Abstract:The compiler is one of the most relied-upon performance tuning tools for program developers. However, due to the limited precision encoding of floating-point numbers, many compiler optimization options can alter the semantics of floating-point calculations, leading to result inconsistency. Locating the program statements that cause compilation optimization-induced result inconsistency is crucial for performance tuning and result reproducibility. The state-of-the-art approach employs precision enhancement-based binary search to locate the code snippets causing result inconsistency but suffers from insufficient support for multi-source localization and low search efficiency. This study proposes a floating-point instruction difference-guided Delta-Debugging localization method, FI3D, which utilizes the backtracking mechanism in Delta-Debugging to better support multi-source problem code localization and exploits the differences in floating-point instruction sequences under different compiler optimization options to guide the localization. FI3D is evaluated using 6 applications from the NPB benchmark, 10 programs from the GNU scientific library, and 2 programs from the floatsmith mixed-precision benchmark. Experimental results demonstrate that FI3D successfully locates the 4 applications where PLiner fails and achieves an average 26.8% performance improvement for the 14 cases successfully located by PLiner.

Abstract:With the increasing adoption of heterogeneous integrated architectures in high-performance computing, it has become essential to harness their potential and explore new strategies for application development. Traditional static compilation methodologies are no longer sufficient to meet the complex computational demands. Therefore, dynamic programming languages, known for their flexibility and efficiency, are gaining prominence. Julia, a modern high-performance language characterized by its JIT compilation mechanism, has demonstrated significant performance in fields such as scientific computing. Targeting the unique features of the Sunway heterogeneous many-core architecture, the ORCJIT engine is introduced, along with an on-chip storage management approach specifically designed for dynamic modes. Based on these advancements, swJulia is developed as a Julia dynamic language compiler tailored for the new generation of the Sunway supercomputer. This compiler not only inherits the flexibility of the Julia compiler but also provides robust support for the SACA many-core programming model and runtime encapsulation. By utilizing the swJulia compilation system, the deployment of the NNQS-Transformer quantum chemistry simulator on the new generation of the Sunway supercomputer is successfully achieved. Comprehensive validation across multiple dimensions demonstrates the efficacy and efficiency of swJulia. Experimental results show exceptional performance in single-threaded benchmark tests and many-core acceleration, significantly improving ultra-large-scale parallel simulations for the NNQS-Transformer quantum chemistry simulator.

Abstract:With the rapid development of autonomous driving technology, the issue of vehicle control takeover has become a prominent research topic. A car equipped with an assisted driving system cannot fully handle all driving scenarios. When the actual driving scenario exceeds the operational design domain of the assisted system, human intervention is still required to control the vehicle and ensure the safe completion of the driving task. Takeover performance is an extremely important metric for evaluating a driver’s performance during the takeover process, which includes takeover reaction time and takeover quality. The takeover reaction time refers to the time from the system’s takeover request to the driver’s control of the steering wheel. The length of the takeover response time not only reflects the driver’s current state but also affects the subsequent handling of complex scenarios. Takeover quality refers to the quality of manual vehicle operation by the driver after regaining control. This study, based on the CARLA driving simulator, constructs 6 typical driving scenarios, simulates the vehicle control takeover process, and collects physiological signals and eye movement data from 31 drivers using a multi-channel acquisition system. Based on the driver’s takeover performance, and regarding International standards, an objective takeover performance evaluation metric is proposed, incorporating the driver’s takeover reaction time, maximum horizontal and vertical accelerations, and minimum collision time, derived from multiple vehicle data. By combining driver data, vehicle data, and scenario data, a deep neural network (DNN) model predicts takeover performance, while the SHAP model analyzes the impact of each feature, improving the model’s interpretability and transparency. The experimental results show that the proposed DNN model outperforms traditional machine learning methods in predicting takeover performance, achieving an accuracy of 92.2% and demonstrating good generalization. The SHAP analysis reveals the impact of key features such as heart rate variability, driving experience, and minimum safe distance on the prediction results. This research provides a theoretical and empirical foundation for the safety optimization and human-computer interaction design of autonomous driving systems and is of great significance for improving the efficiency and safety of human-vehicle cooperation in autonomous driving technology.

