間取り図を用いた賃料予測モデルに関する一検討 @ 2019人工知能全国大会

本研究では，間取り図を用いた賃料の線形回帰モデルを構築し，間取り図を用いない線形回帰モデルとの予測精度の比較を行っている．東京都の賃貸物件9 万件を利用した賃料予測実験より，間取り図にPCA を適用した線形回帰モデルが間取り図を利用しない線形回帰モデルよりも予測精度（RMSE）が高い傾向にあり，間取り図の利用が賃料の予測精度の向上に寄与していると考える．加えて，間取り図からの特徴抽出にはPCAがVGG16 ベースのニューラルネットワークよりもRMSE が高くなることも確認している．

Transcript

1. / 20 JSAI 2019 2019.6.4 ؒऔΓਤΛ༻͍ͨ௞ྉ༧ଌϞσϧʹؔ͢ΔҰݕ౼ !1 ෰෦ ྇య¹ɹԬຊ Ұࢤ¹ɹࣲా ३࢘² ¹ ిؾ௨৴େֶ େֶӃ৘ใཧ޻ֶݚڀՊ ৘ใֶઐ߈ɹ ² ࢈ۀٕज़େֶӃେֶ ࢈ۀٕज़ݚڀՊ ৘ใΞʔΩςΫνϟઐ߈
2. / 20 JSAI 2019 2019.6.4 ௞ି෺݅ͷಛ௃ • ಉ͡෺͕݅ଘࡏ͠ͳ͍ • ؒऔΓɼ֊਺ɼཱ஍ͳͲͷ෺݅ม਺͕௞ྉʹӨڹΛ༩͑Δ ௞ି෺݅ͷՁܾ֨ఆํ๏ [େ໺تɼ2016] • ੵࢉ๏ • ऩӹ෼ੳ๏ • ௞ିࣄྫൺֱ๏ • ର৅෺݅ͷपғͷ෺݅΍ྨࣅ͢Δ෺݅Λࢀরͯ͠௞ྉΛܾఆ • ܾఆࢧԉʹϔυχοΫɾΞϓϩʔνΛ࠾༻ ͸͡Ίʹ !2 େ໺تٱ೭ีɿܧଓ௞ྉؑఆධՁΛ࠶ߟ͢Δɼ ౎ࢢॅ୐ֶɼॅ୐৽ใࣾɼ2016.
3. / 20 JSAI 2019 2019.6.4 ϔυχοΫɾΞϓϩʔν ͋Δ঎඼ͷՁ֨Λͦͷ঎඼ͷม਺ͷՁ஋ʹؔ͢Δू߹ͱ͠ɼ ͦͷ঎඼Ձ֨ͷ༧ଌϞσϧΛઢܗճؼͰߏங͢Δٕज़[౜౉ɼ2016] ɹ طଘͷϔυχοΫɾΞϓϩʔν͸ؒऔΓਤ͕ߟྀ͞Ε͍ͯͳ͍ !3 ˆ yi = ↵1x1 + ↵2x2 + ↵3x3 + · · · + ↵nxn + <latexit sha1_base64="vQx08F43bZHjET/G6F0shHJHj2A=">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</latexit> <latexit sha1_base64="vQx08F43bZHjET/G6F0shHJHj2A=">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</latexit> <latexit sha1_base64="Mhn41z7d/Q2n10aqYFv7sL5P/oA=">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</latexit> ઎༗໘ੵ ཱ஍ ֊਺ ங೥਺ ௞ྉ ౜౉ ޿ࢤɿϔυχοΫɾΞϓϩʔνΛར༻ͨ͠ෆಈ࢈Ձ֨ ࢦ਺ͷਪఆํ๏ͱͦͷ໰୊఺ɼ౎ࢢॅ୐ֶɼvol.2016ɼno. 92ɼpp.17-20ɼ 2016.
