Recent developments have demonstrated the efficacy of enhancing the reliability and reasoning ability of Large Language Models (LLMs) through the integration of Knowledge Graphs (KGs). However, previous works have not fully considered the challenges present in expert domains, including the prevalence of professional terms with abbreviations and ambiguities, knowledge confusion stemming from the mixture of KGs with other retrieval methods, and others, which hinder their deployment in expert question-answering (QA) systems. In this work, we propose a QA framework for expert fields by Synergizing large Language models And knowledge Graphs (SLAG). SLAG is designed to improve the accuracy of KG retrieval and to intelligently coordinate KG retrievers with other retrieval methods through bidirectional enhancement between LLMs and KGs: (1) An LLM-Enhanced KG retriever is proposed to resolve entity ambiguities in queries and to accurately extract pertinent subgraphs. (2) A KG-Enhanced LLM reasoner is designed to distill subgraphs and to flexibly generate answers or explicit queries for cooperative retrievers. We compared SLAG with other retrieval techniques on multiple datasets across diverse expert domains, tasks, and languages. We observed that our approach achieved a relative improvement of 13% over state-of-the-art methods on public benchmarks in terms of the F1 score. Moreover, we successfully deployed SLAG in an industrial financial QA system and outperformed the latest online version. Finally, we provide an open source version of SLAG with code and datasets, which is available at https://anonymous.4open.science/r/SLAG-5611.
Data simulation is an established methodology in data engineering for augmenting limited datasets and safeguarding privacy, particularly for attributed graphs in domains such as finance and social networks. While existing methods successfully replicate topological and attribute distributions (statistical fidelity), we claim that such statistical similarity does not inherently ensure explainability alignment—which we define as the preservation of the underlying rationales for model decisions inherent in the data. Synthetic graphs that are statistically similar but explainability-misaligned can lead downstream models to capture spurious correlations, resulting in unreliable decision-making. To address this, we formalize the task of Explainability-Aligned Attributed Graph Generation (EAGG), which seeks to produce synthetic graphs that satisfy both statistical fidelity and explainability alignment. We propose X-AGG, a tailored framework that distills these predictive rationales into quantitative importance signals and integrates them as explicit conditions into a Variational Autoencoder architecture. By steering the generative process with explainability signals, X-AGG produces synthetic graphs that are intended to be both statistically representative and aligned with the original predictive logic. Experimental results indicate that X-AGG achieves state-of-the-art performance across traditional statistical metrics while demonstrating superior capability in preserving explainability alignment compared to existing baselines.
Spatial co-location pattern mining aims to discover subsets of spatial features that frequently occur in close geographic proximity. However, traditional methods usually discover patterns of all sizes, making them inefficient for mining patterns of a specific size specified by users. To address this issue, we define a problem called k-size Spatial Co-location Pattern (k-SCP) Mining, which focuses on discovering co-location patterns of a user-specified size k. Unlike the general size-wise SCP mining method, the k-SCP mining method directly targets patterns of specific sizes. Thus, it cannot utilize the downward closure property for candidate pruning. For this reason, we propose a novel k-SCP mining framework based on a representation model IC+(k-1)deg, called k-pattern-Miner. The IC+(k-1)deg integrates instance coloring with (k-1)-degree constraint mechanism, enabling the iterative elimination of instances and features that are irrelevant to k-SCPs, which substantially reduces the verification cost in subsequent mining steps. Extensive experiments on both real-world and synthetic datasets validate the efficiency of the k-pattern-Miner. We selected several state-of-the-art SCP mining methods to support k-SCP mining for comparison. Experimental results show that k-pattern-Miner improves the running time by several orders of magnitude compared to existing methods.
Node centrality is one of the most fundamental metrics in network science. One notion of centrality that has garnered significant popularity, due to its ability to capture complex local and global structures, is the resolvent subgraph centrality. For a node i, this measures the weighted sum of the number of closed walks starting and ending at i. We first provide a natural and intuitive generalization of resolvent subgraph centrality to cover the larger class of signed graphs. As our main contribution, we present two scalable approximation algorithms: SOLVERRSC, which leverages a linear equation solver, and SIMPLERSC, which utilizes a novel rooted forest sampling technique. We then provide the algorithms GROUPRSC and NEIGHBORRSC, which achieve better accuracy by extracting richer information from sampled forests. Finally, we conduct an extensive set of experiments to validate the accuracy and efficiency of the proposed algorithms. Our code is available at https://github.com/SunYubo23/SignedRSC.
