A time series is a common form of data that contains a wealth of valuable information. Sequential pattern mining (SPM) methods can be used to analyze this type of series; however, traditional SPM methods mainly focus on analyzing one-dimensional time series, and it is difficult to analyze multivariate time series. More importantly, multivariate time series may interact with each other. To address these issues, this paper explores one-off SPM for multivariate time series analysis (OPM) and presents an OPM-Miner algorithm. At data preprocessing stage, the original multivariate time series is converted to binary inverted dictionaries to avoid the need to scan the original database. At support calculation stage, we introduce strategies for pruning infrequent items and bit index sliding, and develop strategies called Over and Under to further improve the efficiency of support calculation. This paper selected 10 competitive algorithms and 16 real-life multivariate databases to verify the performance of OPM-Miner. The experimental results show that OPM-Miner outperforms other competitive algorithms. More importantly, OPM-Miner can be used for air pollution detection, as it reveals the interactive relationships between various air pollutants, and can predict future trends in air pollutants.

Graph-based Retrieval-Augmented Generation (GraphRAG) mitigates hallucinations in Large Language Models (LLMs) by grounding them in structured knowledge. However, current GraphRAG methods are constrained by a prevailing build-then-reason paradigm, which relies on a static, pre-constructed Knowledge Graph (KG). This paradigm faces two critical challenges. First, the KG's inherent incompleteness often breaks reasoning paths. Second, the graph’s low signal-to-noise ratio introduces distractor facts, presenting query-relevant but misleading knowledge that disrupts the reasoning process. To address these challenges, we argue for a reason-and-construct paradigm and propose Relink, a framework that dynamically builds a query-specific evidence graph. To tackle incompleteness, Relink instantiates required facts from a latent relation pool derived from the original text corpus, repairing broken paths on the fly. To handle misleading or distractor facts, Relink employs a unified, query-aware evaluation strategy that jointly considers candidates from both the KG and latent relations, selecting those most useful for answering the query rather than relying on their pre-existence. This empowers Relink to actively discard distractor facts and construct the most faithful and precise evidence path for each query. Extensive experiments on five Open-Domain Question Answering benchmarks show that Relink achieves significant average improvements of 5.4% in EM and 5.2% in F1 over leading GraphRAG baselines, demonstrating the superiority of our proposed framework.

Sequential recommendation predicts the next item based on user behavior patterns. However, due to the influence of unobserved confounders, there are often significant differences between the distributions of training and testing data, resulting in Out-Of-Distribution (OOD) problems that severely degrade recommendation performance. Existing causality-based sequential recommendation methods typically model the unobserved confounders as confounders and assume that they are fully observable, and use causal intervention to remove confounding. However, in reality, the unobserved confounders is complex and unobservable. In this case, the model may learn spurious correlations, which impair its generalization ability to OOD data. To address this, we propose Causal Multi-Interest Fusion for Distribution-Robust Sequential Recommendation (CURE). Specifically, we use stable long-term and periodic interests, unaffected by confounders, as mediators, and apply front-door adjustment to remove the influence of unobserved confounders. Based on front-door adjustment theory, we derive the formula for the target causal effect and employ a multi-interest modeling method to obtain comprehensive representations for each mediator, fully uncovering the causal pathways. Furthermore, we design four modules to implement the key components of the target causal effect formula. During model optimization, to address the lack of explicit interest labels, we leverage large language models (LLMs) to generate fine-grained interest labels, providing high-quality supervision signals for model training. Experiments conducted across four real-world datasets demonstrate that CURE effectively alleviates OOD issues and significantly improves sequential recommendation performance. We will release the code once accepted.

