(2020.10) 分子のグラフ表現と機械学習: Graph Neural Networks (GNNs) とは？

戣䊛♧㷕⻌嵲麣㣐㷕⻉㷕⿾䘔ⶼ䧭灇瑔䬿挿 81**$3F%% ⴓ㶨ךؚٓؿ邌植ה堣唒㷕统 ͖͕ͨΘ ͍͕ͪ͘ 䎃剢傈抟 傈劤⻉㷕⠓痥㔐$4+⻉㷕ؿؑأة  ر٦ة؟؎ؒٝأך⚅歲׾ך׊ְג׫תׇ׿ַ
https://itakigawa.github.io/

˖ 堣唒㷕统הכ ˖ ⴓ㶨ךؚٓؿ邌植 ˖ 彊⪒帾㾴㷕统ך؝؝ٗ GPSXBSEהCBDLXBSE ˖ (SBQI/FVSBM/FUXPSLTהכ
˖ 荈ⴓדװ׏ג׫׷׋׭ך䞔㜠 ˖ ֶתֽ  剑鵚ך鑧鎘皾ꣲ歲ծⴓ㶨欰䧭ծ✲⵸㷕统ծ 劤傈ךⰻ㺁 (// (SBQI/FVSBM/FUXPSLT דؚٓؿ邌植ׁ׸׋ⴓ㶨׾ Ⰵ⸂ח׃ג堣唒㷕统ׅ׷װ׶倯׾ך׊ְג׫תׇ׿ַ

堣唒㷕统ך⳿殢 ... "ׁ׿ה#ׁ׿ךⱖ溪׾ⴓֽ׷فؚٗٓي׾⡲׶׋ְ "ׁ׿PS#ׁ׿ ⱖ溪 فؚٗٓي ➂꟦ז׵知⽃ח鍑ֽ׷ֽו䩛갫כ荈ⴓד׮铡僇דֹזְ

堣唒㷕统倜׃ְفؚٗٓىؚٝךػٓت؎ي ♧菙暟⡤钠陎 ˑ֮׶ָהֲ˒ J’aime la musique I love music 갈㡮钠陎
堣唒缺鏬馄鍑⫷ ؜٦ي"* 堣唒㷕统㢳ꆀךⰅ⳿⸂ך鋅劤⢽ַ׵فؚٗٓي׾ 荈⹛涸ח⡲䧭ׅ׷ ְְַ־׿ז 䪮遭

➙傈ך⢽ⴓ㶨ך崞䚍װ暟䚍ך✮庠 24"32413 ⴓ㶨ך崞䚍٥暟䚍ך✮庠 CH 3 N N H N H
H 3 C N 0.739 ꣖㹱慬䏝 H 3 C H 3 C NH O N O N O CH3 O N NH 2 O CH3 Br CH3 N H 3 C H N S N O CH3 N OH CH3 CH3 N N N CH3 H 3 C H2 N NH2 Ⰵ⳿⸂ך 鋅劤⢽

ⴓ㶨ךؚٓؿ邌植ה(SBQI/FVSBM/FUXPSLT (//T ⴓ㶨ך崞䚍٥暟䚍ך✮庠 CH 3 N N H N H
H 3 C N 0.739 ꣖㹱慬䏝葿ղװ׶倯֮׷ֽ׸ו➙傈כ(//׾ך׊ְג׫תׇ׿ַ ⴓ㶨圓鸡 ⴓ㶨鎸鶢㶨 )BOTDI谏歊岀זו ⴓ㶨ךؚٓؿ邌植 N O C C C C H H H H H 暴䗙كؙزٕ 害欽ך  堣唒㷕统 (SBQI  /FVSBM  /FUXPSLT ⴓ㶨ؿ؍ٝؖ٦ فٔٝز  &$'1岀זו ✮庠⦼

⹛堣(//כ䎢ְٌتٔذ؍׾窟♧涸ח䪔ֲֿהָדֹ׷ Brc1cncc(Br)c1 4.*-&4 ⻉㷕圓鸡䒭甧⡤圓鸡 ꨵ㶨朐䡾俑㶵⴨ ؚٓؿ  杞纏䎗⡦涸 
挿꧊さ瑞꟦涸ⴓ䋒  نُٔ٦ي 甧⡤ꂁ䏟 1VC$IFN%  ⽃穠さך  㔐鯄ך荈歋䏝 O N N N H NH N N N CH3 CH3 䫑䝤䚍航泊⶝ $*% (//ד䪔ִ׷ٌتٔذ؍כ  杞纏ךչؚٓؿպ״׶䎢ְ / #S #S

