Paper Presentation in ACM RecSys2016, by Jie Yang.

1. Learning Hierarchical Feature Influence for
Recommendation by Recursive Regularisation
Jie Yang, Zhu Sun, Alessandro Bozzon,
Jie Zhang

2. Objective
Fully exploit feature hierarchies for recommendation by
utilising the structured information, and historical data

3. Hierarchy
Feature
Science and
Engineering
- Biology
- Chemistry
- Machine learning
- Natural language
processing
- …
matters
Feature Hierarchies
are everywhere

4. ?
Recommender
Systems
Hierarchy
Feature
Science and
Engineering
- Biology
- Chemistry
- Machine learning
- Natural language
processing
- …
matters
Feature Hierarchies
are everywhere

5. Recommender
Systems
Hierarchy
Feature
Science and
Engineering
- Biology
- Chemistry
- Machine learning
- Natural language
processing
- …
matters
Product Category
Feature Hierarchies
are everywhere

6. Recommender
Systems
Hierarchy
Feature
Science and
Engineering
- Biology
- Chemistry
- Machine learning
- Natural language
processing
- …
matters
Product Category
Article Topic
Feature Hierarchies
are everywhere

7. Recommender
Systems
Hierarchy
Feature
Science and
Engineering
- Biology
- Chemistry
- Machine learning
- Natural language
processing
- …
matters
Product Category
Article Topic
Music Genre
Feature Hierarchies
are everywhere

8. Recommender
Systems
Hierarchy
Feature
Science and
Engineering
- Biology
- Chemistry
- Machine learning
- Natural language
processing
- …
matters
Product Category
Article Topic
Music Genre
Venue Category
…
Feature Hierarchies
are everywhere

9. How can feature hierarchies
be used?

10. Example

11. Structureto be considered

12. Structureto be considered

13. Structureto be considered

14. Structureto be considered
State-of-the-art methods reduce hierarchies to simpler structures
(e.g. flat), which brings severe information loss

15. Datais also important

16. Datais also important

17. Datais also important
The different influence of structured features can only be
learnt from historical user-item interaction data

18. Fully exploit feature hierarchies for recommendation by
utilising the structured information, and historical data
Objective

19. Fully exploit feature hierarchies for recommendation by
utilising the structured information, and historical data
Objective
Solution:
Recursive Regularisation

20. Latent factor model
BasicModel

21. Influence of an isolated feature:
F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
F4= {u1, u2, u3}
u1 u2 u3 u4 u5
Influence (1)
Feature

22. Influence of an isolated feature:
F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
F4= {u1, u2, u3}
u1 u2 u3 u4 u5
Influence (1)
Feature

23. Influence of an isolated feature:
F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
F4= {u1, u2, u3}
u1 u2 u3 u4 u5
E.g. the influence of F4 is the dissimilarity of latent factors of all
users characterised by it, i.e. u1, u2, u3.
Influence (1)
Feature

24. Influence of an isolated feature unit (a parent with its children):
F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
Fu(F5)
Fu(F4) F4= {u1, u2, u3}
u1 u2 u3 u4 u5
Feature
Influence (2)

25. Influence of an isolated feature unit (a parent with its children):
F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
Fu(F5)
Fu(F4) F4= {u1, u2, u3}
u1 u2 u3 u4 u5
Feature
Influence (2)

26. Influence of an isolated feature unit (a parent with its children):
F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
Fu(F5)
Fu(F4) F4= {u1, u2, u3}
u1 u2 u3 u4 u5
Feature
Influence (2)

27. Influence of an isolated feature unit (a parent with its children):
F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
Fu(F5)
Fu(F4) F4= {u1, u2, u3}
u1 u2 u3 u4 u5
Feature
E.g. the influence of feature unit I'(F5) is
Influence (2)

28. Recursive Regularisation
Influence of a feature hierarchy: F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
F4= {u1, u2, u3}
u1 u2 u3 u4 u5