In multi-dimensional classification (MDC), instances are associated with multiple class variables that are assumed in the output space, and each class variable corresponds to one heterogeneous class space and characterizes the objects' semantics from one dimension. Learning from MDC examples poses challenges due to the heterogeneity of class spaces, since the outputs from different class spaces are not directly comparable. Moreover, existing approaches often use identical data representation for all labels in a class, which may lead to suboptimal results as each label might be determined by its own specific characteristics. Critically, the inherent incomparability of raw heterogeneous labels prevents existing methods from effectively capturing label correlations, which are essential for guiding feature learning. In this paper, we propose a novel algorithm named LEAD, i.e., learning Label-spEcific feAtures for multi-Dimensional classification. LEAD first resolves label heterogeneity by transforming the original output space into a unified encoded label space through one-hot label encoding. This critical alignment enables explicit extraction of label correlations from the encoded space. To enhance the reliability of the estimation of label correlations, LEAD then leverages feature-space manifold structures via locally linear embedding, propagating labeling information across similar instances to counteract sparsity. Finally, LEAD jointly learns label-specific feature representations and constructs the classifier through sparse learning while incorporating label correlations. Experimental comparisons on fifteen datasets demonstrate that our proposed method outperforms state-of-the-art multi-dimensional classification methods. The code is available at https://github.com/ZhangZan-source/LEAD.

Inductive Knowledge Graph Reasoning (KGR) aims to discover facts in open-domain KGs containing unknown entities and relations, which poses a challenge for KGR models in comprehending uncertain KG components. Existing studies have proposed Knowledge Graph Foundation Models (KGFMs) that learn structural invariances across KGs to handle this uncertainty. Recently, Large Language Models (LLMs) have demonstrated strong capabilities for open-domain knowledge reasoning. As a result, the latest research has focused on LLM-based KGFMs that integrate LLM knowledge with KG context for inductive KGR. However, the intrinsic knowledge of LLMs may be overshadowed by sparse KG context, leading to LLM knowledge distortion, which can cause irreversible damage to model reasoning. Moreover, existing LLM-based KGR methods still struggle to fully constrain generative hallucinations in LLMs, severely limiting the credibility of reasoning results. To address these limitations, we propose a Knowledge Reasoning Language Model (KRLM) that achieves unified coordination between LLM knowledge and KG context throughout the KGR process. Specifically, we design a Knowledge Reasoning Language (KRL) instruction format and a KRL tokenizer to align LLM knowledge with KG representations. Then, we propose a KRL attention layer that coordinates intrinsic LLM knowledge with additional KG context through a dynamic knowledge memory mechanism. Finally, a structure-aware next-entity predictor is proposed, which strictly constrains the reasoning results within a trustworthy knowledge domain. Extensive experimental results on 25 real-world inductive KGR datasets demonstrate the significant superiority of the proposed KRLM in both zero-shot reasoning and fine-tuning scenarios.

Database parameter automatic tuning is a significant challenge for database administrators (DBAs) in artificial intelligence (AI) enabled database (DB) systems. Optimizing key parameters is crucial for identifying critical interactions among them. Aiming to overcome the disadvantages of existing methods, we propose a collaborative multi-agents model called CMA+DB to automatically tune DB parameters in an effective and efficient fashion. CMA+DB integrates three components including SAPM (Single-Agent Pre-trained Model), MATM (Multi-Agent Joint Training Model), and PJTM (Probability-based Joint Training Model). SAPM applies the deep deterministic policy gradient to explore the impact of one single agent on DB performance, MATM uses multi-agent deep deterministic policy gradients to find agents that collaboratively work to improve DB performance, and PJTM can enhance parameter tuning by important agents based on a probabilistic selection factor. In the CMA+DB model, each agent is responsible for tuning a portion of the parameters, and multiple agents collaborate to recommend the optimal parameter configuration. This hybrid model can expand the number of tunable parameters in order to perform parameter tuning from the aspects of functions and parameter levels (i.e., global, DB, and session level). Experimental results reveal that CMA+DB obtains the fastest convergence performance (when reaching the largest throughput) of 14.83% faster than the state-of-the-art (SOTA) algorithms in the TPC-C benchmark on average. Essentially, after the phase of SAPM model training, CMA+DB outperforms the performance of the SOTA models in throughput. Furthermore, DB performance of CMA+DB can be improved by 1.758% through the phases of MATM and PJTM model training.

