1. Asymmetric Correlation Regularized Matrix Factorization for Web Service Recommendation
Qi Xie1, Shenglin Zhao2, Zibin Zheng3, Jieming Zhu2 and Michael R. Lyu2
1Southwest University for Nationalities, Chengdu, China
2The Chinese University of Hong Kong, Hong Kong, China
3Sun Yat-Sen University, Guangzhou, China

2. Outline Introduction Problem Description
Asymmetric correlation regularized matrix factorization
Experiments
Conclusion & Future Work

3. Introduction
Web Services: Self-described, Loosely-coupled, Highly-dynamic, Cross-domain, Compositional nature
[ref.

4. Introduction s1 s2 s3 s4 s5 u1 u2 u3 Web Service Recommendation
Challenge:
Data sparsity problem
Response time (ms) s1 s2 s3 s4 s5 u1 ?
0.98
0.96 u2
0.89
0.90
0.59 u3
0.88
ri,j
user-service QoS matrix

5. Introduction Related work Collaborative Filtering (CF)
Memory-based: user-based Pearson Correlation Coefficient (PCC), item-based PCC [Zheng2011,Chen2013, Jiang2011]
Model-based: clustering model; latent factor model[Ma2008,Ma2011]
Context information:
location, time

6. Introduction r2,4 r2,6 r2,8 ? (s8, s4) service pair Given 4 users
The correlation among users and services is incorporated in CF approaches to address the data sparsity problem.

7.  Introduction r2,4 r2,6 r2,8 ? (s8, s6) (s8, s4) service pair
4 users
2 users
Then it is natural to suppose that the QoS value of s4 is liable to be a better predictor for service s8 than the QoS value of s6 [Zheng2011, Lemire2015].

8. Introduction
However, these approaches only consider the correlation to improve the computation of similarity.

9. Introduction
The hidden correlation can be transformed into the rank values of users and services.
We define such hidden correlation with an asymmetric matrix (namely Asymmetric Correlation, AC), in which each entry presents the hidden correlation between a user pair or between a service pair.

10. Introduction Motivation Target
To improve the Web service recommendation accuracy and solve the data sparsity problem, hidden correlation should be taken into consideration.
It will be find a different way to choose the optimal similar users or services for target objects.
Can we exploit the asymmetric correlation with matrix factorization (MF) to improve the recommendation performance, when the available data are sparse?
Can we take advantages of both user asymmetric correlation and service asymmetric correlation?

11. Outline Introduction Problem Description
Asymmetric correlation regularized matrix factorization
Experiments
Conclusion & Future Work

12. Problem Description MF
How to predict the missing QoS values of the user-service QoS matrix effectively by exploiting asymmetric correlation among both users and services.
Definition 1 (Co-occurrence criterion)
Definition 2 (Dependent relation)
(b)
(d)
(f) MF
(c)
(e)
(g)

13. Problem Description MF Random walk algorithm
How to predict the missing QoS values of the user-service QoS matrix effectively by exploiting asymmetric correlation among both users and services.
Definition 1 (Co-occurrence criterion)
Definition 2 (Dependent relation)
Random walk algorithm
(b)
(d)
(f) MF
(c)
(e)
(g)

14. Outline Introduction Problem Description
Asymmetric correlation regularized matrix factorization
Experiments
Conclusion & Future Work

15. Asymmetric correlation regularized matrix factorization
Framework overview
Asymmetric Correlation Calculation
Asymmetric Correlation Matrix construction
Asymmetric Correlation Propagation
Asymmetric Correlation based Regularization
User Asymmetric Correlation based Regularization Service Asymmetric Correlation based Regularization
Hybrid Asymmetric Correlation based Regularization

16. A. Asymmetric Correlation Calculation
1. Asymmetric Correlation Matrix Construction
Given an mn user-service QoS matrix
(1)
(2)
(3)

17. A. Asymmetric Correlation Calculation
1. Asymmetric Correlation Matrix Construction
Given an mn user-service QoS matrix
(4)
(5)
Dependent relations can be exploited from correlation matrices by constructing a stochastic matrix.

18. A. Asymmetric Correlation Calculation
2. Asymmetric Correlation Propagation
Given an mn user-service QoS matrix
(6)
(7)
Random walk algorithm

19. B. Asymmetric Correlation based Regularization
3. Hybrid Asymmetric Correlation based Regularization
(8)
Gradient descent algorithm
Target users (services) are liable to have similar Web service invocation experience with popular users (services) which have high rankings of asymmetric correlation.

20. Outline Introduction Problem Description
Asymmetric correlation regularized matrix factorization
Experiments
Conclusion & Future Work

21. Experiments
Dataset
Response time (RT): user-perceived delay of service invocation (s)
339 * 5825 Web service QoS dataset
339 users (Planetlab nodes)
5825 real-world Web services
[ref.

