Yelin Chen, Fanjin Zhang, Suping Sun, Yunhe Pang, Yuanchun Wang, Jian Song, XiaoYan Li, Lei Hou, Shu Zhao, Jie Tang, Juanzi Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zhen Yang, Wei Du, Jie Wang, Wenze Zhou, Xiangfeng Meng, Zhengyang Wang, Suping Sun, Ziwei Du, Haodong Zou, Jie Chen, Yongbin Liu, Shicheng Tan, Jiahao Ying, Shu Zhao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Knowledge Base Question Answering (KBQA) leverages structured knowledge bases to offer superior interpretability and hallucination resistance, making it a critical technology for precise knowledge reasoning.However, the prevailing LLM-based generate-then-execute formulation of semantic parsing is limited by strict syntactic constraints, making it primarily prone to structural deviations that render queries unexecutable, while suffering from semantic deviations that yield incorrect execution results.To address these challenges, we propose the Execution as Verification (EVER) framework, reframing semantic parsing as an iterative, self-correcting reasoning process driven by execution feedback.First, motivated by the insight that query executability serves as a strong proxy for answer correctness, we introduce Fine-Grained Execution-Aware Planning.This mechanism decomposes complex semantic parsing into a sequence of stepwise reasoning processes oriented by executability verification, ensuring high query executability.We further design a Self-Guided Semantic Correction mechanism based on execution result verification, utilizing execution feedback to verify and calibrate semantic deviations, thereby ensuring the semantic correctness of executable queries.Experimental results on the WebQSP and CWQ datasets demonstrate that our method achieves significant improvements in both query executability and answer accuracy, achieving stateof-the-art performance, particularly in complex multi-hop scenarios.Our code is available at https://github.com/ahu-zmh/EVER.

Delay-critical small object detection is crucial for real-time applications, such as autonomous driving and aerial surveillance. While cloud-based approaches incur substantial transmission latency, edge-based detection can avoid such delays; however, it faces its own challenges: limited computational capacity at the edge often leads to unsatisfactory inference delays and low accuracy when employing lightweight models. To tackle these issues, this paper presents an adaptive image batching and slicing framework for delay-sensitive small object detection in distributed edge environments. We first formulate the problem as a mixed-integer nonlinear programming model, which is NP-hard. The proposed scheme then aggregates tasks with tight deadlines into batches processed locally or cooperatively across the edge network, thus meeting stringent timing constraints. For tasks with relatively relaxed deadlines, we introduce an adaptive image slicing strategy that preserves fine-grained pixel information to boost detection accuracy without violating deadline requirements. The framework is implemented on NVIDIA Jetson edge devices and evaluated using real-world datasets. Experimental results show that our approach outperforms state-of-the-art baselines by an average accuracy gain of 0.5%, while improving the task success rate by 19%.

With the proliferation of social media, assessing depression risk through user-generated posts has emerged as a critical and prominent research direction. However, existing methods fail to capture clinically relevant linguistic markers due to ungrounded semantic representations in non-expert models. Therefore, this paper proposes a model, ProDoKE, which leverages prompt chaining integrated with domain knowledge to facilitate in-depth extraction and interpretation of clinically significant linguistic and semantic cues from social media posts. ProDoKE comprises two key modules, consisting of a prompt chaining enhancing module and a capsule fusion module. Following the collection of risky posts, the prompt chaining module systematically integrates domain-specific knowledge with Large Language Models (LLMs). This structured reasoning process guides the LLM to identify relevant clinical symptoms and then infer their underlying emotional states, resulting in a richer and more accurate interpretation. Furthermore, the capsule fusion module incorporates contrastive learning to optimize the alignment of emotional and symptom semantics, while using capsule networks to fortify the model's robustness. Evaluations on the eRisk2017 and eRisk2018 datasets show that ProDoKE achieves state-of-the-art performance, significantly outperforming previous baseline models in F1 scores.

The use of graph-based Retrieval-Augmented Generation (RAG) to retrieve relevant information from an external Knowledge Graph (KG) enables Large Language Model (LLMs) to answer specialty medical questions over private medical documents. However, medical Graph RAG only focuses on improving the response quality of LLMs via enhancing queries indiscriminately with local retrieved information, ignoring the long-tail medical knowledge that LLMs really need for the query, failing to answer global questions on a coarse-grained corpus. In this paper, we design a multi-granularity graph-based RAG framework for specialty medicine called SpecMedRAG, composed of MultiGranularity Knowledge and Multi-Route Retrieval. Specifically, Multi-Granularity Knowledge extracts specialty medical KG from private medical documents and links it with existing general medical KG to build fine-grained KG. Then it generates coarsegrained summaries from fine-grained KG. Multi-Route Retrieval comprises two retrieval paths: Local Retrieval enhanced by Long-tail Weight, which improves the utilization rate of longtail medical knowledge, and Global Retrieval, to reduce the reliance on the local detail of the query results by coarse-grained summaries. SpecMedRAG achieves SOTA results on specialty medical benchmark and demonstrate its universality on general benchmarks, significantly improving the accuracy of LLMs in specialty medical applications.

Speech Emotion Recognition (SER) plays a critical role in enhancing user experience within human-computer interaction. However, existing methods are overwhelmed by temporal domain analysis, overlooking the valuable envelope structures of the frequency domain that are equally important for robust emotion recognition. To overcome this limitation, we propose TF-Mamba, a novel multi-domain framework that captures emotional expressions in both temporal and frequency dimensions. Concretely, we propose a temporal-frequency mamba block to extract temporal- and frequency-aware emotional features, achieving an optimal balance between computational efficiency and model expressiveness. Besides, we design a Complex Metric-Distance Triplet (CMDT) loss to enable the model to capture representative emotional clues for SER. Extensive experiments on the IEMOCAP and MELD datasets show that TF-Mamba surpasses existing methods in terms of model size and latency, providing a more practical solution for future SER applications.

