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
ECサイトにおける閲覧履歴を用いた購買に繋がる行動の変化検出 / Change Detecti...
Search
Hiroka Zaitsu
May 15, 2020
Technology
1
860
ECサイトにおける閲覧履歴を用いた購買に繋がる行動の変化検出 / Change Detection in Behavior Followed by Possible Purchase Using Electronic Commerce Site Browsing History
財津大夏, 三宅悠介
GMOペパボ株式会社 ペパボ研究所
2020.05.15 第49回 情報処理学会 インターネットと運用技術研究会
Hiroka Zaitsu
May 15, 2020
Tweet
Share
More Decks by Hiroka Zaitsu
See All by Hiroka Zaitsu
Vertex AI Matching Engine と CLIP を使って EC サービスの類似画像検索機能を作る / Development of similar image search function for EC services using Vertex AI Matching Engine and CLIP
zaimy
0
640
BigQuery の日本語データを Dataflow と Vertex AI でトピックモデリング / Topic modeling of Japanese data in BigQuery with Dataflow and Vertex AI
zaimy
1
5.3k
データサイエンティストの仕事紹介 / Data Scientist Job Introduction
zaimy
1
560
GMOペパボのサービスと研究開発を支えるデータ基盤の裏側 / Inside Story of Data Infrastructure Supporting GMO Pepabo's Services and R&D
zaimy
1
1.7k
正則化とロジスティック回帰/machine-learning-lecture-regularization-and-logistic-regression
zaimy
0
8.3k
trinity で Cloud Composer に ワークフローを簡単デプロイ / Easy workflow deployment to Cloud Composer with trinity
zaimy
0
830
ハンドメイド作品を対象としたECサイトにおける大量生産品の検出 / Detection of Mass-produced Goods at EC Site to Trade Handmade Goods
zaimy
3
4.6k
キャリアキーノート2018 / Career Keynote 2018
zaimy
1
2.1k
ウェブサービスにおける行動ログ活用基盤を通したデータ駆動マーケティングの実践 / Practice of data driven marketing using behavior log foundation system on web service
