1
本文作者: 楊文 | 2018-02-24 16:25 |
雷鋒網(wǎng)AI科技評論按:近日,Cunchao Tu 和 Yuan Yao 兩位研究者在 GitHub 上總結(jié)發(fā)表了一份關(guān)于網(wǎng)絡(luò)表示學(xué)習(xí)(NRL: network representation learning)和網(wǎng)絡(luò)嵌入研究領(lǐng)域(NE: network embedding)必讀論文清單。這份清單共包含 5 篇綜述論文和 64 篇會議期刊論文。同時(shí)兩位研究者在 GitHub 上發(fā)布了 NE / NERL 的開源工具包 OpenNE。該庫提供了標(biāo)準(zhǔn)的 NE / NRL(網(wǎng)絡(luò)表示學(xué)習(xí))培訓(xùn)和測試框架,目前在 OpenNE 中實(shí)現(xiàn)的模型包括 DeepWalk,LINE,node2vec,GraRep,TADW 和 GCN。
五篇必讀Survey Papers
Representation Learning on Graphs: Methods and Applications.
作者:William L. Hamilton, Rex Ying
論文地址:https://arxiv.org/pdf/1709.05584.pdf
論文摘要:在不同圖(Graph)上的機(jī)器學(xué)習(xí)是一項(xiàng)重要且無處不在的任務(wù),其應(yīng)用范圍從藥物設(shè)計(jì)到社交網(wǎng)絡(luò)中的好友推薦。該領(lǐng)域的主要挑戰(zhàn)是找到一種方法來表示或編碼圖結(jié)構(gòu),以便機(jī)器學(xué)習(xí)模型可以輕松利用它。傳統(tǒng)的機(jī)器學(xué)習(xí)方法依賴于用戶定義的啟發(fā)式方法來提取編碼關(guān)于圖的結(jié)構(gòu)信息的特征(例如,度數(shù)統(tǒng)計(jì)或內(nèi)核函數(shù))。然而,近年來,使用基于深度學(xué)習(xí)和非線性降維的技術(shù),自動學(xué)習(xí)將圖結(jié)構(gòu)編碼為低維嵌入的方法出現(xiàn)了激增。在這里,我們提供了關(guān)于圖形表示學(xué)習(xí)領(lǐng)域進(jìn)展的關(guān)鍵概念回顧,包括基于矩陣分解的方法,基于隨機(jī)游走的算法和圖形卷積網(wǎng)絡(luò)。文中回顧了嵌入單個(gè)節(jié)點(diǎn)的方法以及嵌入整個(gè)(子)圖的方法。為此,制定了一個(gè)統(tǒng)一的框架來描述這些最新的方法,并且強(qiáng)調(diào)了未來工作的一些重要應(yīng)用和方向。
Graph Embedding Techniques, Applications, and Performance: A Survey
作者:Palash Goyal , Emilio Ferrara
論文地址: https://arxiv.org/pdf/1705.02801.pdf
論文摘要:社交網(wǎng)絡(luò),詞語共現(xiàn)網(wǎng)絡(luò)和通信網(wǎng)絡(luò)等圖(Graph)出現(xiàn)在現(xiàn)實(shí)世界的應(yīng)用中。在過去,研究人員已經(jīng)通過很多方法分析它們用來洞察社會結(jié)構(gòu),語言以及不同的交流模式。最近,在向量空間中使用圖節(jié)點(diǎn)表示的方法得到了研究團(tuán)隊(duì)的關(guān)注。在這次調(diào)查中,我們對文獻(xiàn)中提出的各種圖嵌入技術(shù)進(jìn)行了全面和結(jié)構(gòu)化分析。我們首先介紹嵌入任務(wù)及其可擴(kuò)展性,維度選擇,以及要保留的功能等挑戰(zhàn)及其可能的解決方案。然后,我們基于因式分解方法,隨機(jī)游走和深度學(xué)習(xí),提出了三類方法,并給出了每個(gè)類別中具有代表性算法的例子以及各種任務(wù)的表現(xiàn)分析。我們在幾個(gè)常見數(shù)據(jù)集上評估這些最先進(jìn)的方法,并比較它們的性能。我們最終呈現(xiàn)了我們開發(fā)的開源 Python 庫,名為 GEM(圖嵌入方法,可在 https://github.com/palash1992/GEM 上獲得),它提供了所有在統(tǒng)一界面中提到的算法。
A Comprehensive Survey of Graph Embedding: Problems, Techniques and Applications.