As the mainstream approach, LLMs have been widely applied and researched in TableQA tasks. Currently, the core of LLM-based TableQA methods typically include three phases: question decomposition, sub-question TableQA reasoning, and answer verification. However, several challenges remain in this process: i) Sub-questions generated by these methods often exhibit significant gaps with the original question due to critical information overlooked during the LLM's direct decomposition; ii) Verification of answers is typically challenging because LLMs tend to generate optimal responses during self-correct. To address these challenges, we propose a Triple-Inspired Decomposition and vErification (TIDE) strategy, which leverages the structural properties of triples to assist in decomposition and verification in TableQA. The inherent structure of triples (head entity, relation, tail entity) requires the LLM to extract as many entities and relations from the question as possible. Unlike direct decomposition methods that may overlook key information, our transformed sub-questions using triples encompass more critical details. Additionally, this explicit structure facilitates verification. By comparing the triples derived from the answers with those from the question decomposition, we can achieve easier and more straightforward validation than when relying on the LLM's self-correct tendencies. By employing triples alongside established LLM modes, Direct Prompting and Agent modes, TIDE achieves state-of-the-art performance across multiple TableQA datasets, demonstrating the effectiveness of our method.

Anomaly detection (AD) is crucial in industrial production, particularly in the context of few-shot anomaly detection (FSAD), which has gained attention due to the challenges of obtaining labeled samples in real-world scenarios. Existing FSAD methods generate synthetic anomalies from a limited number of normal samples to augment training data. However, the generated anomalies often diverge significantly from actual anomalies, complicating the detection and localization of defects. To address this challenge, we introduce a novel discriminative network named DiscNetAD, which integrates realistic anomaly generation with image pyramid decomposition for FSAD. Our approach employs a diffusion model-based generative strategy to create realistic anomalies from the limited normal samples, and we utilize an image pyramid for multi-scale decomposition, allowing the model to capture fine-grained features across various resolutions. By performing these two operations, more realistic anomalous samples are generated and the model can identify various anomalies. After feature extraction and aggregation on the decomposed images, we apply anomaly discriminators to estimate anomaly scores. Comprehensive evaluations demonstrate that DiscNetAD achieves image-level/pixel-level AUROC of 97.5%/98.1% and image-level/pixel-level AUROC of 89.5%/97.9% on the MVTec AD and VisA datasets, respectively, significantly outperforming current state-of-the-art FSAD methods.

Table Question Answering (TableQA) combines natural language understanding and structured data reasoning, posing challenges in semantic interpretation and logical inference. Recent advances in Large Language Models (LLMs) have improved TableQA performance through Direct Prompting and Agent paradigms. However, these models often rely on spurious correlations, as they tend to overfit to token co-occurrence patterns in pretraining corpora, rather than perform genuine reasoning. To address this issue, we propose Causal Intervention TableQA (CIT), which is based on a structural causal graph and applies front-door adjustment to eliminate bias caused by token co-occurrence. CIT formalizes TableQA as a causal graph and identifies token co-occurrence patterns as confounders. By applying front-door adjustment, CIT guides question variant generation and reasoning to reduce confounding effects. Experiments on multiple benchmarks show that CIT achieves state-of-the-art performance, demonstrating its effectiveness in mitigating bias. Consistent gains across various LLMs further confirm its generalizability.

Depression detection on social media aims to identify depressive tendencies within textual posts, providing timely intervention by the early detection of mental health issues . In predominant approaches, the Pre-trained Language Models(PLMs) are trained solely on public datasets, falling short of vertical scenarios due to insufficient domain-specific and commonsense knowledge . In addition, ambiguous commonsense knowledge could be misleading to PLMs and results in false judgments . Therefore, it poses significant challenges to select commonsense knowledge that is trusted. To address this, we propose CoKE, a model that incorporates trusted commonsense knowledge based on three-way decision theory to enhance depression detection. CoKE comprises three key modules: trusted screening, knowledge generation, and knowledge fusion. First, we utilize psychiatric clinical scales and three-way decision theory to screen out the uncertain domain from the massive user posts. Then, an adaptive framework is applied to generate and refine trusted commonsense knowledge that can explain the true semantics of posts in the uncertain domain. Finally, a dynamic integration of posts with highly trusted knowledge is achieved through a gating mechanism, resulting in embeddings enhanced by trusted commonsense knowledge that are more effective in determining depressive tendencies. We evaluate our model on two prominent datasets, eRisk2017 and eRisk2018, demonstrating its superiority over previous state-of-the-art baseline models .

One of the most popular methods for analyzing the robustness of 3D Deep Neural Networks (DNNs) is the transfer-based adversarial attack method, as it allows to analyze the robustness of an unknown model by generating an adversarial point cloud on an alternative model. However, the adversarial point clouds generated by current methods may overfit the surrogate models that generated them, thus limiting their performance in transfer attacks against different target 3D classifiers. To enhance the transferability of the adversarial point cloud, we propose in this letter an adversarial attack method by Initializing the Transferable Adversarial Noise, which named asITAN. Specifically, we pre-train on the training set a generator capable of generating the adversarial noise with transferability and diversity, and then the noise generated by the generator serves as the initial adversarial noise to be integrated into the iterations of the attack. Extensive experiments on well-recognized benchmark datasets demonstrate that the adversarial point clouds generated by the proposed ITAN could be effectively transferred across unknown 3D classifiers.

Abstract Prompt transfer is a transfer learning method based on prompt tuning, which enhances the parameter performance of prompts in target tasks by transferring source prompt embeddings. Among existing methods, weighted aggregation is effective and possesses the advantages of being lightweight and modular. However, these methods may transfer redundant or irrelevant information from the source prompts to the target prompt, leading to negative impacts. To alleviate this problem, we propose Prompt Contrastive Transformation (PCT), which achieves efficient prompt transfer through prompt contrastive transformation and attentional fusion. PCT transforms the source prompt into task-agnostic embedding and task-specific embeddings through singular value decomposition and contrastive learning, reducing information redundancy among source prompts. The attention module in PCT selects more effective task-specific embeddings and fuses them with task-agnostic embedding into the target prompt. Experimental results show that, despite tuning only 0.035% of task-specific parameters, PCT achieves improvements in prompt transfer for single target task adaptation across various NLP tasks.