29. Recursive Regularisation
Influence of a feature hierarchy: F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
F4= {u1, u2, u3}
u1 u2 u3 u4 u5
E.g. the influence the above hierarchy is:

30. Recursive Regularisation
Influence of a feature hierarchy: F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
F4= {u1, u2, u3}
u1 u2 u3 u4 u5
E.g. the influence the above hierarchy is:

31. Recursive Regularisation
Influence of a feature hierarchy: F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
F4= {u1, u2, u3}
u1 u2 u3 u4 u5
E.g. the influence the above hierarchy is:

32. Recursive Regularisation
F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
F4= {u1, u2, u3}
u1 u2 u3 u4 u5
E.g. the influence the above hierarchy is:
Influence of a feature hierarchy:

33. Recursive Regularisation
F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
F4= {u1, u2, u3}
u1 u2 u3 u4 u5
E.g. the influence the above hierarchy is:
Influence of a feature hierarchy:

34. Recursive Regularisation
F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
F4= {u1, u2, u3}
u1 u2 u3 u4 u5
u1
u2
g5+s5g4- 1
g5+s5g4-
-
-
-
g5
g5
g5
g5
u3
1
g5+s5g4 g5+s5g4 g5 g5
u4 g5 g5 g5 1
1g5 g5 g5u5
u1 u2 u3 u4 u5
With recursive regularisation, we can express the influence of
features on users as a parameterised function of g, and s

35. Recursive Regularisation
F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
F4= {u1, u2, u3}
u1 u2 u3 u4 u5
u1
u2
g5+s5g4- 1
g5+s5g4-
-
-
-
g5
g5
g5
g5
u3
1
g5+s5g4 g5+s5g4 g5 g5
u4 g5 g5 g5 1
1g5 g5 g5u5
u1 u2 u3 u4 u5
With recursive regularisation, we can express the influence of
features on users as a parameterised function of g, and s

36. Recursive Regularisation
F1= {u1, u2}
F3= {u4, u5}
F2= {u3}
F5= {u1, u2, u3, u4, u5}
F4= {u1, u2, u3}
u1 u2 u3 u4 u5
u1
u2
g5+s5g4- 1
g5+s5g4-
-
-
-
g5
g5
g5
g5
u3
1
g5+s5g4 g5+s5g4 g5 g5
u4 g5 g5 g5 1
1g5 g5 g5u5
u1 u2 u3 u4 u5
With recursive regularisation, we can express the influence of
features on users as a parameterised function of g, and s

37. ReMFFramework

38. ReMFFramework

39. ReMFFramework
ReMF can automatically learn feature influence from historical
user-item interaction data

40. ReMFFramework
Remark: ReMF can work with feature hierarchies of both users
and items; can work with imbalanced feature hierarchies; and it
is scalable to large dataset

41. Validation

42. Setup
Users’ foursquare check-ins in 2
platforms, 4 cities, over 3 weeks
Twitter Instagram
Amsterdam Rome Paris London
Datasets

43. Setup
Users’ foursquare check-ins in 2
platforms, 4 cities, over 3 weeks
Twitter Instagram
Amsterdam Rome Paris London
Datasets Evaluation
MAE, RMSE, AUC
5-fold cross-validation

44. Setup
Users’ foursquare check-ins in 2
platforms, 4 cities, over 3 weeks
Twitter Instagram
Amsterdam Rome Paris London
Datasets Evaluation
MAE, RMSE, AUC
5-fold cross-validation
Parameter setting
Grid Search

45. Setup
Users’ foursquare check-ins in 2
platforms, 4 cities, over 3 weeks
Twitter Instagram
Amsterdam Rome Paris London
Datasets Evaluation
MAE, RMSE, AUC
5-fold cross-validation
Parameter setting
Grid Search
Comparison methods
MF
CMF, FM
(generic feature based methods)
TaxMF, HieFM
(w. hierarchical features)

