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
[読み会] TabNet: Attentive Interpretable Tabular L...
Search
mei28
January 05, 2021
0
140
[読み会] TabNet: Attentive Interpretable Tabular Learning
読み会資料
TabNet: Attentive Interpretable Tabular Learning(ICLR, 2020, rejected)
mei28
January 05, 2021
Tweet
Share
More Decks by mei28
See All by mei28
[読み会] “Are You Really Sure?” Understanding the Effects of Human Self-Confidence Calibration in AI-Assisted Decision Making
mei28
0
82
[JSAI'24] 人間の判断根拠は文脈によって異なるのか?〜信頼されるXAIに向けた人間の判断根拠理解〜
mei28
1
490
[CHI'24] Fair Machine Guidance to Enhance Fair Decision Making in Biased People
mei28
0
55
[DEIM2024] 卓球の得点予測における重要要素の分析
mei28
0
37
[Human-AI Decision Making勉強会] 意思決定 with AIは個人vsグループで変わるの?
mei28
0
190
[読み会] Words are All You Need? Language as an Approximation for Human Similality Judgements
mei28
0
36
[参加報告] AAAI'23
mei28
0
88
[計算機構論] Learning Models of Individual Behavior in Chess
mei28
0
70
[計算機構論] Why do tree-based models still outperform deep learning on tabular data?
mei28
0
54
Featured
See All Featured
Keith and Marios Guide to Fast Websites
keithpitt
411
22k
Faster Mobile Websites
deanohume
306
31k
Automating Front-end Workflow
addyosmani
1368
200k
Fontdeck: Realign not Redesign
paulrobertlloyd
83
5.4k
Evolution of real-time – Irina Nazarova, EuRuKo, 2024
irinanazarova
6
570
GitHub's CSS Performance
jonrohan
1030
460k
What's in a price? How to price your products and services
michaelherold
244
12k
How to Create Impact in a Changing Tech Landscape [PerfNow 2023]
tammyeverts
49
2.3k
What’s in a name? Adding method to the madness
productmarketing
PRO
22
3.3k
The Pragmatic Product Professional
lauravandoore
32
6.4k
Helping Users Find Their Own Way: Creating Modern Search Experiences
danielanewman
29
2.4k
Navigating Team Friction
lara
183
15k
Transcript
TabNet: Attentive Interpretable Tabular Learning ಡΈձ@2021/01/05 ༶໌
• ஶऀ • Sercan O. Arik, Tomas Pfister •
Google Cloud AI • ग़య: ArxivͷPreprint • ICLR 2020ͰϦδΣΫτ͞Εͨจ จใ