Abstract:Temporal logic has been extensively applied in domains such as formal verification and robotics control, yet it remains challenging for non-expert users to master. Therefore, the automated extraction of temporal logic formulas from natural language texts is crucial. However, existing efforts are hindered by issues such as sparse sample availability and the ambiguity of natural language semantics, which impede the accurate identification of implicit temporal semantics within natural language texts, thus leading to errors in the translation of the original natural language semantics into temporal logic formulas. To address this issue, a novel method for temporal logic semantic analysis based on a few-shot learning network, termed FSLNets-TLSA, is proposed. This method employs data preprocessing techniques to enhance the temporal semantic logic features of the text. The network architecture consists of an encoder, an induction module, and a relation module, which aim to capture the implicit temporal logic semantic information in the input text. In addition, an enhancement module is incorporated to improve the accuracy of monitoring semantic recognition. The effectiveness of the proposed method is validated through experimental evaluations conducted on three public datasets comprising a total of 3 533 samples, and a comparison with similar tools. The analysis demonstrates an average Accuracy, Recall, and F1-score of 96.55%, 96.29%, and 96.42%, respectively.

Abstract:GUI testing is one of the most important measures to ensure mobile application (App) quality. With the continuous development of the mobile ecosystem, especially the strong rise of the domestic mobile ecosystem, e.g., HarmonyOS, GUI test script recording and replay has become one of the prominent challenges in GUI testing. GUI test scripts must be migrated from traditional mobile platforms to emerging mobile platforms to ensure the reliability of App quality and consistency in user experience across diverse platforms. However, differences in underlying implementations across platforms have created substantial obstacles to the cross-platform migration of mobile App test scripts. This challenge is particularly pronounced in the testing migration for emerging domestic mobile ecosystem platforms. Cross-platform test script recording and replay is essential for maintaining consistency and a high-quality user experience across different platforms and devices. Current state-of-the-art approaches only address the “one-to-one” test event matching situations. However, due to inconsistencies in development practices across platforms, the replay of test events does not always map “one-to-one”; instead, “multiple-to-multiple” mapping situations are common. This means that some test events need to be mapped to a different number of test events to fulfill the same business logic. To address these issues and challenges, this study proposes a cross-platform mobile App test script recording and replay method based on large language model semantic matching (LLMRR). The LLMRR method integrates image matching, text matching, and large language model semantic matching technologies. During the recording phase, user operation information is captured using image segmentation algorithms and saved as recorded test scripts. During the replay phase, corresponding widgets on the replay App page are located using image matching and text matching modules to execute operations. When matching fails, the large language model semantic matching module is invoked for semantic matching, ensuring efficient operation across different platforms. This study presents the first exploration of testing for domestic HarmonyOS Apps, using 20 Apps and a total of 100 test scripts for migration testing across iOS, Android, and HarmonyOS platforms. The effectiveness of the LLMRR method is compared with the current state-of-the-art cross-platform test script recording and replay approaches, LIRAT and MAPIT. The results demonstrate that the LLMRR method exhibits significant advantages in test script recording and replay.

Abstract:Edge servers provide low-latency, high-performance services for mobile intelligent applications. However, due to significant fluctuations in the load on edge servers over time, many edge servers remain idle during periods of low load, and their computational resources are not fully utilized. In contrast to the underutilization of edge servers, computing resources in cloud computing clusters remain relatively scarce for deep learning training tasks as artificial intelligence becomes more widely applied in daily life. Existing cluster scheduling strategies fail to efficiently utilize idle computing resources outside of cloud computing clusters. Effectively utilizing these idle resources can alleviate the resource constraints in cloud computing clusters, thus enabling more deadline-sensitive deep learning training tasks to be completed before their deadlines. To address this issue, this study proposes a cluster scheduling strategy for deadline-sensitive deep learning training tasks, which coordinates the scheduling of cloud computing resources and idle edge computing resources. This strategy fully leverages the performance characteristics of different deep learning tasks and the availability of idle edge server devices, allowing more deadline-sensitive tasks to be completed on time. Simulation results demonstrate that the cloud-edge collaborative scheduling method outperforms other benchmark methods in improving the deadline satisfaction ratio and effectively utilizes idle edge server devices.

