Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
An_Attention-Based_User_Behavior_Modeling_Frame...
Search
Sponsored
·
Your Podcast. Everywhere. Effortlessly.
Share. Educate. Inspire. Entertain. You do you. We'll handle the rest.
→
kishimoto
June 21, 2019
Research
2.2k
1
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
An_Attention-Based_User_Behavior_Modeling_Framework_for_Recommendation.pdf
kishimoto
June 21, 2019
More Decks by kishimoto
See All by kishimoto
LLMと検索エンジン・レコメンドエンジンの協調動作への歩み
kishimotobanana
1
690
ChatGPTを活用して対話的な ファッション検索サービスを作った話
kishimotobanana
0
230
LLMと自社データベースの協調動作による ファッションECへの応用
kishimotobanana
0
1.6k
マイクロアドにおけるCTR予測への取り組み
kishimotobanana
4
10k
Other Decks in Research
See All in Research
[最先端NLP勉強会2026] Checklists Are Better Than Reward Models For Aligning Language Models
nzw0301
1
310
【中間報告】国会議員の立法・政策実務を支える環境を巡る現状と課題
polipoli
0
570
Spatial Active Noise Control Based onSound Field Interpolation Incorporating Physical Constraints
skoyamalab
0
190
SLAMはどこまで解決されたのか?
tomonom
0
1.3k
大規模言語モデルは誰を覚えているか / Who Do Large Language Models Memorize?
upura
0
160
Karkada さんの論文 × 2 の紹介: (1) Closed-Form Training Dynamics Reveal Learned Features and Linear Structure in Word2Vec-like Models, (2) Symmetry in language statistics shapes the geometry of model representations
eumesy
PRO
1
690
論文読み会 SNLP2026 Tau2-Bench: Evaluating Conversational Agents in a Dual-Control Environment
s_mizuki_nlp
0
230
SAKURAONE:An Open Ethernet-based AI HPC System And Its Observed Workload Dynamicsin a Single-Tenant LLM Development Environment
yuukit