Link prediction fails on triples with unseen entities due to sparse representations from missing embeddings. To address this, GGTSM is proposed, an adaptive knowledge distillation framework using a hybrid Graph Convolutional Network - Graph Attention Network architecture for teacher and student models. GGTSM predicts external entities based on known triple structure and relationships. The teacher network learns entity and relation embeddings via supervised training. An adaptive weighting mechanism, using relations and node attention scores generated by teacher, calculates distillation loss weights to guide the student in learning critical graph patterns. During reasoning for triples containing unseen entities, the student performs guided fine-tuning using output of teacher, without external text features. GGTSM employs subgraph analysis to integrate batch-level knowledge, initializing embeddings from adjacent nodes and relations of input triples. Experiments on FB15k_237, WN18RR, and NELL-995 show GGTSM achieves significant gains: an average 0.15 improvement in Mean Reciprocal Rank, and average improvements of 18% (Hits@1), 41% (Hits@3), and 63% (Hits@10) over the baseline. These results confirm adaptive knowledge distillation reduces representation sparsity for unseen entities, enhances generalization, and provides a flexible paradigm for dynamic knowledge graph reasoning.
Time Series Class-Incremental Learning (TSCIL) seeks to continuously learn new tasks from streaming data while retaining previously acquired knowledge. Existing replay-based and exemplar-free methods can alleviate catastrophic forgetting, but they still have limitations: (i) gradients from new tasks interfere with previous knowledge, overwriting old representations; (ii) ignoring the inherent heteroscedasticity of time series data allows high-variance features to dominate model updates, accelerating forgetting; and (iii) task-wise distribution shifts induce representation drift, misaligning representations with prior decision boundaries. Therefore, we propose the Heteroscedastic-aware Analytical Dynamic Adaptation (HADA) framework for TSCIL. HADA mitigates catastrophic forgetting through a two-stage paradigm that uses gradient-free, recursive, closed-form updates. In the pre-training phase, we introduce a layer-wise perturbation consistency regularization that guides the encoder to learn robust and stable temporal representations. During incremental learning, the encoder is frozen, and a heteroscedastic-aware analytical learner downweights statistically unstable dimensions while emphasizing information-rich ones. The learner then performs lightweight closed-form recursive ridge updates, avoiding gradient interference. Experiments on multiple TSCIL benchmarks demonstrate that HADA achieves competitive performance in accuracy, stability, and computational efficiency.
Multi-view unsupervised feature selection (MUFS) aims to extract discriminative low-dimensional representations from multi-source high-dimensional data without class labels' supervision, thereby mitigating the curse of dimensionality. Existing correlation-based methods often fail to capture latent causal relationships, and thus the learned feature subsets might be dominated by spurious correlation features, leading to performance degradation in downstream tasks. However, the absence of ground-truth labels brings significant difficulty in identifying causal features in multi-view unsupervised scenarios. While pseudo-labels offer a potential pathway, directly relying on conventional similarity metrics to construct similarity graphs renders the resulting pseudo-labels highly vulnerable to topological distortions and data noise, thereby severely reducing their reliability and hindering the accurate discovery of causal features within data. To address the challenges, we propose a novel Gravitation and Causality guided Multi-view Unsupervised Feature Selection, named GC-MUFS, which includes three key innovations: (1) physics-inspired gravitational modeling: employing a mass-density coupling mechanism to effectively suppress unreasonable inter-cluster connections, leading to similarity graphs that more faithfully capture the underlying data structure. (2) Low-rank decomposition: isolating primary patterns from redundant perturbations in gravitational similarity graphs to boost pseudo-label reliability. (3) Causality-guided feature optimization: selecting potential causal features for reducing the influence of spurious features with dynamic pseudo-label learning. Experimental results demonstrate that our method outperforms existing MUFS approaches in terms of clustering accuracy and purity.
