$30 off During Our Annual Pro Sale. View Details »
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Turing × atmaCup #18 - 1st Place Solution
Search
Shuhei Goda
December 13, 2024
Technology
0
1.1k
Turing × atmaCup #18 - 1st Place Solution
Turing × atmaCup #18 の表彰式での登壇資料です
https://turing.connpass.com/event/338583/
Shuhei Goda
December 13, 2024
Tweet
Share
More Decks by Shuhei Goda
See All by Shuhei Goda
ジョブマッチングサービスにおける相互推薦システムの応用事例と課題
hakubishin3
3
1.1k
とある事業会社にとっての Kaggler の魅力
hakubishin3
9
3k
課題の解像度が荒かったことで意図した改善ができなかった話
hakubishin3
3
1.1k
Wantedly におけるマッチング体験を最大化させるための推薦システム
hakubishin3
4
1.3k
Recommendation Industry Talks #1 Opening
hakubishin3
1
440
会社訪問アプリ「Wantedly Visit」での シゴトに関する興味選択機能と推薦改善
hakubishin3
0
710
論文紹介: Improving Implicit Feedback-Based Recommendation through Multi-Behavior Alignment(Xin Xin et al., 2023)
hakubishin3
0
690
Feedback Prize - English Language Learning における擬似ラベルの品質向上の取り組み
hakubishin3
1
1.1k
ウォンテッドリーにおける推薦システムのオフライン評価の仕組み
hakubishin3
7
7.3k
Other Decks in Technology
See All in Technology
Microsoft Agent Frameworkの可観測性
tomokusaba
1
110
Lookerで実現するセキュアな外部データ提供
zozotech
PRO
0
200
New Relic 1 年生の振り返りと Cloud Cost Intelligence について #NRUG
play_inc
0
240
Bedrock AgentCore Memoryの新機能 (Episode) を試してみた / try Bedrock AgentCore Memory Episodic functionarity
hoshi7_n
2
1.9k
AI with TiDD
shiraji
1
300
マイクロサービスへの5年間 ぶっちゃけ何をしてどうなったか
joker1007
21
8.2k
AI との良い付き合い方を僕らは誰も知らない
asei
0
270
フルカイテン株式会社 エンジニア向け採用資料
fullkaiten
0
9.9k
『君の名は』と聞く君の名は。 / Your name, you who asks for mine.
nttcom
1
120
さくらのクラウド開発ふりかえり2025
kazeburo
2
1.2k
特別捜査官等研修会
nomizone
0
580