In recent years, multi-view feature selection has demonstrated significant value in fields such as machine learning and data mining, as it can effectively reduce data dimensionality, remove redundant information, and enhance model performance by leveraging complementary information across multiple views. However, existing algorithms primarily focus on the correlations between features and the target variable within each view, while neglecting the potential causal relationships among features across different views, and thus lack interpretability in their feature selection results. To address these limitations, we propose a Multi-View Causal Feature Selection (MVCFS) algorithm to identify key features with causal interpretability across multiple views. Specifically, MVCFS first uncovers the causal relationships among features across different views based on the local causal structures within each view, and then constructs the complete Markov Blanket of the class variable in the multi-view scenario. Experiments conducted on 12 benchmark datasets demonstrate that MVCFS achieves higher classification accuracy than other multi-view feature selection algorithms on most datasets.

In sequential pattern mining (SPM), sequences can be divided into two categories: simple sequences composed of items and general sequences composed of itemsets. Previous one-off SPM methods have mainly discovered patterns from simple sequences. More importantly, although most SPM methods discover all frequent patterns, some of the patterns mined in this way are not of interest to users. To tackle these issues, all items are divided into strong or weak interest, based on the interest level of the user, and we discover one-off weak-gap strong patterns (OSPs) composed of strong interest items from sequences with itemsets, using an effective algorithm called OSP-Miner. At the preparation stage, OSP-Miner creates binomial search lists for patterns of length two, thereby avoiding the redundant subpattern matching processes. At the support calculation stage, OSP-Miner first creates m-1 (m > 2) level nodes based on the binomial search lists, and then employs a depth-first search strategy to calculate the support of the candidate pattern. At the candidate pattern generation stage, I-Join and S-Join algorithms are employed to reduce the number of candidate patterns. Experimental results show that OSP-Miner outperforms competitive algorithms, and a case study demonstrates that OSP-Miner yields better performance in a clustering analysis of driving trajectories.

Cognitive diagnosis (CD) is an important task in the field of intelligent education, aiming to discover the proficiency of students on knowledge concepts with response logs. In applications, different users of the tutoring system demand for a diagnosis of knowledge concepts at different granularities. However, recent methods assume that the concepts are of the same granularity and use explicit correlations between same-granularity concepts to improve the diagnosis performance. If required for diagnosing multi-granularity concepts, these methods will face diminished performance or partial invalidation. To this end, we make the first attempt for multiple-granularity cognitive diagnosis, i.e., diagnosis on coarse- and fine-grained concepts simultaneously. Specifically, in a skillful way, the same-granularity correlations are captured and embedded into concept representations in view of concept semantics and cross-granularity correlations to model the proficiency influence between concepts implicitly. Then, the specific loss for single-granularity diagnosis and the general loss for the consistency of multi-granularity are designed to train the model jointly, achieving multiple-granularity diagnosis. Extensive experiments demonstrate that our method can achieve state-of-the-art accuracy on both coarse- and fine-grained concepts.

Accurately capturing user preferences across diverse domains is a fundamental challenge in cross-domain recommendation (CDR) systems. Recent literature has established that disentangling user preferences into global and domain-specific components significantly enhances recommendation performance. However, existing CDR systems are hindered by two critical challenges: 1) how to align user representations across domains to accommodate the discrepancy of user intents in different domains; and 2) how to mitigate data conflicts caused by cross-domain heterogeneity. To address these limitations, we propose intent-aware contrastive learning for cross-domain recommendation (ICCDR). This framework learns unified user interests across domains while mitigating data conflicts from heterogeneity. The ICCDR framework consists of the following core components: 1) a cross-intent module that dynamically identifies users' consistent intents across domains to guide the transfer of knowledge; and 2) an information bottleneck-based aggregation module designed to mitigate data conflicts by filtering out task-irrelevant information, thereby retaining only the essential latent features required for effective knowledge transfer. Extensive experiments conducted on Amazon datasets demonstrate that the ICCDR framework outperforms traditional baseline methods.