zaimy
8
2.7k
Other Decks in Technology
See All in Technology
技術に触れたり、顔を出そう
maruto
1
130
Docker Desktop で Docker を始めよう
zembutsu
PRO
0
140
GeometryReaderやスクロールを用いた表現と紐解き方
fumiyasac0921
0
100
The future we create with our own MVV
matsukurou
0
1.9k
Unsafe.BitCast のすゝめ。
nenonaninu
0
190
KMP with Crashlytics
sansantech
PRO
0
220
今年一年で頑張ること / What I will do my best this year
pauli
1
210
RubyでKubernetesプログラミング
sat
PRO
4
150
Amazon Q Developerで.NET Frameworkプロジェクトをモダナイズしてみた
kenichirokimura
1
190
comilioとCloudflare、そして未来へと向けて
oliver_diary
6
420
AWS re:Invent 2024 re:Cap Taipei (for Developer): New Launches that facilitate Developer Workflow and Continuous Innovation
dwchiang
0
140
Accessibility Inspectorを活用した アプリのアクセシビリティ向上方法
hinakko
0
170
Featured
See All Featured
Mobile First: as difficult as doing things right
swwweet
222
9k
Bash Introduction
62gerente
610
210k
Large-scale JavaScript Application Architecture
addyosmani
510
110k
XXLCSS - How to scale CSS and keep your sanity
sugarenia
248
1.3M
実際に使うSQLの書き方 徹底解説 / pgcon21j-tutorial
soudai
173
50k
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
30
2.1k
Design and Strategy: How to Deal with People Who Don’t "Get" Design
morganepeng
127
18k
Git: the NoSQL Database
bkeepers
PRO
427
64k
Statistics for Hackers
jakevdp
797
220k
Raft: Consensus for Rubyists
vanstee
137
6.7k
Product Roadmaps are Hard
iamctodd
PRO
50
11k
Practical Tips for Bootstrapping Information Extraction Pipelines