作者:Hongyun Cai, Vincent W. Zheng, Kevin Chen-Chuan Chang
論文地址:https://arxiv.org/pdf/1709.07604.pdf
論文摘要:圖是一種重要的數(shù)據(jù)表示形式,它出現(xiàn)在各種各樣的現(xiàn)實(shí)場景中。有效的圖分析為用戶提供了對數(shù)據(jù)背后知識更深入的了解,從而可以使許多有價(jià)值的應(yīng)用程序受益,如節(jié)點(diǎn)分類,節(jié)點(diǎn)推薦,鏈接預(yù)測等。然而,大多數(shù)圖分析方法仍遭受高計(jì)算和空間成本的限制。圖嵌入是解決圖分析問題的有效且高效的方法,它將圖形數(shù)據(jù)轉(zhuǎn)換為低維空間,其中圖形結(jié)構(gòu)信息和圖形屬性得到最大限度地保留。在這次調(diào)查中,我們對過去關(guān)于圖嵌入的文獻(xiàn)進(jìn)行了全面的回顧。我們首先介紹圖形嵌入的正式定義以及相關(guān)的概念。之后,我們提出了兩種圖嵌入分類法,它們對應(yīng)于不同圖嵌入問題設(shè)置中存在的挑戰(zhàn)以及現(xiàn)有工作如何解決其解決方案中的這些挑戰(zhàn)。最后,我們總結(jié)了圖嵌入的應(yīng)用,并提出了四個(gè)有前景的未來研究方向,包括計(jì)算效率,問題設(shè)置,技術(shù)和應(yīng)用場景。
Network Representation Learning: A Survey.
作者:Daokun Zhang, Jie Yin,Xingquan Zhu, Chengqi Zhang
論文地址:https://arxiv.org/pdf/1801.05852.pdf
論文摘要:隨著信息技術(shù)的廣泛使用,捕捉社交網(wǎng)絡(luò),引文網(wǎng)絡(luò),電信網(wǎng)絡(luò)和生物網(wǎng)絡(luò)等各種學(xué)科網(wǎng)絡(luò)之間的復(fù)雜關(guān)系變得越來越流行。分析這些網(wǎng)絡(luò)揭示了社會生活的不同方面,如社會結(jié)構(gòu),信息傳播和不同的交流模式。然而,大規(guī)模的信息網(wǎng)絡(luò)往往使網(wǎng)絡(luò)分析任務(wù)計(jì)算成本昂貴且棘手。最近,網(wǎng)絡(luò)表示學(xué)習(xí)被提出作為一種新的學(xué)習(xí)范式,通過保留網(wǎng)絡(luò)拓?fù)浣Y(jié)構(gòu),頂點(diǎn)內(nèi)容和其他邊信息將網(wǎng)絡(luò)頂點(diǎn)嵌入低維向量空間。這有助于在新的向量空間中輕松處理原始網(wǎng)絡(luò)以供進(jìn)一步分析。在這次調(diào)查中,我們對數(shù)據(jù)挖掘和機(jī)器學(xué)習(xí)領(lǐng)域中的網(wǎng)絡(luò)表示學(xué)習(xí)的當(dāng)前文獻(xiàn)進(jìn)行了全面的回顧。我們提出了一種新的分類方法,根據(jù)他們所使用的方法和他們保存的網(wǎng)絡(luò)信息來分析和總結(jié)最先進(jìn)的網(wǎng)絡(luò)表示學(xué)習(xí)技術(shù)。最后,為了便于研究這一主題,我們總結(jié)了基準(zhǔn)數(shù)據(jù)集和評估方法,并討論了該領(lǐng)域的未來研究方向。
Network Representation Learning: An Overview.