4. / 20 JSAI 2019 2019.6.4 ϔυχοΫɾΞϓϩʔνΛ༻͍ͨؔ࿈ݚڀ !4 • Ram P. Dahal and Robert K. Grala and Jason S. Gordon and Ian A. Munn and Daniel R. Petrolia and J. Reid Cummingsɿ A hedonic pricing method to estimate the value of waterfronts in the Gulf of MexicoɼUrban Forestry & Urban Greeningɼvol.41ɼpp. 184-194ɼ2019. • Dr. Philipp Deschermeier and Björn Seipeltɿ A Hedonic Rent Index for Student Housing in GermanyɼCologne Institute for economic researchɼpp.1–12ɼ2016ɽ • yung-Jin Jun, Hee-Jae KimɿMeasuring the effect of greenbelt proximity on apartment rents in SeoulɼCitiesɼvol. 62, 2017ɽ ݐ෺ߏ଄ ཱ஍ /ΞΫηε पล؀ڥ ചങܖ໿ ը૾ Ram+ 2019 ◦ ◦ ◦ × × Philipp+ 2016 ◦ ◦ ◦ × × Jun+ 2017 ◦ ◦ ◦ ◦ × ຊݚڀ ◦ ◦ × ◦ ◦ આ໌ม਺ͱͯ͠࢖༻͞Ε͍ͯΔม਺ͷΧςΰϦ
5. / 20 JSAI 2019 2019.6.4 ؒऔΓਤͷಛੑ ؒऔΓن͕֨ಉ͡Ͱ͋ͬͯ΋ɼ෦԰ͷܗঢ়͕ҧ͏΋ͷ͕ଘࡏ ر๬ͷ௞ି෺݅Λ୳͢ࡍʹ͸ؒऔΓਤΛݟΔश׳͕͋Δ[ਗ਼ా+ɼ2017] !5 ਗ਼ాཅ࢘ɼࢁ࡚ढ़඙ɼਡ๚ത඙ɼਗ਼ਫ߂ɿɹ ΤσΟτϦΞϧɿʮෆಈ࢈ͱAI ʯɼਓ޻஌ೳɼ vol.32ɼ2017ɽ ಉ͡ݐ෺ͷؒऔΓن֨KͰؒऔΓਤ͕ҧ͏ྫ
6. / 20 JSAI 2019 2019.6.4 ໨త: ௞ྉ༧ଌʹ͓͚ΔؒऔΓਤͷӨڹΛ໌Β͔ʹ͢Δ खஈ: ௞ྉͷ༧ଌޡࠩΛൺֱ • ؒऔΓਤΛߟྀ͠ͳ͍༧ଌϞσϧʢLR) • ؒऔΓਤΛߟྀͨ͠༧ଌϞσϧ • ઢܗม׵ʹΑΔಛ௃ྔநग़๏ʢPCA-LRʣ • ඇઢܗม׵ʹΑΔಛ௃ྔநग़๏ʢVGG-LR) ݚڀ಺༰ !6
7. / 20 JSAI 2019 2019.6.4 ҰൠతͳϔυχοΫɾΞϓϩʔν • ༧ଌࣜɿ • આ໌ม਺ʹؒऔΓਤΛؚ·ͳ͍ • scikit-learnʢversionɿ0.20.2ʣͷLinearRegressionΫϥε࢖༻ LRʢLinear-Regressionʣ !7 ˆ yi = f(xi) = ↵Txi + <latexit sha1_base64="gAXdu8udRiXbYQWOcxxODvbMMjU=">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</latexit> <latexit