Billion-scale Embedding for E-commerce Recommendation in Alibaba

BQQFBSFEJO,%%`

#BDLHSPVOE 3FDPNNFOEFS4ZTUFNJO5BPCBP 1 billion users x 2 billion items

#BDLHSPVOE 5XPTUBHFSFDPNNFOEBUJPO UPEBZ`TUBML

#BDLHSPVOE 5BPCBP3FDPNNFOEBUJPO"SDIJUFDUVSF 3BOLJOH4FSWJDF1MBUGPSN 5BPCBP1FSTPOBMJ[F1MBUGPSN

"CTUSBDU w 0CKFDUJWF$BMDVMBUFJUFNUPJUFN w 1SPQPTBM w 5PVTFTJEFJOGPSNBUJPO DBUFHPSZ QSJDF
ʜ GPS/PEF7FD w "MMFWJBUFDPMETUBSUQSPCMFN w &YQFSJNFOU w -JOL1SFEJDUJPO "NB[POEBUBTFUBOE UIFJSPXO5BPCBPEBUBTFU w 7JTVBMJ[BUJPO

'SBNFXPSL

3FMBUFE8PSL 'BDUPSJ[BUJPONFUIPET USZUPBQQSPYJNBUFBEKBDFODZNBUSJY &YBNQMF-*/& %FFQMFBSOJOHCBTFENFUIPE 3BOEPNXBMLCBTFENFUIPE FYUSBPSEJOBSZF⒏DJFOU

#VJMEJOH*UFNJUFN(SBQI • Consider sequential information of behavior • D->A->B, B->E,
D->E->F, … • Directed and weighted graph • Session-based • session = 1 hour time window • Noise Removal • Duration of the stay is less than 1 sec. • Over-active user: bought more than 1000 items or clicked 3500 times in three months • Updated item details: in the extreme case, a totally different item for the same identifier.

/PEF&NCFEEJOH • Transition Probability 1. Generate item node sequence 2.
Train skip-gram Based on DeepWalk’s framework

6TJOH4JEF*OGPSNBUJPO Embed other information, too. ঎඼Λಛఆ͢Δ POFIPUϕΫλ ൪໨ͷTJEFJOGPSNBUJPO Λಛఆ͢ΔPOFIPUϕΫλ O൪໨ͷTJEFJOGPSNBUJPO
Λಛఆ͢ΔPOFIPUϕΫλ Simple average weighted average : v-th item’s s-th information

&YQFSJNFOUBM&WBMVBUJPO Task: Link Prediction 1/3 of edeges are randomly removed,
and predict it. Dataset Amazon and Taobao Side information L1 cat, Category, Shop, Brand, City, Spu, Style, Color, Gender, Age, Buying Power, Material

0GGMJOF&WBMVBUJPO OPTJEFJOGPSNBUJPO TJEFJOGPSNBUJPOXJUIXFJHIUFEBWFSBHJOH TJEFJOGPSNBUJPOXJUITJNQMFBWFSBHJOH 0OMJOF&WBMVBUJPO

3FTVMUGPSDPMEJUFNT 8FJHIU7JTVBMJ[BUJPO

&NCFEEJOH7JTVBMJ[BUJPO

4ZTUFN%FQMPZNFOUBOE0QFSBUJPO w 6TFMPHTPGNPOUIT w "GUFSBOUJTQBNQSPDFTTJOH CJMMJPO FOUSJFTSFNBJOFE w (SBQIJTTQMJUJOUPTVCHSBQIT FBDI
HSBQIIBTNJMMJPOOPEFT5PUBM OVNCFSPGHFOFSBUFETFRVFODFJT CJMMJPO w (16XBTVTFEUPUSBJOFNCFEEJOH w -FTTUIBOIPVSTUPEP MPHSFUSJFWBM BOUJTQBNQSPDFTTJOH JUFNHSBQIDPOTUSVDUJPO TFRVFODF HFOFSBUJPO USBJOFNCFEEJOH JJ TJNJMBSJUZDPNQVUBUJPO