松尾研LLM講座2025 応用編Day3「軽量化」 講義資料
aratako
7
4.1k
Featured
See All Featured
Visualizing Your Data: Incorporating Mongo into Loggly Infrastructure
mongodb
48
9.8k
The Cost Of JavaScript in 2023
addyosmani
55
9.4k
How to Build an AI Search Optimization Roadmap - Criteria and Steps to Take #SEOIRL
aleyda
1
1.8k
Making Projects Easy
brettharned
120
6.5k
Agile Actions for Facilitating Distributed Teams - ADO2019
mkilby
0
97
Breaking role norms: Why Content Design is so much more than writing copy - Taylor Woolridge
uxyall
0
120
CoffeeScript is Beautiful & I Never Want to Write Plain JavaScript Again
sstephenson
162
16k
Mobile First: as difficult as doing things right
swwweet
225
10k
The Success of Rails: Ensuring Growth for the Next 100 Years
eileencodes
47
7.9k
Building Flexible Design Systems
yeseniaperezcruz
330
39k
HU Berlin: Industrial-Strength Natural Language Processing with spaCy and Prodigy
inesmontani
PRO
0
110
Designing Dashboards & Data Visualisations in Web Apps
destraynor
231
54k
Transcript
© 2024 Wantedly, Inc. 1st Place Solution Dec. 14 2024
- Shuhei Goda Turing × atmaCup #18 දজࣜ&ৼΓฦΓձ
© 2024 Wantedly, Inc. ໊લɿ ߹ా पฏ Shuhei Goda
ॴଐͱׂɿ ΥϯςουϦʔגࣜձࣾ ɾData Team Manager ɾMachine Learning Tech Lead ɾProduct Manager Kaggle Tierɿ Kaggle Competitions Grandmaster @jy_msc ࣗݾհ https://www.kaggle.com/shuheigoda
© 2024 Wantedly, Inc. Turing × atmaCup #18 ʹ͍ͭͯ •
։࠵ظؒɿ2024/11/15 17:30 ʙ 2024/11/24 18:00 • ࣗಈंͷߦγʔϯͷΧϝϥը૾ं྆ͷঢ়ଶσʔλͳͲ͔Βɺ0.5 ~ 3s ޙͷ ࣗंͷҐஔΛਪఆ͢ΔλεΫʢي༧ଌʣ
© 2024 Wantedly, Inc. ࠓճͷ Public LeaderBoard : 1st Place
🎉 Private LeaderBoard : 1st Place 🎉
© 2024 Wantedly, Inc. λΠϜϥΠϯʢ11/19͔ΒࢀՃʣ
