Text-based person re-identification (TBPReID) aims to retrieve person images from a gallery using natural language descriptions. Despite recent progress, TBPReID remains challenging because image–text pairs in real-world datasets often suffer from noisy correspondence (NC), including ambiguous descriptions, attribute-level inconsistencies, and mismatched image-caption pairs. Such unreliable supervision may force cross-modal models to align semantically inconsistent samples and degrade fine-grained retrieval performance. To address this problem, we propose an Uncertainty-Aware Alignment (UAA) framework for robust TBPReID under NC. The key idea is to model image–text correspondence as a continuous reliability-estimation problem and propagate the estimated reliability into fine-grained alignment. Specifically, we introduce an Uncertainty-Aware Consensus (UAC) mechanism, where consensus denotes the reliability of cross-modal matching inferred from in-batch similarity evidence. UAC constructs Dirichlet-based evidence distributions to estimate epistemic uncertainty and adaptively down-weight unreliable correspondences. We further design a Fine-Grained Attribute Masking (FGAM) module with an uncertainty-weighted ranking loss, which uses uncertainty-derived confidence to suppress ambiguous attributes and strengthen trustworthy identity-related cues. Extensive experiments on CUHK-PEDES, ICFG-PEDES, and RSTPReid demonstrate the effectiveness and robustness of UAA. Under the 50% noise setting, UAA achieves 71.58% R@1 on CUHK-PEDES, yielding a 4.01% improvement over the baseline. Our code is available at https://github.com/pengchengL-hub/UAA.

Drug response prediction driven by artificial intelligence (AI) offers an efficient solution for accelerating the development of precision medicine and personalized treatment. However, existing AI methods are typically limited by high noise levels, heterogeneity, and limited modal data. These limitations decrease model performance and hinder the identification of critical biomarkers. Therefore, we propose a Multimodal Class-Aware Molecular Language (MCML) model for accurate drug response prediction. Specifically, MCML systematically integrates multimodal features of drugs and cell lines and establishes a crossmodal modeling mechanism to achieve deep fusion of multimodal information. Meanwhile, the model dynamically adjusts the contribution weights based on the class features and importance of the samples, effectively alleviating the noise interference inherent in multimodal data. Furthermore, MCML employs self-supervised learning for pre-training to capture potential molecular interaction patterns, enhancing its ability to adapt to data heterogeneity. Experiments performed on cross-scale multiomics datasets and single-cell transcriptomic data indicate that MCML significantly outperforms existing state-of-the-art models in RMSE and MAE scores. These case studies further demonstrate that the MCML model can effectively identify tumor microenvironment characteristics associated with drug resistance, demonstrating its ability to discover relevant biomarkers. Additionally, we performed an interpretability analysis of the model to investigate the impact of key features on the prediction results. This research establishes a new methodological paradigm for multimodal tumor data-driven drug response predictions and offers reliable computational tools for personalized cancer treatment decision making.

Local causal structure learning aims to identify direct causes (parents) and effects (children) of a target variable. Recent advances in this field rely on learning the MB (Markov Blanket) of unidentified variables. However, most existing methods require extensive search to discover spouses during MB learning, and they often need to learn the MB within the PC (parents and children) set of a target variable to identify parents, which becomes computationally expensive when the parent set is large and complex. To address this issue, we propose a novel local causal structure learning algorithm with Efficient Parents Discovery, named EPD. Specifically, EPD introduces an MB discovery subroutine, MBDis, which first identifies some parents of the target variable using the V-structure, and then performs feature selection to exclude candidate spouses that are weakly related to the target variable, thereby reducing the size of the candidate spouse set and accelerating spouse discovery. Additionally, EPD incorporates an IdePC subroutine, which learns the PC sets of the unresolved variables to identify additional parents, reducing the search space of parents. With the proposed MBDis and IdePC subroutines, EPD adopts an MB-by-PC learning strategy; it starts from discovering the MB of the target variable and then learns the PC sets of undetermined variables. This process continues iteratively until the parents and children of the target variable are identified. Using 7 Bayesian networks and 1 real-world data, the experiments have verified the effectiveness of EPD, in comparison with 10 state-of-the-art methods.