Feature selection aims to select the optimal feature subsets from the dataset and has been widely applied in many fields and systems. However, data are not always static, and most of them are unlabeled. Besides, features may be heterogeneous and generated dynamically in practical applications. Therefore, online streaming feature selection was proposed that assumes the features are generated one by one or group by group on the fly while the number of instances remains fixed. This paper focuses on a new practical issue of online unsupervised streaming feature selection where the features are heterogeneous and dynamically generated in groups. Difficulties come from three aspects: the lack of label information, the uncertainty about the feature space, and the dynamic generation of heterogeneous streaming features. To solve this issue, we propose a new online Unsupervised Heterogeneous Group Streaming Feature Selection method named UHGSFS. To handle the problem of heterogeneous streaming features without the feature type information, UHGSFS applies MIC (Maximal Information Coefficient) to evaluate feature relationships without assuming data distribution in advance. To address the challenge of unlabeled information, UHGSFS clusters streaming features by the density based on the Gaussian kernel function and minimizes redundancy by selecting representative features. Extensive experiments were conducted on 13 benchmark datasets, with comprehensive comparisons against state-of-the-art supervised and unsupervised streaming feature selection methods. The experimental results demonstrate that our proposed method achieves comparable or even superior performance relative to supervised streaming feature selection methods.

Multimodal Entity Alignment (MMEA) aims to identify equivalent entities across different data modalities, enabling structural data integration that in turn improves the performance of various large language model applications. To lift the requirement of labeled seed pairs that are difficult to obtain, recent methods shifted to an unsupervised paradigm using pseudo-alignment seeds. However, unsupervised entity alignment in multimodal settings remains underexplored, mainly because the incorporation of multimodal information often results in imbalanced coverage of pseudo-seeds within the knowledge graph. To overcome this, we propose PSQE (Pseudo-Seed Quality Enhancement) to improve the precision and graph coverage balance of pseudo seeds via multimodal information and clustering-resampling. Theoretical analysis reveals the impact of pseudo seeds on existing contrastive learning-based MMEA models. In particular, pseudo seeds can influence the attraction and the repulsion terms in contrastive learning at once, whereas imbalanced graph coverage causes models to prioritize high-density regions, thereby weakening their learning capability for entities in sparse regions. Experimental results validate our theoretical findings and show that PSQE as a plug-and-play module can improve the performance of baselines by considerable margins.

Image clustering is a crucial but open and challenging task in machine learning and computer vision. Deep image clustering methods have made significant advancements in largescale and high-dimensional image datasets, but they only explore clusters according to feature similarity. How to improve the semantic plausibility of these clusters remains a challenging problem. To address this problem, we propose a joint semantic image clustering (SIC) with deep learning framework. Our key idea is to explore semantic clusters from both instance-level and clusterlevel perspectives. At the cluster level, we assume that the highest-confidence partitions possess reasonable semantic divisions; at the instance level, we assume that neighboring samples belong to the same semantic category. With the constraints of the cluster level and instance level, the intracluster compactness and interclass discrepancy are increased. Clustering results heavily rely on the quality of image representations. Therefore, we introduce strong data augmentations with three shared-weight backbone networks to learn the most inherent features. The experimental results on various image datasets demonstrate our framework’s superiority over a wide range of state-of-the-art approaches.

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.

Knowledge Graph Reasoning (KGR) aims to discover latent facts by leveraging the structural evidence available in KGs, posing a challenge to the structural semantic understanding capability of KGR models. Recent studies have demonstrated that Large Language Models (LLMs) can achieve remarkable progress on KGR tasks via flexible in-context learning. However, the inherent representation inconsistency between KG structural context and LLM parametric knowledge remains inadequately addressed. This limitation prevents LLMs from effectively perceiving reasoning evidence that aligns with KG constraints, which undermines both the effectiveness and faithfulness of reasoning. We refer to this problem as reasoning evidence perception drift of LLMs over KGs. To address this problem, we propose a Structure-Internalized Rule Language Model (SIRLM), which centers on structural rule generation to couple the parametric learning of structural knowledge with the faithfulness evaluation of reasoning logic, enabling LLMs to anchor tightly to KG-grounded evidence. Specifically, we first design a Structure-Internalized Rule Generator (SIRG), which incorporates an in-context learning block augmented with a structural relation memory to coordinate structural and parametric knowledge. Furthermore, we equip SIRG with a KG tokenizer based on structural invariance learning and a neuro-symbolic reasoner based on rule-constrained message propagation. These components provide SIRG with learnable structural representations and faithful rule-execution feedback, respectively. Our SIRLM can be seamlessly integrated into standard LLM training paradigms, such as SFT and GRPO. Extensive experiments against 17 state-of-the-art KGR methods on 36 datasets demonstrate the significant superiority of SIRLM.