© 2024 Wantedly, Inc. ϓϩηεʢͬ͘͟Γͱʣ 1.ੳͱํܾΊ
© 2024 Wantedly, Inc. ϓϩηεʢͬ͘͟Γͱʣ 2. 1st-stageͷվળ
© 2024 Wantedly, Inc. ϓϩηεʢͬ͘͟Γͱʣ 3. 2nd-stageͷվળ
© 2024 Wantedly, Inc. ϓϩηεʢͬ͘͟Γͱʣ 4. Ξϯαϯϒϧͷ४උͱ࣮ࢪ
© 2024 Wantedly, Inc. 1. ੳͱํܾΊ
© 2024 Wantedly, Inc. 1. ੳͱํܾΊ EDAϕʔεϥΠϯΞϓϩʔν͔ΒɺΛѲͯ͠ํΛߟ͑Δ • σʔλଟ͘ͳ͘ɺγϯϓϧͳ͕Βςʔϒϧಛྔ͕ڧͦ͏ •
ͱ͍ͬͯɺ༧ଌʹ͓͍ͯը૾ͷใ༗ޮͦ͏ʹݟ͑Δʢৄࡉޙड़ʣ → ଞࢀՃऀͱͷେ͖ͳࠩҟͱͳΓಘΔͷʮը૾ใͷѻ͍ํʯͩͱߟ͑ͨ
© 2024 Wantedly, Inc. 1. ੳͱํܾΊ ਐΊํͱ࣌ؒͷ͍ํΛܾΊΔ • 1st-stage Ϟσϧʢը૾Λѻ͏ϞσϧʣͷվળΛॏతʹΔ
• 2nd-stage ϞσϧʢςʔϒϧಛྔϝΠϯͷϞσϧʣͷվળΛগ͠Δ • ͋ͱϞσϧΛՔ͙࡞ۀΛߦ͍Ξϯαϯϒϧ͢Δ
© 2024 Wantedly, Inc. 1. ੳͱํܾΊ ࣮ݧʹ༻ͨ͠ϕʔεϥΠϯ • tk@tnkcoder ͞Μ͕ެ։ͨ͠ϕʔεϥΠϯϞσϧΛར༻
• [CV 0.2008/LB 0.2017] LightGBM + CNN stacking baseline (LightGBM + CNN) 1st-stage: CNN 2nd-stage: GBDT Image Tabular 1st-stage Predictions Submission
© 2024 Wantedly, Inc. 2. 1st-stageͷվળ
© 2024 Wantedly, Inc. 2-1. Target ͷվળ ֤ Target ͷ࠷େͰׂͬͨ
Target Λֶशɾ༧ଌ͢Δ • ݁ՌɿX ͱ Y ͷ༧ଌੑೳ͕վળɻt ͕খ͍͞΄ͲޮՌ͕େ͖͍ • ղऍɿ༧ଌ࣌Ͱλʔήοτͷεέʔϧ͕େ͖͘ҟͳΔɻεέʔϧΛ߹ΘͤࠐΉ͜ͱͰɺ ֤༧ଌ࣌ͷใΛ·ͱΊͯޮՌతʹֶशͰ͖ΔͷͰͳ͍͔
© 2024 Wantedly, Inc. 2-1. Target ͷվળ ֤༧ଌ࣌ͷՃΛ Auxiliary Target
ͱֶͯ͠शɾ༧ଌ͢Δ • ྫ͑ Target ͷ x_0, x_1 ͔Β vx_1 Λࢉग़͢Δ͜ͱ͕Ͱ͖Δ • ݁Ռͱͯ͠ɺ1st-stage CV: 0.2312 → 0.2288 (-0.0024) ʹվળ • ·ͨɺAuxiliary Target ʹର͢Δ༧ଌΛޙஈͷಛྔͱͯ͠Ճ͢Δ͜ͱͰ ɺ2nd-stage ͷείΞ͕վળʢ2nd-stage CV: 0.1963 → 0.1933ʣ
© 2024 Wantedly, Inc. 2-2. HorizontalFlip ࢥͬͨ͜ͱɿࣗಈं͔ΒࡱӨ͞Εͨը૾ɺਫฏసͤͯ͞ҧײ͕গͳ͍ Ͳ͕ͬͪΦϦδφϧʁ
© 2024 Wantedly, Inc. 2-2. HorizontalFlip ֶश࣌ɾਪ࣌ʹ HorizontalFlip ΛՃ͑Δ •
ֶश࣌ɿp=0.5 Ͱ HorizontalFlip • ਪ࣌ɿΦϦδφϧը૾ͷਪ݁Ռͱਫฏసͨ͠ਪ݁ՌΛฏۉ͢Δ • ͜ΕΒʹΑͬͯείΞ৳ͼΔ͕ɺ1st-stage Ϟσϧʹೖྗ͢Δςʔϒϧಛྔ ͷసΛΕͯ͠·͏ͱείΞ͕ٯʹԼ͕ͬͯ͠·͏ͷͰҙ • 1st-stageͰѻ͏ςʔϒϧಛྔۃྗγϯϓϧʹ͑Δඞཁ͕͋ΔɻΘ Γʹ 2nd-stage ʹෳࡶͳFEΛدͤΔ͜ͱ͕Ͱ͖Δ {“steeringAngleDeg”: 15, “leftBlinker”: True, “rightBlinker”: False} → {“steeringAngleDeg”: -15, “leftBlinker”: False, “rightBlinker”: True}