To meet the challenge of incompleteness within Knowledge Graphs, Knowledge Graph Embedding (KGE) has emerged as the fundamental methodology for predicting the missing link (Link Prediction), by mapping entities and relations as low-dimensional vectors in continuous space. However, current KGE models often struggle with the polysemy issue, where entities exhibit different semantic characteristics depending on the relations in which they participate. Such limitation stems from weak interactions between entities and their relation contexts, leading to low expressiveness in modeling complex structures and resulting in inaccurate predictions. To address this, we propose Contextualized Quaternion Embedding (ConQuatE), a model that enhances the representation learning of entities across multiple semantic dimensions by leveraging quaternion rotation to capture diverse relational contexts. In specific, ConQuatE incorporates contextual cues from various connected relations to enrich the original entity representations. Notably, this is achieved through efficient vector transformations in quaternion space, without any extra information required other than original triples. Experimental results demonstrate that our model outperforms state-of-the-art models for Link Prediction on four widely recognized datasets: FB15k-237, WN18RR, FB15k, and WN18.

Transfer-based adversarial attacks involve generating adversarial point clouds in surrogate models and transferring them to other models to assess 3D model robustness. However, current methods rely too much on surrogate model parameters, limiting transferability. In this work, we use Shapley value to identify positive and negative features, guiding optimization of adversarial noise in feature space. To effectively mislead the 3D classifier, we factorize the adversarial noise into positive and negative noise, with the former keeping the features of the adversarial point cloud close to the negative features, and the latter and the adversarial noise moving it away from the positive features. Finally, a novel adversarial point cloud attack method with Adversarial Noise Factorization is proposed, which is abbreviated asANF. ANF simultaneously optimizes the adversarial noise and its positive and negative noise in the feature space, only relying on partial network parameters, which significantly reduces the reliance on the surrogate model and improves the transferability of the adversarial point cloud. Experiments on well-recognized benchmark datasets show that the transferability of adversarial point clouds generated by ANF could be improved by more than 26.7$\%$on average over state-of-the-art transfer-based adversarial attack methods.

Abstractive summarization has made significant progress in recent years, which aims to generate a concise and coherent summary that contains the most important facts from the source document. Current fine-tuning approaches based on pre-training models typically rely on autoregressive and maximum likelihood estimation, which may result in inconsistent historical distributions generated during the training and inference stages, i.e., exposure bias problem. To alleviate this problem, we propose a hybrid fine-tuning model(HyFit), which combines contrastive learning and reinforcement learning in a diverse sampling space. Firstly, we introduce reparameterization and probability-based sampling methods to generate a set of summary candidates called candidates bank, which improves the diversity and quality of the decoding sampling space and incorporates the potential for uncertainty. Secondly, hybrid fine-tuning with sampled candidates bank, upweighting confident summaries and downweighting unconfident ones. Experiments demonstrate that HyFit significantly outperforms the state-of-the-art models on SAMSum and DialogSum. HyFit also shows good performance on low-resource summarization, on DialogSum dataset, using only approximate 8% of the examples exceed the performance of the base model trained on all examples.

3D adversarial attacks have garnered significant attention in the realm of autonomous driving security due to their high feasibility and multi-view effectiveness. However, existing 3D attacks have limited transferability, primarily due to their overfitting to surrogate models. To address this limitation, we introduce a novel 3D adversarial attack method based on implicit texture modeling, termed ImAdv, against 3D object detection models. Specifically, ImAdv utilizes a positional encoder and a MLP to map the 3D coordinates of an object's surface to the RGB color space, thereby reformulating the object's texture within an implicit framework. This method significantly reduces the parameter number for color modeling, thus mitigating overfitting and improving transferability. Furthermore, we propose two innovative techniques to enhance the transferability, Random Texture Reset (RandReset) and Texture Model Averaging. RandReset randomly restores portions of the adversarial texture, increasing the training set diversity and mitigating overfitting. Texture Model Averaging employs self-ensembling of multiple texture checkpoints during the training phase to reduce overfitting in the final texture model. Comprehensive experiments demonstrate the superiority of our methods, which outperform previous methods by 17.18% in average black-box attack success rate. Additionally, our method shows strong transferability and practicality in zero-shot cross-task attacks and physical attacks.

The primary objective of knowledge tracing (KT) is to evaluate students’ understanding and mastery of knowledge through their responses to exercises, which aids in predicting their future performance. Deep neural networks have been widely applied in the area of knowledge tracing and have demonstrated encouraging results. Nevertheless, in real-world scenarios, there is a substantial amount of noise in students’ response records. These noises may amplify the inherent risk of overfitting in deep neural networks, leading to a decrease in model performance. To address these issues, we introduce a new model called filter knowledge tracing (FKT). This innovative model incorporates a learnable filter into KT to filter out noise information from students’ exercise sequences. We redefine the input paradigm of the data, using learnable filters to perform filtering operations in its frequency domain representation space, effectively removing noise. Additionally, an attention module has been introduced in the FKT model to evaluate the impact of students’ historical interactions on their current knowledge state. To validate our model, we conduct extensive experiments utilizing four publicly available datasets. The results demonstrate that FKT outperforms existing benchmarks, particularly on larger datasets, signifying an improvement in KT performance while effectively reducing the risk of overfitting.

Deploying state-of-the-art object detectors on resource-limited devices presents significant challenges. Knowledge distillation is an efficient and streamlined lightweight technique to improve the accuracy of compact detectors. However, its effectiveness is limited by the redundancy of different types of semantics on the feature map and the closure of same level’s feature distillation. To alleviate this problem, we propose Decoupled and Interactive Distillation, an effective and versatile method to improve knowledge distillation in some complex object detection tasks. The method has two key components. A knowledge decoupled module captures category awareness and localization awareness features. A multi-level feature interaction distillation can aggregate feature distillations from shallow to deep levels, facilitating the collaboration between feature transfers at different levels. The relevant experiments in traffic-related, 3D, rotated object detection have verified the effectiveness of the proposed method, particularly in challenging scenes.