46. Results of ReMF
Amsterdam
Paris
Rome
London
0,05
0,10
0,15
0,20
Log(alpha)
−5 −4 −3 −2 −1 0
Amsterdam
Paris
Rome
London
0,10
0,15
0,20
0,25
Log(alpha)
−5 −4 −3 −2 −1 0
Amsterdam
Paris
Rome
London
0,05
0,10
0,15
0,20
Log(alpha)
−5 −4 −3 −2 −1 0
Amsterdam
Paris
Rome
London
0,10
0,15
0,20
0,25
Log(alpha)
−5 −4 −3 −2 −1 0
MAE - Instagram RMSE - Instagram
MAE - Twitter RMSE - Twitter

47. Results of ReMF
Amsterdam
Paris
Rome
London
0,05
0,10
0,15
0,20
Log(alpha)
−5 −4 −3 −2 −1 0
Amsterdam
Paris
Rome
London
0,10
0,15
0,20
0,25
Log(alpha)
−5 −4 −3 −2 −1 0
Amsterdam
Paris
Rome
London
0,05
0,10
0,15
0,20
Log(alpha)
−5 −4 −3 −2 −1 0
Amsterdam
Paris
Rome
London
0,10
0,15
0,20
0,25
Log(alpha)
−5 −4 −3 −2 −1 0
MAE - Instagram RMSE - Instagram
MAE - Twitter RMSE - Twitter
1
ReMF is robust to parameter setting.

48. ComparativeResults
MAE (RMSE in paper)
two views of each dataset
Views Dataset MF CMF TaxMF FM HieFM ReMF
All
Inst.
Amsterdam .1957 .1564 .1426 .0876 .0822 .0707
Paris .1539 .1550 .1416 .0790 .0830 .0772
Rome .2549 .1584 .1474 .0912 .0885 .0855
London .1799 .1559 .1369 .0834 .0851 .0774
Tw.
Amsterdam .2264 .1606 .1345 .0989 .0942 .0844
Paris .2014 .1714 .1552 .0956 .0935 .0894
Rome .2681 .1713 .1591 .1023 .0996 .0977
London .2176 .1659 .1545 .0931 .0898 .0772
Cold
Start
Inst.
Amsterdam .2938 .1552 .1457 .0924 .0885 .0712
Paris .1939 .1541 .1476 .0849 .0848 .0799
Rome .3840 .1614 .1518 .0952 .0938 .0808
London .3032 .1544 .1415 .0893 .0901 .0791
Tw.
Amsterdam .3261 .1604 .1426 .1006 .0956 .0849
Paris .2439 .1706 .1640 .1012 .0945 .0873
Rome .3922 .1718 .1681 .1073 .1070 .0988
London .3301 .1642 .1587 .0967 .0924 .0756

49. Views Dataset MF CMF TaxMF FM HieFM ReMF
All
Inst.
Amsterdam .1957 .1564 .1426 .0876 .0822 .0707
Paris .1539 .1550 .1416 .0790 .0830 .0772
Rome .2549 .1584 .1474 .0912 .0885 .0855
London .1799 .1559 .1369 .0834 .0851 .0774
Tw.
Amsterdam .2264 .1606 .1345 .0989 .0942 .0844
Paris .2014 .1714 .1552 .0956 .0935 .0894
Rome .2681 .1713 .1591 .1023 .0996 .0977
London .2176 .1659 .1545 .0931 .0898 .0772
Cold
Start
Inst.
Amsterdam .2938 .1552 .1457 .0924 .0885 .0712
Paris .1939 .1541 .1476 .0849 .0848 .0799
Rome .3840 .1614 .1518 .0952 .0938 .0808
London .3032 .1544 .1415 .0893 .0901 .0791
Tw.
Amsterdam .3261 .1604 .1426 .1006 .0956 .0849
Paris .2439 .1706 .1640 .1012 .0945 .0873
Rome .3922 .1718 .1681 .1073 .1070 .0988
London .3301 .1642 .1587 .0967 .0924 .0756
ComparativeResults
MAE (RMSE in paper)
two views of each dataset
7.2% (all)
12.02% (cold)

