ec a modular framework for extensible
play

ec : A Modular Framework for Extensible Op Open enRec and - PowerPoint PPT Presentation

ec : A Modular Framework for Extensible Op Open enRec and Adaptable Recommendation Algorithms Lo Longqi Yang Joshua Gruenstein Cheng-Kang(Andy) Eugene Bagdasaryan Deborah Estrin Hsieh Funders: 1 Pr Prom omising future e of of per


  1. ec : A Modular Framework for Extensible Op Open enRec and Adaptable Recommendation Algorithms Lo Longqi Yang Joshua Gruenstein Cheng-Kang(Andy) Eugene Bagdasaryan Deborah Estrin Hsieh Funders: 1

  2. Pr Prom omising future e of of per erson onali lization on and rec ecom ommen ender er system ems Education Healthcare Social network Media Food and Diet e-Commerce 2

  3. Re Recomme mmendat dation al algorithms hms ar are increasi asingly compl mplex 3

  4. Re Recomme mmendat dation al algorithms hms ar are increasi asingly compl mplex Diverse user feedback signals rating like skip follow … click save watch listen 4

  5. Re Recomme mmendat dation al algorithms hms ar are increasi asingly compl mplex Diverse user feedback signals rating like skip follow … click save watch listen Heterogeneous data streams and context user demographics user social media posts Item descriptions videos images … activities location mood 5

  6. Re Recomme mmendat dation al algorithms hms ar are increasi asingly compl mplex Complex goals Diverse user feedback signals accuracy rating like skip follow … click save watch listen diversity novelty Heterogeneous data streams and context fairness user demographics user social media posts Item descriptions videos images quality … activities location mood interpretability … 6

  7. Bag of algorithms 7

  8. However, current recomme Ho mmendat dation al algorithms hms lac ack k si simpl mplicity an and d mo modu dular arity. … Bag of algorithms 8

  9. O pe penRec Ap Apache e Lic icen ense e 2.0 .0 end and ad apt to various scenarios. o Easy to ex exten adap o Quick experimentation (e.g., model selection) and idea exploration. o Comparable (sometimes even better) performance. Modul Mod ular arity ty 9

  10. Current practice vs. Op Cu Open enRec ec Pr Prior rese search Yo Your research/application cur urrent nt News recommender Music recommender practice pr 10

  11. Cu Current practice vs. Op Open enRec ec Pr Prior rese search Your research/application Yo cur urrent nt News recommender Music recommender pr practice different user feedback signals different data sources tangled implementations 11

  12. Cu Current practice vs. Op Open enRec ec Pr Prior rese search Yo Your research/application cur urrent nt News recommender Music recommender practice pr different user feedback signals different data sources tangled implementations user clicks user clicks Open Op enRec ec user audio twitter news demographics following keywords 12

  13. 1 Ab Abstract action and on and i int nter erface ace 2 Im Implementatio ions 3 Si Simpl ple us use cas ases 4 Takeaways and Future work Ta 13

  14. Abs Abstrac ract entit itie ies in in re recommenda ndatio tion n alg lgorith rithms user (or group) context (environment) item 1 Ab Abstra raction n and nd int nterfa rface 14

  15. Bu Building a a re recommenda ndatio tion n alg lgorith rithm En Entit ity User Context Item 1 Ab Abstra raction n and nd int nterfa rface 15

  16. Bu Building a a re recommenda ndatio tion n alg lgorith rithm Profile Pr Entit En ity User Context Item 1 Ab Abstra raction n and nd int nterfa rface 16

  17. Building a Bu a re recommenda ndatio tion n alg lgorith rithm Profile Pr … … … Entit En ity User Context Item 1 Ab Abstra raction n and nd int nterfa rface 17

  18. Building a Bu a re recommenda ndatio tion n alg lgorith rithm … … … Pr Profile … … … Entit En ity User Context Item 1 Ab Abstra raction n and nd int nterfa rface 18

  19. Building a Bu a re recommenda ndatio tion n alg lgorith rithm … … … Pr Profile … … … Entit En ity User Context Item 1 Ab Abstra raction n and nd int nterfa rface 19

  20. Building a Bu a re recommenda ndatio tion n alg lgorith rithm In Interactio ion … … … Pr Profile … … … Entit En ity User Context Item 1 Ab Abstra raction n and nd int nterfa rface 20

  21. Building a Bu a re recommenda ndatio tion n alg lgorith rithm … In Interactio ion … … … Pr Profile … … … Entit En ity User Context Item 1 Ab Abstra raction n and nd int nterfa rface 21