It has been shown that deep recommendation models are susceptible to adversarial attacks, with this vulnerability potentially leading to significant economic losses in the e-commerce field. However, the robustness of deep recommendation models in response to adversarial attacks has not been systematically investigated. In this article, therefore, we comprehensively evaluate the adversarial robustness of various representative deep models in different settings, aiming to analyze their performance impact under adversarial attacks and compare it with traditional collaborative filtering models. Notably, we examine poisoning attacks under different proportions of fake users and various popularity conditions to understand why certain deep recommendation models perform exceptionally or sub-optimally. On this basis, we further proposed practical robustness improvement strategy for the problems found in the evaluation and fully verified it through rigorous experiments. Key findings include: (1) the sparser the training dataset, the weaker the robustness of a recommendation model’s performance under adversarial attacks; (2) deep recommendation models exhibit greater robustness in recommending popular items under adversarial attacks, while they are more vulnerable when attacked with non-popular items; (3) the robustness of deep recommendation models is not consistently weaker than that of traditional collaborative filtering models across all attack settings. These findings highlight the security concerns in deep recommendation systems and contribute to developing more reliable models.

Feature selection is an essential pre-process component in data mining that aims to select the most relevant features from the target dataset. Datasets are always dynamic in real-world applications, and features may exist in stream mode. Then, online streaming feature selection methods are proposed, which deal with streaming features arriving continuously in real-time. However, most existing algorithms prioritize high accuracy and low time-consumption but overlook the stability of the selected features. Stable feature selection results are crucial for users in practice. For instance, in the medical field, unstable feature selection results can make it challenging for experts to identify the main causative factors of a disease. Motivated by this, this article proposes a new online stable streaming feature selection method via feature aggregation named OSSFS. Specifically, inspired by the cohesive MeanShift approach, OSSFS applies an incremental aggregation strategy to partition the streaming features into multiple hyperellipsoids. Then, we incrementally update and merge these hyperellipsoids with new streaming features. Finally, we select representative features from each hyperellipsoid as the final selected feature subset. Extensive experiments are conducted on several real-world datasets to compare our new method with state-of-the-art competing algorithms in cases of stability and predictive accuracy. Experimental results indicate that OSSFS achieves optimal stability without losing prediction accuracy.

The text-attributed graph (TAG) is one kind of important real-world graph-structured data with each node associated with raw texts. For TAGs, traditional few-shot node classification methods directly conduct training on the pre-processed node features and do not consider the raw texts. The performance is highly dependent on the choice of the feature pre-processing method. In this paper, we propose P2TAG, a framework designed for few-shot node classification on TAGs with graph pre-training and prompting. P2TAG first pre-trains the language model (LM) and graph neural network (GNN) on TAGs with self-supervised loss. To fully utilize the ability of language models, we adapt the masked language modeling objective for our framework. The pre-trained model is then used for the few-shot node classification with a mixed prompt method, which simultaneously considers both text and graph information. We conduct experiments on six real-world TAGs, including paper citation networks and product co-purchasing networks. Experimental results demonstrate that our proposed framework outperforms existing graph few-shot learning methods on these datasets with +18.98% ~ +32.14% improvements.

The existence of noisy labels is inevitable in real-world large-scale corpora. As deep neural networks are notably vulnerable to overfitting on noisy samples, this highlights the importance of the ability of language models to resist noise for efficient training. However, little attention has been paid to alleviating the influence of label noise in natural language processing. To address this problem, we present CoMix, a robust Noise-against training strategy taking advantage of Co-training that deals with textual annotation errors in text classification tasks. In our proposed framework, the original training set is first split into labeled and unlabeled subsets according to a sample partition criteria and then applies label refurbishment on the unlabeled subsets. We implement textual interpolation in hidden space between samples on the updated subsets. Meanwhile, we employ peer diverged networks simultaneously leveraging co-training strategies to avoid the accumulation of confirm bias. Experimental results on three popular text classification benchmarks demonstrate the effectiveness of CoMix in bolstering the network’s resistance to label mislabeling under various noise types and ratios, which also outperforms the state-of-the-art methods.

Feature selection aims to select an optimal minimal feature subset from the original datasets and has become an indispensable preprocessing component before data mining and machine learning, especially in the era of big data. However, features may be generated dynamically and arrive individually over time in practice, which we call streaming features. Most existing streaming feature selection methods assume that all dynamically generated features are the same type or assume we can know the feature type for each new arriving feature in advance, but this is unreasonable and unrealistic. Therefore, this paper first studies a practical issue of Online Heterogeneous Streaming Feature Selection without the feature type information before learning, named OHSFS. Specifically, we first model the streaming feature selection issue as a minimax problem. Then, in terms of MIC (Maximal Information Coefficient), we derive a new metric$MIC_{Gain}$to determine whether a new streaming feature should be selected. To speed up the efficiency of OHSFS, we present the metric$MIC_{Cor}$that can directly discard low correlation features. Finally, extensive experimental results indicate the effectiveness of OHSFS. Moreover, OHSFS is nonparametric and does not need to know the feature type before learning, which aligns with practical application needs.

Recent research has demonstrated that recommendation systems exhibit vulnerability under data poisoning attacks. The primary process of data poisoning attacks involves generating malicious data (i.e., fake users) through surrogate models and injecting the malicious data into the target models’ datasets, thereby manipulating the output results of the target models. However, current methods generating fake users based on gradient descent may cause them to fall into undesired local minimum in the loss landscape and overfitting to the surrogate model, thus limiting the performance of attacking other recommendation models. To address this problem, we propose the reverse optimization algorithm (ReOP), which utilizes the reverse direction of optimization to update fake users, enabling them to steer clear of sharp local minimum in loss landscape and navigate towards the flat local minimum. ReOP makes fake users less sensitive to model changes, alleviates their overfitting to the surrogate model, and thus significantly improves the transferability of fake users. Experimental results demonstrate that ReOP surpasses the state-of-the-art baseline methods, effectively generating fake users with significant attack effects on various target models.