ⴓ㶨邌植 CC1CCNO1 如⯋ 如⯋ 如⯋ 宏稆 *NQMJDJU 宏稆 &YQMJDJU $BOPOJDBM4.*-&4
ⴓ㶨ךؚٓؿ邌植 N O C C C C H H H H H 暴䗙كؙزٕ .0-䕎䒭⾱㶨呌ךYZ[ ⻉㷕穠さ

㹋ꥷךⴓ㶨ؚٓؿر٦ة׾鋅ג׫״ֲ ⴓ㶨$ / 0 'ך穈さַׇ׵䧭׷⾱㶨 )⟃㢩 תדךⴓ㶨 ך㸜㹀%圓鸡ה珏ך暟䚍⦼׾鎘皾 #-:1(
EG Q MFWFM 2.%BUBTFU 3BNBLSJTIOBO Sci Data 1, 140022 (2014). https://doi.org/10.1038/sdata.2014.22

㹋ꥷךⴓ㶨ؚٓؿر٦ة׾鋅ג׫״ֲ N O C C C C H H H
H H 殢湡ך ֿךⴓ㶨ך NPM׾錁㻊

㹋ꥷךⴓ㶨ؚٓؿر٦ة׾鋅ג׫״ֲ Dipole moment Isotropic polarizability Highest occupied molecular orbital (HOMO)
energy Lowest unoccupied molecular orbital (LUMO) energy Gap between HOMO and LUMO Electronic spatial extent Zero point vibrational energy Internal energy at 0K Internal energy at 298.15K Enthalpy at 298.15K Free energy at 298.15K Heat capavity at 298.15K Atomization energy at 0K Atomization energy at 298.15K Atomization enthalpy at 298.15K Atomization free energy at 298.15K Rotational constant A Rotational constant B Rotational constant C N O C C C C H H H H H ֿ׸׵ך%'5鎘皾⦼׾  %'5׃זְד ✮庠ׅ׷ةأؙ ةذח邌爙

㹋ꥷךⴓ㶨ؚٓؿر٦ة׾鋅ג׫״ֲ node[0] node[1] Edge 0 Edge 4 鴟ךぢֹ׮ 罋䣁דֹ׷
0 1

㹋ꥷךⴓ㶨ؚٓؿر٦ة׾鋅ג׫״ֲ 갥挿ך暴䗙كؙزٕ 如⯋ 鴟ך暴䗙كؙزٕ 如⯋ 갥挿ך⡘縧كؙزٕ 如⯋ 갥挿ך⾱㶨殢〾 (one-hot  encoding)
Single Double Triple Aromatic

㹋ꥷךⴓ㶨ؚٓؿر٦ة׾鋅ג׫״ֲ 갥挿ך暴䗙كؙزٕ 如⯋ H C N O F (one-hot  encoding)
0 1 2 3 4 5 6 7 8 9 10 https://pytorch-geometric.readthedocs.io/en/latest/_modules/torch_geometric/datasets/qm9.html ⾱㶨殢〾 BSPNBUJD TQ TQ TQ 穠さׅ׷)ך侧 ֿ׸⟃㢩ח׮ְ׹ְ׹剣׶ִ׷ֿהח岣䠐 N O C C C C H H H H H

(//ָװ׷ֿה 갥挿ך暴䗙كؙزٕ 如⯋ NPMY NPMQPT 갥挿ך⡘縧كؙزٕ 如⯋ NPMFEHF@BUUS NPMFEHF@JOEFY 鴟ך暴䗙كؙزٕ
如⯋ 鴟ה穠さ⾱㶨ل،ך㼎䘔 ⴓ㶨ךؚٓؿ邌植 N O C C C C H H H H H 暴䗙كؙزٕ NPMZ %'5鎘皾⦼ 如⯋ 㔿㹀؟؎ؤך كؙزٕ邌植 3FHSFTTPS  )FBE  .-1זו (// 〳㢌؟؎ؤךⴓ㶨ؚٓؿַ׵չ㔿㹀؟؎ؤךكؙزٕ邌植պפך㢌䳔׾㷕统

