Deep Learning on Graphs: A Survey

IEEE Transactions on Knowledge and Data Engineering, Год журнала: 2020, Номер 34(1), С. 249 - 270

Опубликована: Март 17, 2020

Deep learning has been shown to be successful in a number of domains, ranging from acoustics, images, natural language processing. However, applying deep the ubiquitous graph data is non-trivial because unique characteristics graphs. Recently, substantial research efforts have devoted methods graphs, resulting beneficial advances analysis techniques. In this survey, we comprehensively review different types on We divide existing into five categories based their model architectures and training strategies: recurrent neural networks, convolutional autoencoders, reinforcement learning, adversarial methods. then provide comprehensive overview these systematic manner mainly by following development history. also analyze differences compositions Finally, briefly outline applications which they used discuss potential future directions.

Язык: Английский

---

Deep Closest Point: Learning Representations for Point Cloud Registration

2021 IEEE/CVF International Conference on Computer Vision (ICCV), Год журнала: 2019, Номер unknown, С. 3522 - 3531

Опубликована: Окт. 1, 2019

Point cloud registration is a key problem for computer vision applied to robotics, medical imaging, and other applications. This involves finding rigid transformation from one point into another so that they align. Iterative Closest (ICP) its variants provide simple easily-implemented iterative methods this task, but these algorithms can converge spurious local optima. To address optima difficulties in the ICP pipeline, we propose learning-based method, titled Deep (DCP), inspired by recent techniques natural language processing. Our model consists of three parts: embedding network, an attention-based module combined with pointer generation layer approximate combinatorial matching, differentiable singular value decomposition (SVD) extract final transformation. We train our end-to-end on ModelNet40 dataset show several settings it performs better than ICP, (e.g., Go-ICP, FGR), recently-proposed method PointNetLK. Beyond providing state-of-the-art technique, evaluate suitability learned features transferred unseen objects. also preliminary analysis help understand whether domain-specific and/or global facilitate registration.

Язык: Английский

---

Graph neural network for traffic forecasting: A survey

Expert Systems with Applications, Год журнала: 2022, Номер 207, С. 117921 - 117921

Опубликована: Июнь 23, 2022

Язык: Английский

---

Graph Neural Networks in Recommender Systems: A Survey

ACM Computing Surveys, Год журнала: 2022, Номер 55(5), С. 1 - 37

Опубликована: Май 5, 2022

With the explosive growth of online information, recommender systems play a key role to alleviate such information overload. Due important application value systems, there have always been emerging works in this field. In main challenge is learn effective user/item representations from their interactions and side (if any). Recently, graph neural network (GNN) techniques widely utilized since most essentially has structure GNN superiority representation learning. This article aims provide comprehensive review recent research efforts on GNN-based systems. Specifically, we taxonomy recommendation models according types used tasks. Moreover, systematically analyze challenges applying different data discuss how existing field address these challenges. Furthermore, state new perspectives pertaining development We collect representative papers along with open-source implementations https://github.com/wusw14/GNN-in-RS .

Язык: Английский

---

Diffusion Models: A Comprehensive Survey of Methods and Applications

ACM Computing Surveys, Год журнала: 2023, Номер 56(4), С. 1 - 39

Опубликована: Сен. 30, 2023

Diffusion models have emerged as a powerful new family of deep generative with record-breaking performance in many applications, including image synthesis, video generation, and molecule design. In this survey, we provide an overview the rapidly expanding body work on diffusion models, categorizing research into three key areas: efficient sampling, improved likelihood estimation, handling data special structures. We also discuss potential for combining other enhanced results. further review wide-ranging applications fields spanning from computer vision, natural language processing, temporal modeling, to interdisciplinary scientific disciplines. This survey aims contextualized, in-depth look at state identifying areas focus pointing exploration. Github: https://github.com/YangLing0818/Diffusion-Models-Papers-Survey-Taxonomy

Язык: Английский

---

Deep learning in optical metrology: a review

Light Science & Applications, Год журнала: 2022, Номер 11(1)

Опубликована: Фев. 23, 2022

Abstract With the advances in scientific foundations and technological implementations, optical metrology has become versatile problem-solving backbones manufacturing, fundamental research, engineering applications, such as quality control, nondestructive testing, experimental mechanics, biomedicine. In recent years, deep learning, a subfield of machine is emerging powerful tool to address problems by learning from data, largely driven availability massive datasets, enhanced computational power, fast data storage, novel training algorithms for neural network. It currently promoting increased interests gaining extensive attention its utilization field metrology. Unlike traditional “physics-based” approach, deep-learning-enabled kind “data-driven” which already provided numerous alternative solutions many challenging this with better performances. review, we present an overview current status latest progress deep-learning technologies We first briefly introduce both image-processing basic concepts followed comprehensive review applications various tasks, fringe denoising, phase retrieval, unwrapping, subset correlation, error compensation. The open challenges faced approach are then discussed. Finally, directions future research outlined.