honnibal
PRO
10
860
Transcript
ࡒେՆ, ࡾ༔հ / Pepabo R&D Institute, GMO Pepabo, Inc. 2020.05.15
ୈ49ճ ใॲཧֶձ Πϯλʔωοτͱӡ༻ٕज़ݚڀձ ECαΠτʹ͓͚ΔӾཡཤྺΛ༻͍ͨ ߪങʹܨ͕ΔߦಈͷมԽݕग़
1. ݚڀͷత 2. ՝ 3. ఏҊख๏ 4. ࣮ݧͱߟ 5. ·ͱΊͱࠓޙ
2 ࣍
1. ݚڀͷత
• ECαΠτΛ๚ΕΔϢʔβʔෳͷతΛ࣋ͭ • ྫʣʮΟϯυγϣοϐϯάʯʮͷ୳ࡧʯʮಛఆͷߪങʯͳͲ • ECαΠτͷӡӦऀ͕؍ଌՄೳͳϢʔβʔͷߦಈతʹΑͬͯมԽ͢Δ • ྫʣʮͷݕࡧʯʮͷӾཡʯʮͷߪങʯͳͲ ͷ୳ࡧ͕త ➡
ͷछྨͰݕࡧͯ͠ݕࡧ݁ՌΛϖʔδӾཡ ಛఆͷߪങ͕త ➡ ໊Ͱݕࡧͯ͠ϖʔδΛৄ͘͠Ӿཡ 4 ECαΠτͷϢʔβʔͷతͱߦಈ
• ϢʔβʔͷߦಈͷมԽʹ߹ΘͤͯECαΠτͷγεςϜΛదԠతʹ มԽͤ͞Δ͜ͱͰߪങͷ্͕ظ͞ΕΔ • Λ୳ࡧ͍ͯ͠Δ ➡ ଟ༷ੑͷ͋Δਪનख๏ʹΓସ͑ͯڵຯΛऒ͘ • ಛఆͷߪങΛߦ͓͏ͱ͍ͯ͠Δ ➡
ܾࡁಋઢΛࣔͯ͠ߪങΛଅ͢ • ECαΠτͷγεςϜͷదԠతͳมԽΛ࣮ݱ͢ΔͨΊʹɼ Ϣʔβʔ͕ԿΒ͔ͷߦಈΛऔͬͨޙʹมԽΛݕग़͍ͨ͠ 5 Ϣʔβʔͷߦಈʹ߹ΘͤͨECαΠτͷదԠతͳมԽ
• ECαΠτͷγεςϜͷదԠతͳมԽΛ࣮ݱ͢ΔͨΊʹɼ Ϣʔβʔ͕ԿΒ͔ͷߦಈΛऔͬͨޙʹมԽΛݕग़͍ͨ͠ • Ϣʔβʔ͕औΓ͏ΔߦಈECαΠτ͝ͱʹ༷ʑ • ຊใࠂͰECαΠτʹڞ௨ͷߦಈͱͯ͠ߪങʹܨ͕ΔߦಈͷมԽݕग़ΛఏҊ 6 ࠓճͷใࠂͷൣғ
2. ՝
• ECαΠτ͝ͱʹར༻Մೳͳಛྔͷ͏ͪɼͲΕΛߪങʹܨ͕Δߦಈͷ มԽݕग़ʹ༻͍Δ͖͔͕ະ • ಛྔΛશͯ༻͍ΔਂֶशHMMͳͲͷֶशϕʔεͷख๏͕͋Δ͕ɼ • ࣍ݩ͕૿͑Δ΄ͲඞཁͳαϯϓϧαΠζ͕૿େ͢Δ • Ϟσϧͷ൚ԽੑೳΛ্ͤ͞Δ͜ͱ͕ࠔʹͳΔ •
࣍ݩͷগͳ͍୯७ͳಛྔͰߦಈͷมԽΛݕग़Ͱ͖Δ͜ͱ͕·͍͠ 8 ՝ᶃมԽݕग़ʹ༻͍Δ͖ಛྔ͕ະ
• طଘݚڀʹ͓͚ΔʮϢʔβʔͷతʹରԠ͢ΔӾཡύλʔϯͷྨʯ(*1,2) • ॳظஈ֊ɿΧςΰϦʔϖʔδͱϖʔδΛଟ͘Ӿཡ͢Δ • ߪങͷલɿগͷϖʔδʹӾཡ͕ूத͢Δ • Ϣʔβʔ͝ͱͷ͋ΔظؒͷʮӾཡճʯͱʮͷछྨͷʯ ࣍ݩͷগͳ͍ಛྔʹͳΓ͏Δ *1
Moe, W.W.: Buying, searching, or browsing: Differentiating between online shoppers using in-store navigational clickstream, Journal of Consumer Psychology, Vol.13, Is-sues 1-2, pp.113-123 (2003). *2 Οϥϫϯɾυχɾμϋφ:ใ୳ࡧͷతΛߟྀͨ͠ߪങܾఆϞσϧ,ϚʔέςΟϯάɾαΠΤϯε, Vol.25, No.1,pp.15-35 (2017). 9 طଘݚڀ͔Βͷಛྔͷީิ
• Ϣʔβʔ͝ͱͷ͋ΔظؒͷʮӾཡʯͱʮͷछྨͷʯ ECαΠτϢʔβʔ͝ͱʹಛྔͷ͕औΔൣғʹࠩҟ͕͋Δ • શͯͷϢʔβʔʹֶ͍ͭͯशσʔλΛ४උ͢Δ͜ͱࠔ • ֶशෆཁͳΞϓϩʔνͰߦಈͷมԽΛݕग़͢Δ 10 ՝ᶄڥ͝ͱʹಛྔͷ͕औΔൣғʹࠩҟ͕͋Δ