© 2024 Wantedly, Inc. 2-3. Scene୯Ґ ࢥͬͨ͜ͱɿಉҰγʔϯͷલޙͷࢹ֮తใΛ͏͜ͱͰɺ୯ҰID(t-1.0 ~ t)͚ͩͩ ͱࠔͳ༧ଌͰ͖ΔΑ͏ʹͳΔͷͰʁΑΓظతͳӡసঢ়گͷѲ͕ॏཁ
ྫ1ɿ sec=2.0, t-0.5 sec=2.0, t-1.0 sec=2.0 sec=12.0 12secޙͷใ͔Βɺͦͷ··ਐ͢Ε ྑ͔ͬͨ͜ͱ͕Θ͔Δ
© 2024 Wantedly, Inc. 2-3. Scene୯Ґ ࢥͬͨ͜ͱɿಉҰγʔϯͷલޙͷࢹ֮తใΛ͏͜ͱͰɺ୯ҰID(t-1.0 ~ t)͚ͩͩ ͱࠔͳ༧ଌͰ͖ΔΑ͏ʹͳΔͷͰʁΑΓظతͳӡసঢ়گͷѲ͕ॏཁ
ྫ2ɿ sec=2.0 sec=12.0 12secޙͷใ͔ΒɺࣼΊʹਐΊ ྑ͔ͬͨ͜ͱ͕Θ͔Δ
© 2024 Wantedly, Inc. 2-3. Scene୯Ґ ظͷมԽʢ1ඵ୯Ґʣ ɾٸͳૢ࡞มԽ ɾՃݮ ɾंઢมߋ
ɾӈࠨં ظͷมԽʢ୯Ґʣ ɾߦత ɾӡసελΠϧ ɾతͷܦ࿏
© 2024 Wantedly, Inc. 2-3. Scene୯Ґ ϘτϧωοΫͱͳ͍ͬͯΔʮظมԽʯΛޮՌతʹϞσϧԽ͢Δʹʁ • ϕʔεϥΠϯͰ2nd-stageʹ͓͍ͯલޙͷظใΛߟྀͨ͠༧ଌ͕Մೳͩ ͕ɺ1st-stageʹ͓͍֤ͯID͕ಠཱͳͷͱͯ͠ѻ͏&ܦ࿏༧ଌͷ݁Ռͱͯ͠ͷ
ใΛൖͤ͞ΔܗʹͳΔͷͰඇޮʹݟ͑Δ • 1st-stage ͷNNͷஈ֊ͰɺظͷมԽʹجͮ͘ΛֶशͰ͖ΔΑ͏ʹ͢Δ 1st-stage: CNN 2nd-stage: GBDT sceneA,ID1 1st-stage Predictions 1st-stage: CNN 1st-stage Predictions FE(e.g. shift features) sceneA,ID2 Shared
© 2024 Wantedly, Inc. 2-3. Scene୯Ґ ۩ମతͳΞϓϩʔνɿScene୯ҐͰ 2.5D-CNN + LSTM
CNN 1st-stage Predictions (B×S×N) BiLSTM … Tabular sec=20 MLP sec=120 Scene
© 2024 Wantedly, Inc. 2-3. Scene୯Ґ sec=20 sec=120 Pad Pad
Pad Pad sec=220 sec=320 sec=520 Pad Pad Pad sec=20 sec=120 sec=220 sec=320 sec=420 sec=520 scene=A scene=B scene=C όονͷ࡞Γํ • Scene͝ͱʹ͕͞ҟͳΔͷͰɺ٧ΊͯPadding • αϯϓϧؒͰ࣌ܥྻతͳҐஔ͕ؔҟͳΔͷͰɺscene_sec scene_num ʢsceneͷத ͰԿ൪ʹొͨ͠ID͔ʣΛಛྔͱͯ͠ೖྗ
© 2024 Wantedly, Inc. 2-3. Scene୯Ґ ۩ମతͳΞϓϩʔνɿ2.5D-CNN + LSTM Λ࠾༻͢Δ
• ͜ͷΞʔΩςΫνϟʹมߋ͢Δ͜ͱͰɺCNN୯ମͰ Private 4Ґ૬ͷείΞʹ ྫ1ɿ ྫ2ɿ
© 2024 Wantedly, Inc. 3. 2nd-stageͷվળ
© 2024 Wantedly, Inc. ͍͔ͭ͘ͷಛྔͷՃ ͍ͣΕͦͦ͜͜ͷվળʹد༩ͨ͠ • 1st stage ͷ
target (x_0 ~ z_5) ͷ༧ଌʹՃ͑ͯɺaux target ͷ༧ଌಛྔ ͱͯ͠ར༻͢Δ • 2छྨͷं྆ϞσϧʢϢχαΠΫϧϞσϧͱಈྗֶతόΠγΫϧϞσϧʣͷ༧ଌ ݁ՌΛಛྔͱͯ͠ར༻͢Δ