Abstract:To perform fine-grained vulnerability detection, an ideal model must determine whether software contains vulnerabilities and identify the type of vulnerability (i.e., perform vulnerability classification). A series of deep learning models have demonstrated strong overall performance in vulnerability classification tasks. However, a severe data imbalance exists across different vulnerability types. Many vulnerability types are represented by only a small number of samples (referred to as few-shot types in this study), resulting in poor classification performance and generalization for these few-shot types. To enhance classification performance for these types, VulFewShot is proposed. This contrastive learning-based vulnerability classification framework assigns more weight to few-shot types by bringing samples of the same type closer together while keeping samples from different types further apart. Experimental results show that VulFewShot improves classification performance across all vulnerability types. The smaller the number of samples for a given type, the more significant the improvement. Therefore, VulFewShot improves classification performance for vulnerabilities with limited samples and mitigates the impact of sample size on the learning process.

Abstract:The tuning of database system parameters directly impacts its performance and the utilization of system resources. Relational database management systems typically offer hundreds of parameters that can be adjusted to achieve optimal performance and service capabilities. Database system performance optimization is traditionally carried out manually by experienced database administrators (DBAs). However, due to the characteristics of parameter tuning, such as the large number of parameters, their heterogeneity, and the complex correlations among them, traditional manual methods are inefficient, costly, and lack reusability. To enhance the efficiency of database system performance optimization, automated parameter tuning techniques have become a key focus in the database field. Reinforcement learning, with its ability to interact with the system environment and gradually improve through feedback, has been widely applied in the optimization of complex systems. Some related studies have applied reinforcement learning or its variants to database parameter tuning, but they have relied on single-objective optimization methods. Database system parameter tuning is a multi-objective optimization task, usually performed under resource constraints. Therefore, existing methods have several limitations: (1) transforming the multi-objective optimization problem into a single-objective optimization problem through simple linear transformations requires iterative attempts, making optimizations costly; (2) existing methods cannot adapt to the dynamic changes in database system requirements, limiting their adaptability; (3) reinforcement learning methods used in existing studies are designed for single-objective optimization, and their applications to multi-objective tasks make it difficult to effectively align preferences (the weight coefficients of current objectives) with corresponding optimal strategies, potentially leading to suboptimal solutions; (4) existing research primarily focuses on optimizing throughput and latency, while ignoring resource utilization such as memory. To address these issues, this study proposes a multi-objective deep deterministic policy gradient-based reinforcement learning algorithm (MODDPG). This method is a native multi-objective reinforcement learning approach that does not require transforming the multi-objective task of database system parameters tuning into a single-objective task, enabling it to efficiently adapt to dynamic changes in database system requirements. By improving the reward mechanism of the reinforcement learning algorithm, the alignment between preferences and optimal strategies can be quickly achieved, effectively avoiding suboptimal solutions. Consequently, the training process of the reinforcement learning model can be accelerated, and the efficiency of database system parameter tuning can be improved. To further validate the generality of the proposed method, the multi-objective optimization approach is extended to achieve a collaborative optimization goal of improving both database performance and resource utilization. Experiments using TPC-C and SYSBench benchmarks demonstrate the effectiveness and practicality of the proposed parameter tuning method. The results show significant advantages in terms of model training efficiency and the effectiveness of database parameter tuning.

Abstract:In the field of model-based diagnosis, the system description is first encoded, and all minimal conflict sets are obtained using a mature SAT solver. Finally, the minimal hitting set of the minimal conflict sets is computed as the candidate diagnosis for the equipment to be diagnosed. However, this strategy consumes a significant amount of time, as it is equivalent to solving two NP-hard problems: computing the minimal conflict set and the minimal hitting set. This study re-encodes the description of the circuit system and proposes a novel variant hitting set algorithm, HSDiag, which can directly compute the diagnosis from the encoding. Compared to state-of-the-art diagnosis algorithms that first solve conflict sets and then hitting sets, the efficiency improves by a factor of 5 to 100. As the number of circuit components increases, the encoding clauses increase linearly, while the number of diagnoses increases exponentially. Since solving all conflict sets of large-scale circuits (ISCAS-85) is impractical, the proposed HSDiag algorithm, within the same cutoff time, yields more than twice the number of solutions compared to conflict-set-based diagnosis algorithms. In addition, this study proposes an equivalence class optimization strategy, which further decomposes the conflict set by using the newly proposed set splitting rule, even if the initial conflict set is inseparable. The efficiency of the HSDiag algorithm optimized by equivalence class is improved by more than 2 times in standard Polybox and Fulladder circuits.