1
590
最先端NLP勉強会2026 論文紹介:Reasoning with Sampling: Your Base Model is Smarter Than You Think (ICLR 2026 paper)
kogoro
4
590
[BlackHatAsia2026] Hidden Telemetry: Uncovering TraceLogging ETW Providers You're Not Using (Yet)
asuna_jp
1
700
データサイエンティストの就労意識~2015 → 2026 一般(個人)会員アンケートより
datascientistsociety
PRO
0
760
CVPR2026論文紹介_VLMにとって良いvision encoderとは何か?Rethinking Model Selection in VLM Through the Lens of Gromov-Wasserstein Distance
kobayashi31
1
210
Featured
See All Featured
Leading Effective Engineering Teams in the AI Era
addyosmani
9
2.5k
Google's AI Overviews - The New Search
badams
0
1.6k
Code Review Best Practice
trishagee
74
20k
GraphQLの誤解/rethinking-graphql
sonatard
75
12k
Heart Work Chapter 1 - Part 1
lfama
PRO
8
36k
Exploring the Power of Turbo Streams & Action Cable | RailsConf2023
kevinliebholz
37
6.6k
How to Talk to Developers About Accessibility
jct
2
540
Fight the Zombie Pattern Library - RWD Summit 2016
marcelosomers
234
17k
A Tale of Four Properties
chriscoyier
163
24k
Templates, Plugins, & Blocks: Oh My! Creating the theme that thinks of everything
marktimemedia
31
2.9k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
32
3k
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
141
35k
Transcript
"53BOL"O"UUFOUJPO#BTFE 6TFS#FIBWJPS.PEFMJOH'SBNFXPSL GPS3FDPNNFOEBUJPO 12 / 3/ 6 2 0 /6
6/ 2 0 2 1 # ( # )9 9
จͷ֓ཁ • "UUFOUJPOʹجͮ͘ϢʔβߦಈϞσϦϯάͱͦͷϨίϝϯυ λεΫͷԠ༻ΛఏҊ • 3//$//ϕʔεͷํ๏ΑΓֶशͷऩଋ͕ૣ͘ߴੑೳ • ͬ͘͟Γݴ͑5SBOTGPSNFSͷϨίϝϯυͷద༻ """* ஶऀΒژେֶͱ"MJCBCBͷάϧʔϓ
ಋೖ
*OUSPEVDUJPO • ϢʔβʔࣗͷߦಈʹΑΓදݱՄೳ • ୯ޠ͕ͦͷपғͷίϯςΩετͰදݱͰ͖Δ͜ͱͱಉ༷ • ߦಈܥྻΛΤϯίʔυ͢ΔͨΊʹ3//$//͕සग़ • 3//ฒྻԽࠔ͔ͭظͷґଘؔͷҡ࣋ࠔ •
