive learning and view interaction structural learning framework named SERIES for multi-view clustering.
Contrastive General Graph Matching with Adaptive Augmentation Sampling
Jianyuan Bo (Singapore Management University), Yuan Fang (Singapore Management University)
RecognitionRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Proposes an unsupervised graph matching framework GCGM based on contrastive learning, introducing an adaptive sampling mechanism BiAS to perform graph matching without using any labels or additional information.
Contrastive Learning Drug Response Models from Natural Language Supervision
Kun Li (Wuhan University), Wenbin Hu (Wuhan University)
Drug DiscoveryConvolutional Neural NetworkGraph Neural NetworkTransformerPrompt EngineeringContrastive LearningTextBiomedical Data
🎯 What it does: Propose the CLDR framework, which converts regression labels into text and jointly encodes them with drug-cell information, using contrastive learning to achieve feature learning for drug response prediction.
Contrastive Learning Is Not Optimal for Quasiperiodic Time Series
Adrian Atienza (Technical University of Denmark), Sadasivan Puthusserypady (Technical University of Denmark)
ClassificationContrastive LearningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: Propose a contrastive-free self-supervised learning framework named DEAPS, specifically designed for quasi-periodic time series (e.g., ECG).
Contrastive Representation Learning for Self-Supervised Taxonomy Completion
Yuhang Niu (Nankai University), Xiaojie Yuan (Nankai University)
ClassificationRepresentation LearningTransformerPrompt EngineeringContrastive LearningText
🎯 What it does: Proposes the CoSTC framework, which employs contrastive learning for two-stage pre-training: first performing instance contrast between query-query and position-position within the same view, then conducting query-position contrast matching across views through diversity and difficulty sampling to enhance the quality of representations for self-supervised lexical classification.
Contrastive Transformer Cross-Modal Hashing for Video-Text Retrieval
Xiaobo Shen (Nanjing University of Science and Technology), Yuhui Zheng (Qinghai Normal University)
RetrievalCompressionTransformerContrastive LearningVideoTextBenchmark
🎯 What it does: Propose a contrastive cross-modal hashing model CTCH based on bidirectional Transformer for video-text retrieval.
Contrastive Transformer Masked Image Hashing for Degraded Image Retrieval
Xiaobo Shen (Nanjing University of Science and Technology), Yuhui Zheng (Qinghai Normal University)
RetrievalTransformerContrastive LearningImage
🎯 What it does: Propose an unsupervised contrastive learning combined with mask reconstruction ViT hashing method (CTMIH) to address the image degradation retrieval problem.
Convexity Certificates for Symbolic Tensor Expressions
Paul G. Rump (Friedrich Schiller University Jena), Joachim Giesen (Friedrich Schiller University Jena)
Optimization
🎯 What it does: Proposed a general symbolic Hessian method to prove convexity for differentiable functions with tensor inputs of any order.
Cooperation and Control in Delegation Games
Oliver Sourbut (University of Oxford), Harriet Wood (University of Oxford)
Tabular
🎯 What it does: Systematically studied control and cooperation issues in multi-agent principal games, dividing them into four dimensions: alignment and capability, and providing theoretical and experimental analysis.
Core-Structures-Guided Multi-Modal Classification Neural Architecture Search
Pinhan Fu (Shanxi University), Yuhua Qian (National University of Defense Technology)
ClassificationNeural Architecture SearchMultimodality
🎯 What it does: Propose a core structure guided multi-modal classification neural architecture search (CSG-NAS), which first discovers the core structure through evolutionary search in a simplified core structure subspace, then performs local search for optimal fusion architectures within its neighborhood, and introduces adaptive knowledge inheritance to improve search efficiency.
Correct and Optimal: The Regular Expression Inference Challenge
Mojtaba Valizadeh (University of Sussex), Martin Berger (University of Sussex)
OptimizationTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the Regular Expression Inference (REI) challenge, defining the task of learning minimal regular expressions as a machine learning benchmark, releasing four large-scale datasets and providing baseline models.