Anomaly detection is a crucial problem in data mining and analytics with widespread applications. Existing studies, however, often treat all anomalies equally, overlooking the fact that in many real-world scenarios, anomalies can have varying levels of importance. In this paper, we consider priority-aware anomaly detection. A key challenge in this problem is the overlap between marginal unlabeled instances and anomalies, which leads to high false-positive rates. Most existing models focus on distinguishing between normal and abnormal instances, largely neglecting the potential to distinguish marginal unlabeled data from anomalies. To address this issue, we propose EdgeBoost4PAD (EBPAD), an adaptive strategy that aims to increase the separation between marginal unlabeled instances and labeled anomalies, thus enhancing a model's ability to learn a more discriminative and robust decision boundary. We generalize a distance-based instance partitioning approach to make EBPAD applicable to various semi-supervised anomaly detection models operating in both score and representation spaces. The experimental results demonstrate that EBPAD consistently improves the performance of different anomaly detection models across diverse datasets and settings, yielding average increases of 7% in AUROC and 38.7% in AUPRC under the priority-aware anomaly detection setting. In addition, EBPAD also attains average improvements of 6.4% in AUROC and 36.5% in AUPRC under the priority-unaware setting.
Semi-structured data, encompassing formats such as logs, JSON files, and XML documents, is proliferating due to its flexible, partially organized nature. This growth outpaces advancements in storage technology, creating challenges for data-intensive systems where storage costs and random access performance are critical concerns. Existing compression techniques, primarily designed for general-purpose or block-level data, fail to leverage the inherent structural patterns in semi-structured data, resulting in suboptimal compression ratios or inefficiencies in lookup operations. To address these limitations, we propose Pattern-Based Compression (PBC), which specifically targets patterns in semi-structured data to achieve Pareto-optimality in most cases. Unlike traditional data block-based methods, PBC compresses data on a per-record basis, facilitating rapid random access. Empirical evaluation on real-world and benchmark datasets demonstrates that PBC achieves up to 50% better compression ratios than state-of-the-art methods while maintaining superior random access performance and competitive throughput. We also integrate PBC into a production database system and achieve improvements in both compression ratio and throughput.
Current offline reinforcement learning (ORL) algorithms tend to overfit the training dataset and exhibit poor in-distribution generalization and robustness performance when deployed to real environments, thus compromising their effectiveness. Existing methods typically enhance in-distribution generalization and robustness by leveraging regularization techniques widely used in computer vision. However, due to the high sensitivity of low-level physical signals to distributional shifts, these methods still suffer from notable limitations in in-distribution generalization and robustness, making it difficult to achieve stable performance in complex environments. To address this issue, we theoretically analyze the error bounds of the transition function under random episode interpolation, showing that the interpolation error increases with the distance between states. Based on this insight, we propose a method called $\bf{B}$oundary-$\bf{A}$ware $\bf{D}$ata $\bf{A}$ugmentation (BADA), which leverages neighboring states to construct interpolation boundaries, enabling the generation of synthetic data that more faithfully preserves the original data distribution. We first conduct qualitative studies in a toy environment, showing that BADA generates mixed samples that preserve desirable policy smoothness while faithfully maintaining multimodal value distributions. Extensive experiments on limited offline datasets demonstrate that BADA attains state-of-the-art performance across diverse benchmarks. We further evaluate BADA under noise contamination and both white-box and black-box adversarial attacks, with results demonstrating that the proposed method enhances robustness to state perturbations. Finally, we investigate offline datasets with noisy transition dynamics and find that BADA effectively mitigates performance degradation and improves robustness against transition model perturbations.