$//ฒྻԽՄೳ͕ͩہॴతͳґଘ͔ؔ͠ҡ࣋ෆՄ • 3//$//ͱʹ͍ܥྻͱ͍ܥྻͷ྆ํ্ख͘ѻ͑ͳ͍ • ܥྻͷ͞ʹؔͳ͘ґଘؔΛҡ࣋Ͱ͖Δ"UUFOUJPO͕ొ • ୯ͳΔ"UUFOUJPOͩͱϕΫτϧͷ֤࣍ݩʹ͍ͭͯಉ͡ॏΈͰࢉ • ֤࣍ݩͷҟͳΔlҙຯzΛ্ख͘ѻ͑ͳ͍ .VMUJ)FBE"UUFOUJPOͷ༧ײ
ํ๏
4FMG"UUFOUJPO#BTFE#FIBWJPS.PEFMJOH'SBNFXPSL
4FMG"UUFOUJPO#BTFE#FIBWJPS.PEFMJOH'SBNFXPSL άϧʔϓ ʢ*% ళ*% ʣ ݕࡧ άϧʔϓ ʢݕࡧ τʔΫϯʣ
Ϋʔϙϯडཧ άϧʔϓ ʢΫʔϙϯ*% ళ*% ʣ Ϩίϝϯυ ର *% ళ*% !" #$
4FMG"UUFOUJPO#BTFE#FIBWJPS.PEFMJOH'SBNFXPSL άϧʔϓ ʢ*% ళ*% ʣ ݕࡧ άϧʔϓ ʢݕࡧ τʔΫϯʣ
Ϋʔϙϯडཧ άϧʔϓ ʢΫʔϙϯ*% ళ*% ʣ Ϩίϝϯυ ର *% ళ*% !" #$
4FMG"UUFOUJPO#BTFE#FIBWJPS.PEFMJOH'SBNFXPSL άϧʔϓ ʢ*% ళ*% ʣ ݕࡧ άϧʔϓ ʢݕࡧ τʔΫϯʣ
Ϋʔϙϯडཧ άϧʔϓ ʢΫʔϙϯ*% ళ*% ʣ Ϩίϝϯυ ର *% ళ*%
4FMG"UUFOUJPO#BTFE#FIBWJPS.PEFMJOH'SBNFXPSL !"# άϧʔϓ
ʢ*% ళ*% ʣ ݕࡧ άϧʔϓ ʢݕࡧ τʔΫϯʣ Ϋʔϙϯडཧ άϧʔϓ ʢΫʔϙϯ*% ళ*% ʣ Ϩίϝϯυ ର *% ళ*%
4FMG"UUFOUJPO#BTFE#FIBWJPS.PEFMJOH'SBNFXPSL - άϧʔϓ ʢ*% ళ*% ʣ ݕࡧ άϧʔϓ ʢݕࡧ
τʔΫϯʣ Ϋʔϙϯडཧ άϧʔϓ ʢΫʔϙϯ*% ళ*% ʣ Ϩίϝϯυ ର *% ళ*%
4FMG"UUFOUJPO#BTFE#FIBWJPS.PEFMJOH'SBNFXPSL άϧʔϓ ʢ*% ళ*% ʣ ݕࡧ άϧʔϓ ʢݕࡧ τʔΫϯʣ
Ϋʔϙϯडཧ άϧʔϓ ʢΫʔϙϯ*% ళ*% ʣ Ϩίϝϯυ ର *% ళ*% -
4FMG"UUFOUJPO#BTFE#FIBWJPS.PEFMJOH'SBNFXPSL άϧʔϓ ʢ*% ళ*% ʣ ݕࡧ άϧʔϓ ʢݕࡧ τʔΫϯʣ
Ϋʔϙϯडཧ άϧʔϓ ʢΫʔϙϯ*% ళ*% ʣ Ϩίϝϯυ ର *% ళ*% - DPODBU
4FMG"UUFOUJPO#BTFE#FIBWJPS.PEFMJOH'SBNFXPSL άϧʔϓ ʢ*% ళ*% ʣ ݕࡧ άϧʔϓ ʢݕࡧ τʔΫϯʣ
Ϋʔϙϯडཧ άϧʔϓ ʢΫʔϙϯ*% ళ*% ʣ Ϩίϝϯυ ର *% ళ*% - - - - - - -
4FMG"UUFOUJPO#BTFE#FIBWJPS.PEFMJOH'SBNFXPSL άϧʔϓ ʢ*% ళ*% ʣ ݕࡧ άϧʔϓ ʢݕࡧ τʔΫϯʣ
Ϋʔϙϯडཧ άϧʔϓ ʢΫʔϙϯ*% ళ*% ʣ Ϩίϝϯυ ର *% ళ*%
3BX'FBUVSF4QBDFT • ϢʔβʔͷߦಈΛ{", $, %}ͱͯ͠ఆࣜԽ • Ϣʔβʔ' = { ")
, $) , %) |+ = 1,2, … , /} • 'Λߦಈͷੑ࣭ʹΑΓҟͳΔάϧʔϓʹׂ " $ % ߦಈλΠϓ FH Ӿཡɺߪങɺ ߦಈΦϒδΣΫτ FH \*% ళ*%^ɺ\ݕࡧτʔΫϯ^ɺ ߦಈൃੜ࣌ͷλΠϜελϯϓ 0 = {123 , 124 , … , 125 } 123 ∩ 124 = 7 ' =∪9:9 5 129 άϧʔϓ ݕࡧ άϧʔϓ Ϋʔϙϯडཧ άϧʔϓ