作者:Cunchao Tu, Cheng Yang, Zhiyuan Liu, Maosong Sun
論文地址:http://engine.scichina.com/publisher/scp/journal/SSI/47/8/10.1360/N112017-00145?slug=full%20text
論文摘要:網(wǎng)絡(luò)是表示對象及其關(guān)系的重要方式。網(wǎng)絡(luò)研究中的一個(gè)關(guān)鍵問題是如何正確表示網(wǎng)絡(luò)信息。隨著機(jī)器學(xué)習(xí)的發(fā)展,網(wǎng)絡(luò)頂點(diǎn)的特征學(xué)習(xí)已成為一個(gè)重要的研究領(lǐng)域。網(wǎng)絡(luò)表示學(xué)習(xí)算法將網(wǎng)絡(luò)信息轉(zhuǎn)換為密集的低維實(shí)值向量,可用作現(xiàn)有機(jī)器學(xué)習(xí)算法的輸入。例如,頂點(diǎn)的表示可以被饋送到分類器,例如用于頂點(diǎn)分類的支持向量機(jī)(SVM)。另外,通過將表示作為歐幾里德空間中的點(diǎn),可以將表示用于可視化。網(wǎng)絡(luò)表征學(xué)習(xí)的研究引起了許多研究者的關(guān)注。在這篇文章中,介紹和總結(jié)了近期關(guān)于網(wǎng)絡(luò)表示學(xué)習(xí)的研究工作。
六十四篇期刊會議論文
DeepWalk: Online Learning of Social Representations. KDD 2014.
論文地址:https://arxiv.org/pdf/1403.6652.pdf;
代碼地址:https://github.com/phanein/deepwalk
Learning Latent Representations of Nodes for Classifying in Heterogeneous Social Networks. WSDM 2014.
論文地址:http://webia.lip6.fr/%7Egallinar/gallinari/uploads/Teaching/WSDM2014-jacob.pdf
Non-transitive Hashing with Latent Similarity Componets. KDD 2015.
論文地址:http://media.cs.tsinghua.edu.cn/%7Emultimedia/cuipeng/papers/KDD-NonTransitiveHashing.pdf
GraRep: Learning Graph Representations with Global Structural Information. CIKM 2015.
論文地址:https://www.researchgate.net/publication/301417811_GraRep;
代碼地址:https://github.com/ShelsonCao/GraRep
LINE: Large-scale Information Network Embedding. WWW 2015.
論文地址:https://arxiv.org/pdf/1503.03578.pdf;
代碼地址:https://github.com/tangjianpku/LINE
Network Representation Learning with Rich Text Information. IJCAI 2015.
論文地址:http://thunlp.org/%7Eyangcheng/publications/ijcai15.pdf
代碼地址:https://github.com/tangjianpku/LINE
PTE: Predictive Text Embedding through Large-scale Heterogeneous Text Networks. KDD 2015.
論文地址:https://arxiv.org/pdf/1508.00200.pdf
代碼地址:https://github.com/mnqu/PTE
Heterogeneous Network Embedding via Deep Architectures. KDD 2015.
論文地址:http://www.ifp.illinois.edu/%7Echang87/papers/kdd_2015.pdf
Deep Neural Networks for Learning Graph Representations. AAAI 2016.
論文地址:https://pdfs.semanticscholar.org/1a37/f07606d60df365d74752857e8ce909f700b3.pdf
Asymmetric Transitivity Preserving Graph Embedding. KDD 2016.
論文地址:http://media.cs.tsinghua.edu.cn/%7Emultimedia/cuipeng/papers/hoppe.pdf
Revisiting Semi-supervised Learning with Graph Embeddings. ICML 2016.
論文地址:http://proceedings.mlr.press/v48/yanga16.pdf
node2vec: Scalable Feature Learning for Networks. KDD 2016.
論文地址:http://www.kdd.org/kdd2016/papers/files/rfp0218-groverA.pdf
代碼地址:https://github.com/aditya-grover/node2vec
Max-Margin DeepWalk: Discriminative Learning of Network Representation. IJCAI 2016.
論文地址:http://thunlp.org/%7Etcc/publications/ijcai2016_mmdw.pdf
代碼地址:https://github.com/thunlp/mmdw
Tri-Party Deep Network Representation. IJCAI 2016.
論文地址:https://www.ijcai.org/Proceedings/16/Papers/271.pdf
Discriminative Deep RandomWalk for Network Classification. ACL 2016.