呌הז׷䪮遭♳ך銲⟝ (// 麩ֲ؟؎ؤךⴓ㶨ؚٓؿ׾Ⰵ⸂׃ג׮  ְא׮չ㔿㹀؟؎ؤךكؙزٕ邌植պָ⳿⸂ׁ׸זְהְֽזְ CH 3 N N H N
H H 3 C N H 3 C H 3 C NH O N O N O CH3 O N NH 2 O CH3 Br CH3 N H 3 C H N S N O CH3 N OH CH3 CH3 N N N CH3 H 3 C H2 N NH2 (// (// (// (// 갥挿侧װ 鴟侧ָ殯ז׷ ⴓ㶨ؚٓؿ ְא׮ ずׄ؟؎ؤך كؙزٕ׾⳿⸂

˖ 堣唒㷕统הכ ˖ ⴓ㶨ךؚٓؿ邌植 ˖ 彊⪒帾㾴㷕统ך؝؝ٗ GPSXBSEהCBDLXBSE ˖ (SBQI/FVSBM/FUXPSLTהכ
˖ 荈ⴓדװ׏ג׫׷׋׭ך䞔㜠 ˖ ֶתֽ  剑鵚ך鑧鎘皾ꣲ歲ծⴓ㶨欰䧭ծ✲⵸㷕统ծ 劤傈ךⰻ㺁 (// (SBQI/FVSBM/FUXPSLT דؚٓؿ邌植ׁ׸׋ⴓ㶨׾ Ⰵ⸂ח׃ג堣唒㷕统ׅ׷װ׶倯׾ך׊ְג׫תׇ׿ַ

堣唒㷕统ךٌرٕ 堣唒㷕统ٌرٕ 堣唒㷕统㢳ꆀךⰅ⳿⸂ך鋅劤⢽ַ׵فؚٗٓي׾ 荈⹛涸ח⡲䧭ׅ׷ ְְַ־׿ז 䪮遭㢳侧ךػًٓةך⦼׾⹛ַ׃ג ⳿⸂ָ劄׬⦼חז׷״ֲחׅ׷

ٌرٕך㷕统ػًٓةך⦼׾剑黝חׅ׷ ✮庠⦼ 侄䌌ך⦼ 鋅劤⢽׾ ✮庠 ✮庠⦼ה侄䌌ך⦼ָ鵚בֻ״ֲח ػًٓة׾锃侭

ٌرٕך㷕统ػًٓةך⦼׾剑黝חׅ׷ O⦐ך鋅劤ךⰋ⡤חאְג  ✮庠⦼ה侄䌌ך⦼ָ鵚בֻ״ֲח ػًٓة׾锃侭

ٌرٕך㷕统ػًٓةך⦼׾剑黝חׅ׷ O⦐ך鋅劤ךⰋ⡤חאְג  ✮庠⦼ה侄䌌ך⦼ָ鵚בֻ״ֲח ػًٓة׾锃侭 ꟼ侧խխխխך⦼ָ㼭ֻׁז׷ ״ֲחػًٓة׾锃侭

ٌرٕך㷕统ػًٓةך⦼׾剑黝חׅ׷ O⦐ך鋅劤ךⰋ⡤חאְג  ✮庠⦼ה侄䌌ך⦼ָ鵚בֻ״ֲח ػًٓة׾锃侭 ꟼ侧խխխխך⦼ָ㼭ֻׁז׷ ״ֲחػًٓة׾锃侭 ˖ խ׾ٓٝتي⦼דⴱ劍⻉ ˖ խխխך⦼ָ㼭ֻׁז׷״ֲחխ׾׍׳׏ה׌ֽ㢌ִ׷
ػًٓة锃侭ךװ׶ַ׋ ֻ׶ִַ׃ 㷕统桦  TUFQTJ[F

⺟ꂁ꣬♴岀 (SBEJFOU%FTDFOU ˖ խխխך⦼ָ㼭ֻׁז׷״ֲחխ׾׍׳׏ה׌ֽ㢌ִ׷ ⺟ꂁكؙزٕח嫰⢽ׅ׷ꆀ׾׍׳׏ה䒷ֻה  剑㺔׶ך噰㼭⦼חぢֲַֿהָ׻ַ׏גְ׷ ⺟ꂁكؙزٕ ⨉䗍ⴓ խװ׾ ך㢌⻉ꆀ
㼰׃⹛ַ׃׋הֹך

