Knowledge graph completion (KGC) aims to study the embedding representation to solve the incompleteness of knowledge graphs (KGs). Recently, graph convolutional networks (GCNs) and graph attention networks (GATs) have been widely used in KGC tasks by capturing neighbor information of entities. However, Both GCNs and GATs based KGC models have their limitations, and the best method is to analyze the neighbors of each entity (pre-validating), while this process is prohibitively expensive. Furthermore, the representation quality of the embeddings can affect the aggregation of neighbor information (message passing). To address the above limitations, we propose a novel knowledge graph completion model with mixed geometry message and trainable convolutional attention network named MGTCA. Concretely, the mixed geometry message function generates rich neighbor message by integrating spatially information in the hyperbolic space, hypersphere space and Euclidean space jointly. To complete the autonomous switching of graph neural networks (GNNs) and eliminate the necessity of pre-validating the local structure of KGs, a trainable convolutional attention network is proposed by comprising three types of GNNs in one trainable formulation. Furthermore, a mixed geometry scoring function is proposed, which calculates scores of triples by novel prediction function and similarity function based on different geometric spaces. Extensive experiments on three standard datasets confirm the effectiveness of our innovations, and the performance of MGTCA is significantly improved compared to the state-of-the-art approaches.

Knowledge graph representation learning (KGRL) aims to study the feature representations to solve the incompleteness of knowledge graphs (KGs). Recently, graph neural network (GNN) has achieved satisfactory performance in KGRL tasks. However, many GNN based KGRL approaches fail to capture the various semantic of entities, which may reduce the representation quality of embeddings. Furthermore, most of them treat relations equally when learning embeddings, which will lead to weakening of the specific semantic features. To alleviate the above problems, we propose a novel KGRL method (RGAI) with relation-guided aggregation (RGA) and relation-guided interaction (RGI). To capture rich attributes of entities, the RGA module generates multiple sub-structures according to relation types, and aggregates the semantic information among them. To mine specific semantic features and dig out the different contribution of relations, the RGI module calculates the importance of each relation to the central entity, and generates final entity features through aggregating representations of neighbors. Extensive experiments on standard KG datasets validate the effectiveness of the proposed innovations, and RGAI achieves state-of-the-art performance compared to existing approaches (e.g., compared to state-of-the-art approaches, RGAI improves Hits@10 from 0.548 to 0.562 on FB15K-237, Hits@3 from 0.561 to 0.582 on YAGO3-10, and Hits@1 from 0.449 to 0.455 on WN18RR).

Knowledge Graphs (KGs) typically suffer from incompleteness. A popular approach to solve this problem is multi-hop reasoning through Reinforcement Learning (RL) framework, which is an explainable and effective model to predict missing links in KGs. However, many previous RL-based models use the scoring function of the pre-trained Knowledge Graph Embedding (KGE) methods as the reward function, which will lead to the performance of the model be limited to the KGE methods. Moreover, the agent may reason a meaningless path if it cannot distinguish the different aspect of each entity in different triples. To solve both problems, we propose a multi-hop reasoning model named Ae2KGR, by applying two novel strategies: attention-based exploitation and attention-based exploration. The attention-based exploitation strategy incorporates historical and query information with the neighborhood of the current entity, and dynamically updates the current state during reasoning process to assign different semantic information to the entity to distinguish different triples. The attention-based exploration strategy designs a novel policy network and reward function to dynamically make decisions based on the constantly changing state. Extensive experiments on three standard datasets confirm the effectiveness of our innovations, and the performance of our proposed Ae2KGR is significantly improved compared to the state-of-the-art methods.

