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
Kaggle M5-Forecasting (Walmart)
Search
IHiroaki
July 19, 2020
Programming
2
390
Kaggle M5-Forecasting (Walmart)
先日開催された、Kaggle(M5-Forecasting)の当方のSolution資料です。
IHiroaki
July 19, 2020
Tweet
Share
Other Decks in Programming
See All in Programming
Less waste, more joy, and a lot more green: How Quarkus makes Java better
hollycummins
0
100
Creating a Free Video Ad Network on the Edge
mizoguchicoji
0
120
Macとオーディオ再生 2024/11/02
yusukeito
0
370
Ethereum_.pdf
nekomatu
0
460
Duckdb-Wasmでローカルダッシュボードを作ってみた
nkforwork
0
130
リアーキテクチャxDDD 1年間の取り組みと進化
hsawaji
1
220
色々なIaCツールを実際に触って比較してみる
iriikeita
0
330
Micro Frontends Unmasked Opportunities, Challenges, Alternatives
manfredsteyer
PRO
0
110
3 Effective Rules for Using Signals in Angular
manfredsteyer
PRO
1
100
ピラミッド、アイスクリームコーン、SMURF: 自動テストの最適バランスを求めて / Pyramid Ice-Cream-Cone and SMURF
twada
PRO
10
1.3k
[Do iOS '24] Ship your app on a Friday...and enjoy your weekend!
polpielladev
0
110
Amazon Bedrock Agentsを用いてアプリ開発してみた!
har1101
0
340
Featured
See All Featured
Optimising Largest Contentful Paint
csswizardry
33
2.9k
Helping Users Find Their Own Way: Creating Modern Search Experiences
danielanewman
29
2.3k
Building Adaptive Systems