3. ఏҊख๏
• ᶃߪങʹܨ͕ΔߦಈͷมԽݕग़ʹ༻͍Δ͖ಛྔ͕ະ • ࣍ݩͷগͳ͍୯७ͳಛྔͰߦಈͷมԽΛݕग़Ͱ͖Δ͜ͱ͕·͍͠ • ᶄڥ͝ͱʹಛྔͷ͕औΔൣғʹࠩҟ͕͋Γֶशσʔλͷ४උ͕ࠔ • ֶशෆཁͳΞϓϩʔνͰߦಈͷมԽΛݕग़͢Δ 12 ՝ͷཧ
• ECαΠτͷγεςϜͷదԠతͳมԽΛ࣮ݱ͢ΔͨΊʹɼ Ϣʔβʔ͕ԿΒ͔ͷߦಈΛऔͬͨޙʹมԽΛݕग़͍ͨ͠ • ᶃ࣍ݩͷগͳ͍୯७ͳಛྔΛ༻͍ͯᶄֶशෆཁͳΞϓϩʔνͰ ߪങʹܨ͕ΔߦಈͷมԽݕग़Λߦ͏ • ᶃͷӾཡճʹର͢Δͷଐੑͷछྨͷൺ • ઌߦݚڀΑΓɼ͜ͷߪങʹ͚ͯখ͘͞ͳΔͱԾఆ
• ᶄ౷ܭతԾઆݕఆʹΑΔฏۉͷࠩͷݕఆ 13 ఏҊख๏
• ͷӾཡճʹର͢Δͷଐੑͷछྨͷൺ • Ϣʔβʔ ͷߦಈཤྺ • ʹӾཡ ݕࡧ ͳͲ͕͋Δ •
ͷҙͷҐஔͷΟϯυ Λߟ͑Δ • ୠ͠ɼΟϯυαΠζ ͱ ͔ͭ Λຬͨ͢࠷খͷࣗવ Λ༻͍ͯ u Su = (a1 , a2 , …, al ) a aview asearch Su Wu (t) = (a′ 1 , a′ 2 , a′ 3 , …, at ) w 1 < n < w t − w + n > 0 n a′ 1 = at−w+n a′ 2 = at−w+n+1 a′ 3 = at−w+n+2 14 ಛྔͷఆٛᶃ
• ͷӾཡճʹର͢Δͷଐੑͷछྨͷൺ • ͷҙͷҐஔͷΟϯυ ʹ͓͚Δ • ͷଐੑ ͷछྨʹؔ͢Δू߹ Λ༻͍ͯ ಛྔ
• ͕খ͍͞΄Ͳߪങʹ͔͍ͬͯΔ Su Wu (t) = (a′ 1 , a′ 2 , a′ 3 , …, at ) aview ͷରͱͳͬͨͷଐੑ attr ͷछྨ ͷӾཡ aview ͷճ attr rattr(Wu (t)) = || count(aview) 15 ಛྔͷఆٛᶄ
• Ϣʔβʔɹͷߦಈཤྺ • ͰͷIDʹؔ͢Δಛྔ • ͱ ͷରͷID=1ɼ ͷରͷID=2ͱ͢Δͱ Su =
(asearch 1 , aview 2 , aview 3 , asearch 4 , aview 5 , aview 6 , aview 7 , aview 8 , aview 9 , apurchase 10 ) Wu (5) = (asearch 1 , aview 2 , aview 3 , asearch 4 , aview 5 ) aview 2 aview 3 aview 5 rID(Wu (5)) = || count(aview) = 2 3 16 ಛྔͷྫ u Wu (5)
• ಛྔͷਪҠͷΟϯυ Λߟ͑Δ • ୠ͠ɼΟϯυαΠζ ͱ ͔ͭ Λຬͨ͢࠷খͷࣗવ Λ༻͍ͯ(*) •
ΛҙͷͰೋͨ͠Οϯυ ͱ ʹରͯ͠ ౷ܭతԾઆݕఆʹΑΔฏۉͷࠩͷݕఆΛద༻ • ༗ҙਫ४ Ͱ༗ҙࠩ͋Γͱݟͳͨ͠߹ʹ ͷ࠷ॳͷཁૉΛมԽͱݟͳ͢ * r' ΛٻΊΔࣜΛݚڀใࠂͷ͔࣌Βमਖ਼͍ͯ͠·͢ W′ u (t) = (r′ 1 , r′ 2 , r′ 3 , …, rattr(Wu (t))) w′ 1 < m < w′ t − w′ + m > 0 m r′ 1 = rattr(Wu (t − w′ + m)) r′ 2 = rattr(Wu (t − w′ + m + 1)) r′ 3 = rattr(Wu (t − w′ + m + 2)) W′ u (t) W′ 1 W′ 2 s W′ 2 17 ಛྔͷਪҠΛ༻͍ͨมԽݕग़ͷఆٛᶃ