Sparse Mobile Crowdsensing (SMCS) has emerged as a prominent data engineering paradigm for large-scale, trust-aware environmental data acquisition, enabling diverse data-driven cyber-physical applications. However, the sparsity of the collected sensing data and the presence of untrustworthy workers present key challenges for optimizing data collection and inference accuracy. In particular, malicious workers may submit falsified data, significantly compromising system reliability and the accuracy of data-driven inference. To address these challenges, we propose a Trustworthy Worker Recruitment framework based on Proximal Policy Optimization (PPO-TWR) for trust-aware data collection and inference. The framework evaluates the trustworthiness of workers and leverages high-trust contributions to enhance both data acquisition and inference accuracy in sparse sensing scenarios. The main contributions are threefold: (1) We develop a trust evaluation algorithm that incorporates behavioral scoring to quantitatively evaluate each worker's reliability and contribution. (2) We design a multi-dimensional trust-aware data collection and inference strategy that jointly optimizes data coverage, inference accuracy, trust evaluation, and acquisition cost. (3) To improve the efficiency of PPO training, an action masking mechanism is introduced to mask invalid actions during training, thereby accelerating policy convergence. Extensive simulations demonstrate that the proposed framework significantly improves both data quality and inference accuracy while reducing the cost of data acquisition. Compared to baseline methods, our approach achieves a reduction of 22.23%-57.22% in collection cost and a 12.79%-56.45% improvement in inference accuracy, highlighting its effectiveness in trust-sensitive mobile crowdsensing scenarios.
The aggregation of highly heterogeneous product listings across online webshops suffers from severe scalability limitations and high false positive rates, as identifying duplicates without universal product identifiers typically requires computationally expensive pairwise comparisons. To resolve these specific challenges, this paper introduces two concrete enhancements to Locality-Sensitive Hashing (LSH) within the Multi-component Similarity Method with Pre-selection+ (MSMP+) that significantly improve both overall scalability and duplicate detection precision. First, we substitute the conventional MinHash sketching technique with Fast Similarity Sketching (FSS). FSS provides faster and more precise Jaccard similarity estimates, inherently reducing the time complexity of the initial hashing process. Second, we develop a parametrized amplification strategy leveraging iterative AND/OR operations on LSH functions to substantially minimize both false positive and false negative rates. Crucially, this amplification strategy improves scalability by allowing the algorithm to achieve higher accuracy while using fewer underlying hash functions, thereby directly decreasing the computational overhead of generating candidate pairs. We develop a comprehensive framework for these amplified LSH schemes, deriving optimal parameter configurations by minimizing a weighted error function that balances the trade-off between false positives and false negatives. Empirical evaluations, based on data collected from four webshops, including extensive bootstrap experiments, confirm that combining FSS with amplification consistently outperforms baseline, non-amplified MinHash configurations. To further validate the robustness and generalizability of our approach, we conducted an additional empirical analysis using the established Walmart-Amazon dataset. These supplementary experiments demonstrate that our optimized LSH framework maintains good deduplication precision across different, widely-used product domains. Ultimately, these targeted enhancements enable highly scalable duplicate detection, reflected in improved performance metrics such as pair completeness and pair quality, as well as an increased F1 score.
Query performance prediction is a challenging task in AI-enabled database (DB) research because prediction capability is crucial to optimizing DB operations (e.g., parameter optimization) and resource allocation in modern DB systems. However, existing methods often struggle to meet the flexible requirements for both prediction accuracy and query latency in real-world applications, and focus on predicting the performance of individual queries instead of concurrent queries, which is very universal in large and complex DB tasks. To address these limitations, we propose an efficient prediction model called $C^{3}$-Predictor (Concurrency, Cost factor, and Confidence), which integrates a cost factor model into a graph neural network (GNN) model to reduce the latency of prediction with guaranty of accuracy. First, the cost factor is used to achieve the performance prediction of concurrent queries through a concurrent adjustment mechanism. Second, a confidence estimation algorithm is applied to evaluate the cost factor model, while the confidence mechanism evaluates the reliability of initial predictions to determine whether to activate the GNN model. Third, the GNN model is used to calibrate prediction deviations from the cost factor model, which retains only those features that impact the prediction of concurrent queries to construct graphs, and applies the GNN model to predict query performance in an accurate fashion. Extensive experiments are conducted on varying types of workload. The results demonstrate that $C^{3}$-Predictor outperforms the state-of-the-art (SOTA) models, that is, for different levels of concurrent queries, $C^{3}$-Predictor achieves the lowest prediction error and the lowest prediction latency.