Abstract:Since the currently popular deep learning models are often influenced by the notorious phenomenon known as distribution shift, domain adaptation has been proposed to enhance the generalization of these models, transferring knowledge from labeled source data to unlabeled target data. Existing methods for domain adaptation primarily focus on computer vision tasks, leading to the application of models devised for image data to time series data to address the domain adaptation problem for time series data. Although these methods mitigate distribution shift to some extent, they struggle to effectively extract disentangled domain-invariant representations for time series data, resulting in suboptimal performance. To address this issue, a disentangled invariant and variant latent variable model for time series domain adaptation (DIVV) is proposed. Specifically, a causal generation process for time series data is introduced, where the latent variables are partitioned into domain-specific and domain-invariant latent variables. Based on this data generation process, the identifiability of domain-specific latent variables is established. The DIVV model, built on this identification theory, disentangles domain-specific and domain-invariant latent variables using variational influence and an orthogonal basis alignment module. Finally, the DIVV model leverages domain-invariant representations for time series classification. Experimental results demonstrate that the DIVV model outperforms existing domain adaptation methods for time series data across various benchmark datasets, highlighting its effectiveness in real-world applications.

Abstract:The use of computer technology for intelligent management of genealogy data plays a significant role in inheriting and popularizing Chinese traditional culture. In recent years, with the widespread application of retrieval-augmented large language model (LLM) in the knowledge question-answering (Q&A) field, presenting diverse genealogy scenarios to users through dialogues with LLMs has become a highly anticipated research direction. However, the heterogeneity, autonomy, complexity, and evolution (HACE) characteristics of genealogy data pose challenges for existing knowledge retrieval frameworks to perform comprehensive knowledge reasoning within complex genealogy information. To address this issue, Huaputong, a genealogy Q&A system based on LLMs with knowledge graph reasoning, is proposed. A knowledge graph reasoning framework, suitable for LLM-based genealogy Q&A, is constructed from two aspects: logic reasoning completeness and information filtering accuracy. In terms of the completeness of logic reasoning, knowledge graphs are used as the medium for genealogy knowledge, and a comprehensive set of genealogy reasoning rules based on the Jena framework is proposed to improve the retrieval recall of genealogy knowledge reasoning. For information filtering, scenarios involving name ambiguity and multiple kinship relations in genealogy are considered. A multi-condition matching mechanism based on problem-condition triples and a Dijkstra path ranking algorithm using a max heap are designed to filter redundant retrieval information, thus ensuring accurate prompting for LLMs. Huaputong has been deployed on the Huapu platform, a publicly available intelligent genealogical website, where its effectiveness has been validated using real-world genealogical data.

Abstract:Knowledge graph (KG), with their unique approach to knowledge management and representation capabilities, have been widely applied in various knowledge computing fields, including question answering. However, incomplete information is often present in KG, which undermines their quality and limits the performance of downstream tasks. As a result, knowledge graph completion (KGC) has emerged, aiming to enhance the quality of KG by predicting the missing information in triples using different methods. In recent years, extensive research has been conducted in the field of KGC. This study classifies KGC techniques into three categories based on the number of samples used: zero-shot KGC, few-shot KGC, and multi-shot KGC. To investigate and provide a first-hand reference for the core concepts and current status of KGC research, this study offers a comprehensive review of the latest research advancements in KGC from theoretical research, experimental analysis, and practical applications, such as the Huapu system. The problems and challenges faced by the current KGC technologies are summarized, and potential research directions for the future are discussed.

Abstract:Knowledge graph completion (KGC) models require inductive ability to generalize to new entities as the knowledge graph expands. However, current approaches understand entities only from a local perspective by aggregating neighboring information, failing to capture valuable interconnections between entities across different views. This study argues that global and sequential perspectives are essential for understanding entities beyond the local view by enabling interaction between disconnected and distant entity pairs. More importantly, it emphasizes that the aggregated information must be complementary across different views to avoid redundancy. Therefore, a multi-view framework with the differentiation mechanism is proposed for inductive KGC, aimed at learning complementary entity representations from various perspectives. Specifically, in addition to aggregating neighboring information to obtain the entity’s local representation through R-GCN, an attention-based differentiation mechanism is employed to aggregate complementary information from semantically related entities and entity-related paths, thus obtaining global and sequential representations of the entities. Finally, these representations are fused and used to score the triples. Experimental results demonstrate that the proposed framework consistently outperforms state-of-the-art approaches in the inductive setting. Moreover, it retains competitive performance in the transductive setting.