Язык: Английский

---

Convergence of Edge Computing and Deep Learning: A Comprehensive Survey

IEEE Communications Surveys & Tutorials, Год журнала: 2020, Номер 22(2), С. 869 - 904

Опубликована: Янв. 1, 2020

Ubiquitous sensors and smart devices from factories communities are generating massive amounts of data, ever-increasing computing power is driving the core computation services cloud to edge network. As an important enabler broadly changing people's lives, face recognition ambitious cities, developments artificial intelligence (especially deep learning, DL) based applications thriving. However, due efficiency latency issues, current service architecture hinders vision "providing for every person organization at everywhere". Thus, unleashing DL using resources network near data sources has emerged as a desirable solution. Therefore, intelligence, aiming facilitate deployment by computing, received significant attention. In addition, DL, representative technique can be integrated into frameworks build intelligent dynamic, adaptive maintenance management. With regard mutually beneficial edge, this paper introduces discusses: 1) application scenarios both; 2) practical implementation methods enabling technologies, namely training inference in customized framework; 3) challenges future trends more pervasive fine-grained intelligence. We believe that consolidating information scattered across communication, networking, areas, survey help readers understand connections between technologies while promoting further discussions on fusion i.e., Edge DL.

Язык: Английский

---

Graph neural networks for materials science and chemistry

Communications Materials, Год журнала: 2022, Номер 3(1)

Опубликована: Ноя. 26, 2022

Abstract Machine learning plays an increasingly important role in many areas of chemistry and materials science, being used to predict properties, accelerate simulations, design new structures, synthesis routes materials. Graph neural networks (GNNs) are one the fastest growing classes machine models. They particular relevance for as they directly work on a graph or structural representation molecules therefore have full access all relevant information required characterize In this Review, we provide overview basic principles GNNs, widely datasets, state-of-the-art architectures, followed by discussion wide range recent applications GNNs concluding with road-map further development application GNNs.

Язык: Английский

---

A Survey of Graph Neural Networks for Recommender Systems: Challenges, Methods, and Directions

ACM Transactions on Recommender Systems, Год журнала: 2023, Номер 1(1), С. 1 - 51

Опубликована: Янв. 14, 2023

Recommender system is one of the most important information services on today’s Internet. Recently, graph neural networks have become new state-of-the-art approach to recommender systems. In this survey, we conduct a comprehensive review literature network-based We first introduce background and history development both systems networks. For systems, in general, there are four aspects for categorizing existing works: stage, scenario, objective, application. networks, methods consist two categories: spectral models spatial ones. then discuss motivation applying into mainly consisting high-order connectivity, structural property data enhanced supervision signal. systematically analyze challenges construction, embedding propagation/aggregation, model optimization, computation efficiency. Afterward primarily, provide overview multitude works following taxonomy above. Finally, raise discussions open problems promising future directions area. summarize representative papers along with their code repositories https://github.com/tsinghua-fib-lab/GNN-Recommender-Systems .

Язык: Английский

---

Learning Graph Structures With Transformer for Multivariate Time-Series Anomaly Detection in IoT

IEEE Internet of Things Journal, Год журнала: 2021, Номер 9(12), С. 9179 - 9189

Опубликована: Июль 27, 2021

Many real-world Internet of Things (IoT) systems, which include a variety Internet-connected sensory devices, produce substantial amounts multivariate time-series data. Meanwhile, vital IoT infrastructures, such as smart power grids and water distribution networks are frequently targeted by cyberattacks, making anomaly detection an important study topic. Modeling relatedness is, nevertheless, unavoidable for any efficient effective system, given the intricate topological nonlinear connections that originally unknown among sensors. Furthermore, detecting anomalies in time series is difficult due to their temporal dependency stochasticity. This article presented GTA, new framework involves automatically learning graph structure, convolution, modeling using transformer-based architecture. The connection policy, based on Gumbel-softmax sampling approach learn bidirected links sensors directly, at heart structure. To describe information flow between network nodes, we introduced convolution called influence propagation convolution. In addition, tackle quadratic complexity barrier, suggested multibranch attention mechanism replace original multihead self-attention method. Extensive experiments four publicly available benchmarks further demonstrate superiority our over alternative state arts.

Язык: Английский

---