sha1_base64="gAXdu8udRiXbYQWOcxxODvbMMjU=">AAACwnichVHLShxBFD12otHxNcaN4KbJYDAIUhMCCYIgJossfY0K9qSpbmtmSqsf6a4ZMun0D+QHssgqQgjiZ7jJD2ThJ0h2GcGNi9zpadCo0dt03VOn7rl1qsoJlYw1Yyd9xoOH/QOPBocKwyOjY+PFicebcdCMXFFxAxVE2w6PhZK+qGipldgOI8E9R4ktZ/91d32rJaJYBv6Gboei6vG6L2vS5Zoou1i1GlybbVuai2ZtNrGcQO3GbY+S+SG1E5k+MxevkInFVdjg6bvE8rhuRF6ykaa3qOZMyxGa28USm2dZmDdBOQcl5LESFH/Awi4CuGjCg4APTViBI6ZvB2UwhMRVkRAXEZLZukCKAmmbVCWoghO7T2OdZjs569O82zPO1C7touiPSGlihv1ih6zDfrIjdsou/tsryXp0vbQpOz2tCO3xz1Pr5/eqPMoajUvVnZ41aniVeZXkPcyY7incnr718UtnfWFtJnnKDthv8v+NnbBjOoHfOnO/r4q1ryjQA5SvX/dNsPl8vkx49UVpaTl/ikFM4wlm6b5fYglvsYIK7XuMU/xBx3hj7BnvjbhXavTlmkn8E8anv2ZjsFI=</latexit> <latexit sha1_base64="gAXdu8udRiXbYQWOcxxODvbMMjU=">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</latexit> ͜ͷLRͱؒऔΓਤΛߟྀͨ͠PCA-LRͱVGG-LRΛൺֱ͢Δ f(xi) xi ˆ yi ม਺ ϕΫτϧ ௞ྉ
8. / 20 JSAI 2019 2019.6.4 • ओ੒෼෼ੳʹΑΔಛ௃ྔநग़ • நग़ͨؒ͠औΓਤͷಛ௃ྔͱɹɹɹɹɹɹɹɹɹɹɹɹɹ ม਺ϕΫτϧΛ૊Έ߹ΘͤɼઢܗճؼͰ௞ྉ༧ଌ • ؒऔΓਤͷಛ௃ྔ: 64ɼ128ɼ256ɼ512ɼ1024ɼ 2048࣍ݩͷ6छྨ PCAʢPrincipal Component Analysisʣ-LR !8 • scikit-learnʢversionɿ0.20.2ʣ ͷPCAΫϥε࢖༻ • LRͷઆ໌ม਺ʹؒऔΓਤΛ௥Ճͨ͠Ϟσϧ ಛ ௃ ྔ ओ੒෼ ෼ੳ vi ଐੑ ϕΫτϧ φθ(vi) ui xi ˆ yi ௞ྉ f(xi) x = [u, ✓(v)] <latexit sha1_base64="tbvrKtiF/xOXDGMbwaD2COlchy8=">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</latexit> ม਺
   ϕΫτϧ
9. / 20 JSAI 2019 2019.6.4 VGG-LR !9 • ϛχόονʴ֬཰తޯ഑߱Լ๏Λద༻ • batch size: 100ɼepoch: 100 • ؒऔΓਤͷಛ௃ྔநग़: 64࣍ݩ • ׆ੑԽؔ਺: ReLU • Kerasʢversionɿ2.2.4ʣͷFunction API • ؒऔΓਤΛඇઢม׵Ͱಛ௃நग़͢ΔϞσϧ x = [u, h✓(v)] <latexit sha1_base64="o9TUYv/8nCTnfgkDaEu18FszFcY=">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</latexit> 7(( vi ˆ yi ଐੑ ϕΫτϧ ui ௞ྉ xi f(xi) f(xi) ಛ ௃ ྔ hθ(vi) ม਺ ϕΫτϧ