  22. Bu Building a a re recommenda ndatio tion n alg lgorith rithm Ground-tr Gr truth th in interactio ions Gr Ground-tr truth th in interactio ions … Interactio In ion … … … Pr Profile … … … Entit En ity User Context Item 1 Ab Abstra raction n and nd int nterfa rface 22

  23. Extraction: extra Ex xtract t re repre presenta ntatio tions ns Ground-tr Gr truth th in interactio ions Gr Ground-tr truth th in interactio ions … Interactio In ion data representation Extraction … … … Pr Profile … … … En Entit ity User Context Item 1 Ab Abstra raction n and nd int nterfa rface 23

  24. Fusion: fuse representations Fu Ground-tr Gr truth th in interactio ions Gr Ground-tr truth th in interactio ions … Interactio In ion … … … Pr Profile representation representation Fusion representation … representation … … … Entit En ity User Context Item 1 Ab Abstra raction n and nd int nterfa rface 24

  25. Int Intera ractio tion: n: pre predic dict t clic licks/lik likes/ra ratings tings… Gr Ground-tr truth th in interactio ions Gr Ground-tr truth th in interactio ions … In Interactio ion user representation predicted context representation Interaction interactions item representation … … … Pr Profile … … … Entit En ity User Context Item 1 Abstra Ab raction n and nd int nterfa rface 25

  26. A A hypo pothetic ical al music ic re recommenda ndatio tion n alg lgorith rithm (skip (s ip) (l (lik ike) PairwiseLog PointwiseMSE concat masking latent factor spatial- MFCC CNN MLP LSTM latent factor temporal demographical audio lyrics music id tweets location information user id 1 Ab Abstra raction n and nd int nterfa rface 26

  27. A A hypo pothetic ical al music ic re recommenda ndatio tion n alg lgorith rithm … … … … … … (s (skip ip) (l (lik ike) … … … concat … latent factor … spatial- MFCC CNN MLP latent factor … temporal demographical audio lyrics music id tweets location information user id 1 Ab Abstra raction n and nd int nterfa rface 27

  28. A hypo A pothetic ical al music ic re recommenda ndatio tion n alg lgorith rithm (skip (s ip) (lik (l ike) PairwiseLog PointwiseMSE concat masking latent factor LS LSTM lat atent fac actor latent factor demographical music id mo mood ar artist information user id 1 Ab Abstra raction n and nd int nterfa rface 28

  29. Open Op enRec ec fra framework rk stru ructure re Recommender Utility Sampler R-1 : click logs , text R-2 : watch history , … R-n : posts, and content content visual analysis, … Pairwise topic modeling . and activity detection. sampler Pointwise Module sampler … Extraction Fusion Interaction LSTM Pointwise MSE LF Concatenation Evaluator ResNet MLP Weighted sum Pairwise distance AUC AutoEncoder Average Pointwise MLP Recall@K … … … … 2 Implementation Im ons 29

  30. Inside In e a Re Recomme mmende der Recommender load(…) train(...) serve(…) save(…) build_training_graph() build_serving_graph() train=true train=false build_inputs( train ) build_extractions( train ) build_user_extractions( train ) build_item_extractions( train ) build_extra_extractions( train ) build_fusions( train ) build_default_fusions( train ) build_custom_fusions( train ) … build_interactions( train ) build_default_interactions( train ) build_custom_interactions( train ) if train==true … build_optimizer() 2 Implementation Im ons 30

  31. Inside In e a Re Recomme mmende der Recommender load(…) train(...) serve(…) save(…) build_training_graph() build_serving_graph() train=true train=false build_inputs( train ) build_extractions( train ) build_user_extractions( train ) build_item_extractions( train ) build_extra_extractions( train ) build_fusions( train ) build_default_fusions( train ) build_custom_fusions( train ) … build_interactions( train ) build_default_interactions( train ) build_custom_interactions( train ) if train==true … build_optimizer() 2 Implementation Im ons 31

  32. In Inside e a Mod Module le Extraction data Fusion build_shared_graph() module #1 loss train=True train=False … module #n build_training_graph() outputs Interaction build_serving_graph() user repr. item repr. context repr. 2 Implementation Im ons 32

  33. In Inside e a Mod Module le Extraction data Fusion build_shared_graph() module #1 loss train=True train=False … module #n build_training_graph() outputs Interaction build_serving_graph() user repr. item repr. context repr. 2 Implementation Im ons 33

Recommend


More recommend