Counterfactual User Sequence Synthesis Augmented with Continuous Time Dynamic Preference Modeling for Sequential POI Recommendation
Lianyong Qi (China University of Petroleum (East China)), Wanchun Dou (Nanjing University)
Recommendation SystemRecurrent Neural NetworkContrastive LearningSequentialOrdinary Differential Equation
🎯 What it does: This study proposes a continuous-time dynamic preference model combining GRU with neural ODE, and enhances POI recommendation performance through counterfactual user sequence synthesis.
Couples Can Be Tractable: New Algorithms and Hardness Results for the Hospitals/Residents Problem with Couples
Gergely Csáji (ELTE Eotvs Lor' and University), James Trimble (University of Glasgow)
OptimizationComputational EfficiencyGraph
🎯 What it does: Introduce spousal relationships in the hospital/resident matching problem, propose multiple polynomial-time algorithms and NP-hardness proofs, demonstrating that approximately feasible stable matching can be achieved under sub-responsive, sub-complete preferences with capacity adjustments limited to ±1.
CPa-WAC: Constellation Partitioning-based Scalable Weighted Aggregation Composition for Knowledge Graph Embedding
Sudipta Modak (University of Windsor), Esam Abdel-Raheem (University of Windsor)
Representation LearningGraph Neural NetworkGraph
🎯 What it does: This study proposes a knowledge graph embedding framework CPa-WAC based on Louvain/Leiden constellation partitioning and lightweight GCN, and merges the embeddings of subgraphs through a global decoder to achieve efficient training and high-quality link prediction.
Cross-Domain Feature Augmentation for Domain Generalization
Yingnan Liu (National University of Singapore), Wynne Hsu (National University of Singapore)
Domain AdaptationAuto EncoderImage
🎯 What it does: Propose a cross-domain feature augmentation method called XDomainMix for domain generalization tasks, generating diverse, domain-invariant features by decomposing features into class and domain semantic components.
Cross-Domain Few-Shot Semantic Segmentation via Doubly Matching Transformation
Jiayi Chen (Nanjing University of Aeronautics and Astronautics), Jie Qin (Nanjing University of Aeronautics and Astronautics)
SegmentationDomain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes a Dual Matching Transformation Network (DMTNet) to address the cross-domain few-shot semantic segmentation problem.
Cross-modal Generation and Alignment via Attribute-guided Prompt for Unsupervised Text-based Person Retrieval
Zongyi Li (Huazhong University of Science and Technology), Shijuan Huang (Huazhong University of Science and Technology)
RetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes the GAAP framework, which generates pseudo-text on unlabeled images using attribute-guided prompts and achieves cross-modal retrieval through multi-level alignment.
Cross-Problem Learning for Solving Vehicle Routing Problems
Zhuoyi Lin (Agency for Science, Technology and Research), Senthilnath Jayavelu (Agency for Science, Technology and Research)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningReinforcement LearningGraphBenchmark
🎯 What it does: Studied cross-problem learning by leveraging a pre-trained TSP Transformer backbone and fine-tuning with lightweight problem-specific modules for different VRP variants (such as OP, PCTSP, CVRP) to learn high-quality neural heuristics.
Cross-Scale Domain Adaptation with Comprehensive Information for Pansharpening
Meiqi Gong (Wuhan University), Jiayi Ma (Wuhan University)
Domain AdaptationConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: This paper proposes a cross-scale domain adaptive fusion (pansharpening) method to synthesize high-resolution multispectral images by combining multispectral imagery with full-resolution panchromatic imagery.
Cross-Talk Reduction
Zhong-Qiu Wang (Southern University of Science and Technology), Shinji Watanabe (Carnegie Mellon University)
RestorationConvolutional Neural NetworkAudio
🎯 What it does: This paper proposes a cross-talk reduction task and designs an unsupervised/weakly supervised CTRnet model. By stacking near-field and far-field mixed signals as input, the model uses neural networks to estimate each speaker's near-field speech and predicts and corrects cross-talk through linear convolution.
