AAAI・WSDM 2017論文読み会

Slide 1

Slide 1 text

"""*ɾ84%.࿦จಡΈձ "$PODJTF*OUFHFS-JOFBS1SPHSBNNJOH'PSNVMBUJPO GPS*NQMJDJU4FBSDI3FTVMU%JWFSTJpDBUJPO Recruit Communications Co., Ltd. Kenji Hayashida

Slide 2

Slide 2 text

4&-'*/530%6$5*0/ ̶Recruit Communications Co., Ltd.̶ ໊લྛాݡೋ ৬ྺ֎ࢿܥ4*FS3$0ΤϯδχΞ ۀ຿಺༰ΞυςΫͷ࠷దԽɺίϯςϯπϚʔέɺσʔλύΠϓϥΠϯ झຯϓϩίϯ

Slide 3

Slide 3 text

84%. ̶Recruit Communications Co., Ltd.̶

Slide 4

Slide 4 text

84%. ̶Recruit Communications Co., Ltd.̶ ৄࡉ͸3$0ΞυςΫ-BCϒϩάͰʂ ʮXTENSDPϒϩάʯͰݕࡧ

Slide 5

Slide 5 text

*/530%6$5*0/ Ϣʔβ͕ٻΊ͍ͯΔ৘ใΛఏڙ͢Δͷ͸೉͍͠ ̶Recruit Communications Co., Ltd.̶ )BSSZ1PUUFS ຊ өը ϨϏϡʔ ΩϟϥΫλʔ ʜ

Slide 6

Slide 6 text

*/530%6$5*0/ ̶Recruit Communications Co., Ltd.̶ 43% 4FBSDI3FTVMU%JWFSTJpDBUJPO Ϣʔβ͕ཉ͍͠৘ใΛΧόʔ͢ΔΑ͏ʹଟ༷ͳ݁ՌηοτΛฦ͢ FYQMJDJU43%ɹΫΤϦʹજࡏ͢ΔαϒτϐοΫ͕Θ͔͍ͬͯΔ৔߹ JNQMJDJU43%ɹΫΤϦʹજࡏ͢ΔαϒτϐοΫ͕Θ͔͍ͬͯͳ͍৔߹ ଟ͘ͷ৔߹ɺજࡏαϒτϐοΫ͸Θ͔Βͳ͍ ࠓճ঺հ͢Δ࿦จ͸ɺJNQMJDJU43%ʹϑΥʔΧε

Slide 7

Slide 7 text

*/530%6$5*0/ ̶Recruit Communications Co., Ltd.̶ .BYJNBM.BSHJOBM3FMFWBODF ..3 ..3͸࠷΋୯७ͳख๏ɻ ଞʹ΋༷ʑͳख๏͕ଘࡏ͠ɺओʹҎԼͷ఺͕ҟͳΔɻ ଟ༷ੑΛͲͷΑ͏ʹදݱ͢Δ͔ʁ ద߹ੑͱଟ༷ੑΛͲͷΑ͏ʹόϥϯε͢Δ͔ʁ ݁ՌϦετΛͲͷΑ͏ʹੜ੒͢Δ͔ʁ MMR = arg max Di 2R\S  Sim1(Di, Q) (1 ) max Dj 2S Sim2(Di, Dj) ҎԼͷج४ͰυΩϡϝϯτΛHSFFEZʹબ୒͢Δ ʢ$BSCPOFMMBOE(PMETUFJO ʣ

Slide 8

Slide 8 text

3&-"5&%803, ̶Recruit Communications Co., Ltd.̶ "⒏OJUZ1SPQFSHBUJPOʢ"1ʣ ඪຊʢ୅දతͳΞΠςϜʣΛબͿ͜ͱͰΫϥελϦϯάΛߦ͏ख๏ ը૾෼ྨɺը૾ηάϝϯςʔγϣϯͳͲͰར༻͞ΕΔ U = Uij = ΞΠςϜJͱΞΠςϜKͷྨࣅ౓ ΞΠςϜJ͕ඪຊʹબ͹ΕΔ֬཰ (i = j) (i 6= j) \ ඇඪຊσʔλͱͦΕ͕ଐ͢ΔΫϥελͷඪຊσʔλͷྨࣅ౓ͷ࿨Λ࠷େԽ