鎘皾ؚٓؿה荈⹛䗍ⴓ醱꧟זٌرٕך⺟ꂁך鎘皾堣唒㷕统ٌرٕ 㼭ֻׁ׃׋ְ㽯䏝⺟ꂁكؙزٕ ֿ׸׾וֲ鎘皾ׅ׷ַ הְֲ鑧

鎘皾ؚٓؿ稆怴皾ךさ䧭ꟼ侧ךؚٓؿ邌植 ⢽ ✮庠ٌرٕ 暴ח䠐㄂כזְ5PZ&YBNQMF

⢽ ✮庠ٌرٕ 暴ח䠐㄂כזְ5PZ&YBNQMF さ䧭ꟼ侧鎘皾ؚٓؿ稆怴皾ךさ䧭ꟼ侧ךؚٓؿ邌植

GPSXBSE鎘皾ؚٓؿך㢌侧ח⦼׾إحز׃ג׫׷ 5 0.2 0.3 0.4 1 -2 ֿךהֹխխװխך⦼כ

GPSXBSE鎘皾ؚٓؿך㢌侧ח⦼׾إحز׃ג׫׷ 5 0.2 0.3 0.4 1 -2 ֿךהֹխխװխך⦼כ 1.4

GPSXBSE鎘皾ؚٓؿך㢌侧ח⦼׾إحز׃ג׫׷ 5 0.2 0.3 0.4 1 -2 ֿךהֹխխװխך⦼כ 0.17 1.4

1.4

1.0 1.4

1.0 0.95 1.4

1.0 0.95 8.7 1.4

GPSXBSE 5 0.2 0.3 0.4 1 -2 ֿךהֹխխװխך⦼כ 0.17 0.05
1.0 0.95 8.7 1.4

GPSXBSEדאְדחװ׷ֿה⨉䗍ⴓ⦼׮鎘皾 5 0.2 0.3 0.4 1 -2 0.17 0.05 1.0
0.95 8.7 1.4 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

0.95 8.7 1.4 1 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

0.95 8.7 1.4 1 -0.99 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

0.95 8.7 1.4 1 -0.99 0.3 0.17 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

0.95 8.7 1.4 1 -0.99 0.3 0.17 5 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

0.95 8.7 1.4 1 -0.99 0.3 0.17 5 -1 1 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

0.95 8.7 1.4 1 -0.99 0.3 0.17 5 -1 1 5.9 ⺟ꂁ׾鎘皾ׅ׷㢌侧כאְדחぐ稆怴皾ך⨉䗍ⴓ⦼׮鎘皾׃גֶֻ

GPSXBSEדאְדחװ׷ֿה⨉䗍ⴓ⦼׮鎘皾さ䧭ח⢪ִ׷稆怴皾כ✮׭寸׭ג׉׸׊׸ך䗍ⴓⰕ䒭׮䭯׏גֶֻ ˖ ⸇幾⛦ꤐ  UPSDIBEE UPSDITVC UPSDINVM UPSDIEJW ˖
♲錬ꟼ侧  UPSDITJO UPSDIDPT UPSDIUBO UPSDIBTJO  ˖ 䭷侧٥㼎侧٥ץֹ  UPSDIQPX UPSDIMPH UPSDITRVBSF UPSDITRSU ˖ 遤⴨怴皾  UPSDINBUNVM UPSDIEPU UPSDIJOWFSTF UPSDIEFU ˖ 乼⡲  UPSDITVN UPSDINFBO UPSDINBY UPSDIBSHNJO 㢌侧ךさ䧭ח⢪ִ׷稆怴皾ך⢽ https://pytorch.org/docs/stable/torch.html#math-operations

GPSXBSE https://pytorch.org/tutorials/beginner/nlp/pytorch_tutorial.html 5 0.2 0.3 0.4 1 -2 0.17 0.05
1.0 0.95 8.7 1.4

劤䔲ח鎘皾׃׋ְךכ⺟ꂁ׾鎘皾ׅ׷㢌侧חꟼׅ׷⨉䗍ⴓ ⺟ꂁ كؙزٕ 荈⹛䗍ⴓ ٔغ٦أٌ٦س #BDL1SPQBHBUJPO 5 0.2 0.3
0.4 1 -2 0.17 0.05 1.0 0.95 8.7 1.4 ׾㼰׃⹛ַ׃׋הֹך ך㢌⻉ꆀ ⺟ꂁ岀ָ⢪ְ׋ְ