Cross-View Contrastive Fusion for Enhanced Molecular Property Prediction
Yan Zheng (University of Electronic Science and Technology of China), Lifang He (Lehigh University)
Representation LearningDrug DiscoveryRecurrent Neural NetworkGraph Neural NetworkContrastive LearningGraphSequentialBiomedical Data
🎯 What it does: Propose a cross-perspective contrastive fusion framework MolFuse for molecular property prediction.
Cross-View Diversity Embedded Consensus Learning for Multi-View Clustering
Chong Peng (Qingdao University), Qiang Cheng (University of Kentucky)
OptimizationRepresentation LearningImageBenchmark
🎯 What it does: Propose a new multi-view clustering method called CCL-MVC, which constructs a cross-sequential neighbor tensor, recovers a low-rank essential tensor, and embeds cross-view diversity to learn a consensus affinity matrix for clustering.
Cutting the Black Box: Conceptual Interpretation of a Deep Neural Net with Multi-Modal Embeddings and Multi-Criteria Decision Aid
Nicolas Atienza (Thales Research and Technology), Michele Sebag (Universite Paris Saclay)
Explainability and InterpretabilityKnowledge DistillationConvolutional Neural NetworkTransformerVision Language ModelAuto EncoderImage
🎯 What it does: Through the CB2 framework, pixel-level depth visual models are mapped to a concept space, and an interpretable 2-additive hierarchical Choquet integral (MCDA) student model is trained to explain the teacher model's decisions.
D3ETR: Decoder Distillation for Detection Transformer
Xiaokang Chen (Peking University), Gang Zeng (Peking University)
Object DetectionKnowledge DistillationTransformerImageBenchmark
🎯 What it does: Perform knowledge distillation on the decoder of a DETR-based detector, proposing the MixMatcher matching strategy and implementing Decoder Distillation (DETR).
DANCE: Dual-View Distribution Alignment for Dataset Condensation
Hansong Zhang (Chinese Academy of Sciences), Shiming Ge (Chinese Academy of Sciences)
Data SynthesisKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Propose a dual-view distribution alignment (DANCE) method for dataset distillation, enhancing distribution matching effectiveness by introducing pseudo long-term distribution alignment (PLTDA) and distribution calibration.
DarkFed: A Data-Free Backdoor Attack in Federated Learning
Minghui Li (Huazhong University of Science and Technology), Yichen Wang (National Engineering Research Center for Big Data Technology and System)
Federated LearningAdversarial AttackImage
🎯 What it does: Propose DarkFed, an attack scheme that achieves backdoor injection in federated learning without real data, utilizing shadow datasets and attribute simulation techniques to successfully implant backdoors in simulated clients even without task-related data.
Data Complexity in Expressive Description Logics with Path Expressions
Bartosz Bednarczyk (Technische Universitaet Dresden)
🎯 What it does: Investigated the satisfiability problem of ZOIQ and its decidable sublanguages (ZOQ, ZOI, ZIQ) under ABox changes only, proving that their data complexity is NP-complete, and the satisfiability problem of rooted queries has a complexity of coNEXPTIME-complete.
DBPNet: Dual-Branch Parallel Network with Temporal-Frequency Fusion for Auditory Attention Detection
Qinke Ni (Anhui University), Zhao Lv (Anhui University)
ClassificationConvolutional Neural NetworkTransformerBiomedical Data
🎯 What it does: Propose the DBPNet dual-branch parallel network, integrating time-domain and frequency-domain features to achieve auditory attention decoding.
