RecSys Multi-Stack Ensemble for Job Recommendation, Pumpkin-Pie

Politecnico di Milano Team Pumpkin-Pie09-15-2016

Team Pumpkin-Pie

T. Carpi

M. Edemanti

E. Kamberoski

E. Sacchi

R. Pagano

P. Cremonesi

M. Quadrana


Outline• Collaborative Filtering • Content-Based • Interactions and Impressions • Multi-Stack Ensemble

• Linear Ensemble • Evaluation Score Ensemble

• Conclusions


Collaborative-Filtering

IDF value as a rate for the job



User-based



Item-based


Content-BasedConcept-based

IDF value as a weight for the tag


Content-Based

Concept-based joint User-Item similarity


Interactions & Impressions

Click

ReplyBookmark

Already clicked jobs are likely to be clicked again


Interactions & Impressions

A B C D E F

B C DA

B D AC

Dataset

Filtering-step

Reordering-step


Multi-Stack Ensemble

Past Interactions

Ens CF

Past Impressions

UBCF IntInt

UBCF IntImp

UBCF ImpImp

UBCF ImpInt

IBCF IntInt

IBCF ImpImp

CBJUIS CBIS

Baseline

Ens CF+CB

Ens CB

Final Ensemble



Ens CF

UBCF IntInt

UBCF IntImp

UBCF ImpImp

UBCF ImpInt

IBCF IntInt

IBCF ImpImp

CBJUIS CBIS

Ens CB

Stack Level 1



Ens CF

UBCF IntInt

UBCF IntImp

UBCF ImpImp

UBCF ImpInt

IBCF IntInt

IBCF ImpImp

CBJUIS CBIS

Ens CF+CB

Ens CB

Stack Level 2



Past Interactions

Ens CF

Past Impressions

UBCF IntInt

UBCF IntImp

UBCF ImpImp

UBCF ImpInt

IBCF IntInt

IBCF ImpImp

CBJUIS CBIS

Baseline

Ens CF+CB

Ens CB

Final Ensemble

Stack Level 3


Linear Ensemble

weight for algorithm a

rank of item i in algorithm a

decay for algorithm a


red 1.9990

pink 1.9980

yellow 1.9970

blue 1.9960

green 1.9985

orange 1.9970

purple 1.9955

red 1.9940

Algorithm A Weight 2 Decay 0.001

Algorithm B Weight 2 Decay 0.0015

Linear Ensemble

1

2

3

4

1

2

3

4


pink 1.9980yellow 1.9970

blue 1.9960

green 1.9985

orange 1.9970

purple 1.9955

Algorithm A Weight 2 Decay 0.001

Algorithm B Weight 2 Decay 0.0015

Linear Ensemble

red 3.9930


Linear Ensemble

Past Interactions

Ens CF

Past Impressions

UBCF IntInt

UBCF IntImp

UBCF ImpImp

UBCF ImpInt

IBCF IntInt

IBCF ImpImp

CBJUIS CBIS

Baseline

Ens CF+CB

Ens CB

Final Ensemble


Evaluation-Score Ensembleleaderbord score

# of elements


Evaluation-Score Ensemble

red 7.5660

pink 7.5660

purple 5.5660

orange 5.5660

yellow 9.4575

green 9.4575

blue 6.9575

grey 6.9575

Algorithm A l_a 200k n_a 1 Mln Weight 0.2

Algorithm B l_b 200k n_b 800k Weight 0.25



red 7.5660pink 7.5660

purple 5.5660orange 5.5660

yellow 9.4575green 9.4575

blue 6.9575grey 6.9575

Algorithm A l_a 200k n_a 1 Mln Weight 0.2

Algorithm B l_b 200k n_b 800k Weight 0.25



Past Interactions

Ens CF

Past Impressions

UBCF IntInt

UBCF IntImp

UBCF ImpImp

UBCF ImpInt

IBCF IntInt

IBCF ImpImp

CBJUIS CBIS

Baseline

Ens CF+CB

Ens CB

Final Ensemble


Conclusions

Competition promotes the algorithms that learn which are the best items among the ones

recommended by the Xing platform


Conclusions


4th Place

Team Pumpkin-Pie

RecSys Multi-Stack Ensemble for Job Recommendation, Pumpkin-Pie

Data & Analytics

Transcript of RecSys Multi-Stack Ensemble for Job Recommendation, Pumpkin-Pie