The reliability and robustness of 3D object detection play an instrumental role in the practical deployment of autonomous driving systems. Despite previous research indicating that adversarial examples can negatively affect 3D object detection models, leading to misinterpretations of the environment, these models still maintain the capability to detect the majority of objects within adversarially manipulated point clouds. To further probe into the adversarial robustness of these models, we propose an effective adversarial attack method named IoU-S attack in this paper. We meticulously formulate the adversarial loss to adversely affect the decision-making behavior (such as localization, etc.) of 3D object detection, thereby compromising its ability to accurately interpret the environment. Owing to the significant relevance of this adversarial loss to 3D object detection tasks, we have integrated the IoU-S attack into three attack paradigms: point cloud perturbation, detachment, and attachment. Comprehensive experiments on the widely accepted nuScenes dataset illustrate that the IoU-S attack outperforms existing attack methods in both white-box and black-box scenarios (https://github.com/haichen-ber/IoU-S-Attack). It reinforces its potential to serve as a valuable method in understanding and enhancing the robustness of 3D object detection models against adversarial attacks.

The recommendation performance of deep collaborative filtering models drops sharply under imperceptible adversarial perturbations. Some methods promote the robustness of recommendation systems by adversarial training. However, these methods only study shallow models and lack the exploration of deep models. Furthermore, the way these methods add adversarial noise to the weight parameters of users and items is not fully applicable to deep collaborative filtering models, because the adversarial noise is not sufficient to fully affect its network structure with multiple hidden layers. In this article, we propose a novel adversarial training framework, Random Layer-wise Adversarial Training (RAT), which trains a robust deep collaborative filtering model via adversarial noise propagation. Specifically, we inject adversarial noise into the output of the hidden layer in a random layer-wise manner. The adversarial noise propagates forward from the injected position to obtain more flexible model parameters during the adversarial training process. We validate the effectiveness of RAT on multilayer perceptron (MLP) and implement RAT on MLP-based and convolutional neural networks-based deep collaborative filtering models. Experiments on three publicly available datasets show that the deep collaborative filtering model trained by RAT not only defends against adversarial noise but also guarantees recommendation performance.

3D object detection is an essential perception task in autonomous driving to understand the environments. The Bird's-Eye-View (BEV) representations have significantly improved the performance of 3D detectors with camera inputs on popular benchmarks. However, there still lacks a systematic understanding of the robustness of these vision-dependent BEV models, which is closely related to the safety of autonomous driving systems. In this paper, we evaluate the natural and adversarial robustness of various representative models under extensive settings, to fully understand their behaviors influenced by explicit BEV features compared with those without BEV. In addition to the classic settings, we propose a 3D consistent patch attack by applying adversarial patches in the 3D space to guarantee the spatiotemporal consistency, which is more realistic for the scenario of autonomous driving. With substantial experiments, we draw several findings: 1) BEV models tend to be more stable than previous methods under different natural conditions and common corruptions due to the expressive spatial representations; 2) BEV models are more vulnerable to adversarial noises, mainly caused by the redundant BEV features; 3) Camera-LiDARfusion models have superior performance under different settings with multi-modal inputs, but BEV fusion model is still vulnerable to adversarial noises of both point cloud and image. These findings alert the safety issue in the applications of BEV detectors and could facilitate the development of more robust models.

Emoji are often used in social media to enrich users’ emotions, and they play an important role in the task of social media sentiment analysis. In practice, researchers are more likely to consider emoji as special symbols and treat them separately from the text. Some existing methods use emoji as a dictionary for matching or converting emoji into text. However, these methods disregard the relationship between emoji and context, blue and they do not reflect the emotions that users are expected to express. It is challenging to incorporate the original emotions of emoji in social media sentiment analysis. In this article, we propose the EPE model: Emoji Pre-trained feature Enhanced sentiment analysis. Specifically, we collected 8 million tweets and selected 5 million tweets with pre-trained emoji with context using the BERT model. We labeled 20,000 tweets as a three-category dataset and used Bi-LSTM with an attention layer to extract text features. Emoji were retained as key emotion information and combined with text features in the final layer as a connected vector for final prediction. Experimental results with our dataset showed that the proposed EPE model achieved better performance than other baseline models.

Considering the expensive annotation in Named Entity Recognition (NER ), Cross-domain NER enables NER in low-resource target domains with few or without labeled data, by transferring the knowledge of high-resource domains. However, the discrepancy between different domains causes the domain shift problem and hampers the performance of cross-domain NER in low-resource scenarios. In this article, we first propose an adversarial adaptive augmentation, where we integrate the adversarial strategy into a multi-task learner to augment and qualify domain adaptive data. We extract domain-invariant features of the adaptive data to bridge the cross-domain gap and alleviate the label-sparsity problem simultaneously. Therefore, another important component in this article is the progressive domain-invariant feature distillation framework. A multi-grained MMD (Maximum Mean Discrepancy) approach in the framework to extract the multi-level domain invariant features and enable knowledge transfer across domains through the adversarial adaptive data. Advanced Knowledge Distillation (KD) schema processes progressively domain adaptation through the powerful pre-trained language models and multi-level domain invariant features. Extensive comparative experiments over four English and two Chinese benchmarks show the importance of adversarial augmentation and effective adaptation from high-resource domains to low-resource target domains. Comparison with two vanilla and four latest baselines indicates the state-of-the-art performance and superiority confronted with both zero-resource and minimal-resource scenarios.

Emotion recognition in conversation (ERC) is challenging because the conversation takes place in real time and the speakers interact with each other. However, existing methods ignore the dynamic characteristics of interaction between speakers, and the problem of long-range context propagation still exists. In this article, we propose a dynamic interaction emotion unit to solve the preceding problems on the transcription of the conversation. First, we propose a main influence interval search algorithm to provide a dynamic interaction interval for each utterance. Then, we utilize the speaker-aware influence module and the two-stream context module to capture the dynamic interaction and the contextual information from this interval. Furthermore, to obtain the speaker state representation rich in emotional information, we propose a novel dynamic routing algorithm to fuse the preceding information. These well-integrated state representations also enable our model to capture contextual information at a longer distance. Experiments on multiple datasets demonstrate the effectiveness of the proposed method.