50. Views Dataset MF CMF TaxMF FM HieFM ReMF
All
Inst.
Amsterdam .1957 .1564 .1426 .0876 .0822 .0707
Paris .1539 .1550 .1416 .0790 .0830 .0772
Rome .2549 .1584 .1474 .0912 .0885 .0855
London .1799 .1559 .1369 .0834 .0851 .0774
Tw.
Amsterdam .2264 .1606 .1345 .0989 .0942 .0844
Paris .2014 .1714 .1552 .0956 .0935 .0894
Rome .2681 .1713 .1591 .1023 .0996 .0977
London .2176 .1659 .1545 .0931 .0898 .0772
Cold
Start
Inst.
Amsterdam .2938 .1552 .1457 .0924 .0885 .0712
Paris .1939 .1541 .1476 .0849 .0848 .0799
Rome .3840 .1614 .1518 .0952 .0938 .0808
London .3032 .1544 .1415 .0893 .0901 .0791
Tw.
Amsterdam .3261 .1604 .1426 .1006 .0956 .0849
Paris .2439 .1706 .1640 .1012 .0945 .0873
Rome .3922 .1718 .1681 .1073 .1070 .0988
London .3301 .1642 .1587 .0967 .0924 .0756
ComparativeResults
MAE (RMSE in paper)
two views of each dataset
ReMF consistently outperform the comparative methods.
7.2% (all)
12.02% (cold)

51. Views Dataset MF CMF TaxMF FM HieFM ReMF
All
Inst.
Amsterdam .1957 .1564 .1426 .0876 .0822 .0707
Paris .1539 .1550 .1416 .0790 .0830 .0772
Rome .2549 .1584 .1474 .0912 .0885 .0855
London .1799 .1559 .1369 .0834 .0851 .0774
Tw.
Amsterdam .2264 .1606 .1345 .0989 .0942 .0844
Paris .2014 .1714 .1552 .0956 .0935 .0894
Rome .2681 .1713 .1591 .1023 .0996 .0977
London .2176 .1659 .1545 .0931 .0898 .0772
Cold
Start
Inst.
Amsterdam .2938 .1552 .1457 .0924 .0885 .0712
Paris .1939 .1541 .1476 .0849 .0848 .0799
Rome .3840 .1614 .1518 .0952 .0938 .0808
London .3032 .1544 .1415 .0893 .0901 .0791
Tw.
Amsterdam .3261 .1604 .1426 .1006 .0956 .0849
Paris .2439 .1706 .1640 .1012 .0945 .0873
Rome .3922 .1718 .1681 .1073 .1070 .0988
London .3301 .1642 .1587 .0967 .0924 .0756
ComparativeResults
MAE (RMSE in paper)
two views of each dataset
ReMF consistently outperform the comparative methods.ReMF achieves higher performance in coping with the cold start
problem compared to the state-of-the-art methods.
7.2% (all)
12.02% (cold)

52. CMF
TaxMF
SoReg
FM
HieFM
ReMF
0,5
0,6
0,7
0,8
0,9
Amsterdam Paris Rome London
CMF
TaxMF
SoReg
FM
HieFM
ReMF
0,5
0,6
0,7
0,8
Amsterdam Paris Rome London
Instagram Twitter
ComparativeResults
AUC

53. ReMF consistently outperform the comparative methods in terms of
ranking performance.
CMF
TaxMF
SoReg
FM
HieFM
ReMF
0,5
0,6
0,7
0,8
0,9
Amsterdam Paris Rome London
CMF
TaxMF
SoReg
FM
HieFM
ReMF
0,5
0,6
0,7
0,8
Amsterdam Paris Rome London
Instagram Twitter
ComparativeResults
AUC

54. Take away
Recursive Regularisation, as a parameterised regularisation
function, can better exploit feature hierarchy for recommendation
by learning structured feature influence from data

55. Thank you!
j.yang-3@tudelft.nl