Recently, an enormous amount of research has emerged on multimodal knowledge graph completion (MKGC), which seeks to extract knowledge from multimodal data and predict the most plausible missing facts to complete a given multimodal knowledge graph (MKG). However, existing MKGC approaches largely ignore that visual information may introduce noise and lead to uncertainty when adding them to the traditional KG embeddings due to the contribution of each associated image to entity is different in diverse link scenarios. Moreover, treating each triple independently when learning entity embeddings leads to local structural and the whole graph information missing. To address these challenges, we propose a novel link aware fusion and aggregation based multimodal knowledge graph completion model named LAFA, which is composed of link aware fusion module and link aware aggregation module. The link aware fusion module alleviates noise of irrelevant visual information by calculating the importance between an entity and its associated images in different link scenarios, and fuses the visual and structural embeddings according to the importance through our proposed modality embedding fusion mechanism. The link aware aggregation module assigns neighbor structural information to a given central entity by calculating the importance between the entity and its neighbors, and aggregating the fused embeddings through linear combination according to the importance. Extensive experiments on standard datasets validate that LAFA can obtain state-of-the-art performance.

Recently, graph convolutional networks (GCNs) and graph attention networks (GATs) have been used extensively in knowledge graph completion (KGC), which aims to solve the incompleteness of knowledge graphs (KGs). However, both GCNs and GATs have limitations in the KGC task, and the best method is analyzing the neighbors of each entity (pre-validating), while this process is prohibitively expensive. Furthermore, relations in KGs have specific semantics and should be considered when aggregating neighbor information (message passing). To address the above limitations, we propose a learnable convolutional attention network for knowledge graph completion named LCA-KGC. LCA-KGC introduces a knowledge graph convolutional attention network using a convolution operation before the attention mechanism to ensure structural information acquisition and avoid redundant information stacking. Furthermore, to complete the autonomous switching of GNNs types and eliminate the necessity of pre-validating the local structure of KGs, LCA-KGC designs a learnable knowledge graph convolutional attention network by comprising three types of GNNs in one learnable formulation. Moreover, a learnable message function is proposed to emphasize relational semantics when aggregating neighbor information. Extensive experiments on standard KG datasets validate the effectiveness of the proposed innovations, and LCA-KGC achieves state-of-the-art (SOTA) performance compared to existing approaches (e.g., compared to SOTA approaches, LCA-KGC improves MRR from 0.360 to 0.372 on FB15k-237 dataset, and Hits@3 from 0.561 to 0.581 on YAGO3-10 dataset).

At present, knowledge graph completion (KGC) is mainly divided into structure-based methods and language-based methods, which characterize the structural information and semantic information of knowledge graphs, respectively. Though existing works have developed methods to integrate both information, we argue their end-to-end training manner suffers discrepancy, compatibility, resources redundancy issues. Therefore, we propose a novel two-stage training paradigm for tackling KGC task, i.e. information adaptation and refinement (KGCIAR). Specifically, KGCIAR has two stages, 1) adaptation and 2) refinement. In the adaptation stage, we fine-tune the PLM with the input of descriptive information and supervised by the KG structural information. In the second refinement stage, we freeze the adapted PLM model and infer the description embeddings of entities and relations. Then, those embeddings are leveraged as the entity/relation initial embeddings. Finally, we train a lightweight KGC model. Moreover, we devise two novel objectives for knowledge adaptation, which are self-supervised adaptation and structure-aware contrastive adaptation. Furthermore, we systematically compare the performance of different lightweight KGC models for information refinement. The experiments on KGC task and various variants analyses demonstrate that KGCIAR is effective in harnessing both structure and language information in KG.

