1. 1 RegionKNN: A Scalable Hybrid Collaborative Filtering Algorithm for Personalized Web Service Recommendation Xi Chen, Xudong Liu, Zicheng Huang, and Hailong Sun School of Computer Science and Engineering Beihang University Beijing, China

2. 2 Outline Introduction Motivation RegionKNN Algorithm Experiments Conclusion and Future Work

3. 3 1. Introduction

4. 4 Introduction Current situation –More than 25,000 public available services (seekda.com) –About 200,000 related documents Goal of service recommendation –Optimal QoS –User preference Current method: Collaborative Filtering (CF) –predict and recommend the potential favorite items for a particular user by using rating data collected from similar users. If Alice and Bob both like X and Alice likes Y then Bob is more likely to like Y Problems –Characteristics of QoS are neglected –Online performance need to be improved

5. 5 2. Motivation

6. 6 A Motivating Scenario Some QoS properties (e.g. availability, response time) highly correlate to users’ physical locations. Email Filtering WS Email Filtering WS

7. 7 3. RegionKNN Algorithm

8. 8 What’s RegionKNN Hybrid CF Algorithm –recommend web services with optimal QoS to the active user with consideration of the region factor Two phases of RegionKNN –Region model building (offline) Region-sensitive services identification Region aggregation –Service recommendation (online) (modified KNN) Neighbor selection QoS Prediction I take response time/round trip time (RTT) as an example to describe our algorithm

9. 9 3.1 Region model

10. 10 Region Model Region –a group of users who are closely located with each other and have similar RTT values u5u5 u 19 u 2 u 22, u 8 u 1, u 3 Service A Service B Service X

11. 11 Input Dataset User-Service RTT Matrix: m services, n users The set of non-zero RTTs of service s {R 1 (s), R 2 (s),…, R k (s)} collected from all users is a sample from population R. s1s1 s2s2 … smsm u1u1 0245 … 20078 u2u2 2023342 …539 …………… unun 03040 … 498 RTT is much longer than others

12. 12 Region-sensitive Services Identification To estimate the mean μ and the standard deviation σ of R, we use: Median: the numeric value separating the higher half of a sample from the lower half. e.g. {120, 128, 200, 250, 258, 2000, 3500} median = 250 MAD: the Median of the Absolute Deviations from the sample's median. e.g. {120, 128, 200, 250, 258, 2000, 3500}  {8, 50, 122, 130, 1750, 2250} MAD = 130

13. 13 Region-Sensitive Service –Let R = {R 1 (s), R 2 (s),…, R k (s)} be the set of RTTs of service s provided by users from all regions. Service s is a sensitive service to region M iff {120, 128, 200, 250, 258, 2000, 3500} u 1 u 3 u 5 u 19 u 2 u 22 u 8 u5u5 u 19 u 2 u 22 u 8 u 1, u3 Service A Region-sensitive services Identification

14. 14 Definition Region Sensitivity Sensitive Region –Region M is a sensitive region iff regSen >λ. Region center – the median vector of all the RTT vectors provided by users in a region

15. 15 Region Aggregation Why? –Users only provide limited number of QoS values, the sparse dataset always leads to poor recommendation. How? –It treats users with similar IP addresses as a region at the outset – In each iteration, the two most similar and non- sensitive regions are selected and aggregated, if their similarity exceeds threshold μ. –It executes at most N-1 steps (N is the number of regions at the outset), in case that all regions are non- sensitive, extremely correlates to each other and finally aggregates into one region.

16. 16 Region Similarity The similarity between region M and N is measured by the similarity of the two centers. Similarity by Pearson Correlation Coefficient (PCC) s1s1 s2s2 s3s3 s4s4 s5s5 cmcm 12500 cncn 00513 By PCC, the similarity is of the two regions is 1

17. 17 Region Similarity PCC often overestimates the similarity when the two regions have few co-invoked services. To adjust it, we use: s1s1 s2s2 s3s3 s4s4 s5s5 cmcm 12500 cncn 00513 By adjustment, the similarity of the two regions is 0.2

18. 18 3.2 Service Recommendation

19. 19 Neighbor Selection Neighbors: users with similar QoS experiences Advantages of region-based neighbor selection –Do not need to search the entire dataset, thousands of users are clustered into a certain number of regions –The feature of the group of users in a region is represented by the region center

20. 20 QoS Prediction To calculate the RTT prediction for the active user u and service s i Get the active user’s IP address and find the region the user belongs to. If no appropriate region is found, the active user will be treated as a member of a new region. Identify whether service s i is sensitive to the specific region. If it is region-sensitive, then the prediction is generated from the region center:

21. 21 QoS Prediction (cont.) Otherwise, use adjusted PCC to compute the similarity between the active user and each region center that has evaluated service s i, and find up to k most similar centers {c 1, c 2,…, c k }. If the active user’s region center has the RTT value of s i, i.e., the prediction is computed using the equation:

22. 22 QoS Prediction (cont.) Otherwise, Previous CF-based web service recommendation algorithms use the following equation, to predict the missing QoS value. This equation is based on the assumption that each user’s rating range is subjective and comparatively fixed, while it is not applicable in our context.

23. 23 Time complexity Model building (offline) –The time complexity of region aggregation algorithm is O(N 2 logN), and N is the number of regions at the outset. QoS prediction (online) – Let l be the number of regions, m the number of web services, and n the number of users. In the online part, O(l) similarity weight calculations are needed, each of which takes O(m) time. Therefore, the online time- complexity is O(lm)≈O(m). Previous user-based CF algorithm has O(mn) online time complexity.

24. 24 4. Experiments

25. 25 Experiments Dataset –a subset of WSRec with 300,000 RTT records –3000 users –100 services Evaluation Metric – R u (s) denotes the actual RTT of web service s given by user u – denotes the predicted one – L denotes the number of tested services Dataset: http://www.wsdream.net

26. 26 MAE Performance

27. 27 Impact of λ and μ

28. 28 Impact of neighborhood size K

29. 29 Impact of Data Sparsity

30. 30 5.Conclustion and Future Work

31. 31 Conclusion and Future Work Conclusion –a new region model for clustering users and identifying region-sensitive web services –a hybrid model-based and memory-based CF algorithm for web service recommendation, which significantly improves the recommendation accuracy –We demonstrate RegionKNN’s scalability advantage over traditional CF algorithms via time-complexity analysis Future Work –Investigation of more QoS properties and their variation with time –Internal relations between QoS properties

32. 32