Job Recommendation with Hawkes Process W. Xiao, X. Xu, K. Liang, J. - - PowerPoint PPT Presentation

job recommendation with hawkes process
SMART_READER_LITE
LIVE PREVIEW

Job Recommendation with Hawkes Process W. Xiao, X. Xu, K. Liang, J. - - PowerPoint PPT Presentation

Job Recommendation with Hawkes Process W. Xiao, X. Xu, K. Liang, J. Mao, and J. Wang OneSearch Team, Alibaba Group Boston, MA, USA Sep. 15, 2016 ED DBED


slide-1
SLIDE 1

Job Recommendation with Hawkes Process

  • W. Xiao, X. Xu, K. Liang, J. Mao, and J. Wang

OneSearch Team, Alibaba Group Boston, MA, USA

  • Sep. 15, 2016
slide-2
SLIDE 2

ED

DBED B.DDD.DD EB.ED

Ø 6AMHH.MIARKMG Ø .AOPMA-CEAAMEC Ø AMECOKGREOD/,: Ø AIATBKMKHOMO Ø :AILKMH,ANMELOKMREOD0RGANMKANN

  • CEDCCECCC

2

slide-3
SLIDE 3

DBED

3

slide-4
SLIDE 4

DBED .B B

4

.B

Ø P6 AMDKB)HE/MKPL Ø 6AAMD MKFAO 1OAHHECAO,OHOBKMIBKM16:

slide-5
SLIDE 5

DBED .B B

5

B!C

  • RecSys2015(Vienna)
  • The Second Place
  • KDD Cup2016 (San Francisco)
  • The Second Place (Phase 1)

Boston

  • RecSys2016
  • The First Place

Wenming Xiao Xiao Xu Kang Liang Junkang Mao Jun Wang

slide-6
SLIDE 6

B.DDD.DD

6

slide-7
SLIDE 7

B.DDD

7

C

Ø /EAODA ODAGATONGENOK MAKIIAHENOKBFKLKNONRDEDODAPNAMNIECDOEOAMO REODEODAASORAAG

FEDDB

slide-8
SLIDE 8

B

8

D .D D

  • )OEA PMEC OANO

1OAI EILEO OMA 1OAI LMKBEHEC 1OAI NA . NAM NA . NAM1OAI MKBEHA EIEHMEOT EMRENA AMEC

Two OCA KAHEC

  • NAIHA AMEC
  • :AILKMH,ANMELOKM

NAMLMKBEHEC

slide-9
SLIDE 9

EB.ED

9

slide-10
SLIDE 10

FBBGB

10

BB-BGB

Ø :REODMEKPNIKAHNOKNKMEAMNAEBKMIOEK Ø -NAIHA/,:REODODAELPOBMKIEEOEHMGEC Ø 0RGANMKANN NAHBASEOECOKLOPMAOAILKMHLOOAMN

  • Ø /AAMOAODAMECDOHENOKBFK

LKNOECOKMAKIIAE ODAMECDOOEIEC

slide-11
SLIDE 11

DEBB

11

IBDBCDCDEBC

Ø :DMAANEBAOPMAOACKMEAN(PNAMEOAIPNAMEOAIEOAMOEK BAOPMAN Ø -SOAABAOPMANREODBPMODAMHTNENNOOENOEN Ø )CCMACOEKENODAMAIATBKMNLMNEOT Ø :EIAREKRNACIAOOEK

7 days 14 days 21 days 28 days 3 days 2 1

slide-12
SLIDE 12

DEBCBGDF

12

ratio: ui/u, ui/i trend: item interaction counts similarity(interaction&profiling) : u2u, i2i, u2i poi: item, user interaction user&item profile: 0/1 user/item/ui interaction statistics: (time window) 0/1, count, distinct count, days…… conditional probability: impression to interaction, interaction to interaction

Low Level High Level

user: last day if click/imp item: last day click counts user-item: user click counts for the item in last week item impression trend item earliest and last interaction time ……

slide-13
SLIDE 13

BD-GD

13

ABBCCBB

Ø ANEPHNCMEAO Ø ,EBBAMAOHKNNBPOEKNBKMKKNOEC Ø KEORENAEMRENAIAODK

slide-14
SLIDE 14

B

14

Train Data Train Model

28 days 5 days 28 days 6 days 28 days 7 days 28 days 8 days

Stage 1 LR XGBOOST GBRT GBDT

Pass Reject

Stage 2 GBDT GBDT GBDT …… Ensemble

Top N Output

slide-15
SLIDE 15

!.DBD

15

!CDBED ECB C DBED I BG DEBC

Ø ,AIKCMLDEO APOEKACMAAMAAMHAAH RKMGASLAMEAAAO Ø 1OAMOEKN AORAA FK MKHAN FK OEOHA KM OC

BBD

Ø )LLHT,) OK BE ODA HOAO OKLE IKAH Ø )LLHTGIAN OKCAOPNAMHPNOAMN Ø NA 255 OK CAO ODA OKL MANPHO

slide-16
SLIDE 16

ABDCDIGDGCBCC

16

. HDADABCCBBEBBDFD

Ø :DAKEOEKHLMKEHEOTODOODAASOPNAMEOAIAAO DLLANOOEIA Ø 0KIKCAKPNLKEOLMKANN

User-item event Model, Du, etal. 2015

baseline intensity temporal dependency user-item event

slide-17
SLIDE 17

G-GCBCC

17

Ø /AAMHEUAODAIKAHECODAAAOKBNECHAPNAMEOAILEMOK HHPNAMEOAILEMN Ø KRMGNNPILOEK(KODPNAMNEOAINAOACKMEUA EOKHEIEOAPIAMHPNOAMN baseline intensity matrix self exciting matrix exponential form

slide-18
SLIDE 18
  • GDABDCDI

18

Ø 1OAIMAKIIAOEKREODOAILKMHEOANEOT ,PAOH& Ø :AILKMHEOANEOTNEOEKHBAOPMAN

slide-19
SLIDE 19
  • CEDCEDEBB

19

Ø ANPHONEATN Ø 6CKECOPEAN(

q NOPTODAEILOKBODAOAILKMHLOOAMN q KIEAKODKOASOPHOAILKMHEBKMIOEK q EANOECOANAHBASEOECNAHBKMMAOECLMKANN

slide-20
SLIDE 20

E

20