Structure-activity relationship (SAR) modeling of natural products presents a persistent methodological challenge: datasets typically contain fewer than 100 compounds, which restricts the use of data-hungry deep learning models, while conventional QSAR approaches lack mechanistic interpretability and function as black boxes. We present MK-Ensemble, a systematic four-stage optimization framework for interpretable, fragment-based SAR modeling with small-sample natural product datasets. The framework integrates multi-kernel support vector regression, hybrid molecular representations, adversarial domain adaptation, and stacking ensemble learning with strictly nested cross-validation. As a validation case study, we applied the framework to a curated dataset of 91 antioxidant compounds–comprising 24 steroidal saponins from Polygonatum cyrtonema and 67 structurally diverse reference compounds–with 128 activity records across DPPH, ABTS, and FRAP assays. We additionally performed applicability domain characterization via Williams plots and descriptor-space distance analysis, Y-randomization testing (500 permutations), and rigorous statistical model comparison using corrected resampled t-tests and Bayesian correlated t-tests. The Stacking Ensemble achieved R^2 = 0.846 (95

Remote sensing image-text retrieval (RSITR) provides intuitive access to large-scale agricultural data by aligning visual observations with natural language. However, transferring generic Vision-Language Pre-training (VLP) models (CLIP) to Farm Remote Sensing (FRS) remains challenging. Repetitive textures, high inter-class similarity, and large intra-class variation can cause semantic confusion and intra-modal over-clustering. Moreover, retrieval-oriented FRS benchmarks remain limited, while full fine-tuning is costly and prone to overfitting. To support evaluation and adaptation, we construct FRS-M\textsuperscript{2}Bench by pairing multi-source farmland imagery with LLM-curated and manually verified descriptions across diverse crops, parcel geometries, and temporal variations. We further propose Multi-modal Migration-and-Matching Alignment (M\textsuperscript{2}Align), a parameter-efficient framework built on a frozen CLIP backbone. M\textsuperscript{2}Align combines a Structured Multimodal Adapter (SMA) for structured cross-modal interaction with a Dual-Consistency-aware Contrastive loss (DC\textsuperscript{2}on). DC\textsuperscript{2}on couples cross-modal contrastive alignment with an adaptive triplet-based intra-modal uniformity constraint to mitigate over-clustering. Experiments on FRS-M\textsuperscript{2}Bench show that M\textsuperscript{2}Align achieves the best overall retrieval performance among the compared methods while reducing trainable parameters by over 96\% relative to full fine-tuning. Qualitative results further show more discriminative and uniformly distributed embeddings, reducing confusion among visually similar agricultural scenes.

Signed bipartite link-sign prediction aims to infer the polarity of candidate interactions between two disjoint node sets from observed positive and negative relations. Existing methods face three key challenges in this setting: conventional triangle-based structural balance is incompatible with bipartite topology, node-centric aggregation may obscure interaction-specific structural evidence, and severe sign imbalance is often accompanied by substantial intra-class heterogeneity. To address these challenges, we propose a Balance-Aware Prototypical Graph Transformer (BAPGT), which formulates signed bipartite link-sign prediction as query-conditioned interaction-level reasoning. BAPGT represents observed interactions as signed edge tokens and introduces each candidate interaction as an unknown-sign query token. It then retrieves a bounded query-relevant context without explicitly constructing a complete line graph. To adapt structural balance to bipartite topology, BAPGT extracts query-specific polarity evidence from odd-length signed paths and local compatibility evidence from signed butterflies, i.e., four-node bipartite motifs, and incorporates both signals into sparse Transformer attention. Furthermore, a multi-prototype sign decoder captures heterogeneous patterns within positive and negative interactions, while a confidence-weighted balanced-path consistency objective reduces the influence of weak or contradictory structural evidence. Extensive experiments on five benchmark signed bipartite datasets demonstrate that BAPGT consistently outperforms representative unsigned graph, signed graph, signed bipartite graph, and Transformer-based baselines. Ablation, structural, sparsity, and scalability analyses further verify the effectiveness and efficiency of the proposed components.