#FIBWJPS&NCFEEJOH4QBDFT • άϧʔϓ!"#ʹ͓͍ͯ $% , '% , (% Λ&NCFEEJOH )#%
= +# ($% , '% , (% ) $% '% (% ී௨ʹ&NCFEEJOH ී௨ʹ&NCFEEJOHͯ݁͠߹ʢ \*% ళ*%^ͳͲෳଘࡏ͢ΔͷͰʣ ΧςΰϦԽͯ͠&NCFEEJOH ߦಈൃੜ࣌ͷλΠϜελϯϓ͔ΒϨίϝϯυ࣌·ͰͷܦաͷϨϯδ͝ͱʹΧςΰϦʹม 0, 1 , 1,2 , 2,4 , … , [24, 2456) 0 1 2 7 + 1
#FIBWJPS&NCFEEJOH4QBDFT !"# = %&'" (# + *((+!," - '!.+%/01%" /#
+ *((+!," 2(4# ) จͰՃࢉ͕ͩ ࣮Ͱ݁߹ ͜ͷϨΠϠʔͷ࠷ऴతͳग़ྗ 6 = {!89: , !89< , … , !89> } !89@ = .(A.4/# (!"# ) ֤άϧʔϓͰ!"#ͷϕΫτϧͷTIBQFҟͳΔ ಉ͡ಛྔͷ&NCFEEJOHڞ༗ ͨͩ࣌ؒ͠ͷ&NCFEEJOHڞ༗͠ͳ͍
-BUFOU4FNBOUJD4QBDFT • ֤ߦಈΛ,ݸͷજࡏతҙຯۭؒʹӨ • ͭ·Γ.VMUJ)FBE"UUFOUJPOΛߦ͏ • Ϣʔβʔͷߦಈ!Λάϧʔϓ͝ͱʹҟͳΔӨؔ"#$ Λ༻͍ͯ ݻఆͷϕΫτϧʹม •
4ΛͦΕͧΕͷજࡏతҙຯۭؒʹӨ % = '()'*+ , ("#. /01. , "#3 /013 , … , "#5 (/015 )) %7 = "87 (%) 5SBOTGPSNFSͷ .VMUJ)FBE"UUFOUJPOͷ9:ʹ૬
-BUFOU4FNBOUJD4QBDFT • !"# !$% จͰ୯ҰͷEFOTFMBZFSͰ׆ੑԽ͕ؔ3F-6 • ֤TIBQFΛཧ &'(( &) *'((
શߦಈ *% άϧʔϓ+ͷߦಈ ݻఆͷϕΫτϧͷ࣍ݩ ֤જࡏతҙຯۭؒͷ࣍ݩ *% = *'(( -
4FMG"UUFOUJPO-BZFS • ,ݸͷ֤જࡏతҙຯۭؒͰߦಈؒͷ4FMG"UUFOUJPO
4FMG"UUFOUJPO-BZFS • จͷهड़Ͱ !" = $%&'()*()(," , ,: /" ))
ۭؒ1ʹ͓͚Δ"UUFOUJPOXFJHIUϕΫτϧͷߦྻ ) ," , ,: /" = ," 2" ,3 ۭؒ1ʹ͓͚Δ"UUFOUJPOϕΫτϧͷߦྻ 4" = !" 567 (,) ࠷ऴతͳग़ྗ 4 = 589:; (<%=<)' > (4> , 4? , … , 4" )) Ͱ্࣮݁ہͬͯΔ͜ͱ 567 , 589:; TJOHMFEFOTFMBZFS XJUI3F-6
%PXOTUSFBN"QQMJDBUJPO/FUXPSL • ͜͜·ͰϢʔβʔߦಈϞσϧ͕ • ͜ͷจͰϨίϝϯυλεΫΛߟ͑Δ ͬͪ͜ଆͷ
%PXOTUSFBN"QQMJDBUJPO/FUXPSL • Ϩίϝϯυରͷߦಈͱ4FMG"UUFOUJPOޙͷߦಈؒͰ7BOJMMB "UUFOUJPO • ͭ·ΓϢʔβʔߦಈଆΛΤϯίʔμɺϨίϝϯυରߦಈଆΛ σίʔμͱߟ͑ͨ4PVSDF5BSHFU"UUFOUJPO Τϯίʔμ σίʔμ
%PXOTUSFBN"QQMJDBUJPO/FUXPSL • Ϩίϝϯυରߦಈͷ&NCFEEJOHͰ͋Δ!"ʹରͯ͠ɺ͜Ε· Ͱͱ΄΅ಉ͡ԋࢉΛߦ͏ • ݻఆϕΫτϧͷมͱ.VMUJ)FBE"UUFOUJPO ℎ" = %&' "