鲲鹏920 CPU是2021年面世、全球第一款基于7纳米制造工艺的ARM 64位CPU,该CPU内置一个名为KAEzip的硬件加速引擎,其核心是一个硬件压缩算法,能通过硬件提升压缩与解压缩性能.相关研究表明,压缩算法的硬化与传统软件压缩算法相比具备明显性能优势.但大数据领域中的基础性系统软件都无法识别和使用这类算法.因此研究评估硬件压缩算法在大数据环境下的性能,发现揭示制约这类算法性能的关键因素以及可能存在的缺陷具有重要意义.为此,本文首先提出一种基于"生产-消费"模型的Spark任务性能模型,形式化地表示多维资源、压缩算法和Spark任务性能之间的内在关系,从理论上分析揭示出Spark下影响压缩算法性能的关键因素.然后提出一种三层架构支持Spark识别使用硬件压缩算法.这种分层架构为进一步调优硬件压缩算法在Spark中的性能提供了灵活性,也能复用到其他大数据系统软件.在此基础上本文以KAEzip 为实验对象,使用经典Spark 基准测试程序全面评估它在Spark中的性能,结合性能模型分析挖掘制约KAEzip性能的关键因素与根源.对KAEzip的测试表明:(1)硬件压缩算法可有效提升Spark性能.比如,KAEzip比snappy有最多13.8%的压缩性能优势、最多7%的解压优势和最多5.7%的实际应用场景下的性能优势;(2)磁盘的数据传输率与硬件压缩算法性能之间的不匹配是制约硬件压缩算法性能的重要因素;(3)压缩算法在Spark中的运行机制更易导致CPU的数据处理能力与硬件压缩算法性能不匹配,也制约着硬件压缩算法的性能.测试结果也表明KAEzip 在压缩小数据时会导致数据膨胀问题.为此,本文扩展三层架构分析揭示出导致该问题的根源,并结合压缩算法在Spark中的运行机制提出一种优化方法.硬件压缩算法作为压缩算法领域的新研究方向,本文的研究工作不仅可广泛用于优化内置于CPU中的硬件压缩算法在Spark下的性能,也有助于持续演化完善KAEzip和鲲鹏920 CPU.

Recently, a large amount of work has emerged for knowledge graph completion (KGC), which aims to reason over known facts and to infer the missing links. Meanwhile, contrastive learning has been applied to the KGC tasks, which can improve the representation quality of entities and relations. However, existing KGC approaches tend to improve their performance with high-dimensional embeddings and complex models, which make them suffer from large storage space and high training costs. Furthermore, contrastive loss with single positive sample learns little structural and semantic information in knowledge graphs due to the complex relation types. To address these challenges, we propose a novel knowledge graph completion model named ConKGC with the embedding dimension scaling and a relation-aware multi-positive contrastive loss. In order to achieve both space consumption reduction and model performance improvement, a new scoring function is proposed to map the raw low-dimensional embeddings of entities and relations to high-dimensional embedding space, and predict low-dimensional tail entities with latent semantic information of high-dimensional embeddings. In addition, ConKGC designs a multiple weak positive samples based contrastive loss under different relation types to maintain two important training targets, Alignment and Uniformity. This loss function and few parameters of the model ensure that ConKGC performs best and has fast convergence speed. Extensive experiments on three standard datasets confirm the effectiveness of our innovations, and the performance of ConKGC is significantly improved compared to the state-of-the-art methods.

Recently, multi-head Graph Attention Networks (GATs) have achieved satisfactory performance in Knowledge Graph Embedding (KGE) tasks by imposing attention mechanism in local information. However, existing GATs based KGE approaches update entities with few neighbors is difficult to obtain structured semantic information, and these methods only use relations to model the local pairwise importance of entities, which result in missing semantic information of entity embedding. Meanwhile, different entities may have the same position in vector space, which result in poor performance of the model. To this end, we propose a contrastive knowledge graph embedding model named HADC with hierarchical attention network and dynamic completion. HADC dynamically adds the neighbors of entities to complement its local structural information, incorporates both entities' and relations' importance in any given entity's neighborhood, and proposes a contrastive learning-based loss function to distinguish the position of positive and negative samples in vector space. Different experiments on three standard datasets confirm the effectiveness of our innovations, and the performance of our proposed HADC is significantly improved compared to the state-of-the-art methods.