• ςʔϒϧσʔλ͚ͷDNNϞσϧ • ܾఆͱNNϞσϧͷ͍͍ͱ͜औΓΛࢦͨ͠ख๏ • ղऍੑ + ਫ਼ ͷ্͕ୡͰ͖ͨɽ
֓ཁ ͲΜͳจʁ
• DNNͷϞσϧ͕ಛʹը૾,ݴޠ,ԻͷͰSOTAͰ͋Δɽ • KaggleͳͲͷੳίϯϖͰॳΊʹܾఆϕʔεͷख๏͕ओྲྀ • ղऍੑ͕ߴ͍͔Β ং ݚڀഎܠ
• ͳΜͰςʔϒϧσʔλʹରͯ͠ɼਂֶशΛऔΓೖΕ͍ͨͷ͔ʁ • େنͳσʔληοτʹ͍ͨͯ͠ɼਂֶशʹΑ্͕ͬͯظͰ͖Δ ͔Β • Deep Learning Scaling
is Predictable, Empirically.(Hestness et al., 2017) ং ݚڀഎܠ
• ςʔϒϧσʔλʹରͯ͠NNϞσϧΛ͏3ͭͷϝϦοτ 1. ෳͷσʔλΛޮΑ͘ΤϯίʔσΟϯάͰ͖Δ 2. ಛྔΤϯδχΞϦϯάͷखؒΛݮΒͤΔ 3. End-to-endͰѻ͏͜ͱ͕Ͱ͖Δɽ ং
ݚڀഎܠ
• σʔλͷલॲཧΛߦΘͣʹend-to-endͰͷֶशΛߦ͑Δɽ • ஞ࣍ҙΛ༻͍Δ͜ͱͰղऍੑͷߴ͍Ϟσϧʹͳ͍ͬͯΔɽ • Local interpretability: ೖྗಛͷॏཁ •
Global interpretability: ֤ಛྔ͕Ϟσϧʹରͯ͠Ͳͷ͘Β͍Өڹ͔ͨ͠ ং ఏҊख๏ͷߩݙ
• DNN+DT • ஞ࣍ҙΛ༻͍ͯɼಛબΛߦ͍ಛΛೖΕࠐΜͰ͍Δɽ • Tree-based learning • ಛબʹDNNΛ༻͍͍ͯΔɽ
• Feature Selection • ίϯύΫτͳදݱ͕Ͱ͖ͨɽ ؔ࿈ݚڀ
• Attentive transformer • ಛྔʹରͯ͠͏MaskͷֶशΛߦ͏ɽ • Feature transformer •
ಛྔͷมɼ࣍εςοϓʹ͏ͷΛܾΊΔɽ ఏҊख๏ ॏཁͳύʔπ
• ͜ΕҎ߱ग़ͯ͘Δ εςοϓ1,2,...ʹରԠ͍ͯ͠Δ i ఏҊख๏ શମͷߏ
• • : աڈͷMͰΘΕ͍ͯΔ͔ʁʹΑͬͯ มΘΔॏΈ(࣮Ͱར༻੍ݶΈ͍ͨͳͷ) • Sparsemax: softmaxʹࣅͨ׆ੑԽؔ M[i]
= sparsemax(P[i] ⋅ hi (a[i − 1])) P[i] ఏҊख๏ Attentive Transformer: ϚεΫͷֶशΛߦ͏ɽ
• SoftmaxΑΓૄʹͳΓ͍͢ ͔ΒɼॏཁͳಛྔΛऔΓग़ ͍͢͠ ίϥϜ SparseMax (Andre et al.,
2016)
• ɼa࣍ͷεςοϓʹճ͞ΕΔ [d[i], a[i]] = fi (M[i] ⋅ f)
ఏҊख๏ Feature Transformer: ೖྗΛม͠ɼ࣍ʹ͏ͷΛܾΊΔ
• ֤εςοϓ Λूܭ ͯ͠࠷ऴతͳ༧ଌʹ ༻͍Δ d[i] ఏҊख๏ ࠷ऴ༧ଌ
• ಛྔͷॏཁϚεΫΛͬͯܭࢉ͢Δ • ؆୯ʹܭࢉ͢ΔͨΊɼϚεΫͰͳ͘ಛྔΛ༻͍Δ ɹɹɹ ɹˠͲͷαϯϓϧ͕ॏཁ͔ʁ • → ಛྔͷॏཁ
ηb [i] = Nd ∑ c ReLU(db,c [i]) Magg−b,j = ∑Ns teps i=1 ηb [i]Mb,j [i] ∑D j=1 ∑Nsteps i=1 ηb [i]Mb,j [i] ఏҊख๏ ղऍੑʹ͍ͭͯ
• Feature selection͕֤εςοϓʹରԠ ఏҊख๏ ಛྔબͷΠϝʔδ
• ֤ϚεΫʹΑͬͯ࡞ΒΕΔಛྔ͕ذʹରԠ͍ͯ͠Δɽ ఏҊख๏ Ͳ͕ܾ͜ఆΆ͍ͷʁ
• ର߅ख๏: • ޯϒʔεςΟϯάܥ: LightGBM, XGBoost, CatBoost • NNϞσϧ
• ͳʹͰൺΔ͔ʁ • ςετσʔλʹର͢Δaccuracy • ϞσϧͷαΠζ ࣮ݧ ࣮ݧઃఆ
• ࣮σʔλ(ForestCoverType)Ͱର߅ख๏ΑΓਫ਼͕ྑ͔ͬͨɽ ࣮ݧ݁Ռ ਫ਼ʹؔͯ͠
• ϞσϧαΠζ͕ܰྔͰਫ਼͕͍͍ɽ ࣮ݧ݁Ռ ϞσϧαΠζʹؔͯ͠
࣮ݧ݁Ռ ղऍੑʹ͍ͭͯ • ͷ݁ՌΛՄࢹԽ • ߦ͕αϯϓϧɼྻ͕ಛྔ • ന͍ͱ͜Ζ͕ಛྔͱͯ͠ॏཁ ͱஅͨ͠ͱ͜Ζ
ηb [i]
• ஞ࣍ҙΛߦ͏͜ͱͰɼॏཁͳಛྔબΛߦͳ͍ͬͯΔɽ • ϚεΫΛ༻͍Δ͜ͱͰղऍੑͷߴ͍Ϟσϧʹͳͬͨɽ • ༷ʑͳྖҬͷςʔϒϧσʔλͰੑೳΛൃشͰ͖Δ͜ͱΛࣔ͠ ͨɽ ·ͱΊ
• Accuracy: 0.81, ROC-AUC: 0.78 ͓·͚ TitanicσʔληοτͰTabNetΛ༡ΜͰΈͨɽ
͓·͚ LightGBM vs NN model vs TabNet LightGBM NN
model • TabNet: Accuracy: 0.81, ROC-AUC: 0.78 https://github.com/mei28/playground_python/blob/main/notebooks/titanic.ipynb ϋΠύϥॳظͷ··Ͱ νϡʔχϯάΛߦͳ͍ͬͯͳ͍