The attacker's malicious behavior of injecting well-designed adversarial examples (i.e., fake users) into recommender systems will severely affect the security of systems. It's difficult to fully obtain details of victim recommendation models (i.e., black-box model) in practical recommendation scenarios, using the transferability of adversarial examples to achieve black-box attacks is still an effective way. At present, adversarial examples generated by existing gradient-based methods are prone to drop into local minima, making it impossible to achieve the expected attack effect and reducing the transferability. In this article, we propose an attack algorithm that enhances the transferability of adversarial examples based on the Nesterov Momentum for Recommendation Systems (ETANRS). With white-box recommendation surrogate models, we utilize Nesterov momentum to generate better adversarial examples, then inject them into black-box victim models to attack. We utilize the accumulated gradients and pre-determine the update direction of the gradients to keep the optimal value from being lost, thus enhancing the transferability of the adversarial examples. Experimental results demonstrate that our method is better than state-of-the-art gradient-based attack algorithms, which affect recommendation performance.

3D semi-supervised medical image segmentation is extremely essential in computer-aided diagnosis, which can reduce the time-consuming task of performing annotation. The challenges with current 3D semi-supervised segmentation algorithms includes the methods, limited attention to volume-wise context information, their inability to generate accurate pseudo labels and a failure to capture important details during data augmentation. This article proposes a dual uncertainty-guided mixing consistency network for accurate 3D semi-supervised segmentation, which can solve the above challenges. The proposed network consists of a Contrastive Training Module which improves the quality of augmented images by retaining the invariance of data augmentation between original data and their augmentations. The Dual Uncertainty Strategy calculates dual uncertainty between two different models to select a more confident area for subsequent segmentation. The Mixing Volume Consistency Module that guides the consistency between mixing before and after segmentation for final segmentation, uses dual uncertainty and can fully learn volume-wise context information. Results from evaluative experiments on brain tumor and left atrial segmentation shows that the proposed method outperforms state-of-the-art 3D semi-supervised methods as confirmed by quantitative and qualitative analysis on datasets. This effectively demonstrates that this study has the potential to become a medical tool for accurate segmentation. Code is available at:https://github.com/yang6277/DUMC.

Recent studies have shown that Graph Neural Networks (GNNs) are vulnerable to well-designed and imperceptible adversarial attack. Attacks utilizing gradient information are widely used in the field of attack due to their simplicity and efficiency. However, several challenges are faced by gradient-based attacks: 1) Generate perturbations use white-box attacks (i.e., requiring access to the full knowledge of the model), which is not practical in the real world; 2) It is easy to drop into local optima; and 3) The perturbation budget is not limited and might be detected even if the number of modified edges is small. Faced with the above challenges, this article proposes a black-box adversarial attack method, named NAG-R, which consists of two modules known asNesterovAcceleratedGradient attack module andRewiring optimization module. Specifically, inspired by adversarial attacks on images, the first module generates perturbations by introducing Nesterov Accelerated Gradient (NAG) to avoid falling into local optima. The second module keeps the fundamental properties of the graph (e.g., the total degree of the graph) unchanged through a rewiring operation, thus ensuring that perturbations are imperceptible. Intensive experiments show that our method has significant attack success and transferability over existing state-of-the-art gradient-based attack methods.

Network representation learning, also called network embedding, aiming to learn low dimensional vectors for nodes while preserving essential properties of the network, benefits plenty of practical applications. However, how to do representation learning on the network quickly and effectively is a meaningful and challenging task, especially for the attributed networks. In this paper, we propose HANE, a Hierarchical Attributed Network Embedding framework, which is a fast and effective method by quickly constructing a hierarchical attributed network of different granularities to learn nodes representations. Specifically, for an attributed network, HANE first builds a hierarchy of successively smaller attributed network from fine to coarse by the fast granulation strategy fusing topological structure and node attributes. After using any unsupervised network embedding method to learn nodes representations of the coarsest network, HANE refines the nodes representations of the hierarchical attributed network from coarse to fine. HANE improves the speed of network representation learning while maintaining its performance and the representation learning method of the coarsest network is flexible. We conduct extensive evaluations for the proposed framework HANE on six datasets and two benchmark applications. Experimental results demonstrate that HANE achieves significant improvements over previous state-of-the-art network embedding methods in efficiency and effectiveness.

Streaming data mining can be applied in many practical applications, such as social media, market analysis, and sensor networks. Most previous efforts assume that all training instances except for the novel class have been completely labeled for novel class detection in streaming data. However, a more realistic situation is that only a few instances in the data stream are labeled. In addition, most existing algorithms are potentially dependent on the strong cohesion between known classes or the greater separation between novel class and known classes in the feature space. Unfortunately, this potential dependence is usually not an inherent characteristic of streaming data. Therefore, to classify data streams and detect novel classes, the proposed algorithm should satisfy: 1) it can handle any degree of separation between novel class and known classes (both easy and difficult novel class detection) and 2) it can use limited labeled instances to build algorithm models. In this article, we tackle these issues by a new framework called semisupervised streaming learning for difficult novel class detection (SSLDN), which consists of three major components: an effective novel class detector based on random trees, a classifier by using the information of nearest neighbors, and an efficient updating process. Empirical studies on several datasets validate that SSLDN can accurately handle different degrees of separation between the novel and known classes in semisupervised streaming data.

Network representation learning, also called network embedding, aiming to learn low dimensional vectors for nodes while preserving essential properties of the network, such as structural similarity, attribute similarity, etc. The low-dimensional vector of the node can be used as the input of the machine learning algorithm and applied to a lot of downstream tasks, such as node classification and link prediction, benefits plenty of practical applications.