DCDet: Dynamic Cross-based 3D Object Detector
Shuai Liu (Sun Yat-sen University), Kai Huang (Sun Yat-sen University)
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: Proposed a dynamic label assignment (DCLA) based on cross-shaped regions and a rotation-weighted IoU (RWIoU) loss, constructing a single-stage 3D object detection framework named DCDet;
Deciphering the Projection Head: Representation Evaluation Self-supervised Learning
Jiajun Ma (Hong Kong University of Science and Technology), Wenjia Wang (Hong Kong University of Science and Technology)
Representation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Analyze the role of projection heads in self-supervised learning, revealing that they mainly achieve uniformity, and propose Representation Evaluation Design (RED), which enhances downstream task performance and robustness to unseen augmentations and OOD (out-of-distribution) by establishing shortcut connections between the projection head and representation vectors and reweighting sample similarity.
Decoupled Invariant Attention Network for Multivariate Time-series Forecasting
Haihua Xu (University of Macau), Pengyang Wang (University of Macau)
Graph Neural NetworkTransformerTime Series
🎯 What it does: Proposed the Decoupled Invariant Attention Network (DIAN), which decouples invariant and variant patterns in multivariate time series through a variable-invariant attention mechanism in spatial and temporal dimensions, and enhances prediction robustness by generating intervention samples via a time intervention mechanism.
Decoupling Breaks Data Barriers: A Decoupled Pre-training Framework for Multi-intent Spoken Language Understanding
Libo Qin (Central South University), Wanxiang Che
TransformerLarge Language ModelSupervised Fine-TuningTextAudio
🎯 What it does: Proposes a decoupled pre-training framework DPF for multi-intent speech understanding, conducting two-stage pre-training using a large amount of unlabeled multi-intent data.
Deep Embedding Clustering Driven by Sample Stability
Zhanwen Cheng (Shanxi University), Yuhua Qian (Shanxi University)
Representation LearningConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: Propose a sample stability-driven deep embedded clustering method called DECS, eliminating the dependence on pseudo-labels in traditional clustering;
Deep Frequency Derivative Learning for Non-stationary Time Series Forecasting
Wei Fan (University of Oxford), Ning An (Hefei University of Technology)
Representation LearningConvolutional Neural NetworkTime Series
🎯 What it does: This paper proposes a frequency-domain based framework for non-stationary time series prediction called DERITS, which applies frequency derivative transformation (FDT) to compute first or higher-order derivatives across the entire spectrum, achieving a more stationary frequency-domain representation. Based on this, an adaptive Fourier convolution network (OFCN) is designed, employing a parallel stacked structure to realize multi-order derivative feature fusion, ultimately restoring prediction results in the time domain.
Deep Hierarchical Graph Alignment Kernels
Shuhao Tang (Tongji University), Wei Ye (Tongji University)
ClassificationRepresentation LearningLarge Language ModelGraph
🎯 What it does: Proposed Deep Hierarchical Graph Alignment Kernels (DHGAK), achieving more accurate graph similarity computation by clustering and aligning hierarchical substructures in a deep embedding space.
Deep Multi-Dimensional Classification with Pairwise Dimension-Specific Features
Teng Huang (Southeast University), Min-Ling Zhang (Southeast University)
ClassificationImageTextBiomedical Data
🎯 What it does: Propose a deep multi-dimensional classification method called PIST, which predicts by learning dual-dimensional features and considering dependencies between class variables.
Deep Neural Networks via Complex Network Theory: A Perspective
Emanuele La Malfa (University of Oxford), Vito Latora (University of Catania)
Explainability and InterpretabilityConvolutional Neural NetworkRecurrent Neural NetworkImageReview/Survey Paper
🎯 What it does: Proposed a unified complex network theory (CNT) framework to treat deep neural networks (DNNs) as graphs and analyze their structure and learning dynamics through newly defined metrics that account for the influence of input data;
Delocate: Detection and Localization for Deepfake Videos with Randomly-Located Tampered Traces
Juan Hu (Hunan University), Mike Zheng Shou (National University of Singapore)
Anomaly DetectionMeta LearningTransformerAuto EncoderContrastive LearningVideo
🎯 What it does: Designed and implemented a two-stage model called Delocate, which first uses self-supervised random masking of facial regions to restore real facial consistency, and then employs a mapping module combined with an Encoder-Decoder for localization learning, enabling both detection of Deepfake videos and localization of forged regions.