Precisely identifying protein-binding locations is essential for advancing protein engineering and pharmaceutical discovery. However, due to the intrinsically hierarchical nature of protein data, extracting deep representations from such data remains a difficult task. Furthermore, issues such as class imbalance and uneven topological distribution of interaction sites limit further improvements in prediction performance. To tackle these issues, we present HGRL-PPIS, a novel hierarchical graph representation learning approach for predicting protein-protein interaction sites. Specifically, our framework first constructs a three-tier graph encompassing a protein-level graph, residue-level and atomic-level graphs. A Transformer-based module is then employed to enable efficient cross-hierarchical information propagation. To mitigate the class imbalance caused by the sparse and uneven distribution of binding sites, we introduce a prototypical graph neural network training strategy that balances data distribution and enhances prediction accuracy. Our experiments reveal that HGRL-PPIS achieves superior performance over competing methods on multiple benchmark datasets, enabling more reliable detection of protein–protein binding residues.

Multi-modal test-time adaptation (MM-TTA) tackles distribution shifts in multi-modal data by adapting models with only unlabeled target samples, without accessing source data. Most existing MM-TTA methods rely on attention-based fusion to re-balance cross-modal contributions, which works well when at least one modality remains relatively clean. However, under simultaneous shifts across modalities, no reliable anchor exists and attention scores tend to lose discriminative power, yielding brittle fusion and degraded representations. To address this challenge, we propose Adapt-Then-Fuse (ATF), a new MM-TTA framework that performs instance-level adaptation prior to multi-modal integration. ATF comprises two components: Instance-Aware Mixture-of-Experts Adapter (IAMA) and Stratified Entropy Modulation (SEM). IAMA equips each modality with a pool of lightweight expert adapters and uses a context-conditioned router to compute per-instance mixing weights, enabling flexible token modulation that can accommodate diverse shift patterns. After per-modality modulation, the refined representations are fused to produce predictions. SEM further stabilizes adaptation by partitioning test instances into reliability strata and applying stratum-specific entropy objectives, reducing the adverse effects of overconfidence and noisy pseudo-signals during optimization. Extensive experiments on challenging multi-modal corruption benchmarks verify the effectiveness of ATF, delivering consistent improvements over prior methods, with gains of 3.4% and 2.0% on VGGSound-MC and Kinetics50-MC, respectively.

Continuous Test-Time Adaptation (CTTA) aims to adapt a pre-trained model to continuously evolving unlabeled target domain. Recent advances in this field predominantly adopt pseudo-label–guided self-supervised learning under teacher–student paradigms. However, under dynamic distribution shifts, such methods often suffer from noisy pseudo-labels, which leads to error accumulation and catastrophic forgetting. To alleviate this issue, we propose a novel Low-Frequency Modulation (LFM) algorithm for CTTA. Specifically, we introduce a low-frequency-guided sample selection strategy that filters out samples dominated by high-frequency components, guided by posterior prediction discrepancy. To further enrich the target distribution and reduce the model’s reliance on high-frequency cues, we employ a cross-class mixed-frequency augmentation method that selectively injects high-frequency components from other classes while preserving low-frequency content. In addition, we introduce a frequency-aware domain knowledge allocation strategy, which dynamically distributes domain-specific and domain-shared knowledge to maintain model stability in continuously evolving environments. Extensive experiments on four widely used benchmarks demonstrate the effectiveness of LFM in both classification and segmentation CTTA tasks.