Automated and accurate segmentations of left atrium (LA) and atrial scars from late gadolinium-enhanced cardiac magnetic resonance (LGE CMR) images are in high demand for quantifying atrial scars. The previous quantification of atrial scars relies on a two-phase segmentation for LA and atrial scars due to their large volume difference (unbalanced atrial targets). In this paper, we propose an inter-cascade generative adversarial network, namely JAS-GAN, to segment the unbalanced atrial targets from LGE CMR images automatically and accurately in an end-to-end way. Firstly, JAS-GAN investigates an adaptive attention cascade to automatically correlate the segmentation tasks of the unbalanced atrial targets. The adaptive attention cascade mainly models the inclusion relationship of the two unbalanced atrial targets, where the estimated LA acts as the attention map to adaptively focus on the small atrial scars roughly. Then, an adversarial regularization is applied to the segmentation tasks of the unbalanced atrial targets for making a consistent optimization. It mainly forces the estimated joint distribution of LA and atrial scars to match the real ones. We evaluated the performance of our JAS-GAN on a 3D LGE CMR dataset with 192 scans. Compared with the state-of-the-art methods, our proposed approach yielded better segmentation performance (Average Dice Similarity Coefficient (DSC) values of 0.946 and 0.821 for LA and atrial scars, respectively), which indicated the effectiveness of our proposed approach for segmenting unbalanced atrial targets.

Feature selection is one of the core concepts in machine learning, which hugely impacts the model’s performance. For some real-world applications, features may exist in a stream mode that arrives one by one over time, while we cannot know the exact number of features before learning. Online streaming feature selection aims at selecting optimal stream features at each timestamp on the fly. Without the global information of the entire feature space, most of the existing methods select stream features in terms of individual feature information or the comparison of features in pairs. This article proposes a new online scalable streaming feature selection framework from the dynamic decision perspective that is scalable on running time and selected features by dynamic threshold adjustment. Regarding the philosophy of “Thinking-in-Threes”, we classify each new arrival feature as selecting, discarding, or delaying, aiming at minimizing the overall decision risks. With the dynamic updating of global statistical information, we add the selecting features into the candidate feature subset, ignore the discarding features, cache the delaying features into the undetermined feature subset, and wait for more information. Meanwhile, we perform the redundancy analysis for the candidate features and uncertainty analysis for the undetermined features. Extensive experiments on eleven real-world datasets demonstrate the efficiency and scalability of our new framework compared with state-of-the-art algorithms.

Name disambiguation is a key and also a very tough problem in many online systems such as social search and academic search. Despite considerable research, a critical issue that has not been systematically studied isdisambiguation on the fly— to complete the disambiguation in the real-time. This is very challenging, as the disambiguation algorithm must be accurate, efficient, and error tolerance. In this paper, we propose a novel framework — CONNA — to train a matching component and a decision component jointly via reinforcement learning. The matching component is responsible for finding the top matched candidate for the given paper, and the decision component is responsible for deciding on assigning the top matched person or creating a new person. The two components are intertwined and can be bootstrapped via jointly training. Empirically, we evaluate CONNA on two name disambiguation datasets. Experimental results show that the proposed framework can achieve a 1.21-19.84 percent improvement on F1-score using joint training of the matching and the decision components. The proposed CONNA has been successfully deployed on AMiner — a large online academic search system.

Cross-domain Named Entity Recognition (NER) transfers the NER knowledge from high-resource domains to the low-resource target domain.Due to limited labeled resources and domain shift, cross-domain NER is a challenging task.To address these challenges, we propose a progressive domain adaptation Knowledge Distillation (KD) approach -PDALN.It achieves superior domain adaptability by employing three components: (1) Adaptive data augmentation techniques, which alleviate cross-domain gap and label sparsity simultaneously; (2) Multi-level Domain invariant features, derived from a multigrained MMD (Maximum Mean Discrepancy) approach, to enable knowledge transfer across domains; (3) Advanced KD schema, which progressively enables powerful pre-trained language models to perform domain adaptation.Extensive experiments on four benchmarks show that PDALN can effectively adapt highresource domains to low-resource target domains, even if they are diverse in terms and writing styles.Comparison with other baselines indicates the state-of-the-art performance of PDALN.

Overlapping communities exist in real networks, where the communities represent hierarchical community structures, such as schools and government departments. A non-binary tree allows a vertex to belong to multiple communities to obtain a more realistic overlapping community structure. It is challenging to select appropriate leaf vertices and construct a hierarchical tree that considers a large amount of structural information. In this paper, we propose a non-binary hierarchical tree overlapping community detection based on multi-dimensional similarity. The multi-dimensional similarity fully considers the local structure characteristics between vertices to calculate the similarity between vertices. First, we construct a similarity matrix based on the first and second-order neighbor vertices and select a leaf vertex. Second, we expand the leaf vertex based on the principle of maximum community density and construct a non-binary tree. Finally, we choose the layer with the largest overlapping modularity as the result of community division. Experiments on real-world networks demonstrate that our proposed algorithm is superior to other representative algorithms in terms of the quality of overlapping community detection.

The estimation of multitype cardiac indices from cardiac magnetic resonance imaging (MRI) and computed tomography (CT) images attracts great attention because of its clinical potential for comprehensive function assessment. However, the most exiting model can only work in one imaging modality (MRI or CT) without transferable capability. In this article, we propose the multitask learning method with the reverse inferring for estimating multitype cardiac indices in MRI and CT. Different from the existing forward inferring methods, our method builds a reverse mapping network that maps the multitype cardiac indices to cardiac images. The task dependencies are then learned and shared to multitask learning networks using an adversarial training approach. Finally, we transfer the parameters learned from MRI to CT. A series of experiments were conducted in which we first optimized the performance of our framework via ten-fold cross-validation of over 2900 cardiac MRI images. Then, the fine-tuned network was run on an independent data set with 2360 cardiac CT images. The results of all the experiments conducted on the proposed adversarial reverse mapping show excellent performance in estimating multitype cardiac indices.