論文地址:http://www.aclweb.org/anthology/P16-1095
Structural Deep Network Embedding. KDD 2016.
論文地址:http://media.cs.tsinghua.edu.cn/%7Emultimedia/cuipeng/papers/SDNE.pdf
Structural Neighborhood Based Classification of Nodes in a Network. KDD 2016.
論文地址:http://www.kdd.org/kdd2016/papers/files/Paper_679.pdf
Community Preserving Network Embedding. AAAI 2017.
論文地址:http://media.cs.tsinghua.edu.cn/%7Emultimedia/cuipeng/papers/NE-Community.pdf
Semi-supervised Classification with Graph Convolutional Networks. ICLR 2017.
論文地址:https://arxiv.org/pdf/1609.02907.pdf
代碼地址:https://github.com/tkipf/gcn
CANE: Context-Aware Network Embedding for Relation Modeling. ACL 2017.
論文地址:http://thunlp.org/%7Etcc/publications/acl2017_cane.pdf
代碼地址:https://github.com/thunlp/cane
Fast Network Embedding Enhancement via High Order Proximity Approximation. IJCAI 2017.
論文地址:http://thunlp.org/%7Etcc/publications/ijcai2017_neu.pdf
代碼地址:https://github.com/thunlp/neu
TransNet: Translation-Based Network Representation Learning for Social Relation Extraction.IJCAI 2017.
論文地址:http://thunlp.org/%7Etcc/publications/ijcai2017_transnet.pdf
代碼地址:https://github.com/thunlp/transnet
metapath2vec: Scalable Representation Learning for Heterogeneous Networks. KDD 2017.
論文地址:https://www3.nd.edu/%7Edial/publications/dong2017metapath2vec.pdf
代碼地址:https://ericdongyx.github.io/metapath2vec/m2v.html
Learning from Labeled and Unlabeled Vertices in Networks.KDD 2017.
論文地址:https://dl.acm.org/citation.cfm?id=3098142
Unsupervised Feature Selection in Signed Social Networks. KDD 2017.
論文地址:http://www.public.asu.edu/%7Ejundongl/paper/KDD17_SignedFS.pdf
struc2vec: Learning Node Representations from Structural Identity. KDD 2017.
論文地址:https://arxiv.org/pdf/1704.03165.pdf
代碼地址:https://github.com/leoribeiro/struc2vec
Label Informed Attributed Network Embedding. WSDM 2017.
論文地址:http://people.tamu.edu/%7Exhuang/Xiao_WSDM17.pdf
代碼地址:https://github.com/xhuang31/LANE
Accelerated Attributed Network Embedding. SDM 2017.
論文地址:http://www.public.asu.edu/%7Ejundongl/paper/SDM17_AANE.pdf
代碼地址:https://github.com/xhuang31/AANE_Python
Inductive Representation Learning on Large Graphs. NIPS 2017.
論文地址:https://arxiv.org/pdf/1706.02216.pdf
代碼地址:https://github.com/williamleif/GraphSAGE
Variation Autoencoder Based Network Representation Learning for Classification. ACL 2017.
論文地址:http://aclweb.org/anthology/P17-3010
Preserving Proximity and Global Ranking for Node Embedding. NIPS 2017.
論文地址:https://papers.nips.cc/paper/7110-prune-preserving-proximity-and-global-ranking-for-network-embedding.pdf
Learning Graph Embeddings with Embedding Propagation. NIPS 2017.
論文地址:https://arxiv.org/pdf/1710.03059.pdf
CIKM 2017
Name Disambiguation in Anonymized Graphs using Network Embedding.
論文地址:https://arxiv.org/pdf/1702.02287.pdf
Enhancing the Network Embedding Quality with Structural Similarity.
論文地址:http://www.cis.pku.edu.cn/faculty/system/zhangyan/papers/CIKM2017-lts.pdf
Attributed Signed Network Embedding.
論文地址:http://www.public.asu.edu/%7Eswang187/publications/SNEA.pdf
Attributed Network Embedding for Learning in a Dynamic Environment.
論文地址:https://arxiv.org/pdf/1706.01860.pdf
HIN2Vec: Explore Meta-paths in Heterogeneous Information Networks for Representation Learning.