CBDLXBSE さ䧭ꟼ侧ך䗍ⴓך鸬ꓲ䖒״׶ծ鎘皾ؚٓؿ׾⢪׏ג知⽃ח鎘皾דֹ׷ 5.9 5.9 5.9 1 5 -1 0.17 -1
0.3 -0.99 1 29.5 -1.00 1.75 1 -0.99 0.3 0.17 5 -1 1 5.9 = = = × × × × × × × ×

فؚٗٓيהכ瑔噰涸חכ稆怴皾ך⡦׵ַךさ䧭ꟼ侧 GPSXBSE CBDLXBSE 荈⹛䗍ⴓ 5 0.2 0.3 0.4 1 -2
0.17 0.05 1.0 0.95 8.7 1.4 1 -0.99 0.3 0.17 5 -1 1 5.9 ✮庠⦼ MPTT⦼ ⺟ꂁكؙزٕ

⺟ꂁكؙزٕ 帾㾴㷕统ך؝؝ٗ䗍ⴓ〳腉זفؚٗٓىؚٝ ו׿זفؚٗٓي׮稆怴皾ךさ䧭ꟼ侧鎘皾ؚٓؿ חז׏גִׁ ְ׸ל荈⹛䗍ⴓ⺟ꂁ꣬♴岀דػًٓة׾剑黝חדֹ׷ 㷕统桦  TUFQTJ[F 荈⹛䗍ⴓד鎘皾 فؚٗٓيך♶然㹀ז皘䨽כػًٓةח׃גֶֹ䖓ד堣唒㷕统

劤傈ךⰻ㺁 ˖ 堣唒㷕统הכ ˖ ⴓ㶨ךؚٓؿ邌植 ˖ 彊⪒帾㾴㷕统ך؝؝ٗ GPSXBSEהCBDLXBSE ˖
(SBQI/FVSBM/FUXPSLTהכ ˖ 荈ⴓדװ׏ג׫׷׋׭ך䞔㜠 ˖ ֶתֽ  剑鵚ך鑧鎘皾ꣲ歲ծⴓ㶨欰䧭ծ✲⵸㷕统ծ (// (SBQI/FVSBM/FUXPSLT דؚٓؿ邌植ׁ׸׋ⴓ㶨׾ Ⰵ⸂ח׃ג堣唒㷕统ׅ׷װ׶倯׾ך׊ְג׫תׇ׿ַ

(//♴鎸ך㢌䳔׾⡦׵ַךさ䧭ꟼ侧דرؠ؎ٝ׃׋׮ך 갥挿ך暴䗙كؙزٕ 如⯋ NPMY NPMQPT 갥挿ך⡘縧كؙزٕ 如⯋ NPMFEHF@BUUS NPMFEHF@JOEFY 鴟ך暴䗙كؙزٕ

㛇劤麣Ⱗ.-1 㢳㾴ػ٦إفزٗٝ x =  x1 x2 x1 x2 1
1 w0 ji w00 i wkj 1 y import torch.nn as nn nn.Sequential(nn.Linear(2, 3), nn.ReLU(), nn.Linear(3, 2), nn.ReLu(), nn.Linear(2, 1)) UPSDIOO-JOFBS ˖ JO@GFBUVSFT ˖ PVU@GFBUVSFT ˖ CJBT 5SVF'BMTF Ⰵ⸂؟؎ؤ ⳿⸂؟؎ؤ CJBTך剣搀 1 Ⰵ⸂  ؟؎ؤ CJBT ⳿⸂  ؟؎ؤ

.FTTBHF1BTTJOH PS/FJHICPSIPPE"HHSFHBUJPO N O C C C C H H
H H H N O C C C C H H H H H ぐ暴䗙كؙزٕך刿倜 MLPͳͲ (sum, mean or max) 갫殢ח⣛㶷׃זְ꧊秈乼⡲ MLPͳͲ UPSDI@HFPNFUSJDOO .FTTBHF1BTTJOH Ԯ ԯ ԰ Ԯ ԯ ԰

.FTTBHF1BTTJOH PS/FJHICPSIPPE"HHSFHBUJPO 1 2 3 4 5 6 0 1
0 0 0 0 1 0 1 1 1 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 0 1 0 0 0 0 1 0 1 1 1 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 0 1 0 0 0 0 1 0 1 1 1 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 0 1 0 0 0 0 1 0 1 1 1 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 = = = 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6 1 3 4 5 2 + + + ꦄ䱸遤⴨זו׾⢪׏ג׮ ꧊秈ׅ׷乼⡲כ剅ֽ׷