Traditional feature selection methods assume that all data instances and features are known before learning. However, it is not the case in many real-world applications that we are more likely faced with data streams or feature streams or both. Feature streams are defined as features that flow in one by one over time, whereas the number of training examples remains fixed. Existing streaming feature selection methods focus on removing irrelevant and redundant features and selecting the most relevant features, but they ignore the interaction between features. A feature might have little correlation with the target concept by itself, but, when it is combined with some other features, they can be strongly correlated with the target concept. In other words, the interactive features contribute to the target concept as an integer greater than the sum of individuals. Nevertheless, most of the existing streaming feature selection methods treat features individually, but it is necessary to consider the interaction between features. In this article, we focus on the problem of feature interaction in feature streams and propose a new streaming feature selection method that can select features to interact with each other, named Streaming Feature Selection considering Feature Interaction (SFS-FI). With the formal definition of feature interaction, we design a new metric named interaction gain that can measure the interaction degree between the new arriving feature and the selected feature subset. Besides, we analyzed and demonstrated the relationship between feature relevance and feature interaction. Extensive experiments conducted on 14 real-world microarray data sets indicate the efficiency of our new method.

Quantification of coronary artery stenosis on X-ray angiography (XRA) images is of great importance during the intraoperative treatment of coronary artery disease. It serves to quantify the coronary artery stenosis by estimating the clinical morphological indices, which are essential in clinical decision making. However, stenosis quantification is still a challenging task due to the overlapping, diversity and small-size region of the stenosis in the XRA images. While efforts have been devoted to stenosis quantification through low-level features, these methods have difficulty in learning the real mapping from these features to the stenosis indices. These methods are still cumbersome and unreliable for the intraoperative procedures due to their two-phase quantification, which depends on the results of segmentation or reconstruction of the coronary artery. In this work, we are proposing a hierarchical attentive multi-view learning model (HEAL) to achieve a direct quantification of coronary artery stenosis, without the intermediate segmentation or reconstruction. We have designed a multi-view learning model to learn more complementary information of the stenosis from different views. For this purpose, an intra-view hierarchical attentive block is proposed to learn the discriminative information of stenosis. Additionally, a stenosis representation learning module is developed to extract the multi-scale features from the keyframe perspective for considering the clinical workflow. Finally, the morphological indices are directly estimated based on the multi-view feature embedding. Extensive experiment studies on clinical multi-manufacturer dataset consisting of 228 subjects show the superiority of our HEAL against nine comparing methods, including direct quantification methods and multi-view learning methods. The experimental results demonstrate the better clinical agreement between the ground truth and the prediction, which endows our proposed method with a great potential for the efficient intraoperative treatment of coronary artery disease.

The paper-reviewer recommendation task is of significant academic importance for conference chairs and journal editors. It aims to recommend appropriate experts in a discipline to comment on the quality of papers of others in that discipline. How to effectively and accurately recommend reviewers for the submitted papers is a meaningful and still tough task. Generally, the relationship between a paper and a reviewer often depends on the semantic expressions of them. Creating a more expressive representation can make the peer-review process more robust and less arbitrary. So the representations of a paper and a reviewer are very important for the paper-reviewer recommendation. Actually, a reviewer or a paper often belongs to multiple research fields, which increases difficulty in paper-reviewer recommendation. In this article, we propose a Multi-Label Classification method using a HIErarchical and transPArent Representation named Hiepar-MLC . First, we introduce HIErarchical and transPArent Representation (Hiepar) to express the semantic information of the reviewer and the paper. Hiepar is learned from a two-level bidirectional gated recurrent unit based network applying the attention mechanism. It is capable of capturing the two-level hierarchical information (word-sentence-document) and highlighting the elements in reviewers or papers to support the labels. This word-sentence-document information mirrors the hierarchical structure of a reviewer or a paper and captures the exact semantics of them. Then we transform the paper-reviewer recommendation problem into a multi-level classification issue, whose multiple research labels exactly guide the learning process. It is flexible in that we can select any multi-label classification method to solve the paper-reviewer recommendation problem. Further, we propose a simple multi-label-based reviewer assignment (MLBRA) strategy to select the appropriate reviewers. It is interesting in that we also explore the paper-reviewer recommendation in the coarse-grain granularity. Extensive experiments on the real-world dataset consisting of the papers in the ACM Digital Library show that Hiepar-MLC achieves better label prediction performance than the existing representation alternatives. In addition, with the MLBRA strategy, we show the effectiveness and the feasibility of our transformation from paper-reviewer recommendation to multi-label classification.

Citation recommendation is of great value in scientific research during the era of big scholarly data. The task deals with different object types (e.g., paper, author, etc.) and relation types, which naturally constitute a heterogeneous information network (HIN). In order to capture semantic relations and attribute values on relations, we propose a Weighted Heterogeneous Information Network Containing Semantic Linking algorithm (WHIN-CSL) to recommend references. Firstly, we construct a weighted HIN consisting of two kinds of vertexes (papers and authors) and four kinds of relations (semantic linking, citing, writing and co-author). Secondly, we use the network representation learning method to obtain feature representation of each vertex, then we can compute the similarity between vertexes. Finally, we recommend references through linear combination among multimodal similarities. Experimental results on two real-world datasets show that WHIN-CSL achieves better performance because of the flexibly integrating information with the help of weighted HIN containing semantic linking.

The optimal path finding problem in weighted edge networks is an old and interesting one in many fields. There were many well-known algorithms to deal with that issue. But they were confronted with the high computational complexity while the network becoming larger. We present a hierarchical quotient space model based algorithm that reduces the computational complexity. The basic idea is the following. The nodes of a given network are partitioned with respect to the weights of their adjacent edges. We construct a variety of coarser versions of the given network with new nodes corresponding to the blocks of partitions at various levels of granularity. They are called the quotient spaces (networks) of the original network. The construction of the (sub- )optimal path is then done incrementally, throughout the hierarchy of quotient networks. Since each version of the network is much simpler than the original one, especially of the coarsest spaces, the computational complexity is reduced. In this paper, we present the basic principles of the algorithm and its experimental comparison to other well-known algorithms.