Remote Sensing Image-Text Retrieval (RSITR) has been a prominent research topic on remote sensing (RS) as a branch of cross-modal image-text retrieval. A central challenge in RSITR lies in bridging the substantial semantic gaps between heterogeneous modalities, especially in complex RS scenes containing multiple objects, diverse attributes, and intricate spatial relations. Existing alignment strategies often overlook the hierarchical nature of these semantics and fail to maintain consistency across multiple granularities, which limits retrieval accuracy. To address this, we introduce scene graph to learn hierarchical fine-grained cross-modal interactions for RSITR. Specifically, we propose a Hierarchical Scene-Graph Consistency Learning Framework (HSGC) that masks object-, attribute-, and relation- level phrases based on the scene graph parsed from the textual annotation, providing structured semantic guidance for alignment. To fully align fine-grained cross-modal features, an Object-Attribute-Relation (OAR) inference module is designed, which narrows the semantic gap between text and image features at each level. Meanwhile, we use a Multi-Grained Semantic Alignment (MGSA) module to facilitate coarse- and fine-grained semantic alignment by constraining the distribution of image-text feature similarity. Both quantitative and qualitative experimental results on the RSICD and RSITMD benchmark datasets highlight the superiority of our HSGC over 26 baselines.

Accurate late-season rice–weed mapping is particularly important for identifying residual weed escapes that directly influence targeted herbicide application and final yield outcomes, enabling site-specific management and reducing unnecessary chemical inputs. While deep learning approaches perform well at early growth stages, they frequently generate unstable predictions and overestimate residual weeds under late-season conditions. This study presents an operational, knowledge-guided refinement framework for late-season rice–weed segmentation from UAV RGB orthomosaics. A compact U-Net is used to produce an initial semantic prediction, which is subsequently refined using interpretable agronomic and photographic priors. The refinement stage applies vegetation gating with a dark-pixel safeguard to suppress water and non-vegetated surfaces, followed by the integration of soft chromatic cues and crop-row-informed geometric constraints to stabilize class boundaries while preserving plausible inter-row weed occurrences. The framework is evaluated using a late-season UAV orthomosaic and co-registered reference labels from an irrigated rice paddy in Anqing, China, tiled into leakage-safe training and validation subsets with three classes (rice, weed, and other). On independent test tiles, the refined outputs improve spatial consistency relative to appearance-only baselines, achieving an overall accuracy of 0.974, a macro-F1 score of 0.937, and a mean intersection-over-union of 0.883. The workflow operates directly on georeferenced UAV imagery and produces GIS-ready classification and probability layers in GeoTIFF format, supporting large-area mapping and integration into agricultural decision-support systems. Scripts for the refinement stage are released to facilitate reproducibility and comparison and are available at https://github.com/mnasarahmad/WeedGraphaNet; we also outline extensions toward multi-site validation and learned priors for glare and shadow handling.

Drug-drug interactions (DDIs) are a major cause of adverse drug events in clinical practice, especially under polypharmacy settings where patients receive multiple medications simultaneously. Reliable computational prediction of DDIs is therefore essential for improving medication safety and supporting clinical decision-making. Despite recent advances in computational DDI prediction, existing methods often struggle to jointly model multi-granularity pharmacological semantics and stereochemical molecular characteristics, limiting their ability to generalize to previously unseen drugs under cold-start scenarios. To address these limitations, we propose DSMV-DDI, a multimodal representation learning framework for drug-drug interaction prediction that integrates biomedical knowledge graph topology, chemical substructure features, dual-level pharmacological semantic representations, and stereochemical molecular visual representations derived from three-dimensional molecular conformations. In particular, the proposed dual-level semantic strategy jointly characterizes interaction-level pharmacological associations and intrinsic single-drug functional properties, enabling complementary modeling of pharmacological information across different semantic granularities. Furthermore, molecular visual representation learning captures geometric and spatial characteristics beyond topology-based molecular representations, improving generalization to topologically unseen drugs. Extensive experiments on real-world DDI datasets demonstrate that DSMV-DDI outperforms state-of-the-art methods, achieving an accuracy of 0.967 and an AUPR of 0.992 under the conventional setting. The proposed framework also maintains strong performance under both partial and complete cold-start settings. Ablation analyses show that dual-level pharmacological semantics contribute most to overall performance, while molecular visual representations provide complementary geometric information that further improves prediction accuracy.