Multi-hop reasoning is an effective and explainable approach to predicting missing facts in Knowledge Graphs (KGs). It usually adopts the Reinforcement Learning (RL) framework and searches over the KG to find an evidential path. However, there are few RL based reasoning methods in Knowledge Forests (KFs), the existing approach is usually to transfer multi-hop reasoning methods in the KG to the KF. However, this way often leads to a long reasoning path with a lot of redundant information. And the existing reasoning methods tend to capture vector representations from local structures, which will lead to get meaningless paths. In order to solve these two problems, we present, in this paper, a deep reinforcement learning based model named by MultiKR for KFs and KGs, which reduces redundant information in the path by learning global vector representations. In addition, we construct a novel reward function that guides the agent's choice of actions by considering accuracy, diversity and rationality to get valid and short path. We test our model on two KF datasets, that is, COMPUTER SCIENCE (CS) and MEDICINE (MED), and two KG datasets FB15K-237 and NELL-995. Extensive experiments show that our model is effective and competitive with many current state-of-the-art methods, and also performs well in practice.

Kubernetes makes it easier to automate deployment and scale containerized applications to achieve a near-native performance. However, there is still a lack of systematic performance studies on how Spark applications perform on Kubernetes. In this paper, we first propose a model to capture the execution behavior of tasks, stages, and jobs, and present an implementation of a prototype system based on the model. The system is then used to collect and analyze various types of performance and system metrics, such as execution time and CPU utilization. Second, with the use of various Spark applications, we evaluate the performance of Spark on Kubernetes by comparing it with its baseline, i.e., Spark on bare metal. Based on the comparison and leveraging the system, we locate what stages suffer from the performance loss of these applications on Kubernetes, and then reveal the root causes of the loss by analyzing their work-flows, execution time and costs of system resources. Through extensive measurements, we find that Spark on Kubernetes falls behind its baseline in the range of - 2.9 % to 83.9%. There are several root causes of the performance loss and benefits of Spark on Kubernetes. First, data locality deterioration by pods is a crucial root cause of the loss. To address the problem, we propose an approach to schedule tasks by taking both data locality and the utilization of executors into account. Experiments show that this approach increases the performance of Spark on Kubernetes by up to 32.2%. Second, the lower CPU usages of executors are another root cause of the performance loss, even if they have an equivalent CPU configuration on both Kubernetes and bare metal. In contrast, with the same memory configuration, executors use more memory on Kubernetes than on bare metal, contributing to the performance benefit of Spark on Kubernetes in some stages. Our research efforts in this paper benefit developers and researchers when they make valuable decisions on deploying Spark applications on Kubernetes for a better performance.

开源指令集架构RISC-V具有高性能、模块化、简易性和易拓展等优势,在物联网、云计算等领域的应用日渐广泛,其向量拓展部分V模块更是很好地支持了矩阵数值计算.稀疏矩阵向量乘法SpM V作为矩阵数值计算的一个重要组成部分,具有深刻的研究意义与价值.利用RISC-V指令集的向量可配置性和寻址特性,分别对基于CSR、ELLPACK和HYB压缩格式存储的稀疏矩阵向量乘法进行向量化.同时,考虑稀疏矩阵极度稀疏和每行非零元素数量波动较大的情况,通过压缩非零元素密度低的行向量的存储、调整HYB分割阈值等手段,改进了HYB存储格式,显著改善了计算效率和存储效率.

针对阅卷时由于卷宗内容繁多导致的信息过载和知识迷航的问题,提出面向随案电子卷宗的知识森林自动构建方法,以主题分面树以及主题间的认知关系作为卷宗的知识化表示.首先,对卷宗进行碎片化的预处理,从而将不同类型的文书分类并划分为多个单一主题的碎片;然后,针对不同碎片采取不同的信息抽取方法,并利用知识融合将同义信息进行合并,之后利用本体结构和规则构建主题分面树并抽取主题关系;最后,将知识森林构建的主题分面树和主题关系在数据库中存储起来,从而实现知识森林的可视化.实验结果表明,该方法可以较为完整、准确地展示卷宗信息,组织分散的知识碎片和复杂的卷宗主题,使得选择部分卷宗主题和少量卷宗碎片实现阅卷目标成为可能,减轻了通过全面浏览卷宗内容来完成阅卷任务的负担.