Low-rank tensor decomposition has attracted growing attention in multivariate data analysis owning to its potentials of capturing latent representations of multivariate data correlating various data attributes. However, existing studies still have limited capabilities to represent the nonlinear relationships within multivariate data. Moreover, their performance tends to degrade under the circumstance of intensive noise embedded in the source data as well as improper setting of initial tensor rank. To tackle the challenges, this study proposes a Deep Bayesian Low-rank Tensor Decomposition (DBLTD) framework to handle the multivariate data: 1) DBLTD first establishes a probabilistic tensor decomposition model and imposes a Gaussian-Gamma sparsity prior over factor matrices and noise tensor to induce the sparse components of multidomain factors and separate the noise from the source data, respectively, 2) A generative convolution network is designed on the basis of multiple convolution blocks to produce the distributions of deep factors along with deep parameters in a self-training manner, 3) The framework develops a deep variational Bayesian inference algorithm with guaranteed convergence for model training, where the loss function is constituted of reconstruction error and sparse constraint term over deep factors and parameters. Experimental results on diverse multivariate datasets indicate that DBLTD is effective to learn the tensor rank and reduce the impact of various noises during multivariate data decomposition. Furthermore, DBLTD has demonstrated remarkable advantages in decomposition based medical applications like blind EEG denoising, structural EEG feature construction and MRI image restoration.
Two-phase locking (2PL) is a fundamental and widely used concurrency control protocol. It regulates concur rent access to database data by following a specific sequence of lock acquisition and release during transaction execution, thereby ensuring transaction isolation. However, in strict 2PL, transactions must wait for conflicting transactions to commit and release their locks, which reduces concurrency and system throughput. We have observed that this issue is exacerbated in high-contented workloads at Tencent, where lock contention can severely degrade system performance. While existing optimiza tions demonstrate some effectiveness in high-contention scenarios, their performance remains insufficient because they suffer from lock contention and hotspot access waits. This paper presents optimizations in lock management imple mented in Tencent's database, TXSQL, with a particular focus on high-contention scenarios. First, we discuss our motivations and the journey toward general lock optimization, which includes lightweight lock management, a copy-free active transaction list, and queue locking mechanisms that effectively enhance concurrency. Second, we introduce a hotspot-aware approach that enables certain highly conflicting transactions to switch to a group locking method, which groups conflicting transactions at a specific hotspot and allows them to execute serially in an uncommitted state within a conflict group without the need for locking, thereby reducing lock contention. Our evaluation shows that under high-contented workloads, TXSQL achieves performance improvements of up to 6.5x and up to 22.3x compared to state-of-the-art methods and systems, respectively.
The unsupervised representation learning of the multiplex network has attracted significant attention due to its powerful ability to model multiple relation types between nodes. Existing methods generally learn representations with the original graph structure but disregard the fact that the prior graph structure is inevitably incomplete, which may mislead the feature aggregation process. Therefore, improving the graph structure is of importance for multiplex network unsupervised representation. Unfortunately, heterogeneous graph structure optimization in multiplex networks is still less explored and encounters the challenges of how to jointly consider multiple relation types to supplement the missing edges in each relation. To address this issues, we formulate each relation type as one view and propose the Multiplex network Unsupervised Representation model with Adaptive multi-view graph structure learning (MURA). To improve the graph structure, we develop a multi-view low-rank self-representation tensor learning module to optimize the heterogeneous graph structure which well explores the global community property. Besides, a local neighborhood manifold structure learning module is employed to capture the essential local connection. With the refined multiplex graph, we further propose a similarity regularization term to enhance the multiplex node representations. Comprehensive experiments on three downstream tasks confirm the superiority of our MURA among five datasets when compared to the state-of-the-art rivals.