(!" ) *+ = %,- (ℎ" ) .+ = */01234 3 *+ , 6; 8+ %9- (6) :; " = %<=>?@@= (./A.31 B (.B , .C , … , .+ )) ߦಈάϧʔϓͷΠϯσοΫεͷஔ %<=>?@@= TJOHMFEFOTFMBZFS XJUI3F-6
%PXOTUSFBN"QQMJDBUJPO/FUXPSL • QPJOUXJTFͷ༧ଌϞσϧͰߦಈ!͕औΒΕΔ֬Λ༧ଌ • ଛࣦؔʹγάϞΠυͷΫϩεΤϯτϩϐʔ − # $,& '$ ()*+(-
ℎ$ , /& $ ) + 1 − '$ log(1 − +(- ℎ$ , /& $ )) - ϥϯΩϯάؔʢจͰ୯७ͳੵʣ
࣮ݧ
%BUBTFU • "NB[PO%BUBTFU • Ϣʔβʔͷʹର͢ΔϨϏϡʔσʔλ • ߦಈͷάϧʔϓ͚ͭͩʢάϧʔϓʣ • ಛྔϢʔβʔ*%ɺ*%ɺΧςΰϦ*%ɺλΠϜελϯϓ •
Ϣʔβͷߦಈ("# , "% , … , "' )͕͋Δͱֶͯ͠शɾςετσʔλҎԼ ("# ) "% ("# , "% ) ") ("# , "% , … , "'*% ) "'*# *OQVU QSFEJDU ֶश ςετ ("# , "% , … , "'*# ) "' *OQVU QSFEJDU ˞OFHBUJWFTBNQMF QPTJUJWFTBNQMFҎ֎ͷΛϥϯμϜͰબ
%BUBTFU • 5BPCBP%BUBTFU • "MJCBCB͕ӡӦ͢Δ&$αΠτͷϩά • ߦಈͷάϧʔϓͭʢɺݕࡧɺΫʔϙϯडཧʣ • άϧʔϓ •
*%ɺళ*%ɺϒϥϯυ*%ɺΧςΰϦ*%ɺλΠϜελϯϓ • ߦಈλΠϓӾཡɺϚʔΫɺߪೖ • ݕࡧάϧʔϓ • ݕࡧτʔΫϯɺλΠϜελϯϓ • ߦಈλΠϓͭ • Ϋʔϙϯडཧάϧʔϓ • Ϋʔϙϯ*%ɺళ*%ɺΫʔϙϯλΠϓɺλΠϜελϯϓ • ߦಈλΠϓͭ
$PNQFUJUPST)ZQFSQBSBNFUFST • $PNQFUJUPST • #13.'#BZFTJBO1FSTPOBMJ[FE3BOLJOH • #J-45. • #J-45. "UUFOUJPO
• $// 1PPMJOH • )ZQFSQBSBNFUFST • શͯͷ&NCFEEJOHͷ࣍ݩ • શͯͷӅΕͷ࣍ݩ • જࡏతҙຯۭؒͷۭؒʢ.VMUJ)FBEͷIFBEʣ • όοναΠζ • ਖ਼ଇԽFͷ-ϊϧϜ • 0QUJNJ[FS4(%ɺֶशΛ͔Β࢝ΊͯͰEFDBZ
&WBMVBUJPO.FUSJDT • BWFSBHFVTFS"6$ !"# = 1 "&'() * +∈-./01 1
2+ 3 2+ 4 * 5∈67 8 * 5∈67 9 :(<+,5 > <+,? )
3FTVMUTPO4JOHMFUZQF#FIBWJPS%BUBTFU • "NB[PO%BUBTFU "6$"53BOL͕࠷ྑ ࣌ؒͷΧςΰϦ͝ͱͷ7BOJMMB"UUFOUJPOXFJHIUͷฏۉ ͕࣌ؒܦա͢ΔʹͭΕXFJHIU͕খ͘͞ 3//$//ͷ༻͕ग़དྷ͍ͯΔ "53BOLֶ͕श͕ૣ͘ऩଋ
3FTVMUTPO.VMUJUZQF#FIBWJPS%BUBTFU • 5BPCBP%BUBTFU "53BOL͕ྑ͍͕ɺதͰશͯͷߦಈάϧʔϓΛ༻͍ͯ ߦಈϞσϦϯάΛߦͬͨϞσϧ "53BOLBMMPOF ͕࠷ྑ
$BTF4UVEZ • 4FMG"UUFOUJPOͱ7BOJMMB"UUFOUJPOͷޮՌΛ֬ೝ ߦಈ Ϩίϝϯυ ର
$BTF4UVEZ • 4FMG"UUFOUJPOͷޮՌ ಛఆͷྻʢ7BMVFʣ͕ߴ͍XFJHIU ๅੴಉ࢜Ͱਖ਼ํܗΛܗɺ͓ͦΒ͘ΧςΰϦͷྨࣅੑΛݟ͍ͯΔ
$BTF4UVEZ • 7BOJMMB"UUFOUJPOͷޮՌ ؔ͋Γͦ͏ͳߦಈͷΈڧௐ શͯͷߦಈͰฏۉత