Container-based virtualization techniques are becoming an alternative to traditional virtual machines, due to less overhead and better scaling. As one of the most widely used open-source container orchestration systems, Kubernetes provides a built-in mechanism, that is, horizontal pod autoscaler (HPA), for dynamic resource provisioning. By default, scaling pods only based on CPU utilization, a single performance metric, HPA may create more pods than actually needed. Through extensive measurements of a containerized n-tier application benchmark, RUBBoS, we find that excessive pods consume more CPU and memory and even deteriorate response times of applications, due to interference. Furthermore, a Kubernetes service does not balance incoming requests among old pods and new pods created by HPA, due to stateful HTTP. In this paper, we propose a bi-metric approach to scaling pods by taking into account both CPU utilization and utilization of a thread pool, which is a kind of important soft resource in Httpd and Tomcat. Our approach collects the utilization of CPU and memory of pods. Meanwhile, it makes use of ELBA, a milli-bottleneck detector, to calculate queue lengths of Httpd and Tomcat pods and then evaluate the utilization of their thread pools. Based on the utilization of both CPU and thread pools, our approach could scale up less replicas of Httpd and Tomcat pods, contributing to a reduction of hardware resource utilization. At the same time, our approach leverages preStop hook along with liveness and readiness probes to relieve load imbalance among old Tomcat pods and new ones. Based on the containerized RUBBoS, our experimental results show that the proposed approach could not only reduce the usage of CPU and memory by as much as 14% and 24% when compared with HPA, but also relieve the load imbalance to reduce average response time of requests by as much as 80%. Our approach also demonstrates that it is better to scale pods by multiple metrics rather than a single one.

Integrating lexicon knowledge into character-based methods can improve the performance of neural network models for Chinese named entity recognition (NER). For example, Lattice LSTM [1]and WC-LSTM [2] perform well on several public Chinese NER datasets. However, the directed acyclic graph (DAG) structure makes lattice LSTM challenging to train on minibatch. In addition, the Lattice LSTM and WC-LSTM only incorporate the word-level semantics into the representation of the first or last character in each word. The inside characters that the word contain are ignored. Besides, they have difficulty in dealing with the conflicts between potential words in the lexicon. This work proposes an attention- based hierarchical meta-embedding method (AHME) to incorporate lexicon knowledge into Chinese NER to alleviate the above limitations. The proposed model can incorporate the word boundary information into character representation and deal with conflicts between potential incorporated words. The experimental results on four datasets show that our method outperforms state-of-the-art baselines.

竞技类体育赛事和游戏等一般都依赖于等级分系统进行评价,现有等级分系统存在对局信息未得到充分利用问题.针对围棋对局数据的时间跨度大及包含让子棋局特点,基于经典成对数据比较模型(Bradley-Terry模型)构建神经网络等级分模型(NN-Rating).通过历史衰减方法提高模型时效性,同时借鉴主场优势特性扩展模型处理让子棋局.在真实围棋比赛数据上的实验结果及其分析表明,NN-Rating模型有良好的客观性和稳定性,相比较Elo、Trueskill和Whole-History Rating(WHR)算法具有更高的预测准确性.