Abstract:To address the issues in current OWL representation learning methods, which lack the ability to jointly represent complex semantic information across both the concept layer and the instance layer, an OWL representation learning approach using multi-semantic views of concepts, properties, and instances is proposed. The proposed method adopts a three-stage architecture including multi-semantic views partitioning, semantic-aware self-supervised post-training, and joint multi-task representation learning. First, MSV-KRL optimizes the mapping strategy from OWL to RDF graphs based on OWL2Vec*, and five fine-grained semantic view partitioning strategies are proposed. Subsequently, serialized post-training data is generated through the random walk and annotated attribute replacement strategy. The self-supervised post-training of the pre-trained model is then carried out to enhance adaptability to multi-semantic views. Finally, by employing a multi-task learning strategy, the complex semantic representation learning of concepts, properties, and instances in OWL graphs is achieved through joint optimization loss of multi-semantic view prediction tasks. Experimental results demonstrate that MSV-KRL outperforms baseline representation learning methods on multiple benchmarks. MSV-KRL can be adapted to multiple language models, significantly improving the knowledge representation capability of OWL’s complex semantics.

Abstract:Quality issues, such as errors or deficiencies in triplets, become increasingly prominent in knowledge graphs, severely affecting the credibility of downstream applications. Accuracy evaluation is crucial for building confidence in the use and optimization of knowledge graphs. An embedding-model-based method is proposed to reduce reliance on manually labeled data and to achieve scalable automatic evaluation. Triplet verification is formulated as an automated threshold selection problem, with three threshold selection strategies proposed to enhance the robustness of the evaluation. In addition, triplet importance indicators are incorporated to place greater emphasis on critical triplets, with importance scores defined based on network structure and relationship semantics. Experiments are conducted to analyze and compare the impact on performance from various perspectives, such as embedding model capacity, knowledge graph sparsity, and triplet importance definition. The results demonstrate that, compared to existing automated evaluation methods, the proposed method can significantly reduce evaluation errors by nearly 30% in zero-shot conditions, particularly on datasets of dense graphs with high error rates.

Abstract:Session-based recommendation aims to predict the next item a user will interact with based on a series of items. Most existing session-based recommender systems do not fully utilize the temporal interval information between items within a session, affecting the accuracy of recommendations. In recent years, graph neural networks have gained significant attention in session-based recommendation due to their strong ability to model complex relationships. However, session-based recommendations that rely solely on graph neural networks overlook the hidden high-order relationships between sessions, resulting in less rich information. In addition, data sparsity has always been a phenomenon in recommender systems, and contrastive learning is often employed to address this issue. However, most contrastive learning frameworks lack strong generalization capabilities due to their singular form. Based on this, a session-based recommendation model combined with self-supervised learning is proposed. First, the model utilizes the temporal interval information between items within user sessions to perform data augmentation, enriching the information within the sessions to improve recommendation accuracy. Second, a dual-view encoder is constructed, combining a hypergraph convolutional network encoder and a Transformer encoder to capture the hidden high-order relationships between sessions from multiple perspectives, thus enhancing the diversity of recommendations. Finally, the model integrates the augmented intra-session information, the multi-viewed inter-session information, and the original session information for contrastive learning to strengthen the model’s generalization ability. Comparisons with 11 existing classic models on 4 datasets show that the proposed model is feasible and efficient, with average improvements of 5.96% and 5.89% on HR and NDCG metrics, respectively.