Interactive Recommendation (IR) has attracted considerable attention in recent years. IR agents commonly utilize Deep Reinforcement Learning (DRL) to capture users' unique dynamic interests and demands. However, it's challenging to train DRL agents in online recommender systems due to the large action space and the sample inefficiency problem. We discover that the principle of tackling this challenge is to integrate neighborhood collaborative information into the representation. To implement this idea and address the challenge, we propose CPIR (Contrastive Paradigm for Representation Learning in Interactive Recommendation), which comprises three Contrastive Learning (CL) approaches, a data sampling mechanism, and an agent training mechanism. The CL methods start from the perspectives of sequence alignment, interest partial order, and interest community collaborative filtering separately. The recommender agent can optimize itself more efficiently on top of the extracted representations. The data sampling and agent training mechanism ensure the stability and consistency of the multi-task training in IR. Moreover, we have explained the effectiveness of the CPIR both theoretically and experimentally. Extensive experiments conducted on Virtual-Taobao and ML-1M have demonstrated superior improvements in sample efficiency during the training of the DRL-based IR agent.
Sequential recommendation aims to predict the next item that aligns with user preferences based on their historical interaction sequence. Traditional sequential recommendation methods can be seen as understanding-based approaches, where the next relevant item is determined by analyzing user historical interactions and rating patterns. Recently, diffusion-based models have emerged as a promising paradigm, focusing on learning data distributions rather than explicitly mining sequential patterns. However, existing diffusion-based methods still face two limitations. Firstly, they often map discrete target items into continuous spaces through transformations, failing to accurately model target items in a way that reflects future user preferences. Secondly, existing conditionally guided diffusion models rely heavily on explicit conditions derived from intra-sequence patterns while neglecting inter-sequence collaborative signals, which hinders the robustness of user preference modeling. To bridge the gap between collaborative signals and diffusion models, we propose DCDRec, a Dual Collaborative Signal-Guided Diffusion Recommendation model. Specifically, for target item representation, we employ a cross-attention based encoder to obtain context-aware target item embeddings. For conditional guidance modeling, we incorporate social homophily theory and item-item affinity into the conditional generation process, introducing a dual collaborative signal-guided denoising mechanism to generate new items. Extensive experiments demonstrate the effectiveness of DCDRec and its superiority over state-of-the-art methods.
Continuous subgraph matching (CSM) is a critical task for analyzing dynamic graphs and has a wide range of applications, such as merchant fraud detection, cyber-attack hunting, and rumor detection. Although many efficient CSM algorithms have been recently proposed, they are mainly designed to process a single query. However, in some application scenarios, multi-query oriented continuous subgraph matching (MQCSM) is more critical. To our knowledge, the two existing solutions to MQCSM are outdated due to unsatisfactory performance. In this paper, we propose MQ-Match, an efficient approach for MQCSM. Specifically, we design a compact index structure, called candidate classification graph (CCG), to maintain local matching relations in the data graph and prune invalid can-didate expansions. We further develop a computation sharing incremental matching method based on shared matching trees, where common structures among query graphs are merged and processed once for each graph update. To reduce redundant search caused by automorphisms, we introduce an optimization approach MQ-Match∗. Specifically, we introduce an exclusion set based optimization that prunes symmetric search branches and directly generates symmetric results. Extensive experiments on real datasets show that MQ-Match∗ achieves 2.0x-37280.7x speedup over existing methods with much less memory in most settings, while scaling to billion-edge graph.
Text embedding has emerged as a pivotal technique in natural language processing, facilitating the effective understanding and processing of textual information by machines. With the continuous advancement of data-driven methods like large language models (LLMs), text embeddings have become richer and of higher quality. However, researchers have identified limitations in purely data-driven methods, which may lack interpretability and logical consistency. Conversely, purely knowledge-driven methods require extensive manual effort from experts to design rules, leading to low efficiency. To overcome these challenges, researchers have explored integrating data-driven and knowledge-driven methods, termed Data and Knowledge Dual-driven Text Embedding (DKDTE). In this paper, we introduce a novel taxonomy categorizing existing text embedding methods into three primary categories, namely, knowledge-driven approaches, data-driven approaches, and dual-driven approaches. We provide formal definitions of text embeddings with distinctions in input granularity, a dedicated overview of application tasks, evaluation benchmarks (including MTEB, BEIR, and AIR-Bench), and real-world applications. We offer a comprehensive comparison of representative methods across categories and discuss the latest advances including LLM-based embedding methods, instruction-tuned embedding paradigms, and multimodal knowledge integration. We also identify promising future research directions, including debiasing, exploring diverse knowledge sources, and data contamination mitigation.