keathley
38
2.3k
Let's Do A Bunch of Simple Stuff to Make Websites Faster
chriscoyier
506
140k
Testing 201, or: Great Expectations
jmmastey
38
7.1k
A Philosophy of Restraint
colly
203
16k
Fantastic passwords and where to find them - at NoRuKo
philnash
50
2.9k
Why Our Code Smells
bkeepers
PRO
334
57k
How GitHub (no longer) Works
holman
310
140k
Creating an realtime collaboration tool: Agile Flush - .NET Oxford
marcduiker
25
1.8k
Statistics for Hackers
jakevdp
796
220k
The Language of Interfaces
destraynor
154
24k
Transcript
LBHHMFOBNF*)JSPBLJ .'PSFDBTUJOH "DDVSBDZ6ODFSUBJOUZ
࣍ɿ 1. ࣗݾհ 2. ݁Ռ 3. ࠓճͷऔΓΈͱߟ͑ 4. Ϟσϧ֓ཁ 5.
σʔλ୳ࡧ 6. ಛྔબ 7. Ϟσϧৄࡉ 8. লͱ՝
̍ɽࣗݾհ
̎ɽ݁Ռ ίϯϕͷ֓ཁͪ͜ΒΛࢀরɿhttps://www.kaggle.com/c/m5-forecasting-accuracy/overview ίϯϕͷ֓ཁͪ͜ΒΛࢀরɿhttps://www.kaggle.com/c/m5-forecasting-uncertainty/overview
̏ɽࠓճͷऔΓΈ ͱߟ͑ ʻऔΓΈʼ ɾॳίϯϖɻ ɾ3݄த०ʙ6݄ͷίϯϖऴྃ·Ͱͷ̏ϲ݄΄΅ٳΈͳ͠ͰରԠɻ ɾҰฏۉ̍̎ʙ̍̒࣌ؒΛίϯϖʹ๋͛Δɻ ʻߟ͑ʼ Accuracyɿ ɾ༧ଌΛͬͨಛྔٴͼલͷ༧ଌΛ༻͍ͨཌͷ༧ଌʢ࠶ؼతΞϓϩʔνʣߦΘͳ͍ɻʢಛʹ࠶ؼత Ξϓϩʔν̎ɺ̏ͷ༧ଌͳΒ༗ޮ͔͠Εͳ͍͕̎̔ͷ༧ଌͩͱޡࠩͷੵ͕େ͖͘ͳΓ͗͢ΔՄೳੑ͕͋
Δɻʣ ɾલͷ28ؒTrainDataͱͯ͠༻͢ΔɻʢաֶशɺֶशෆͷڪΕ͕͋Δ͜ͱ͔ΒҙΛ͍ϞσϧΛ࡞͢ Δඞཁ͕͋Δɻʣ Uncertaintyɿ ɾAccuracyͰͷ࠷ऴఏग़ΛҐͷ̑̌ˋͱ͢Δɻ ɾAccuracyϞσϧʹ͓͚ΔValidationظؒͷ࣮ͱ༧ଌͱͷֹࠩΛෆ࣮֬ੑͱͯ͠༻͢Δɻ ɾΑͬͯAccuracyʹ͓͍ͯ൚Խੑೳͷߴ͍Ϟσϧͷ࡞͕ॏཁͱͳΔɻ