10. / 20 JSAI 2019 2019.6.4 VGG16ͷߏ଄ !10 Karen Simonyanɼand Andrew Zissermanɿ Very Deep Convolutional Networks for Large-Scale Image RecognitionɼComputing Research Repositoryɼvol.abs/1409.1556ɼ2014. VGG16 [Karen+ɼ2014]: ΦοΫεϑΥʔυେֶͷVGGνʔϜʹΑͬͯ։ൃ͞Εͨ ֶशࡁΈͷχϡʔϥϧωοτϫʔΫ 7 × 7 × 512 13 224 × 224 × 3 1 × 1000 1 × 4096 14 × 14 × 512 28 × 28 × 512 ɾ ɾ ɾ ૚ 224 × 224 × 3 28 × 28 × 512 11 1 × 64 ɾ ɾ ɾ ૚ ࢖༻͢ΔVGG16 • ޙΖͷ5૚Λআڈ • 64࣍ݩͷશ݁߹૚௥Ճ • ޙΖ3૚ͷΈֶश
11. / 20 JSAI 2019 2019.6.4 ࣮ݧ؀ڥ LIFULL HOME’Sσʔληοτʢ2015೥9݄࣌఺ʣ • ௞ି෺݅σʔλʢ70ม਺ɼ533ສ݅ʣ • 120×120ϐΫηϧͷը૾σʔλʢ8,300ສϑΝΠϧʣ ࢖༻͢ΔܭࢉػʢOS: Ubuntu 18.04.̍ʣ • CPUɿXeon(R) CPU E5-2650 v3 @ 2.30GHz • GPUɿGeForce GTX 1080 Ti 11GB • ϝϞϦɿ64GB !11
12. / 20 JSAI 2019 2019.6.4 ెาڑ཭ɼཱ஍ɼ֊਺ʢ஍্ʣɼ֊਺ʢ஍Լʣɼࢢொଜ۠ɼ෦԰֊਺ɼ ܖ໿ظؒɼݐ෺ߏ଄ɼறं৔ྉۚɼ৽ஙɾະೖډϑϥάɼ઎༗໘ੵɼɹ ங೥਺ɼؒऔΓਤ ࢖༻௞ି෺݅σʔλʢ14ม਺ɼ9ສ݅ʣ • ౦ژ౎ͷσʔλΛ࢖༻ʢ㲎 47౎ಓ෎ݝͰ࠷΋σʔλྔ͕ଟ͍ͨΊʣ • ໨తม਺ɿ௞ྉ + ڞӹඅ • આ໌ม਺ɿ • ؒऔΓਤɿάϨʔεέʔϧม׵ σʔληοτ෼ׂ • ࢖༻௞ି෺݅σʔλΛؒऔΓن֨Λج४ʹKɼRɼDKɼLDɼLDKͷ5छྨʹ෼ׂ • LD͸σʔλ਺͕35݅ͷͨΊ࢖༻͠ͳ͍ • 20%Λςετσʔλʹɼ80%Λ։ൃ༻σʔλʹ෼ׂ σʔληοτ !12
13. / 20 JSAI 2019 2019.6.4 ࣮ݧ֓ཁ !13 1. ϋΠύʔύϥϝʔλνϡʔχϯά • PCA-LRʹ͓͚Δద੾ͳؒऔΓਤͷ࣍ݩ਺Λ୳ࡧ • 64ɼ128ɼ256ɼ512ɼ1024ɼ2048࣍ݩͷ6छྨ • ։ൃ༻σʔλʹ10-fold cross-validationΛద༻ • ධՁࢦඪɿฏํฏۉೋ৐ޡࠩʢRMSE: Root Mean Square Errorʣ 2. ςετσʔλʹର͢ΔLRͱVGG-LRɼPCA-LRͷධՁ 1. RMSEʹΑΔධՁͱֶश࣌ؒ 2. ༧ଌ஋ͷՄࢹԽ
14. 1. ϋΠύʔύϥϝʔλνϡʔχϯά DK LDK K R
15. / 20 JSAI 2019 2019.6.4 2-1ɽRMSEʹΑΔධՁͱֶश࣌ؒ !15 σʔλ਺ LR PCA-LR VGG-LR K 1,712 11,602 11,689 16,862 R 7,525 9,217 8,727 12,541 DK 3,548 11,156 10,957 16,467 LDK 6,842 20,121 19,646 22,878 σʔλ਺ LR PCA-LR VGG-LR K 9,205 0.06 11 2,897 R 39,985 0.08 48 14,566 DK 18,975 0.05 23 6,471 LDK 36,595 0.07 45 13,026 ςετσʔλʹର͢ΔRMSE[ԁ] ։ൃ༻σʔλͷֶश࣌ؒ[ඵ]