Delve into Base-Novel Confusion: Redundancy Exploration for Few-Shot Class-Incremental Learning
Haichen Zhou (Huazhong University of Science and Technology), Kui Xiao (Hubei University)
ClassificationRepresentation LearningMeta LearningImage
🎯 What it does: To address the confusion between base and new classes in few-shot incremental learning, the Redundancy Disentanglement and Integration (RDI) method is proposed, which separates and reintegrates label-irrelevant redundancies in the feature space to compress the base class space and reserve a buffer for new classes.
Denoising Diffusion-Augmented Hybrid Video Anomaly Detection via Reconstructing Noised Frames
Kai Cheng (Fudan University), Rui Feng (Fudan University)
Anomaly DetectionTransformerDiffusion modelAuto EncoderOptical FlowVideo
🎯 What it does: Designed and implemented the DHVAD framework, combining the denoising diffusion reconstruction unit (DRU) with the frame prediction unit (FPU) for unsupervised video anomaly detection.
Denoising-Aware Contrastive Learning for Noisy Time Series
Shuang Zhou (Hong Kong Polytechnic University), Korris Chung (Hong Kong Polytechnic University)
ClassificationRepresentation LearningTransformerAuto EncoderContrastive LearningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: Propose an end-to-end denoising-aware contrastive learning framework DECL, which automatically selects an appropriate denoising method for each noisy time series and suppresses noise through contrastive learning in representation learning.
DenseKoopman: A Plug-and-Play Framework for Dense Pedestrian Trajectory Prediction
Xianbang Li (Beihang University), Liang Xu (Beihang University)
Autonomous DrivingConvolutional Neural NetworkRecurrent Neural NetworkVideo
🎯 What it does: Propose the DenseKoopman framework, which utilizes the Koopman operator to map nonlinear pedestrian trajectories in dense scenes to a high-dimensional linear space, enabling linearized trajectory prediction;
Deriving Provably Correct Explanations for Decision Trees: The Impact of Domain Theories
Gilles Audemard (University of Artois), Nicolas Szczepanski (University of Artois)
Explainability and Interpretability
🎯 What it does: This paper studies the computational complexity of calculating local explanations for decision trees (including inductive and contrastive explanations) under domain theories (i.e., logical relationships describing Boolean conditions in decision trees).
Design a Win-Win Strategy That Is Fair to Both Service Providers and Tasks When Rejection Is Not an Option
Yohai Trabelsi (Bar-Ilan University), Sarit Kraus (Bar-Ilan University)
Autonomous DrivingOptimizationTabularTime Series
🎯 What it does: Proposed two online matching models based on minimax fairness, and designed algorithms using linear programming and heuristics to achieve fair allocation between service providers and tasks without allowing task rejections.
Designing Behavior-Aware AI to Improve the Human-AI Team Performance in AI-Assisted Decision Making
Syed Hasan Amin Mahmood (Purdue University), Ming Yin (Purdue University)
Reinforcement Learning from Human FeedbackConvolutional Neural NetworkImageTabular
🎯 What it does: Studied how to train AI models by considering humans' confidence levels in AI-assisted decision-making, proposing a behavior-aware AI approach to improve the overall accuracy of human-AI team decisions.
Detecting and Understanding Vulnerabilities in Language Models via Mechanistic Interpretability
Jorge García-Carrasco (University of Alicante), Juan Trujillo (University of Alicante)
Explainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Propose a framework based on mechanism interpretability, combined with adversarial sample generation, to systematically locate and understand adversarial vulnerability components in language models for specific tasks.
Detector Collapse: Backdooring Object Detection to Catastrophic Overload or Blindness in the Physical World
Hangtao Zhang (Huazhong University of Science and Technology), Chao Chen (RMIT University)
Object DetectionAdversarial AttackConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Proposed a novel backdoor attack framework for object detection called Detector Collapse (DC), which implants backdoors during training through two strategies, SPONGE and BLINDING, causing the attacked model to globally fail or make all targets disappear when triggered.