̐ɽϞσϧ֓ཁ "DDVSBDZ 6ODFSUBJOUZ Ϟσϧɿ LightGBMͷΈΛ༻ Ϟσϧߏ : 28Λਖ਼֬ʹ༧ଌ͢ΔͨΊʹ1ຖʹݸผͷϞσϧ Λ࡞ɻ·ͨϝϞϦͷ͋Γɺstore_idຖʹϞ
σϧΛׂɻ߹ܭ 28 day × 10 id = 280 models ॏཁͳಛྔ: ಛྔʹؔͯ͋͠·Γಛผͳͷͳ͘ඪ४త ͳͷͷΈͱͳͬͨɻ ex) Basic Lagʢmean, max, ,min, std, medianʣ Average Encoding ʢ֤Ϩϕϧຖʣ IDʢTrainDataʹͯ༩͑ΒΕͨIDʣ ֶश࣌ؒɿ 8ʙ9ʢՄೳͳݶΓϦεΫΛഉআ্ͨ͠Ͱͷ࣌ ؒʣ ※ֶश࣌ؒΛॖ͢ΔͨΊͷํ๏ɻʢ༧ଌ͕গ͠ߥ͘ͳΔ͕ͦ͜·Ͱ μϝʔδ͕ͳ͍ͷʣ ɾLearningRateΛେ͖͘͠ɺnum_iterΛݮΒ͢ɻʢlr0.03ͳΒ iter500~700ఔʣ ɾBasicLagಛྔΛআ͢Δɻʢಛʹmulti_2, 3, 5, ʣ ɾstore_id୯ҐϞσϧΛͳ͘͢ɻʢͨͩ͠ಛྔΛेݮΒ͞ͳ͍ͱϝ ϞϦͷൃੜʣ Ϟσϧ : AccuracyΛ࡞͢Δࡍʹ༻ͨ͠Model Λ༻ɻ ࢉग़ํ๏ : Ґͷ͏ͪ̑̌ˋʹؔͯ͠Accuracyͷ Final SubmissionΛ͏ɻ ͦͷଞ̔ʹؔͯ͠Accuracyʹͯࢉग़ͨ͠ Validationظؒʹ͓͚Δ࣮ͱ༧ଌͷࠩΛෆ֬ ࣮ੑͱ͠ɺల։͢Δɻ
̑. σʔλ୳ࡧ ച্ݸͷϓϩοτʢ߹ܭʣ Ұݟ͢Δͱશମʹͬͯ ্ঢͰ͋ΔΑ͏ʹݟ ͑Δɻ ຖͷొΞΠςϜ ຖʹΞΠςϜ͕Ճ͞Ε͓ͯΓ Totalͷ্ঢͷཁҼͱͳ͍ͬͯΔ͜ͱ ͕ఆ͞ΕΔɻ
30490 ʢ̍ʣτϨϯυ ্ਤɿຖͷച্ݸͷ߹ܭਪҠ ԼਤɿຖͷΞΠςϜొਪҠ ্ਤΛݟΔͱҰݟ௨ظʹΘͨͬͯ૿Ճ͠ ͍ͯΔΑ͏ʹݟ͑Δ͕ԼਤͰΞΠςϜ͕ ʑొ͞Ε͍ͯΔ͜ͱ͕Θ͔Δɻ Αͬͯ͜ΕΒͷ৽͘͠ೖͬͨΞΠςϜʹ ΑΓ্ঢ͕ݟΒΕΔ͜ͱ͕ߟ͑Β Εɺ͜ͷ߹্ਤͰΛଊ͑Δ͜ͱ ͕Ͱ͖ͳ͍ɻ Αͬͯ࣍ʹΞΠςϜొผʢച্։࢝ ʣͷຖͷച্ݸͷ߹ܭਪҠΛݟͯ ΈΔɻ
̑. σʔλ୳ࡧ ച্։࢝ผͷച্ݸͷϓϩοτ ਤɿച্։࢝ผͷചΓ্͛ݸͷ߹ܭਪ Ҡ Ͳͷਤʹ͓͍ͯ2015લ·Ͱݮগ ʹ͋Δͷʹ͔͔ΘΒͣɺ2015ޙ͔ Β2016ʹ͔͚ͯ૿Ճ͍ͯ͠Δ͜ͱ͕Θ͔ Δɻ ͜ΕԿ͔͠ΒτϨϯυ͕มΘͬͨ͜ͱΛ
ද͍ͯ͠ΔՄೳੑ͕͋ΓValidationͷऔΔظ ؒϞσϧͷߏஙํ๏ʹؾΛ͚ͭΔඞཁ͕ ͋Δɻ ͔͠͠ɺاۀଆͷԿ͔ࢼ࡞ʹΑΔͷͳͷ ͔ɺফඅτϨϯυʹΑΔͷͳͷ͔͕ෆ໌ Ͱ͋ΓɺࠓճͷίϯϖΛߟ͑Δ্Ͱ͍͠ ͱ͜Ζͱͳͬͨɻ ʢ̍ʣτϨϯυ 2011 2012 2013 2014 2015 2016
̑. σʔλ୳ࡧ ਤɿ28ຖͷച্ݸͷ߹ܭਪҠʢάϥϑ store_idຖ͓Αͼച্։࢝ຖͰ͋Δʣ 28ؒʹ͓͚Δ߹ܭച্ݸͷਪҠͲ͏ มಈ͍ͯ͠Δͷ͔ΛݟͨάϥϑͰ͋Δ͕ɺ Γधཁ͋ΔఔҰఆͰ͋Δ͜ͱ ͔Β͔ɺٸܹͳ্ঢͷ͋ͱͷ28͋Δఔ ͑ΒΕௐ͞Ε͍ͯΔΑ͏ʹݟ͑Δɻ xʹ̓̌PublicLBظؒͰ͋Δ͕ଟ͘ͷάϥ