© 2024 Wantedly, Inc. 4. Ξϯαϯϒϧ
© 2024 Wantedly, Inc. Ξϯαϯϒϧ ༷ʑͳόοΫϘʔϯͰϞσϧΛ࡞ͬͯ Weighted Average • جຊతʹϞσϧΛ૿͢΄ͲείΞ͕େ͖͘৳ͼΔɻ࠷ऴతʹ11ݸࠞͥͨɻ
• ͬͨόοΫϘʔϯɿresnext, efficientnet, resnet, swin-transformer ͳͲ
© 2024 Wantedly, Inc. ࠷ऴ݁Ռ
© 2024 Wantedly, Inc. ֤ϞσϧͷύϑΥʔϚϯε model cv public private private
ॱҐ ɹsingle 1st stage 0.1906 0.1958 0.1808 4Ґ ɹsingle 2nd stage 0.1883 0.1928 0.1785 1Ґ ɹensemble 0.1792 0.1885 0.1754 1Ґ
© 2024 Wantedly, Inc. ϕʔεϥΠϯʹൺͯ͏·͍͘͘Α͏ʹͳͬͨྫ - ظతͳঢ়گѲ͕ޮ͍͍ͯΔ าಓʹಥͬࠐ ·ͳ͘ͳͬͨ
นʹಥͬࠐ ·ͳ͘ͳͬͨ ରंઢʹ ৵ೖ͠ͳ͘ͳͬͨ ΨʔυϨʔϧ ʹಥͬࠐ·ͳ ͘ͳͬͨ
© 2024 Wantedly, Inc. ૬มΘΒͣ͏·͍͔͘ͳ͍ྫ - ͦͷʹ͓͚Δঢ়گѲ͕ͳ͔ͳ͔͍͠ ࠨ͔Β ं͕ग़͖ͯͨ
τϥοΫͰ ৴߸͕ݟ͑ͳ͍ ETCϨʔϯ ঃߦ͠ͳ͍ͱ ͍͚ͳ͍ ԣஅาಓۙ͘ʹ ௨ߦਓ͍ͳ͍
© 2024 Wantedly, Inc. ຊίϯϖʹର͢ΔऔΓΈํʹ͍ͭͯ
© 2024 Wantedly, Inc. എܠ ࢠͲ͕ੜ·Ε͔ͯΒɺॳΊͯͷσʔλੳίϯϖͷࢀՃ ύύKagglerʹͳΓ·ͨ͠
© 2024 Wantedly, Inc. എܠ ͔ͤͬ͘ࢀՃ͢ΔͳΒPrizeݍʹೖΓ͍ͨ… Ͱ • ͕ͬͭΓίʔυΛॻ͚Δͷɺൺֱత͘৸ͯ͘ΕΔਂͷΈ •
։࠵ظؒͷલͱޙՈఉͷ༻ࣄͰ1த͕࣌ؒऔΕͳ͍ ͋Μ·Γ࣌ؒऔΕͳ͍ɺͲ͏͠Α͏
© 2024 Wantedly, Inc. Ͳ͏औΓΉ͖͔ Do everything
© 2024 Wantedly, Inc. Ͳ͏औΓΉ͖͔ Do everything Δ͜ͱɾΒͳ͍ ͜ͱΛܾΊΔ
© 2024 Wantedly, Inc. ελϯε Δ͜ͱ • ڝ૪༏ҐͱͳΔٕज़՝ʢղܾ͖͍͢ʣΛਪఆ͠ɺͦΕʹṌ͚ͯऔΓΉ • ֎ΕͨΒૉʹఘΊΔɺΘΜͪΌΜϗʔϜϥϯͶΒ͍
Βͳ͍͜ͱ • ࡉ͔͍վળɺϋΠύϥνϡʔχϯάͳͲ • ܭࢉϦιʔε͕ۭ͍͍ͯͯɺͳΜͱͳ͘Ͱ࣮ݧΛճ͞ͳ͍Α͏ʹ͢Δ
© 2024 Wantedly, Inc. Ͳ͏͍͏՝Λղ͖͔͘Λઃఆ͢Δ Ͳ͏͍͏͍ʢnot Ξϓϩʔνʣ͕ࠩผԽϙΠϯτʹͳΔͷ͔ߟ͑Δ • ΞΠσΞΛεϙοτతʹݕূ͢ΔΑΓɺूத͢Δ͖՝Λઃఆͯ͠ਂ΅ͬ ͨ΄͏͕ɺదͳΞϓϩʔνʹͨͲΓண͖͍͢
ੳޙʹઃఆͨ͠՝ QɿલޙͷγʔϯͷมԽظͷΛ֫ಘ͢Δͷʹ༗ޮ͔ʁ Qɿӡస࣌ͷঢ়گ༧ଌʹͲͷΑ͏ʹӨڹ͢Δͷ͔ʁʢྫ͑ߴಓ࿏ͩͱʁʣ
© 2024 Wantedly, Inc. ੜAIػೳͰ࣮ݧεϐʔυΛૣ͘͢Δ ࣮ݧαΠΫϧ͕ैདྷͷ1/2~1/3ͷ࣌ؒͰճͤΔΑ͏ʹ • ࣮ݧͷઃܭ͔ΒݕূʢσόοάʣʹࢸΔ·Ͱͷ࣌ؒͷେ෯ͳॖ • ΊΜͲ͘͞…
ͱ͍͏৺ཧతϋʔυϧΛେ෯ʹԼ͛Δʢਖ਼͜Ε͕େ͖͍ʣ • ྫ͑ɺID୯Ґ→Scene୯ҐͷมߋɺมߋՕॴ͕ଟͯ͘ਏ͍ • Ͳ͏ઃܭ͢Δ͖͔ɺͲ͏͍͏มߋՕॴ͕͋Δ͔ɺͲ͏࣮͢Δ͔Λ͑ ͯΒ͏ɻͦͯ͠ίέͨΒσόοάͷࡐྉΛΒ͏ ΞΠσΞͷ ݕ౼ ઃܭ ࣮ ݕূ