Deep reinforcement learning has achieved significant success in various domains. However, it still faces a huge challenge when learning multiple tasks in sequence. This is because the interaction in a complex setting involves continual learning that results in the change in data distributions over time. A continual learning system should ensure that the agent acquires new knowledge without forgetting the previous one. However, catastrophic forgetting may occur as the new experience can overwrite previous experience due to limited memory size. The dual experience replay algorithm which retains previous experience is widely applied to reduce forgetting, but it cannot be applied in scalable tasks when the memory size is constrained. To alleviate the constrained by the memory size, we propose a new continual reinforcement learning algorithm called Self-generated Long-term Experience Replay (SLER). Our method is different from the standard dual experience replay algorithm, which uses short-term experience replay to retain current task experience, and the long-term experience replay retains all past tasks' experience to achieve continual learning. In this paper, we first trained an environment sample model called Experience Replay Mode (ERM) to generate the simulated state sequence of the previous tasks for knowledge retention. Then combined the ERM with the experience of the new task to generate the simulation experience all previous tasks to alleviate forgetting. Our method can effectively decrease the requirement of memory size in multiple tasks, reinforcement learning. We show that our method in StarCraft II and the GridWorld environments performs better than the state-of-the-art deep learning method and achieve a comparable result to the dual experience replay method, which retains the experience of all the tasks.

Performance estimation for executing graph algorithms on the distributed systems, especially with sacrificing the accuracy of results to improve the runtime performance, is a pre-requisite for optimizing system parameters to achieve an applicable tradeoff between the runtime and the inaccuracy of results. This paper presents an experimental approach that predicts the runtime and the inaccuracy of conducting graph algorithms on the BSP-based distributed graph computing systems to optimize system parameters by using an artificial neural network (ANN) model. It samples different scales of subgraphs from the complete input graphs and executes the underlying algorithm on each subgraph to capture its characteristics. Then it essentially predicts the performance of executing the underlying algorithm on the complete graph by learning the scalability that how the runtime and the inaccuracy of results vary with different scales of graphs with an ANN network, which is trained off-line based on the captured characteristics of the subgraphs. We conducted all experimental studies on the Amazon EC2 Cloud. The experimental results demonstrate that the prediction approach can effectively predict the runtime with a relative error rate under 8% averagely and the inaccuracy of results with a relative error rate under 25% averagely compared to the actual performance results.

Deceptive games are games that utilize the reward structure to keep the agent away from the global optimization and have been grown up to become a huge challenge in the field of deep reinforcement learning intelligent exploration. Most of the cutting-edge exploration approaches, such as count-based and curiosity-driven, even with intrinsic motivation, which achieves better performance in the sparse reward game, still easily fall into local optimal traps in the deceptive game. To address this shortfall, we introduce a further exploration approach called Maximum Entropy Explore (MEE). Based on entropy rewards and the off-policy actor-critic reinforcement learning algorithm, we divided the agent exploration policy into two independent parts, namely, the target policy and the explorer policy. The explorer policy, taking the maximum entropy of the target policy as the optimization goal, is used to interact with the environment and generated trajectories for the target policy. The target policy regards the maximization of external reward as the optimization goal to achieve the global solution. To alleviate the catastrophic forgetting problem which leads to the training of the agent not stabilized during the off-policy exploration phrase, the optimal experience replay is applied. An on-policy mode switch trick is used to validly prevent the unstable and diverge which caused by the deadly triad. We conduct experiments comparing our approach with state-of-the-art deep reinforcement learning algorithm and exploration methods in the grid world and StarCraft II environments with deceptive reward. The experiment indicates that the MME approach sets out to be in the present paper effectively avoids the deceptive reward trap and learns the global optimal strategy.

Kubernetes makes it easier to automate deployment and scale containerized applications to achieve near-native performance in cloud environment. However, there still lacks a systematic comparison study on how Spark applications perform between on the bare metal and on Kubernetes. In this paper, we focus on the performance evaluation of these applications running on the two environments by a series of experiments. Based on these experiments, we locate what stages cause their performance gap and reveal out root causes to the gap by analysing work-flows of these Spark applications and their resource costs. Through extensive measurements, we find out that Spark on the bare metal almost always contribute to better performance when compared with Spark on Kubernetes. More CPU usage of executors and better data locality on the bare metal are the root causes to the gap. By contrast, Spark on Kubernetes also has some its advantages over Spark on the bare metal in terms of disk W-IOPs. The research work in this paper can help practitioners and researchers to make more informed decisions on tuning their cloud environment and configuring the big data applications, so as to achieve better performance and higher resources utilization.