論文地址:http://shichuan.org/hin/topic/Embedding/2017.%20CIKM%20HIN2Vec.pdf
From Properties to Links: Deep Network Embedding on Incomplete Graphs.
論文地址:https://dl.acm.org/citation.cfm?id=3132975&dl=ACM&coll=DL
An Attention-based Collaboration Framework for Multi-View Network Representation Learning.
論文地址:https://arxiv.org/pdf/1709.06636.pdf
On Embedding Uncertain Graphs.
論文地址:http://i.cs.hku.hk/%7Ezphuang/pub/CIKM17.pdf
Multi-view Clustering with Graph Embedding for Connectome Analysis.
論文地址:https://www.cs.uic.edu/%7Eclu/doc/cikm17_mcge.pdf
Learning Node Embeddings in Interaction Graphs.
論文地址:https://web.cs.wpi.edu/%7Exkong/publications/papers/cikm17.pdf
Learning Community Embedding with Community Detection and Node Embedding on Graphs.
論文地址:http://sentic.net/community-embedding.pdf
代碼地址:https://github.com/andompesta/ComE
WSDM 2018
Network Embedding as Matrix Factorization: Unifying DeepWalk, LINE, PTE, and node2vec.
論文地址:https://arxiv.org/pdf/1710.02971.pdf
Exploring Expert Cognition for Attributed Network Embedding.
論文地址:http://people.tamu.edu/~xhuang/Xiao_WSDM18.pdf
SHINE: Signed Heterogeneous Information Network Embedding for Sentiment Link Prediction.
論文地址:https://arxiv.org/pdf/1712.00732.pdf
Multidimensional Network Embedding with Hierarchical Structures.
論文地址:http://cse.msu.edu/~mayao4/downloads/Multidimensional_Network_Embedding_with_Hierarchical_Structure.pdf
Curriculum Learning for Heterogeneous Star Network Embedding via Deep Reinforcement Learning.
論文地址:https://dl.acm.org/citation.cfm?id=3159711&dl=ACM&coll=DL
AAAI 2018
Adversarial Network Embedding.
論文地址:https://arxiv.org/pdf/1711.07838.pdf
COSINE: Community-Preserving Social Network Embedding from Information Diffusion Cascades.
Dynamic Network Embedding by Modeling Triadic Closure Process.
論文地址:http://yangy.org/works/dynamictriad/dynamic_triad.pdf
Multi-facet Network Embedding: Beyond the General Solution of Detection and Representation.
RSDNE: Exploring Relaxed Similarity and Dissimilarity from Completely-imbalanced Labels for Network Embedding.
Link Prediction via Subgraph Embedding-Based Convex Matrix Completion.
Generative Adversarial Network based Heterogeneous Bibliographic Network Representation for Personalized Citation Recommendation.
DepthLGP: Learning Embeddings of Out-of-Sample Nodes in Dynamic Networks.
論文地址:http://media.cs.tsinghua.edu.cn/%7Emultimedia/cuipeng/papers/DepthLGP.pdf
Structural Deep Embedding for Hyper-Networks.
TIMERS: Error-Bounded SVD Restart on Dynamic Networks.
Community Detection in Attributed Graphs: An Embedding Approach.
Bernoulli Embeddings for Graphs.
論文地址:http://sumitbhatia.net/papers/aaai18.pdf
Distance-aware DAG Embedding for Proximity Search on Heterogeneous Graphs.
GraphGAN: Graph Representation Learning with Generative Adversarial Nets.
論文地址:https://arxiv.org/pdf/1711.08267.pdf
HARP: Hierarchical Representation Learning for Networks.
論文地址:https://arxiv.org/pdf/1706.07845.pdf
代碼地址:https://arxiv.org/pdf/1706.07845.pdf
Representation Learning for Scale-free Networks.
雷鋒網(wǎng)整理編譯
更多內(nèi)容歡迎關(guān)注雷鋒網(wǎng)AI科技評論
雷峰網(wǎng)原創(chuàng)文章,未經(jīng)授權(quán)禁止轉(zhuǎn)載。詳情見轉(zhuǎn)載須知。