Slide 9

Slide 9 text

3&-"5&%803, ̶Recruit Communications Co., Ltd.̶ %FTJSBCMF'BDJMJUZ1MBDFNFOU %'1 ࢪઃ഑ஔ໰୊ΛԠ༻ͨ͠Ϟσϧʢ;VDDPOFUBM ʣ ϝλώϡʔϦεςΟοΫͳΞϓϩʔνʢIJMMDMJNCJOHʣͰܭࢉ ຊݚڀͷCBTFMJOFϞσϧͷͭ

Slide 10

Slide 10 text

130104&%.&5)0% ̶Recruit Communications Co., Ltd.̶ *-1*%ʢJOUFHFSMJOFBSQSPHSBNNJOHBQQSPBDIGPSJNQMJDJU43%ʣ

Slide 11

Slide 11 text

130104&%.&5)0% ̶Recruit Communications Co., Ltd.̶ *-1*% YJJυΩϡϝϯτEJ͕ඪຊʹબ͹Ε͔ͨʁ YJK J㱠KυΩϡϝϯτEJ͕EKΛඪຊʹબΜ͔ͩʁ

Slide 12

Slide 12 text

130104&%.&5)0% ̶Recruit Communications Co., Ltd.̶ *-1*% ద߹ੑΛ࠷େԽ

Slide 13

Slide 13 text

130104&%.&5)0% ̶Recruit Communications Co., Ltd.̶ *-1*% ୅දੑΛ࠷େԽ

Slide 14

Slide 14 text

130104&%.&5)0% ̶Recruit Communications Co., Ltd.̶ *-1*% NυΩϡϝϯταΠζͱLඪຊαΠζͷେ͖͞Λௐ੔ʢNLʣ

Slide 15

Slide 15 text

130104&%.&5)0% ̶Recruit Communications Co., Ltd.̶ *-1*% ඪຊʹબ͹ΕΔυΩϡϝϯτ͸Lݸ

Slide 16

Slide 16 text

130104&%.&5)0% ̶Recruit Communications Co., Ltd.̶ *-1*% υΩϡϝϯτEJͷඪຊ͸͚ͭͩ

Slide 17

Slide 17 text

130104&%.&5)0% ̶Recruit Communications Co., Ltd.̶ *-1*% EJ͕EKΛඪຊʹબΜͩͳΒɺEK͸ඪຊυΩϡϝϯτͰ͋Δ

Slide 18

Slide 18 text

130104&%.&5)0% ̶Recruit Communications Co., Ltd.̶ %'1͸*-1*%ͷಛघέʔεͱߟ͑Δ͜ͱ͕Ͱ͖Δ *-1*%Ͱදݱ͢Δ͜ͱͰ࠷ద஋ΛܭࢉՄೳ ͦͷଞͷ*NQMJDJU43%ͷख๏ʢ..3 .51ͳͲʣ΋ಉ༷ʹ*-1*%ͰදݱՄೳ *-1*% %'1

Slide 19

Slide 19 text

&91&3*.&/54 ̶Recruit Communications Co., Ltd.̶ 53&$8FC 5SBDL #. 43% *-1*% %'1 53&$8FC5SBDL ΫΤϦ GBDFUFERVFSZෳ਺ͷҟͳΔαϒτϐοΫΛ࣋ͭΫΤϦ BNCJHVPVTRVFSZෳ਺ͷ·ͬͨ͘ҟͳΔղऍΛ࣋ͭΫΤϦ JOJUJBMSFUSJFWBM DIPPTF999EPDVNFOUT DIPPTFEPDVNFOUT

Slide 20

Slide 20 text

&91&3*.&/54 ̶Recruit Communications Co., Ltd.̶ ͢΂ͯͷΫΤϦʹରͯ͠ɺ%*-1*%㱢%%'1 %'1Ͱ͸࠷ద஋͕͑ΒΕͯͳ͍͜ͱΛ࣮ূ υΩϡϝϯταΠζ͕૿͑Δ΄Ͳࠩ͸ݦஶ ࠷దੑͷݕূ ͓ࣜΑͼ ࣜͷЕɺ ͔ࣜΒLͷ܎਺Λআ͘ʢଟ༷ੑʹϑΥʔΧεʣ ಉ͡ܗͷ࠷దԽ໰୊ʹͯ͠ɺ*-1*%ͱ%'1ͷ࠷ద஋Λൺֱ ˞3&'&3&/$&4<>ΑΓҾ༻

Slide 21

Slide 21 text

&91&3*.&/54 ̶Recruit Communications Co., Ltd.̶ ֤Ϟσϧͷੑೳൺֱ *-1*%͸ଞͷϕʔεϥΠϯख๏ͱൺ΂ͯੑೳ͕Α͍ ଟ͘ͷվળ͸౷ܭతʹ༗ҙ ˞3&'&3&/$&4<>ΑΓҾ༻