Knowledge Graph Completion (KGC) aims to infer missing entities for incomplete triples. Traditional embedding-based methods rely solely on the graph structural information, making limited use of textual semantics. Although emerging text-based methods, i.e., utilizing large language models (LLMs) to learn semantic information, can solve the aforementioned problems to some extent. They still face limitations: 1) Due to the extensive scale of knowledge graphs, prompt information is often lengthy, making it difficult for LLMs to focus on key information when processing long texts; 2) When dealing with semantically similar entities, LLMs often struggle to capture subtle differences between them, leading to insufficient discriminative capability and resulting in confusion. To address these challenges, we propose a reasoning-enhanced and interaction-corrective framework based on large language models for knowledge graph completion (ReaCo-KGC). First, a two-stage prompting mechanism is designed to enable LLMs to extract key information and reasoning processes from in-context information, i.e., adjacent triples, helping LLMs focus on the core content. In addition, a multi-agent re-ranking component that applies a turn-taking summarization strategy is proposed to refine the results and resolve confusion caused by semantic similarity. Moreover, we reduce the candidate set size effectively by using a lightweight model to eliminate irrelevant entities. Experiments on benchmark datasets FB15k-237 and WN18RR demonstrate the superior performance of the proposed framework, validating its effectiveness.

Weed invasion poses a significant threat to global rice production, causing substantial yield losses and environmental degradation from excessive herbicide use. Unmanned Aerial Vehicles (UAVs), combined with advanced remote sensing and deep learning techniques, offer a transformative approach for precise weed and rice classification, supporting site-specific weed management. This review not only synthesizes recent advancements in deep learning methods using UAV-acquired data, diverse vegetation indices, and multiple sensor modalities (RGB, multispectral, hyperspectral, thermal, and LiDAR) but also provides a critical perspective on the evolution of model architectures, highlighting key trends and challenges in real-world agricultural applications. We discuss persistent issues, including data scarcity, limited model generalizability across varying environmental conditions, and the computational demands for real-time deployment. Furthermore, we propose future research directions informed by our perspective on the field’s development, emphasizing synthetic data generation via generative adversarial networks, advanced attention mechanisms, and the integration of UAVs with ground-based robotic platforms to enable more autonomous, efficient, and sustainable agricultural practices. This review thus offers both a comprehensive synthesis and a forward-looking viewpoint on advancing UAV-based precision weed management in rice cultivation. By integrating these insights, we provide a roadmap for translating UAV-based weed detection from experimental research to scalable, field-ready solutions.

Image-text matching has been a prominent research topic at the intersection of computer vision and natural language processing. Recently, image-text matching based on fine-grained paradigm has achieved remark able progress, yet, it still faces challenges such as insufficient phrase-level representations and inconsistencies in multi-granularity semantic alignment. To address these limitations, we propose Triple-level Scene Graph Relation Reasoning (called TSGR2), a novel framework that enhances fine-granularity semantic alignment through multi-level relation reasoning and adaptive language masking strategies. Specifically, first, we propose a scene graph-based semantic parsing module that decomposes textual annotations into object-level, attribute-level, and relation-level representations, enriching masked text modeling. Second, we design a triple-level relation reasoning module to capture fine-granularity semantic nuances by aligning corresponding elements across differ ent levels. Additionally, a multi-level semantic matching mechanism bridges the semantic gap between images and text by ensuring comprehensive alignment from coarse-to-fine granularity. Extensive experiments on bench mark datasets, Flickr30K and MSCOCO, demonstrate that our TSGR2 outperforms 23 state-of-the-art baselines, showcasing its effectiveness in both quantitative metrics and qualitative analysis.