• ౷ܭతԾઆݕఆʹΑΔฏۉͷࠩͷݕఆʹ Welch ͷ ݕఆΛ༻͍Δ • Student ͷ ݕఆͷվྑ •
ࢄ͕͍͜͠ͱΛԾఆ͠ͳ͍ • ͷΈʹରԠ͕Մೳ • ඪຊͷࢄ͕͘͠ͳ͍߹ʹൣʹରԠ͠͏Δ t t 18 ಛྔͷਪҠΛ༻͍ͨมԽݕग़ͷఆٛᶄ
• ͷͱ͖ ͷ֤ʹ Welch ͷ ݕఆΛద༻ • ͱ ͷͰ༗ҙࠩ͋Γͱݟͳͨ͠߹ ͷ࣌ࠁ
ΛมԽͱݟͳ͢ W′ u (t) = (r′ 1 , r′ 2 , r′ 3 , r′ 4 , r′ 5 ) W′ 1 = (r′ 1 ) W′ 2 = (r′ 2 , r′ 3 , r′ 4 , r′ 5 ) W′ 1 = (r′ 1 , r′ 2 ) W′ 2 = (r′ 3 , r′ 4 , r′ 5 ) W′ 1 = (r′ 1 , r′ 2 , r′ 3 ) W′ 2 = (r′ 4 , r′ 5 ) W′ 1 = (r′ 1 , r′ 2 , r′ 3 , r′ 4 ) W′ 2 = (r′ 5 ) t W′ 1 = (r′ 1 , r′ 2 ) W′ 2 = (r′ 3 , r′ 4 , r′ 5 ) r′ 3 = rattr(Wu (t − w′ + m + 2)) t 19 ಛྔͷਪҠΛ༻͍ͨมԽݕग़ͷྫ
4. ࣮ݧͱߟ
• ࣮ࡍͷECαΠτͷӾཡཤྺʹ͓͚ΔఏҊख๏ͷ༗ޮੑͷݕূ • GMOϖύϘגࣜձࣾͷӡӦ͢ΔECαΠτʮminneʯͷӾཡཤྺʹద༻ͨ͠ 1. ϋΠύʔύϥϝʔλͷݕ౼ 2. ఏҊख๏ʹదͨ͠࡞ଐੑͷߟ 3. ݸผͷϢʔβʔʹର͢ΔมԽݕग़ͷ݁Ռͷ֬ೝ
• ECαΠτͷߦಈੳʹ༻͍ΒΕΔӅΕϚϧίϑϞσϧͱͷਫ਼ͷൺֱ • ܭࢉ࣌ؒͷ֬ೝ ࣮ݧͷతͱํ๏ 21
• ECαΠτʮminneʯͷϓϩμΫγϣϯڥʹ͓͚ΔӾཡཤྺ • 20203݄10͔࣌Β24࣌·Ͱͷσʔλ • Ӿཡཤྺ ͷܥྻ ͷ 96,984 Ϣʔβʔ
• ൺֱͷͨΊߪങΛߦͬͨϢʔβʔͱߦΘͳ͔ͬͨϢʔβʔʹׂ • ࡞ʹඥͮ͘4ͭͷଐੑͰ࣮ݧ • ࡞IDɼ࡞ͷग़ऀIDɼ࡞ͷΧςΰϦάϧʔϓɼ࡞ͷΧςΰϦ Su l ≥ 6 σʔληοτ 22
• ΧςΰϦάϧʔϓ • ྫʣʮϑΝογϣϯʯΧςΰϦάϧʔϓͷΧςΰϦ • TγϟπɼϫϯϐʔεɼτοϓεɼίʔτɼεΧʔτ ͳͲ ࡞ଐੑ - ࡞ͷΧςΰϦάϧʔϓͱΧςΰϦ
23
ϋΠύʔύϥϝʔλͷݕ౼ • Ӿཡཤྺ͔ΒಛྔͷΛٻΊΔࡍͷΟϯυͷ෯ Λ {5,10} Ͱ࣮ݧ • ಛྔͷͷมԽΛݕग़͢ΔࡍͷΟϯυͷ෯ Λ {3,5}
Ͱ࣮ݧ • ߪങϢʔβʔʹؔͯ͠ΑΓଟ͘ͷมԽΛݕग़͠ɼඇߪങϢʔβʔʹؔͯ͠ গͳ͍มԽΛݕग़ͨ͠ ͱ ΛҎ߱ͷ࣮ݧʹ༻͍ͨ • ༗ҙਫ४ • ׳ྫతͳͱͯ͠ Λ༻͍ͨ w w′ w = 10 w′ = 5 s s = 0.05 24
• ࡞ଐੑ͝ͱͷಛྔͷͷਪҠΛശͻ͛ਤͰ֬ೝ • ྫ ఏҊख๏ʹద͢Δ࡞ଐੑͷߟ 25 • ԣ࣠ɿ࣌ܥྻ • ॎ࣠ɿಛྔͷ
• ശͷ্ɿୈࡾ࢛Ґ • ശͷԼɿୈҰ࢛Ґ • ശͷதͷԣઢɿதԝ • ͻ͛ͷ্ɿୈࡾ࢛Ґʴ࢛Ґൣғͷ1.5ഒ • ͻ͛ͷԼɿୈҰ࢛Ґ−࢛Ґൣғͷ1.5ഒ • ͻ͛ͷ্Լͷɿ֎Ε • ͍ॎઢɿதԝʹରͯ͠ఏҊख๏Λద༻ͯ͠ݕग़ͨ͠มԽ