Determining Winners in Elections with Absent Votes
Qishen Han (Rensselaer Polytechnic Institute), Lirong Xia (Rensselaer Polytechnic Institute)
🎯 What it does: Studies the computational complexity of determining whether a candidate can become a winner when votes are missing under the top-ℓ / up-to-L truncated voting framework
DFMDA-Net: Dense Fusion and Multi-dimension Aggregation Network for Image Restoration
Huibin Yan (Shenzhen University), Shuoyao Wang (Shenzhen University)
RestorationConvolutional Neural NetworkTransformerImageBenchmark
🎯 What it does: This paper proposes an image restoration network based on a U-shaped structure, named DFMDA-Net. The core innovation lies in integrating a dense fusion mechanism (ICI) for same-layer features of the Encoder-Decoder and a multi-dimensional aggregation module (MDA) to achieve efficient information fusion and multi-scale feature integration.
DGCD: An Adaptive Denoising GNN for Group-level Cognitive Diagnosis
Haiping Ma (Anhui University), Hengshu Zhu (Anhui University)
Representation LearningGraph Neural NetworkAuto EncoderGraph
🎯 What it does: Construct a group-student-exercise (GSE) heterogeneous graph, leverage graph neural networks to learn high-order representations of groups, and integrate an adaptive denoising module and entropy-weighted balancing module to achieve group-level cognitive diagnosis.
DGR: A General Graph Desmoothing Framework for Recommendation via Global and Local Perspectives
Leilei Ding (University of Science and Technology of China), Yanyong Zhang (University of Science and Technology of China)
Recommendation SystemGraph Neural NetworkSupervised Fine-TuningContrastive LearningGraph
🎯 What it does: Proposes a general graph smoothing elimination framework called DGR to address the over-smoothing problem in GCN-based recommendation systems.
Dialogue Cross-Enhanced Central Engagement Attention Model for Real-Time Engagement Estimation
Jun Yu (University of Science and Technology of China), Peng Chang (PAII Inc)
RecognitionTransformerMultimodality
🎯 What it does: Propose a Dialogue Cross-Enhanced CEAM (Dialogue Cross-Enhanced Center Attention Model) and a central sliding window method to achieve accurate estimation of participation in real-time two-person interactions.
DiffStega: Towards Universal Training-Free Coverless Image Steganography with Diffusion Models
Yiwei Yang (Shanghai Jiao Tong University), Guangtao Zhai (Shanghai Jiao Tong University)
GenerationSafty and PrivacyDiffusion modelImage
🎯 What it does: This paper proposes DiffStega, an unencapsulated image steganography method based on diffusion models, which generates a reference image using a preset password and achieves encryption and decryption through noise flipping.
Diffusion Mask-Driven Visual-language Tracking
Guangtong Zhang (Guangxi Normal University), Shuxiang Song (Guangxi Normal University)
Object TrackingTransformerVision Language ModelDiffusion modelMultimodality
🎯 What it does: Propose a vision-language tracker called DMTrack based on diffusion models, which utilizes diffusion models to convert multimodal features into pixel-level target masks, thereby improving tracking errors caused by inaccurate initial language descriptions.
DifTraj: Diffusion Inspired by Intrinsic Intention and Extrinsic Interaction for Multi-Modal Trajectory Prediction
Yanghong Liu (Wuhan University), Mang Ye (Wuhan University)
Autonomous DrivingRecurrent Neural NetworkGraph Neural NetworkTransformerDiffusion modelTime SeriesSequential
🎯 What it does: Proposes DifTraj, an end-to-end diffusion model framework for simultaneously modeling intrinsic intent (goal points) and extrinsic interactions (social constraints) in human trajectory prediction, and improves multimodal generation quality through sample consistency loss.