.ZUIPVHIU w "OUJTQBNQSPDFTTJOHNBZCFVTFGVM w 7JTVBMJ[BUJPOPGTJEFJOGPSNBUJPOJTJOUFSFTUJOH w 5SBJONJYUVSFSBUJPGPSFBDIJUFN 4PVOETUPPJOF⒏DJFOU w
)PXXFMMXBTUIFDPMETUBSUQSPCMFNTPMWFE 5PPBNCJHVPVT w -JOLQSFEJDUJPOJTUPPFBTZCFDBVTFXFLOPXEBUBJTWFSZTQBSTF w 8IZOPUDPNQBSFUP%//CBTFENFUIPE NBZCFUPPJOF⒏DJFOUGPSEFQMPZNFOU

Billion-scale Embedding for E-commerce Recommen...

Billion-scale Embedding for E-commerce Recommendation in Alibaba

Hayato Maki

More Decks by Hayato Maki

Other Decks in Research

Featured

Transcript

BQQFBSFEJO,%%`

#BDLHSPVOE 3FDPNNFOEFS4ZTUFNJO5BPCBP 1 billion users x 2 billion items

#BDLHSPVOE 5XPTUBHFSFDPNNFOEBUJPO UPEBZ`TUBML

#BDLHSPVOE 5BPCBP3FDPNNFOEBUJPO"SDIJUFDUVSF 3BOLJOH4FSWJDF1MBUGPSN 5BPCBP1FSTPOBMJ[F1MBUGPSN

"CTUSBDU w 0CKFDUJWF$BMDVMBUFJUFNUPJUFN w 1SPQPTBM w 5PVTFTJEFJOGPSNBUJPO DBUFHPSZ QSJDF

'SBNFXPSL

3FMBUFE8PSL 'BDUPSJ[BUJPONFUIPET USZUPBQQSPYJNBUFBEKBDFODZNBUSJY &YBNQMF-*/& %FFQMFBSOJOHCBTFENFUIPE 3BOEPNXBMLCBTFENFUIPE FYUSBPSEJOBSZF⒏DJFOU

#VJMEJOH*UFNJUFN(SBQI • Consider sequential information of behavior • D->A->B, B->E,

/PEF&NCFEEJOH • Transition Probability 1. Generate item node sequence 2.

6TJOH4JEF*OGPSNBUJPO Embed other information, too. ঎඼Λಛఆ͢Δ POFIPUϕΫλ ൪໨ͷTJEFJOGPSNBUJPO Λಛఆ͢ΔPOFIPUϕΫλ O൪໨ͷTJEFJOGPSNBUJPO

&YQFSJNFOUBM&WBMVBUJPO Task: Link Prediction 1/3 of edeges are randomly removed,

0GGMJOF&WBMVBUJPO OPTJEFJOGPSNBUJPO TJEFJOGPSNBUJPOXJUIXFJHIUFEBWFSBHJOH TJEFJOGPSNBUJPOXJUITJNQMFBWFSBHJOH 0OMJOF&WBMVBUJPO

3FTVMUGPSDPMEJUFNT 8FJHIU7JTVBMJ[BUJPO

&NCFEEJOH7JTVBMJ[BUJPO

4ZTUFN%FQMPZNFOUBOE0QFSBUJPO w 6TFMPHTPGNPOUIT w "GUFSBOUJTQBNQSPDFTTJOH CJMMJPO FOUSJFTSFNBJOFE w (SBQIJTTQMJUJOUPTVCHSBQIT FBDI

.ZUIPVHIU w "OUJTQBNQSPDFTTJOHNBZCFVTFGVM w 7JTVBMJ[BUJPOPGTJEFJOGPSNBUJPOJTJOUFSFTUJOH w 5SBJONJYUVSFSBUJPGPSFBDIJUFN 4PVOETUPPJOF⒏DJFOU w