Remote Sensing Image-Text Retrieval (RSITR) aims to retrieve corresponding images from text descriptions and vice versa, a task challenged by modality heterogeneity, diverse scenes, and intricate semantic alignment demands in geospatial contexts. Existing RSITR approaches often suffer from cross-modal semantic ambiguities including spatial, temporal-distribution, and label-signal ambiguities. To address these limitations, we propose a novel False Negative Informed Semantic Refinement (FNSR) framework for RSITR to obtain temporally stable and semantically reliable multi-modal representations under inherently ambiguous label signals. Specifically, the Local Feature Refinement (LFR) unit enhances local feature informativeness by leveraging Transformer-derived correlation weights to filter redundant local features. Then, the Momentum-Augmented Similarity Learning (MASL) module utilizes temporal knowledge via dynamically updated feature queues to enhance the FN refinement. Furthermore, the False Negative Refinement (FNR) mechanism detects FNs via similarity distribution analysis and mitigates their impact with a tailored loss function, aligning them closer to anchors. Extensive experiments validate FNSR’s effectiveness on RSITR datasets (RSICD, RSITMD, UCM-Captions), it achieves state-of-the-art R@1 scores of 14.94%, 24.23%, 18.13% (text-to-image, t2i) and 18.68%, 28.82%, 22.78% (image-to-text, i2t), respectively.

Causal discovery is a critical problem with broad applications across diverse research disciplines, where determining causal directions remains a major challenge. Most bi-variate causal direction discovery methods strongly depend on assumptions about data distributions and lack scalability to multivariate causal direction discovery. To this end, we propose Related Conditional Residual Proportion Causal Discovery (RPCD), a novel method for inferring causal directions from joint observation. Our general idea is that under the presence of causality without confounding, selection bias and feedback, the prediction error in the true causal direction is smaller. Specifically, we calculate the conditional expectation of the Related Conditional Residual Proportion (RCRP) under the aforementioned assumptions to distinguish the cause from the effect. In addition to bi-variate causal direction discovery, we extend RPCD to multivariate causal discovery by embedding it within constraint-based methods, which enables efficient causal direction discovery in complex causal structures. We theoretically analyze the conditions under which RCRP can reliably identify causal directions in both bi-variate and multivariate scenarios. To evaluate RPCD’s effectiveness, we conduct extensive experiments on synthetic and real-world datasets. Extensive experiments demonstrate that RPCD achieves state-of-the-art accuracy (0.84 average) on benchmark bi-variate datasets and significantly improves orientation accuracy in multivariate causal discovery. Our source code is available at https://github.com/CSF819/RPCD.

Deep learning methods have been extensively used for drug-drug interaction (DDI) prediction, aiding the development of effective and safe combination therapies. Most studies focus on either the internal molecular structure or external contextual information of individual drugs to improve feature diversity and validity. However, the latent similarities between drug pairs, which are essential for accurate predictions, have largely been overlooked. Therefore, we propose an interpretable predictive approach for graph embedding called PINGE, which relies solely on the interaction network of drugs. Specifically, we constrain the joint features of drug pairs to their interactions, allowing those with similar types to achieve cosine similarity. This similarity in direction helps the joint features converge to the same class during prediction. Additionally, each known drug can link to multiple others, enhancing its diversity. Extensive experiments demonstrate that PINGE outperforms current advanced prediction methods on both KEGG and Drugbank datasets, achieving improvements of 0.7% and 2.4% in ACC while providing network structure-based explanations for predictions. Furthermore, PINGE surpasses advanced baselines by 1% and 1.1% in AUC on the human drug-target dataset and HuRI protein-protein interaction dataset, showcasing excellent versatility.

To address low planning efficiency, poor trajectory smoothness, and limited robustness of unmanned forklifts for handling sick and dead livestock, a Smooth Adaptive Bi-objective Optimization Ant Colony Algorithm (SABO-ACO) is proposed. The method integrates an A* heuristic and turning cost suppression to balance path efficiency and smoothness, while a dynamic pheromone boundary and improved max-min pheromone update mitigate local optima. Deadlock detection and grid backoff mechanisms enhance reliability in narrow and concave obstacle regions. Simulation in a $20 \times 20$ grid shows that SABO-ACO reduces average path length by 32.9% versus artificial potential field and 12.8% versus RRT. Indoor tests on an Ackermann unmanned vehicle demonstrate a 17.441 m trajectory with smoothness 0.0291 and RMS command tracking error 0.1510 rad/s. Results confirm its efficiency, smoothness, robustness, and practical deployability in unstructured environments.