ఏҊख๏ʹద͢Δ࡞ଐੑ ߪങϢʔβʔ ඇߪങϢʔβʔ ࡞*% ࡞ͷग़ऀ*% 26 • ߪങϢʔβʔɿಛྔͷ͕Լ͕ΔʹมԽΛݕग़ • ඇߪങϢʔβʔɿ΄΅มԽΛݕग़͍ͯ͠ͳ͍ʢߦಈͷॳظಛྔͷͷมಈ͕େ͖͍ͨΊ1Օॴݕग़ʣ
➡ ఏҊख๏ͷಛྔʹ༻͍Δ࡞ଐੑͱͯ͠ద͍ͯ͠Δ
ఏҊख๏ʹద͞ͳ͍࡞ଐੑ ߪങϢʔβʔ ඇߪങϢʔβʔ ࡞ͷΧςΰϦάϧʔϓ ࡞ͷΧςΰϦ 27 • ߪങϢʔβʔͱඇߪങϢʔβʔͷ྆ํͰ࣌ܥྻͷॳظʹಛྔͷ͕Լ͕ΓɼͦͷޙมԽ͠ͳ͘ͳΔ • minne
ͰΧςΰϦͷߜΓࠐΈ͕ߪങͷ༗ແͱؔͳ͘ߦಈͷॳظʹߦΘΕΔ ➡ ఏҊख๏ͷಛྔʹ༻͍Δ࡞ଐੑͱͯ͠ద͍ͯ͠ͳ͍
ӅΕϚϧίϑϞσϧʢHMMʣͱͷൺֱᶃ • ݸผͷϢʔβʔʹର͢Δਫ਼ͷݕ౼ • Ϟσϧͷग़ྗΛ༧ଌϥϕϧʮߪങϢʔβʔʯʹϚοϐϯά͢Δ • ఏҊख๏ɿมԽΛݕग़ͨ͠߹ • HMMɿӅΕঢ়ଶ2ͷ͏ͪಛྔͷͷฏۉ͕͍ঢ়ଶʹભҠͨ͠߹ •
HMMͷϞσϧͷߏஙͷͨΊσʔληοτΛ9:1ʹׂ • ܇࿅σʔλɿ87,285Ϣʔβʔ • ςετσʔλɿ9,523Ϣʔβʔ 28
ӅΕϚϧίϑϞσϧʢHMMʣͱͷൺֱᶄ • ఏҊख๏ΑΓHMMͷํ͕ੵۃతʹʮߪങϢʔβʔʯͷϥϕϧΛ͚ͨ ࡞IDΛಛྔʹ༻͍ͨ߹ͷࠞಉߦྻ ਖ਼ղϥϕϧ ߪങ ඇߪങ ༧ଌϥϕϧ ఏҊख๏ ߪങ
526 4551 ඇߪങ 201 4245 HMM ߪങ 662 5571 ඇߪങ 65 3225 ࡞ͷग़ऀIDΛಛྔʹ༻͍ͨ߹ͷࠞಉߦྻ ਖ਼ղϥϕϧ ߪങ ඇߪങ ༧ଌϥϕϧ ఏҊख๏ ߪങ 483 5719 ඇߪങ 244 3077 HMM ߪങ 679 7047 ඇߪങ 48 1749 29
ӅΕϚϧίϑϞσϧʢHMMʣͱͷൺֱᶅ • ఏҊख๏ • ਅͷඇߪങϢʔβʔʹର͢Δਫ਼͕ߴ͍ • ِཅੑʹରِͯ͠ӄੑ͕͍ • ߪങʹܨ͕ΔϢʔβʔͷߦಈͷมԽݕग़ͷతʹԊ͍ͬͯΔ •
HMM • ਅͷߪങϢʔβʔʹର͢Δਫ਼͕ߴ͍ • ʮߪങ͠ͳ͔ͬͨʯʹϚοϐϯά͞ΕΔӅΕঢ়ଶͷ͕ฏۉ1.0ɼඪ४ภࠩ1.16*10−8ͱͳͬͯ ͓Γɼ͔ᷮͰಛྔͷ͕ݮগ͢Δͱʮߪങͨ͠ʯӅΕঢ়ଶʹભҠ͍ͯͨ͠ 30
ܭࢉ࣌ؒ • 3.1GHz ΫΞουίΞ Intel Core i7 Λར༻͢ΔධՁڥʹ͓͍ͯɼΟϯυ ͋ͨΓͷܭࢉ࣌ؒ1.71ϛϦඵʙ1.75ϛϦඵ
• ΣϒαΠτͷಡΈࠐΈ࣌ؒ1,000ϛϦඵະຬ͕·͍͠ͱ͞Ε͓ͯΓɼఏ Ҋख๏ʹΑΔมԽݕग़ʹֻ͔Δ࣌ؒेʹখ͍͞ W′ u (t) 31
5. ·ͱΊͱࠓޙ
·ͱΊ • ߪങʹܨ͕ΔϢʔβʔͷߦಈͷมԽݕग़ • Ӿཡཤྺ͔ΒಛྔΛ࡞ͯ͠౷ܭతԾઆݕఆʹΑͬͯมԽݕग़Λߦ͏ • ࣮ࡍͷECαΠτͷσʔλΛ༻͍ͯಛྔʹ༻͍Δଐੑͷݕ౼ͱਫ਼͓Α ͼܭࢉ࣌ؒͷ֬ೝΛߦͬͨ • HMMͱͷൺֱͰඇߪങϢʔβʔʹؔ͢Δਫ਼ʹ্ؔͯ͠ճΓɼࣄલͷֶश
͕ෆཁ 33
ࠓޙʹ͍ͭͯ • ఏҊख๏ͷਫ਼ͷվળ • ಛྔͷ͕มԽ͢Δࡍͷਖ਼ෛํͷϞσϧͷΈࠐΈ • ಛྔͷͷมಈ͕େ͖͍ظؒͷআ֎ͳͲ • ܭࢉ࣌ؒͷॖ •
มԽݕग़ʹ༻͍ΔΟϯυΛ֤ཁૉͰׂͤͣҰՕॴͰׂ͢Δ • খඪຊʹରͯ͠ؤ݈ͳ౷ܭతԾઆݕఆͷख๏ͷݕ౼ 34