Slide 22

Slide 22 text

&91&3*.&/54 ̶Recruit Communications Co., Ltd.̶ ֤Ϟσϧͷੑೳൺֱ *-1*%͸ଞͷϕʔεϥΠϯख๏ͱൺ΂ͯੑೳ͕Α͍ ଟ͘ͷվળ͸౷ܭతʹ༗ҙ ˞3&'&3&/$&4<>ΑΓҾ༻

Slide 23

Slide 23 text

&91&3*.&/54 ̶Recruit Communications Co., Ltd.̶ ܭࢉ଎౓ Ұൠʹ੔਺ઢܗܭը໰୊͸/1ࠔ೉Ͱ͋ Δ ༷ʑͳ෼ࢬݶఆ๏ͷΞϧΰϦζϜ͕ఏҊ ͞Ε͍ͯΔ ࠷৽ͷιϧόʔΛ࢖͑͹े෼ߴ଎ʹܭࢉ Մೳ ੔਺ઢܗܭըʹ͍ͭͯ ಉ༷ͷఆࣜԽͰ͋Δ͕ɺܭࢉίετ͸ *-1*%ͷํ͕%'1ΑΓߴ͍ *-1*%Ͱ͸෼ࢬݶఆ๏Ͱ࠷ద஋ΛٻΊ ͍ͯΔ͜ͱΛߟྀ͢Ε͹ೲಘͰ͖Δ ܭࢉ଎౓ͷൺֱ ˞3&'&3&/$&4<>ΑΓҾ༻

Slide 24

Slide 24 text

$0/$-64*0/4"/%'6563&5"4, ̶Recruit Communications Co., Ltd.̶ JNQMJDJU43%ͷ໰୊Λղͨ͘Ίͷ৽͍͠ϞσϧΛఏҊ NݸͷυΩϡϝϯτ͔ΒLݸͷඪຊΛબͿॲཧͱͯ͠ఆࣜԽ ੔਺ઢܗܭը໰୊ͱͯ͠ఆࣜԽ͢Δ͜ͱͰɺ໨తؔ਺ͷ࠷ద஋Λ ܭࢉՄೳ JNQMJDJU43%ͱಉ༷ͷΞϓϦέʔγϣϯʢྫϨίϝϯμʔγες Ϝʣʹ΋৽͍͠ࢹ఺Λఏڙ ύϥϝʔλLͷνϡʔχϯάʹ͍ͭͯௐ͕ࠪඞཁʢΫΤϦʹΑͬͯҟ ͳΔ͸ͣʣ Ϋϥελ಺ͷυΩϡϝϯτͷ૬ؔɺΫϥελಉ࢜ͷ૬ؔͳͲΛߟྀ ͨ͠ϞσϦϯάͷվྑ

Slide 25

Slide 25 text

"11&/%*953&$8FC5SBDL ̶Recruit Communications Co., Ltd.̶ 53&$ʢ5FYU3&USJFWBM$POGFSFODFʣͷίϯϖ<> 8FCݕࡧٕज़ͷ୳ٻͱධՁΛߦ͏ σʔληοτͷઆ໌<> BNCJHVPVTRVFSZෳ਺ͷશ͘ผͷղऍ͕Ͱ͖ΔΫΤϦ ྫUIFDVSSFOU 5IF$VSSFOUͱ͍͏ϥδΦ൪૊ͷϖʔδʁ χϡʔδϟʔδʔͷ5IF$VSSFOUOFXTQBQFSʁ αϯΞϯτχΦͷ5IF$VSSFOU.BHB[JOFʁ GBDFUFERVFSZෳ਺ͷαϒτϐοΫΛ࣋ͭෆ໌ྎͳΫΤϦ ྫWPMWP 7PMWPͷϗʔϜϖʔδʁ 7PMWP9$467ͷϨϏϡʔʁ 7PMWPͷσΟʔϥʔʁ

Slide 26

Slide 26 text

"11&/%*9#. ̶Recruit Communications Co., Ltd.̶ score ( D, Q ) = n X i=1 IDF( qi) f ( qi, D )( k1 + 1) f ( qi, D ) + k1(1 b + b |D| avgdl ) f(qi, D) avgdl IDF ( qi) = log N n ( qi) + 0 . 5 n ( qi) + 0 . 5 ୯ޠRJͷυΩϡϝϯτ%಺ͷUFSNGSFRVFODZ υΩϡϝϯτͷฏۉ୯ޠ਺ N n(qi) υΩϡϝϯτ਺ ୯ޠRJΛؚΉυΩϡϝϯτ਺