ϑͰٸܹͳ্ঢΛԋ͍ͯ͡Δɻ ΑͬͯݟͨͰ༧͢ΔʹɺPrivateظؒͷ 28ؒͷ߹ܭച্ݸPublicLBظؒʹൺ ͯݮগ͢ΔՄೳੑ͕͋Δఔ͋Δ͜ͱ͕ ૾Ͱ͖Δɻ ʢ͜Εʹؔͯ͠LagಛྔͷRollingʹͯ Ϟσϧʹ৫ΓࠐΊΔ͔ʁʣ ̎̔ຖͷച্ݸͷϓϩοτʢstore_idຖʣ ʢ̍ʣτϨϯυ
̑. σʔλ୳ࡧ ̎̔ຖͷച্ݸͷϓϩοτʢstore_idຖʣ ʢ̍ʣτϨϯυ
̑. σʔλ୳ࡧ ਤɿ֤ΞΠςϜʹ͓͚Δ͍Ζ͍Ζͳθϩ ͷύλʔϯΛάϥϑԽͨ͠ͷɻ DiscussionͰθϩύλʔϯʹର͢Δҙ ݟ͕ඇৗʹଟ͔ͬͨͱࢥ͏ɻ ࠓճͷ࣌ܥྻʹଟ͘ͷθϩ͕͋Δ͕ઓ ུతɺඞવతͳθϩ͕ଟؚ͘·Ε͍ͯ ͨɻ اۀʹࡏݿઓུɺઓུ͕͋ΓͦΕ
ΒຖมΘΓ͏ΔɻͦͷͨΊࡏݿઓ ུɺઓུ͕Θ͔Βͳ͍ঢ়ଶͰθϩύ λʔϯΛ༧ଌ͢Δ͜ͱͦΕͳΓʹϦε Ϋ͕͋Δͱײ͡Δɻ ·ͨࡏݿΕͨ·ͨ·ച্͕ͳ͔ͬͨ ͳͲͷθϩΛ༧͢Δʹͯ͠ධՁࢦඪ ্1ͷζϨڐ͞Εͳ͍͜ͱ͔Βɺ ΓθϩύλʔϯΛ༧͢ΔϦεΫେ ͖͍ɻ ࡏݿઓུɺઓུΛΒͳ͍ঢ়ଶͰθϩύλʔϯΛ༧ ͖͢Ͱͳ͍ʁ ? Change strategy? Irregular Long term ʢ̎ʣ͍Ζ͍Ζͳθϩύλʔϯ
̒. ಛྔબ ॏཁͳಛྔ ɾجຊతͳLagಛྔ ɹˠstore_id × item_idʹ͓͚ΔLagಛྔ ɹˠstore_id × item_id͔༵ͭ୯Ґʹ͓͚ΔLagಛྔ
ɾฏۉ ɹˠstore_id × item_id, state_id × item_id, item_idʹ͓͚Δ༵୯Ґͷฏۉʢ݄ʙʣ ɹˠstore_id × item_id, state_id × item_id, item_idʹ͓͚Δ୯Ґͷฏۉʢ̍ʙ̏̍ʣ ɾՁ֨มಈ ɾTrainDataʹͯ༩͑ΒΕͨID ࢼ͕ͨ͠͏·͍͔͘ͳ͔ͬͨಛྔ ɾ༧ଌΛ༻ͨ͠ಛྔ(ച্θϩύλʔϯΛԽͨ͠ಛetc…) ɾΫϥελϦϯάʹΑΔ৽ͨͳΧςΰϦ͚ʢྨࣅɺิʣ ɾ֎෦σʔλ etc…..
̒. ಛྔબ pred_day1 1ͷϞσϧͱ28ͷϞσϧॏཁ ͕ߴ͍ಛྔ͕͔ͳΓҟͳΔɻ 1ʹ͍ۙ΄ͲLagܥ͕ߴ͘ɺ28ʹ ۙͮ͘΄ͲฏۉIDͳͲͷΑΓҰൠԽ ͞Εͨಛྔͷॏཁ্͕͕Δɻ ϞσϧΛ28ݸʹ͚Δ͖ࠜڌʹͳ Δɻ
※ಛྔ໊ͷઆ໌࣍ͷεϥΠυ Feature Importance Plot - Top 20 pred_day28
̒. ಛྔબ ಛྔ໊ͷઆ໌ • sales_residual_diff_28_roll_365 : Targetʢৄࡉ࣍ͷεϥΠυʣ • multi_5_sales_residual_diff_28_roll_365_shift_1_roll_4_mean :