In partial multi-label learning (PML), each instance is associated with a candidate label set containing both ground-truth and noisy labels. While existing methods predominantly disambiguate labels by eliminating noise from candidate sets, they often overlook the high-precision information inherent in non-candidate labels - which reliably indicate label absence. To address this, we propose Partial Multi-label learning via Competitive Disambiguation (PMCD). Our approach constructs contrastive prototypes: positive prototypes from candidate labels and negative prototypes from non-candidate labels. By optimizing the distance loss between instances and these prototypes, PMCD simultaneously derives a label confidence matrix (quantifying label presence) and a complementary label confidence matrix (quantifying label absence). These matrices exhibit an intrinsic counteractive relationship, enabling the highly reliable complementary matrix (derived from accurate non-candidate labels) to iteratively refine the label confidence matrix through adversarial optimization. This iterative competitive disambiguation mechanism progressively enhances prediction accuracy. Extensive experiments on synthetic and real-world datasets demonstrate that PMCD significantly outperforms state-of-the-art methods in recovering ground-truth labels from noisy candidate sets.

Heterogeneous graph neural networks (HGNNs) have demonstrated exceptional capabilities in modeling complex relationships for recommendation tasks. Their integration with contrastive learning (CL) has recently garnered significant attention due to its ability to effectively capture both structural and semantic features, while leveraging unlabeled data to improve robustness. However, existing methods face two key challenges: 1) noise in metapath-based message passing weakens user and item representations, especially with sparse data and 2) popularity bias in heterogeneous graphs causes popular items to dominate, while less popular “tail” items suffer from limited connections and relations. To address these challenges, we propose heterogeneous debiasing CL (HDCL), which combines heterogeneous graph debiasing with a neighborhood-aggregated strategy to enhance recommendation accuracy and diversity. Specifically, HDCL employs a neighborhood-aggregated CL (NACL) algorithm, which utilizes k-nearest neighbor features to enhance training signals and mitigate noise in sparse data settings. Additionally, HDCL introduces a hierarchical clustering-based debiasing (HCD) mechanism that dynamically distinguishes head nodes from tail nodes, refining the embedding space and improving recommendations for long-tail items. Extensive experiments on public datasets demonstrate that HDCL consistently outperforms state-of-the-art methods in terms of Recall and NDCG. The source code for the model implementation is available at the link https://github.com/Jhcodeno1/HDCL

Join order selection is one of the most challenging problems in query optimization and plays an essential role in providing high query performance in Big Data management. Currently, researchers have applied deep reinforcement learning methods, for example, Rejoin and DQ, to join order selection in order to obtain high query performance. However, Rejoin and DQ cannot capture the structural characteristic of the join tree, which may lead to similar encoding structure for different execution plans. To tackle these challenges, we propose a new learning optimizer called ACJoin (asynchronous advantage Actor-Critic for multi-table Join order selection). ACJoin employs a new encoding method to capture the structural characteristics of the join tree through integrating GRU (Gated Recurrent Unit). In particular, ACJoin can distinguish different execution plans. It uses A3C (Asynchronous Advantage Actor-Critic) to guide the join order selection and reduce the time taken to find the best query plan with the minimum cost. Compared with existing search strategies, ACJoin can find the globally optimal solution with efficient and stable query performance. Extensive experiments are conducted on the real JOB and the synthetic TPC-H datasets. The results show that ACJoin outperforms the state-of-the-art join order selection methods and DRL Deep Reinforcement Learning)-based methods in cost and latency.