22. Experiments Metrics MAE (Mean Absolute Error):
NMAE (Normalized Mean Absolute Error):
A smaller MAE or NMAE value means a better performance

23. Performance Comparison
Experiments
Performance Comparison
Compared approaches:
UPCC: user-based CF approach using Pearson Correlation Coefficient (PCC) [Shao2007]
IPCC: item-based CF approach using PCC [Zheng2011]
WSRec: WSRec is a model that integrates both UPCC and
IPCC systematically [Zheng2011]
PMF: PMF namely probabilistic matrix factorization is a state-of- the-art MF model for recommendation task [Salakhutdinov2008]
Our approaches:
UACR: User Asymmetric Correlation based Regularization
SACR: Service Asymmetric Correlation based Regularization
HACR: Hybrid Asymmetric Correlation based Regularization

24. Performance Comparison
Experiments
Performance Comparison
Compared approaches:
In order to keep consistence with the real-world QoS data sparsity, we randomly choose different number of entries from 5825 Web services, named Given Service (GS).
1= 2=30, 1= 2 = 30, density=10%, dimensionality=10, GS from 1000 to 3000 with a step value of 1000.

25. Experiments Impact of Top-K
In our asymmetric correlation regularized matrix factorization approaches, the parameter top-K is employed to control the number of popular entries which include users and services.
1= 2=30, 1= 2 = 30 , GS=1000, dimensionality=10, density=10%, and top-K from 10 to 100 with a step size of 10.

26. Experiments Impact of Training Matrix Density
means how many historical data of user-service QoS matrix we use
1= 2=30, 1= 2 = , =30, top-K=70, GS=1000, GS=3000, dimensionality=10 and the density from 1% to 10% with a step of 1%.

27. Outline Introduction Problem Description
Asymmetric correlation regularized matrix factorization
Experiments
Conclusion & Future Work

28. Conclusion & Future Work
In this paper, we employed asymmetric correlation among users and services to enhance the prediction accuracy in Web service recommendation
UACR, SACR, and HACR
Future Work
Continue to optimize our models in terms of stability and prediction accuracy
more QoS properties will be considered
other contextual features of Web services (e.g., location, time, etc.) can be integrated

29. Thank you!
Q&A

30. Problem Description MF Definition 1 (Co-occurrence criterion)
Definition 2 (Dependent relation)
Co-occurrence criterion describes the co-invoked services which are invoked by user pairs or co-invoking users who invoke service pairs, shown in Fig. 2(b) and Fig. 2(c) respectively.
(f)
(b)
(d) MF
(c)
(e)
(g)

31. Problem Description MF Definition 1 (Co-occurrence criterion)
Definition 2 (Dependent relation)
Dependent relation contains not only the co-occurrence criterion of each pair but also the co-occurrence criterion of the other pairs. Dependent relation can be obtained by constructing the stochastic matrix illustrated in Fig. 2(d) and Fig. 2(e) respectively.
(f)
(b)
(d) MF
(c)
(e)
(g)

32. Experiments Impact of Top-K
In our asymmetric correlation regularized matrix factorization approaches, the parameter top-K is employed to control the number of popular entries which include users and services.
1= 2=30, 1= 2 = 30 , GS=1000, dimensionality=10, density=5%, density=10%, and top-K from 10 to 100 with a step size of 10.

33. Experiments Impact of Beta
In our approaches, 1 and 2 are the regularization parameters to determine how much the asymmetric correlation of users and services influences to the objective functions respectively.
1= 2=30, 1= 2 = , top-K=70, GS=1000, dimensionality=10, density=5% and density=10% respectively, and  from 10 to 100 with a step value of 10.

34. Impact of Dimensionality
Experiments
Impact of Dimensionality
dimensionality controls how many latent features related to MF.
1= 2=30, 1= 2 = , =30, top-K=70, GS=1000, density=5% and density=10%
respectively. And we also tune the dimensionality from 10 to 100 with a step size of 10.

35. B. Asymmetric Correlation based Regularization
1. User Asymmetric Correlation based Regularization
(8)
(9)

36. B. Asymmetric Correlation based Regularization
1. User Asymmetric Correlation based Regularization
(8)
(10)

37. B. Asymmetric Correlation based Regularization
2. Service Asymmetric Correlation based Regularization
(11)
(12)

38. B. Asymmetric Correlation based Regularization
2. Service Asymmetric Correlation based Regularization
(11)
(13)

39. B. Asymmetric Correlation based Regularization
3. Hybrid Asymmetric Correlation based Regularization
(14)

40. B. Asymmetric Correlation based Regularization
3. Hybrid Asymmetric Correlation based Regularization
(15)