Code: df[“Target_shift_1”] = df.groupby([“id”])[“Target”].transform(lambda x : x.shift(1)) df.groupby([“id”, “multi_5”])[“Target_shift_1”].transform(lambda x: x.rolling(4).mean()) • private_sales_residual_diff_28_roll_365_enc_week(day)_LEVEL12_mean: privateɿϓϥΠϕʔτظؒͷલ·ͰͷσʔλΛ༻͢Δɻ enc_week(day)_LEVEL12_meanɿLEVEL12ͷ༵()ͷฏۉചΓ্͛ • sell_price_minority12 : sell_priceͷগୈҰҐͱೋҐ ex) 10.58345 => 58 • id_serial : ֤ID୯Ґʹઃఆͨ͠0 ~ 30489ͷ࿈൪
̓. Ϟσϧৄࡉ <Accuracy> TARGET = TARGET - TARGET.shift(28).rolling(365) ʢ̍ʣτϨϯυআڈ ܾఆܥͷϞσϧΛ͏߹ɺকདྷ༧ଌ
Λ͢ΔʹτϨϯυΛ͘ඞཁ͕͋Δͱ ͍ͬͨ༰͕Discussionʹ͋ͬͨΑ͏ ʹࠓճ༩͑ΒΕͨσʔλͷτϨϯυΛऔ Γআ͘͜ͱʹͨ͠ɻ ͔͠͠ɺػցֶशͳͲͰ༧ͨ͠༧ଌ ΛτϨϯυআڈͷࡐྉͱͯ͠͏͜ͱ ϦεΫ͕͋ΔͨΊ༻ͨ͘͠ͳ͔ͬͨɻ ࣮ࡍ༧ଌʹΑΔτϨϯυͷআڈࢼ͠ ͕ͨτϨϯυʹͯΊΔࣜʹΑΓɺ কདྷͷ༧ଌʹେ͖ͳ͕ࠩ͋ͬͨɻ ͦͷͨΊ࣮ΛͬͨআڈΛߟ͑ΔதͰ Ұ൪҆ఆ͍ͯͨ͠TARGET͔Β TARGET.shift(28)rolling(365)Λݮͨ͡ ͷΛTARGETͱ͢Δ͜ͱͱͨ͠ɻ ͔͠͠ɺ࣮ΛͬͨͨΊશʹτϨϯ υΛऔΓআ͚͓ͯΒͣޮՌݶఆతͰ ͋ͬͨͱײ͍ͯ͡Δɻ ͨͩखݩͰݕূ͢ΔݶΓ̎̔ؒͷ༧ଌ ͷ͏ͪޙʢ28͍ۙͷ༧ଌʣʹͳΔ ʹͭΕτϨϯυআڈޙͷํ͕҆ఆੑ͕ߴ ͔ͬͨɻ TARGET TARGET.shift(28).rolling(365) TARGET - TARGET.shift(28).rolling(365)
̓. Ϟσϧৄࡉ <Accuracy> lightgbm.Datasets( x_train, y_train, weight = myweight )
ʢ̎ʣweight objective : regression ධՁࢦඪͰ͋ΔWRMSSEΛೋͨ͠ͷ ͷޯΛܭࢉ͠Λlightgbm.Datasets ͷWEIGHTͱͯͨ͠͠ɻ WEIGHT^2÷SCALED͋Β͔͡Ί42840 ݸΛܭࢉ͓͖ͯ͠30490ΞΠςϜʹల։ ͦ͠ͷ߹ܭͱͨ͠ɻ 42840 1 30490 12Ϩϕϧ 30490 1 શϨϕϧʢ42840ݸʣͷʢWeight^2 ÷ ScaledʣΛ ܭࢉ͢Δɻ 30490ΞΠςϜ×12Ϩϕϧʹม 30490Ҏ֎ͷΞΠςϜΛ֤IDΧςΰϦຖʹׂΓ ৼΔɻ Ϩϕϧํʹ߹ܭΛࢉग़͢Δɻ