.FTTBHF1BTTJOH PS/FJHICPSIPPE"HHSFHBUJPO ˖ UPSDI@HFPNFUSJDOO($/$POW ,JQGBOE8FMMJOH *$.- https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html ˖ UPSDI@HFPNFUSJDOO4"(&$POW )BNJMUPOFUBM
/*14 PS ˖ UPSDI@HFPNFUSJDOO(SBQI$POW .PSSJTFUBM """* ˖ UPSDI@HFPNFUSJDOO(*/$POW 9VFUBM *$-3

.FTTBHF1BTTJOH PS/FJHICPSIPPE"HHSFHBUJPO ˖ UPSDI@HFPNFUSJDOO4DI/FU 4DI»UUFUBM /*14 ˖ UPSDI@HFPNFUSJDOO("5$POW 7FMJčLPWJćFUBM *$-3
https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html YZ[דך騃ꨄ׾鴟ꅾ׫ח׃ג꧊秈儗ח崞欽ׅ׷ ꧊秈乼⡲כءٝفٕ׌ָ،ذٝءّٝ堣圓דꅾ׫׾黝䘔涸ח 銲稆琎 ]]DPODBU

.FTTBHF1BTTJOH PS/FJHICPSIPPE"HHSFHBUJPO ˖ UPSDI@HFPNFUSJDOO%JNF/FU ,MJDQFSBFUBM *$-3 YZ[ח㼎׃ג⻉㷕涸זسً؎ٝ濼陎׾崞欽׃ג乼⡲׾⡲׶鴥׫תֻ׏׋⢽

.FTTBHF1BTTJOH PS/FJHICPSIPPE"HHSFHBUJPO 剑ⴱח⢽ח⳿׃׋2.ر٦ةדכ䖞勻䩛岀׾⮚馉ׅ׷✮庠礵䏝

3FBEPVUؚٓؿח㼎ׅ׷㔿㹀؟؎ؤ邌植׾䖤׷ N O C C C C H H H
H H 3FBEPVU ˖ UPSDI@HFPNFUSJDOO(MPCBM"UUFOUJPO 㾴侧 ։ ˖ UPSDI@HFPNFUSJDOO4FU4FU ˖ UPSDI@HFPNFUSJDOOHMPCBM@BEE@QPPM  UPSDI@HFPNFUSJDOOHMPCBM@NBY@QPPM  UPSDI@HFPNFUSJDOOHMPCBM@NFBO@QPPM ˖ UPSDI@HFPNFUSJDOOHMPCBM@TPSU@QPPM (sum, mean or max) 갫殢ח⣛㶷׃זְ꧊秈乼⡲ 剑䖓ך銲稆דTPSU׃גزحفL׾DPODBU

(//♴鎸ך㢌䳔׾⡦׵ַךさ䧭ꟼ侧דرؠ؎ٝ׃׋׮ך 갥挿ך暴䗙كؙزٕ 如⯋ NPMY NPMQPT 갥挿ך⡘縧كؙزٕ 如⯋ NPMFEHF@BUUS NPMFEHF@JOEFY 鴟ך暴䗙كؙزٕ

1Z5PSDI(FPNFUSJDך⢽ (PPHMF$PMBC

ٌرٕהGPSXBSE㹀纏 ⳿⸂׾鎘皾ׅ׷GPSXBSE 3FBEPVU .FTTBHF 1BTTJOH Y㾴 3FHSFTTPS )FBE .-1

㷕统ٌرٕךػًٓة׾剑黝חׅ׷ ؚٓؿ׾CBUDI@TJ[F ֿ באGFFE ٌرٕחⰅ׸ג✮庠׾䖤׷ MPTT׾鎘皾׃CBDLXBSE ⺟ꂁ꣬♴ ػًٓة刿倜 ה׶ִ֮׆FQPDI 侧涰䗳銲