16. LR PCA-LR VGG-LR ؒऔΓن֨R LR PCA-LR VGG-LR ؒऔΓن֨K 2-2ɽ༧ଌ஋ͷՄࢹԽ
17. / 20 JSAI 2019 2019.6.4 2-2ɽྦྷੵີ౓ؔ਺ʢ1/2ʣ !17 K R = |1 − ̂ y y | ԣ࣠ɿerror rate ॎ࣠ɿerror rate ҎԼͷޡࠩʹऩ·͍ͬͯΔσʔλ਺ͷׂ߹ y : ਅ஋ ̂ y :༧ଌ஋
18. / 20 JSAI 2019 2019.6.4 DK LDK !18 2-2ɽྦྷੵີ౓ؔ਺ʢ2/2ʣ
19. / 20 JSAI 2019 2019.6.4 LIFULL HOME’Sσʔληοτ֎ͷ෺݅༧ଌ !19 ؒऔΓਤ ߏ଄ ؒऔΓ ن֨ ௞ྉ LR PCA-LR VGG-LR ࣗ୐ RC K 65,000 61,361 62,545 83,695 ֶੜྈ RC K 47,700 56,931 49,716 64,526 A୐ ໦଄ K 74,000 66,695 68,701 58,561 B୐ మے LDK 128,000 130,988 122,887 181,913
20. / 20 JSAI 2019 2019.6.4 ͓ΘΓʹ ໨తɿ௞ྉ༧ଌʹ͓͚ΔؒऔΓਤͷӨڹΛ໌Β͔ʹ͢Δ ࣮ݧ݁Ռ ؒऔΓਤΛߟྀ͢Δ͜ͱͰ༧ଌޡ͕ࠩվળ͢Δ • ಛ௃ྔநग़ʹओ੒෼෼ੳΛ༻͍Δ͜ͱͰRMSE͕վળ͢Δ • ؒऔΓਤͷಛ௃ྔ͸1024ɼ2048࣍ݩͷߴ࣍ݩ͕๬·͍͠ • 1024ɼ2048࣍ݩͷؒऔΓਤͷಛ௃ྔͰؒऔΓਤͷେ෦෼ͷಛ௃ྔΛɹ நग़Ͱ͖Δ • ಛ௃ྔநग़ʹVGGΛ༻͍Δ͜ͱͰɼerror rate͕0.1ҎԼͷ෺݅਺͕૿͑Δ • ༧ଌޡࠩͷ͹Β͖͕ͭ՝୊ • ܭࢉίετ͕՝୊ !20
21. / 20 JSAI 2019 2019.6.4 PCAʹΑΔ࣍ݩผͷؒऔΓਤͷಛ௃ྔ !21 ݕূσʔλʢLDKʣʹର͢Δ֤࣍ݩ਺͔Βͷ෮ݩਤ ࣍ݩ ࣍ݩ ࣍ݩ ࣍ݩ ࣍ݩ ࣍ݩ ೖྗը૾ 1024࣍ݩͷಛ௃ྔͰݩͷը૾ΛೝࣝͰ͖Δఔ౓ʹ෮ݩͰ͖͍ͯΔ
22. / 20 JSAI 2019 2019.6.4 LDKͷςετσʔλʹର͢Δ༧ଌ஋ͷՄࢹԽ !22 LR PCA-LR VGG-LR
23. / 20 JSAI 2019 2019.6.4 ؒऔΓਤ ௞ྉ ݐ෺ߏ଄ ங೥਺ ෦԰֊਺ ཱ஍ ࣄྫ̍ 146,000 ໦଄ 13 1 ࿅അ۠ ࣄྫ̎ 140,000 ໦଄ 8 1 ੈా୩۠ ࣄྫ̏ 237,000 SRC 17 4 ઍ୅ా۠ !23 ςετσʔλ਺ͷ31.6%͕VGG16ͷΈͰߴ͍༧ଌਫ਼౓Λ͍ࣔͯ͠Δ LDKʹ͓͚ΔVGG16ͷΈͰ༧ଌਫ਼౓͕ߴ͍ ෺݅ࣄྫ
24. / 20 JSAI 2019 2019.6.4 ؒऔΓਤ ௞ྉ ݐ෺ߏ଄ ங೥਺ ෦԰֊਺ ཱ஍ ࣄྫ̍ 236,000 SRC 4 10 ৽॓۠ ࣄྫ̎ 120,000 ܰྔమࠎ 4 1 ෎தࢢ ࣄྫ̏ 128,000 ܰྔమࠎ 0 2 খۚҪࢢ !24 ςετσʔλ਺ͷ10.1%͕PCA-LRͷΈͰߴ͍༧ଌਫ਼౓Λ͍ࣔͯ͠Δ LDKʹ͓͚ΔPCA-LRͷΈͰ༧ଌਫ਼౓͕ߴ͍ ෺݅ࣄྫ
25. / 20 JSAI 2019 2019.6.4 ؒऔΓਤ ௞ྉ ݐ෺ߏ଄ ங೥਺ ෦԰֊਺ ཱ஍ ࣄྫ̍ 114,000 RC 20 2 ߐށ઒۠ ࣄྫ̎ 171,000 RC 0 2 ඼઒۠ ࣄྫ̏ 82,000 ໦଄ 28 2 ࡾୋࢢ !25 ςετσʔλ਺ͷ36.9%͕྆ํͷϞσϧͰ௿͍༧ଌਫ਼౓Λ͍ࣔͯ͠Δ LDKʹ͓͚Δ྆ํͷϞσϧͰ༧ଌਫ਼౓͕௿͍ ෺݅ࣄྫ