Abstract:This study investigates the task of clickbait detection in social media posts. These posts often employ deceptive headlines or thumbnails to mislead readers into clicking on irrelevant or undesirable content, thus enabling widespread dissemination and generating commercial benefits such as increased clicks. To evade detection, malicious creators frequently disguise clickbait posts as legitimate ones, using techniques such as adding irrelevant or misleading content to deceive the detector. Detecting such posts requires a detailed analysis and complex multi-step reasoning using commonsense knowledge to identify inconsistencies. However, existing methods typically treat a post as a simple text span and feed it into a neural network for classification, neglecting the analysis of inherent false details, which leads to misjudgments. Moreover, these black-box models lack explainability. To address this issue, a new question-guided detector is proposed, which systematically analyzes the details through a doubt-then-verify approach to uncover potential inconsistencies and falsehoods. Specifically, a multi-modal retrieval-augmented technique is used to extract detailed clues from the content of the post, followed by questioning each clue. To ensure thorough verification of facts and their complex relationships, both simple matching questions and deep commonsense reasoning questions with varying levels of complexity are employed. Each question yields a plausible answer from the post, but the answer may be fabricated or inaccurate. Therefore, an open-domain QA model is utilized for cross-verification, leveraging external knowledge to derive a more reliable answer. When discrepancies are found between answers, the post is likely to contain false content. This inconsistency serves as a valuable feature and can be combined with other multi-modal features indicative of clickbait, improving the discriminative power of the detection model. By breaking down the complex clickbait detection task into a series of question-guided verification steps, inconspicuous inconsistencies can be identified to explain the underlying reasons for clickbait. Extensive experiments on three popular datasets demonstrate the effectiveness of the proposed approach.

Abstract:Existing deep learning-based point cloud registration methods primarily focus on feature extraction and feature matching. However, the exploration of local and global graph structures during the feature extraction stage remains insufficient, and the investigation of difference information during the matching process is also limited. To address these issues, this study proposes a point cloud registration method based on local-global dynamic graph learning and complementary fusion. Specifically, the dynamic offset-based local graph learning module characterizes the underlying graph structure in the feature space by constructing proxy points that contain both geometric and semantic information, leading to more discriminative local features. In addition, a dynamic attention-based global graph learning module is designed, which adaptively adjusts attention weights based on the relationships between points, effectively capturing long-range dependencies in the point cloud. To further enhance the correspondence between the two point clouds, the attention-driven complementary fusion module utilizes the cross-attention mechanism to extract similar and distinctive information, while applying the self-attention mechanism to refine the relationships between features. Experimental results demonstrate that the proposed method achieves optimal registration performance on public datasets while maintaining acceptable computational efficiency.

Abstract:Smart contracts are computer programs running on blockchain platforms, which extend the functionality of the blockchain and enable complex applications. However, the potential security vulnerabilities of smart contracts can lead to significant financial losses. Symbolic execution-based security vulnerability detection methods offer advantages such as high accuracy and the ability to generate test cases that can reproduce vulnerabilities. Nevertheless, as the code size increases, symbolic execution faces challenges such as path explosion and excessive constraint-solving overhead. To address those issues, a novel approach for detecting smart contract security vulnerabilities through target-guided symbolic execution is proposed. First, vulnerable statements identified by static analysis tools or manually are treated as targets. The statements that depend on these target statements are analyzed, and the transaction sequence is augmented with symbolic constraints for the relevant variables. Second, the control flow graph (CFG) is constructed based on the bytecode of smart contracts, with the basic blocks containing the target statements and the dependentstatements located. The CFG is then pruned to generate guidance information. Third, path exploration in symbolic execution is optimized by reducing the number of basic blocks to be analyzed and reducing the time required for solving path constraints. With the guidance information, vulnerabilities are efficiently detected, and test cases capable of reproducing the vulnerabilities are generated. Based on this approach, a prototype tool named Smart-Target is developed. Experiments conducted on the SB Curated dataset in comparison with the symbolic execution tool, Mythril, demonstrate that Smart-Target reduces time overheads by 60.76% and 92.16% in vulnerability detection and replication scenarios, respectively. In addition, the analysis of target statementdependencies enhances vulnerability detection capability by identifying 22.02% more security vulnerabilities.

Abstract:With the widespread promotion of smart mobility, there has been increasing attention on the application of vehicular ad hoc network (VANET) in data collection. However, due to the high-speed movement of vehicles and the unpredictability of their trajectories, traditional position-based greedy forwarding strategies struggle to meet the data transmission demands of highly dynamic VANET. To address this issue, an intelligent routing algorithm driven by historical traffic data for VANET (HTD-IR) is proposed. First, an optimal forwarding table for path selection is obtained through an offline learning method based on historical traffic flow information. Then, using an online V2V transmission mechanism based on Markov prediction, the next reliable vehicle is selected according to the vehicle’s motion state. Finally, this study compares HTD-IR with other routing protocols in simulations. The results demonstrate that HTD-IR outperforms in terms of packet delivery ratio, average end-to-end delay, network yield, average successful packet transmission cost, and online computation time complexity.