岣䠐 ˖ 2.ךةأؙך㜥さכקרYZ[ YQPT ך䞔㜠ָ䗳銲׌ָծ  ⽃秪ז(//כYFEHF@BUUS׮⢪׻זְךד黝㹅何㢌ָ䗳銲 ˖ ⢽ִלYFEHF@BUUSח⾱㶨꟦騃ꨄ׾⸇ִג$POWד崞欽 ˖
׉ךתת⢪ֲ㜥さכ4DI/FUװ%JNF/FUָぢְגְ׷ ˖ Ⱅ䒭؝٦س؟ٝفٕ׮ثؑحؙ  IUUQTHJUIVCDPNSVTUZTQZUPSDI@HFPNFUSJDUSFFNBTUFSFYBNQMFT

0QFO(SBQI#FODINBSL https://ogb.stanford.edu

(SBQI/FVSBM/FUXPSLT׾濼׷׋׭ךאךا٦أ https://ai.tencent.com/ailab/ml/KDD-Deep-Graph-Learning.html • KDD2020 Tutorial  Deep Graph Learning: Foundations, Advances
and Applications http://cse.msu.edu/~mayao4/dlg_book/index.html • Book "Deep Learning on Graphs" • CS224W: Machine Learning with Graphs (Stanford) http://web.stanford.edu/class/cs224w/ https://youtu.be/-UjytpbqX4A

1ZUPSDI(FPNFUSJD׾⢪׏׋عٝؤؔٝ $48 https://colab.research.google.com/drive/1DIQm9rOx2mT1bZETEeVUThxcrP1RKqAn

,BHHMF$)".14 YZ[ָ銲׷ةأؙ https://www.kaggle.com/c/champs-scalar-coupling/discussion/93972

*$-3"'BJS$PNQBSJTPOTPG(//T https://github.com/diningphil/gnn-comparison

(//כ涪㾜鷿♳ך䪮遭ד植㖈鹌遤䕎ד灇瑔ׁ׸גְ׷ ˖ ✲⵸㷕统ָדֹזְ ˖ %FFQחדֹזְ ˖ 㷕统דֹ׷邌植חꣲ歲ָ֮׷ Generalization and Representational
Limits of Graph Neural Networks Garg, Jegelka, Jaakkola (ICML2020) PairNorm: Tackling Oversmoothing in GNNs Zhao, Akoglu (ICLR2020) Towards Deeper Graph Neural Networks with Differentiable Group Normalization Zhou, Huang, Li, Zha, Chen, Hu (NeurIPS2020) Strategies for Pre-training Graph Neural Networks Hu, Liu, Gomes, Zitnik, Liang, Pande, Leskovec (ICLR2020) GROVER: Self-Supervised Message Passing Transformer on Large-scale Molecular Graphs Rong, Bian, Xu, Xie, Wei, Huang, Huang (NeurIPS2020) DeepGCNs: Can GCNs Go as Deep as CNNs? Li, Müller, Thabet, Ghanem (ICCV2019)

➙傈ך鑧ך䘔欽٥涪㾜 ˖ ⻉㷕⿾䘔ך✮庠٥鷞さ䧭 ˖ ⴓ㶨ך欰䧭 ˖ (SBQI1PPMJOH Path Integral Based
Convolution and Pooling for Graph Neural Networks Ma, Xuan, Guang Wang, Li, Liò (NeurIPS2020) Hierarchical Generation of Molecular Graphs using Structural Motifs Jin, Barzilay, Jaakkola (ICML2020) GraphAF: a Flow-based Autoregressive Model for Molecular Graph Generation Shi, Xu, Zhu, Zhang, Zhang, Tang (ICLR2020) RetroXpert: Decompose Retrosynthesis Prediction like A Chemist Yan, Ding, Zhao, Zheng, Yang, Yu, Huang (NeurIPS2020) A Graph to Graphs Framework for Retrosynthesis Prediction Shi, Xu, Guo, Zhang, Tang (ICML2020) Rethinking Pooling in Graph Neural Networks Mesquita, Souza, Kaski (NeurIPS2020)

1VSFMZ%BUB%SJWFOךꣲ歲WT$IFNBUJDB

תה׭ ˖ 堣唒㷕统הכ ˖ ⴓ㶨ךؚٓؿ邌植 ˖ 彊⪒帾㾴㷕统ך؝؝ٗ GPSXBSEהCBDLXBSE ˖

(2020.10) 分子のグラフ表現と機械学習: Graph Neural Networks (GNNs) とは？

(2020.10) 分子のグラフ表現と機械学習: Graph Neural Networks (GNNs) とは？

More Decks by itakigawa

Featured

Transcript