̓. Ϟσϧৄࡉ <Accuracy> ʢ̏ʣΠςϨʔγϣϯճ ֶश࣌ؒΛߟ͑Ε LearningRate→0.03 Iter→ 500 ~ 700
ͰΑ͔͕ͬͨstore_idຖ·ͨظؒʹ ΑͬͯऩଋͷλΠϛϯάͷζϨ͕͢ ͜͠େ͖͔ͬͨͷͰࠓճίϯϖͱ ͍͏͜ͱ͋Γɺ LearningRate→0.01 Iter→ 1200 & 1500(Blend) Λ࠾༻ͨ͠ɻ
̓. Ϟσϧৄࡉ <Accuracy> ʢ̐ʣ day-by-day Ϟσϧ Γ1ͷϞσϧͷํ͕είΞ͕͔ͳ Γྑ͘ͳ͍ͬͯΔɻ ಛʹ̍ʙ̏ͷӨڹ͕େ͖͘ɺਫ਼Λ ٻΊΔͳΒ̎̔Ϟσϧॏཁͱײ͡Δɻ
0.016
̓. Ϟσϧৄࡉ <Accuracy> • ݕূظؒ ʢݕূظؒ̍ʣ2016-04-25 ~ 2016-05-22 : score
0.53(Public LB) ʢݕূظؒ̎ʣ2016-03-28 ~ 2016-04-24 : score 0.51 ʢݕূظؒ̏ʣ2016-02-29 ~ 2016-03-27 : score 0.60 ʢςετظؒʣ2016-05-23 ~ 2016-06-19 : score 0.576 (Private LB) ɹɹ=>ݕূظؒʹؔͯ͠ຖʹΞΠςϜ͕࣍ʑʹೖ͞Ε͍ͯΔͨΊɺ·ͨۙʹτϨϯυ͕มΘͬͯɹ ɹɹɹɹ͍ΔՄೳੑ͕͋Δ͜ͱ͔ΒͳΔ͘લΛͬͨɻ • ύϥϝʔλʔ store_idʹΑͬͯগ͠มߋɻ • ϝτϦοΫ ϊʔτϒοΫΛࢀߟʹ࡞ʢߦྻܭࢉΛ༻͍ͯ͠ΔͨΊܭࢉ͕͍ʣ ɹ (https://www.kaggle.com/girmdshinsei/for-japanese-beginner-with-wrmsse-in-lgbm) • ࠶ؼతΞϓϩʔνɺͷ༻ͳ͠ • ޙॲཧͳ͠ ʢ̑ʣ ͦͷଞ
̓. Ϟσϧৄࡉ <Uncertainty> ʢ̍ʣ̑̌ˋͷࢉग़ ̑̌ˋɺM5 - Accuracy ʹ͓͚Δ࠷ऴఏग़ͱ͢Δɻ ·ͨɺߟ͑ํͱͯ͠ Accuracyͷ༧ଌϞσϧʹؔͯ͠
ݕূظؒͷWRMSSEɹ㲈ɹςετظؒͷWRMSSE ͳΒ ݕূظؒͷޡࠩʢෆ࣮֬ੑʣɹ㲈ɹςετظؒͷޡࠩʢෆ࣮֬ੑʣ Accuracyͷ༧ଌϞσϧ͕ҰൠԽ͞Ε͍ͯɺAccuracyͷϞσϧͦͷ ··ෆ࣮֬ੑͱͯ͑͠Δɻ
̓. Ϟσϧৄࡉ <Uncertainty> ʢ̎ʣෆ࣮֬ྖҬͷࢉग़ํ๏ʢ̑̌ˋҎ֎ͷࢉग़ʣ Accuracyͷ࠷ऴఏग़Λࢉग़ͨ͠ϞσϧΛ༻ͯ͠ ݕূظؒʹ͓͚Δޡࠩʹʛ࣮ʔ༧ଌʛΛͱΓɺޡࠩΛঢॱʹฒΔ ࠓճݕূظؒΛ3ͭઃఆͨͨ͠Ί߹ܭ̎̔ˎ̏ʹ̔̐ݸͷޡ͕ࠩੜ͡Δɻ ex) diff =