Slide 27

Slide 27 text

"11&/%*9.15 ̶Recruit Communications Co., Ltd.̶ .PSEFSO1PSUGPMJP5IFSPZ ద߹ੑͷظ଴஋ͱ෼ࢄ͓Αͼ͢Ͱʹબ୒͞ΕͨυΩϡϝϯτͱͷ૬ ؔΛߟྀͨ͠Ϟσϧ ҎԼͷ஋͕ߴ͍΋ͷΛHSFFEZʹબ୒͍ͯ͘͠ E(dk) w ⇢ik k υΩϡϝϯτͷظ଴ద߹ੑ υΩϡϝϯτͷద߹ੑͷඪ४ภࠩ ϥϯΫݻ༗ͷॏΈ υΩϡϝϯτEJͱELͷ૬ؔ܎਺

Slide 28

Slide 28 text

"11&/%*9FYQFDUFEDBMM!L ̶Recruit Communications Co., Ltd.̶ ΞΠςϜTJͷTVCUPQJD෼෍ ΫΤϦͷTVCUPQJD෼෍ ద߹ੑʢద߹ɺෆద߹ʣ ҎԼͷ஋͕ߴ͍΋ͷΛHSFFEZʹબ୒͍ͯ͘͠ ˞3&'&3&/$&4<>ΑΓҾ༻

Slide 29

Slide 29 text

"11&/%*9O&33*" ̶Recruit Communications Co., Ltd.̶ &YQFDUFE3FDJQSPDBM3BOL &33 ద߹Ϩϕϧ ήΠϯ *OUFOU"XBSF&YQFDUFE3FDJQSPDBM3BOL &33*" ΫΤϦͷҙਤΛ࢖ͬͯظ଴஋Λܭࢉ SFDJQSPDBMSBOL S൪໨ͰॳΊͯҙਤJʹద߹͢ ΔυΩϡϝϯτ͕͑ΒΕͨ

Slide 30

Slide 30 text

"11&/%*9ЋO%$( ̶Recruit Communications Co., Ltd.̶ OPSNBMJTFE%JTDPVOUFE$VNVMBUJWF(BJO O%$( g(r) g⇤(r) ϥϯΫSͷυΩϡϝϯτͷήΠϯʢద߹ϨϕϧʹԠܾͯ͡·Δʣ ཧ૝తʹϥϯΩϯά͞Εͨͱ͖ͷϥϯΫSͷυΩϡϝϯτͷήΠϯ ద߹Ϩϕϧ ήΠϯ ЋO%$( ήΠϯͷ୅ΘΓʹOPWFMUZCJBTFEήΠϯΛ༻͍Δ ϥϯΫSͷυΩϡϝϯτ͸ҙਤJʹద߹͢Δ ϥϯΫSͷυΩϡϝϯτ͸ҙਤJʹద߹͠ͳ͍ ͨͩ͠ ΫΤϦ͕࣋ͭҙਤͷू߹ʹରͯ͠ద߹ੑΛܭࢉ ↵ ύϥϝʔλʢ͕Α͘࢖ΘΕΔʣ

Slide 31

Slide 31 text

3&'&3&/$&4 ̶Recruit Communications Co., Ltd.̶ <>l"$PODJTF*OUFHFS-JOFBS1SPHSBNNJOH'PSNVMBUJPOGPS*NQMJDJU4FBSDI 3FTVMU%JWFSTJpDBUJPOz )BJUBP:V "EBN+BUPXU 3PJ#MBODP )JEFP+PIP +PFNPO+PTF -POH$IFO 'BKJF:VBO <>5FYU3&USJFWBM$POGFSFODF 53&$ )PNF1BHF <>l0WFSWJFXPGUIF53&$8FC5SBDLz $IBSMFT-"$MBSLF /JDL $SBTXFMM *BO4PCPSPG <>l)PX*OUVJUJWF"SF%JWFSTJpFE4FBSDI.FUSJDT $PODPSEBODF5FTU3FTVMUT GPSUIF%JWFSTJUZ6.FBTVSFTz 5FUTVZB4BLBJ <>%JWFSTFSFUSJFWBMWJBHSFFEZPQUJNJ[BUJPOPGFYQFDUFEDBMM!LJOBMBUFOU TVCUPQJDSFMFWBODFNPEFM 44BOOFS 4(VP 5(SBFQFM 4,IBSB[NJ 4 ,BSJNJ