Explicit Vs Implicit"Graph"Representation Graph Embedding ge ISOMap [Tenenbaum et al. Science 001 g° LLE RoWeis and Saul, Science 00l Laplacian Eigen Map[ Belkin et al., NIPS 01 o○● (t-SNE [Maaten and Hinton, JMLR08 Deepwalk perozzi et al KDD'14 AUSun Jane LINE [Tang et al., WWW15] (childOf Node 2vec Grover and Leskovec, KDD'16 Mary born Pati Knowledge Graph Embedding John TransE [ Bordes et al., NIPS 13 bornIn TransH [Wang et al., AAAl14 TransR [Lin et al. AAAl15 Mai PathEmbedding Guu et al. and Lin et al. EMNLP 15 Austin Miami ATran B Jane Patti
Explicit vs. Implicit “Graph” Representation • Graph Embedding – ISOMap [Tenenbaum et al., Science’00] – LLE [Roweis and Saul, Science’00] – Laplacian EigenMap [Belkin et al., NIPS’01] – (t)-SNE [Maaten and Hinton, JMLR’08] – Deepwalk [Perozzi et al., KDD’14] – LINE [Tang et al., WWW’15] – Node2vec [Grover and Leskovec, KDD’16] • Knowledge Graph Embedding – TransE [Bordes et al., NIPS’13] – TransH [Wang et al., AAAI’14] – TransR [Lin et al., AAAI’15] – PathEmbedding [Guu et al., and Lin et al., EMNLP’15] – ATranB… 11 John Jane Mary Patti bornIn childOf Joh n Patti Mary Jane Miami Austin bornIn
Explicit Vs Implicit Representation Representation Implicit Explicit Flat/Homogenous LDA, word2vec ESA Graph/Heterogeneous TransE This talk From meta-path to meta-graphs Semi-supervised learning Jiang et al., JCAI'17 Recommendation zhao et al. KDD'17 Benefits Have explicit semantics explainable Knowledge discovery Resolve different kinds of ambiguity
• From meta-path to meta-graphs – Semi-supervised learning [Jiang et al., IJCAI’17] – Recommendation [Zhao et al., KDD’17] • Benefits – Have explicit semantics • Explainable • Knowledge discovery – Resolve different kinds of ambiguity Explicit vs. Implicit Representation Representation Implicit Explicit Flat/Homogenous LDA, word2vec ESA Graph/Heterogeneous TransE 12 This talk
What semantics Can hin Provide? Fel On Feb 10, 2007, Obama announced Obama his candidacy for President of the field, United States in front of the Old State Old State llinois candidacy Capitol located in Springfield, Illinois Document Bush Feh. I ▲ Location Bush portrayed himself as a traced announeed compassionate conservative, implying ■Date he was more suitable than other Politician States compassionate Republicans to go to lead the United Republican States Bush HIN network-schema: network with multiple object types and or multiple link types Document Word A Headquarter RunBy CE Named Location Organization CEO Entity Named RumBusiness Type 1( Entity Founded Mailing Type 3 Address Windward Founder Industry Named Named Entity Type 2 Entity Type 13
What Semantics Can HIN Provide? … … … … Document Word Named Entity Type 1 Named Entity Type 2 Named Entity Type 3 Named Entity Type T HIN network-schema: network with multiple object types and/or multiple link types. 13
Meta-path, Commuting Matrix, and Path Sim Meta-path defined over network schema Sun et al., VLDB'1 Document Word g Entit Document Contain, word contain Document T Named Entity 3 Named Entity Named Type 2 Commuting matrix Type T e.g., document->word binary occurrence matrix: W Un-normalized similarity: WTW: dot product Overall normalization Path Sim [Sun et al., VLDB'111 Individual normalization Path Ranking Algorithm [Lao et al ML10 EMNLP 11
Meta-path, Commuting Matrix, and PathSim • Meta-path defined over network schema. – [Sun et al., VLDB’11] – E.g., • Commuting matrix: – e.g., document->word binary occurrence matrix: 𝑊 • Un-normalized similarity: 𝑊𝑇𝑊: dot product • Overall normalization: PathSim [Sun et al., VLDB’11] • Individual normalization: Path Ranking Algorithm [Lao et al., ML’10, EMNLP’11] 14 Document word Document Contains Contains
What distinct semantics Can hin Provide? The semantics of entities and their relations Obama On Feb 10, 2007, obama announced his candidacy for President of the United States in old State candidacy front of the Old State Capitol located in Springfield, Illinois President Feb 10 United Bush portrayed himself as a compassionate SIntes conservative, implying he was more suitable Republican than other Republicans to go to lead the United States. What can context cover New york vs. New york times What cannot? George Washington"VS. Washington Higher order relations Affiliation In Affiliation In Contains Contains Document Basketball NBA Basketball DOcument Contains Contains Document 9 Basketball Olympi Basketball - Document
What Distinct Semantics Can HIN Provide? 15 • The semantics of entities and their relations • What can context cover? • What cannot? – Higher order relations ``New York'' vs. ``New York Times'' ``George Washington'' vs. ``Washington'' Document Basketball NBA Basketball Document Contains Contains Affiliation In Affiliation In Document Basketball Olympics Basketball Document Contains Contains