[0.5, 0.7, 1.4, 1.6, 1.7, 2.2, 2.6 ɾɾɾ 8.2, 8.5] ̔̐ <EJ⒎@DPVOU> <EJ⒎> ̐̎ ̔̍ άϥϑԽ ̑̒ ̔̐ 99.5% 0.5%ͷෆ࣮֬ੑ A B C D 97.5% 2.5%ͷෆ࣮֬ੑ 75.0% 25.0%ͷෆ࣮֬ੑ 83.5% 16.5%ͷෆ࣮֬ੑ 50.0% ʔ D ʹ 0.5% 50.0% ʔ C ʹ 2.5% 50.0% ʔ B ʹ 16.5% 50.0% ʔ A ʹ 25.0% Accuracyͷ࠷ऴఏग़ ʹ 50.0% 50.0% ʴ A ʹ 75.0% 50.0% ʴ B ʹ 83.5% 50.0% ʴ C ʹ 97.5% 50.0% ʴ D ʹ 99.5% ঢॱԽͨ͠ޡࠩͷ͏ͪ̎̑ˋɺ̓̑ˋʹ͋ͨΔޡࠩʢ̐̎൪ͷޡࠩʣΛ̑̌ˋ͔Β૿ݮͤͨ͞ͷΛ̎̑ˋɺ̓̑ˋͱ ͠ɺଞಉ༷ʹల։͢Δɻ ※͜ͷޡ͕ࠩ͜ͷϞσϧʹ͓͚Δෆ࣮֬ੑͱͳΔ 5SVF 1SFE
̓. Ϟσϧৄࡉ <Uncertainty> ࠓճݕূظؒΛ̎̔×̏Ͱߦͳ͕ͬͨ ຊདྷ֎Ε͕͋ͬͨ߹ͷճආߟ͑ Δͱഒͷ̎̔×̒͋ͬͨํ͕Α͔ͬͨ ͱײ͡Δɻ ͕͔͔ͨͩ࣌ؒΓ͗͢ΔͨΊɺaccuracy ͷϞσϧΛΑΓ্ܰͨ͘͠Ͱਫ਼Λग़͢ ͜ͱ͕͍Ζ͍ΖͳҙຯͰͷվળͷ༨ͱ
ͳΔɻʢࠓޙͷ՝ʣ ࠓճίϯϖͰͷݕূظؒͷ༧ଌʹ earlystop=100, lr =0.08ͱ͠গ͠ߥͷઃ ఆͰߦ͍ͬͯΔɻʢaccuracyଆͷաֶ शɺֶशෆϦεΫରࡦɻʣ ʢ̏ʣ༧ଌຖͷෆ࣮֬ੑ ༧ଌʹԠͯ͡ෆ࣮֬ੑͷେ͖͞ҟͳΔɻ ࠓճAccuracyʹ͓͍ͯຖͷϞσϧ(̎̔Ϟσϧ)Λ࡞͓ͯ͠Γɺ ਫ਼̍ͷϞσϧͷํ͕̎̔ͷϞσϧΑΓྑ͘ͳΔɻ ͦͷͨΊෆ࣮֬ੑʹ͓͍ͯ̎̔ϞσϧͦΕͧΕʹ͓͚ΔޡࠩʢલϖʔδʣΛࢉग़͠ɺల։͢Δ͜ͱ͕·͍͠ɻ ※දͷAɺBɺCɺDલϖʔδͷͦΕΒͱಉ͡ҙຯ߹͍ɻ
̔. লͱ՝ ֶश࣌ؒɿ ͬͱݕূΛ͏·͘ΕɺείΞΛ΄ͱΜͲམͱͣ͞ʹֶश࣌ؒΛେ෯ʹ͘Ͱ͖ͨͱࢥ͏ɻ ɾಛྔΛݮΒͯ͠ɺstore_id୯ҐͷϞσϧΛͳ͘͢ɻ ɾLearningRateͱIterationճͷௐ Etc Validationͷେࣄ͞ɿ ίϯϖং൫ɺPublicLBͷείΞʹؾΛऔΒΕ͗ͯ͢ɺޙ͔Βߟ͑ΕΔ͖Ͱͳ͍͜ͱʹ࣌ؒΛ͔͚ͯ͠ ·ͬͨɻ͜ͷίϯϖͰValidationͷେ͞Λ௧ײͰ͖ͨ͜ͱΑ͔ͬͨɻ
େͳσʔλͷॲཧɿ ಛʹং൫ϝϞϦͷ੍ݶͷதͲ͏Δ͔Ͱ͔ͳΓ࿑ྗΛͬͨɻػցֶशҎલʹࢄॲཧσʔλܕͳͲ ͬͱษڧ͠ͳ͚Ε͍͚ͳ͍͜ͱ͕ͨ͘͞Μ͋Δɻ ධՁࢦඪͷཧղɿ ·ͣॳΊʹධՁࢦඪͷཧղΛਂΊͳ͚Ε͍͚ͳ͍͜ͱΛ௧ײͨ͠ɻॳධՁࢦඪͷཧղ͕ᐆດͷ··ਐ ΜͰ͍ͨͨΊɺΔ͖Ͱͳ͍͜ͱΛଟ͍ͬͯͨ͘ɻධՁࢦඪʹΑͬͯ࡞Δ͖Ϟσϧ͕େ͖͘ҟͳΔ͜ͱ ͕Θ͔ͬͨɻ