Abstract:The reproducibility of scientific research results is a fundamental guarantee for the reliability of scientific research and the cornerstone of scientific and technological advancement. However, the research community is currently facing a serious reproducibility crisis, with many research results published in top journals and conferences being irreproducible. In the field of data science, the reproducibility of research results faces challenges such as heterogeneous research data from multiple sources, complex computational processes, and intricate computational environments. To address these issues, this study proposes ReproLink, a reproducibility-oriented research data management system. ReproLink constructs a unified model of research data, abstracting it into research data objects that consist of three elements: identifier, attribute set, and data entity. Through fine-grained modeling of the reproduction process, ReproLink establishes a precise method for describing multi-step, complex reproduction processes. By integrating code and operating environment modeling, ReproLink eliminates the uncertainties caused by different environments affecting code execution. Performance tests and case studies show that ReproLink performs well with data scales up to one million records, demonstrating practical value in real-world scenarios such as paper reproduction and data provenance tracking. The technical architecture of ReproLink has been integrated into Conow Software, the only integrated comprehensive management and service platform in China specifically designed for scientific research institutes, supporting the reproducibility needs of hundreds of such institutes across the country.

Abstract:Non-uniform memory access (NUMA) is the mainstream memory access architecture for state-of-the-art multicore and multi-way processor platforms. Reducing the latency of cross-NUMA node accesses during queries is a key issue for modern in-memory database query optimization techniques. Due to the differences in NUMA architectures and NUMA latency across various processors, NUMA optimization techniques should be combined with hardware characteristics. This study focuses on the in-memory foreign key join algorithm, which has high cost and strong locality of data dependency in in-memory databases, and explores different NUMA optimization techniques, including NUMA-conscious and NUMA-oblivious implementations, on five platforms featuring ARM, Intel CLX/ICX, and AMD Zen2/Zen3 processors. The study also compares the performance of the algorithms across different processor platforms with strategies such as data storage, data partitioning, and join intermediate result caching. Experimental results show that the NUMA-conscious optimization strategy requires the integration of both software and hardware. Radix Join demonstrates neutral sensitivity to NUMA latency, with NUMA optimization gains constantly around 30%. The NPO algorithm shows higher sensitivity to NUMA latency, with NUMA optimization gains ranging from 38% to 57%. The Vector Join algorithm is sensitive to NUMA latency, but the impact is relatively minor, with NUMA optimization gains varying from 1% to 25%. For algorithm performance characteristics, cache efficiency influences the Vector Join performance more than NUMA latency. NUMA-conscious optimization techniques show significant differences on ARM platforms, while the differences are minimal on x86 platforms. The less complex NUMA-oblivious algorithms exhibit greater generality. Given hardware trends, reducing NUMA latency can effectively reduce performance gaps in NUMA-conscious optimization techniques, simplify join algorithm complexity, and improve join operation performance.

Abstract:Semi-supervised semantic segmentation methods typically employ various data augmentation schemes to ensure differentiation in the input of network branches, enabling mutual self-supervision. While successful, this approach faces several issues: 1) insufficient diversity in feature extraction leads to feature signal assimilation during inference; 2) inadequate diversity in supervision signals results in the assimilation of loss learning. These issues cause network branches to converge on similar solutions, degrading the functionality of multi-branch networks. To address these issues, a cross semi-supervised semantic segmentation method based on differential feature extraction is proposed. First, a differential feature extraction strategy is employed, ensuring that branches focus on distinct information, such as texture, semantics, and shapes, thus reducing reliance on data augmentation. Second, a cross-fusion pseudo-labeling method is introduced, where branches alternately generate neighboring pixel fusion pseudo-labels, enhancing the diversity of supervision signals and guiding branches toward different solutions. Experimental results demonstrate this method achieves excellent performance on the Pascal VOC 2012 and Cityscapes validation datasets, with scores of 80.2% and 76.8%, outperforming the latest methods by 0.3% and 1.3%, respectively.