1. 1 Shooting Stars in the Sky:An Online Algorithm for Skyline Queries 作者： Donald Kossmann Frank Ramsak Steffen Rost 報告：黃士維

2. 2 Outline Introduction An Online Algorithm for two-dimention Skylines The NN Algorithm for d-dimention Skylines Performance Experiments Conclusion

3. 3 Introduction Skyline Bold points Skyline Queries Online Skyline Computation Related work

4. 4 Introduction-Skyline Skyline queries ask for a set of interesting points from a potentially large set of data points. -ex:a cheaper,nearer,better food hotel. Skyline point-also called “bold point” -a point dominate another point in at least one dimension.

5. 5 Introduction- Bold points The bold points respresent this hotels which part of Skyline.

6. 6 Skyline Queries Skyline queries can involve more than 2 dimensions and they could depend on current position of a user. -to find a near,good food,cheap restaurant,but the user moves in the same time. If the user moves on,the Skyline should be re-computed for this user.

7. 7 Online Skyline Computation-1 An online algorithm would probably take much longer than a batch-oriented algorithm to produce the full Skyline,but an online algorithm would produce a subset of the Skyline very quickly.

8. 8 Online Skyline Computation-2 Six properties of an online algorithm: 1. The first result should be return instantaneously. 2. The algorithm should produce the full Skyline. 3. The algorithm should only return pionts which are part of the Skyline. 4. The algorithm should be fair. 5. The algorithm should be well-controled. 6. The algorithm should be universal in Skyline queries and DB.

9. 9 Related work All these algorithms work in a batch-oriented way. Tan et al. proposed two progressive Skyline algorithms. - Bitmap algorithm, the order in which Skyline points are returned depends on the clustering of the data. -B-tree algorithm, the order in which points are returned depends on the value distribution of the data. Both algorithms require that all interesting properties are materialized in the database.

10. 10 An Online Algorithm for two- dimention Skylines Three asumptions Basic Observations An example Discussion

11. 11 Three asumptions All values are positive real numbers. There are no duplicates in the data set. We try to fine minimal points.

12. 12 Basic Observations D is 2-dimentional data set. n ∈ D,is a nearest neighbor of O = (0, 0) according to f. - n is in the Skyline Observation 1: n must be part of the Skyline. Observation 2: n is in the Skyline of D.

13. 13 Basic Observations If we partition the data set, then it is sufficient to look for Skyline points using nearest neighbor search in each region separately. This observation gives rise to a divide & conquer algorithm using nearest neighbor search.

14. 14 An example-first step Region 1:corner sign,smaller y. Region 2:circle sign,smaller x, Region 3:triangle sign,greater x,y.

15. 15 An example-second step Divide & conquer algorithm. In this way, the algorithm continues until all Skyline points are retrieved.

16. 16 Algorithm Description

17. 17 Discussion-1 Six requirements: We believe that Skyline queries that involve two dimensions are the most common case. The algorithm will find all points of the Skyline. All nearest neighbors found by the NN algorithm are part of the Skyline.

18. 18 Discussion-2 The NN algorithm produces results from the whole range of results very quickly. The NN algorithm can continue to produce Skyline points using the new distance function without any adaptions. The NN algorithm is universal.

19. 19 The NN Algorithm for d-dimention Skylines Introduction Laisser-faire Propagate Merge Fine-grained Partitioning Hybrid Approaches

20. 20 Introduction-1 Find NN and compute

21. 21 Introduction-2 Two observations from Section 2.1 can be generalized to three and more dimensions. Region 1 (xn, ∞, ∞) Region 2 (∞, yn, ∞) Region 3 (∞, ∞, zn) Etc… p will be produced by the NN algorithm twice.

22. 22 Laisser-faire If the nearest neighbor p is found in the hash table, it is a duplicate and it is not output. The big disadvantage of this algorithm is that it results in a great deal of wasted work.

23. 23 Propagate We can do this by “propagation”. Whenever the NN algorithm finds a Skyline point, it scans the whole to-do list in order to find those regions in the to-do list that contain that point. The big advantage of this technique is that it completely avoids wasted work to find duplicates

24. 24 Merge Assume that two regions: a = (a1, a2,..., ad ) b = (b1, b2,..., bd) are in the to-do list. We can merge these two regions into a single region:a ⊕ b = (max(a1, b1), max(a2, b2),..., max(ad, bd)) A particular situation arises, if a supersedes b, a = a ⊕ b and thus b can be simply discarded from the to-do list.

25. 25 Fine-grained Partitioning In Figure 5, for instance, we could partition into 8 nonoverlapping regions of which 6 regions would be relevant for further processing. Implementing this approach, however, results in a sharp growth of the number of regions in the to-do list.

26. 26 Hybrid Approaches As mentioned above, merge can be combined with both the laisser-faire and the propagate approach. Another option would be to start and propagate duplicates until the to-do list has reached a certain size.

27. 27 Performance Experiments Experimental Environment Points in database Variant algorithms Two-dimensional Skyline Queries Quality of Results Comparing Algorithm Variants

28. 28 Experimental Environment 167MHz processor 128MB of main memory OS:Solaris 8 17.1G HD Program language: C++

29. 29 Points in database 100,000 points VS 1million points Points are generated using one of the following three value distributions: Corr:points which are good in one dimension tend to be good in otherdimensions, too. Anti:points which are good in one dimension are bad in at least one other dimension. Indep:points are generated using a uniform distribution.

30. 30 Variant algorithms NN : R-tree [BKSS90] in order to carry out nearest neighbor search. D&C : Kung’s divide & conquer algorithm extended by m-way partitioning and “Early Sky-line”. BNL : The block-nested-loops algorithm with a self- organizing list. Bitmap : scans the database and uses bitmaps in order to detect whether a point is part of the Skyline. B-tree : used a light-weight implementation of this algorithm that does not require the use of extended B-trees.

31. 31 Two-dimensional Skyline Queries

32. 32 Two-dimensional Skyline Queries It can be seen that the NN algorithm scales best. The general trends were the same: (1) The NN algorithm was the clear overall winner. (2) The Bitmap algorithm was in the same ball park as the BNL and D&C algorithms. (3) The B-tree algorithm showed good performance for the correlated and independent databases, but terrible performance for the anti-correlated database.

33. 33 Quality of Results Figure 6a shows the full Skyline for the anti- correlated database for d = 2. Figure 6b shows the first 10 points returned by the NN algorithm (without user interaction). Figure 6c shows the first 10 points returned by the B-tree algorithm. It can be seen that the NN algorithm gives a good big picture of the Skyline.

34. 34 Quality of Results Computing the full Skyline takes a long time,regardless what algorithm is used. an online algorithm is particularly important for interactive applications in this scenario: (1) toselect relevant points from the Skyline (rather than flooding the user with results) (2) to give the firstanswers quickly.

35. 35 Quality of Results B-tree algorithm produces the results very quickly. If the user gives preferences, the B-tree algorithm cannot adapt well. The Bitmap algorithm is not competitive in either aspect: (1) It produces Skyline points at approximately the same rate as the NN algorithm. (2) It is no match to the NN algorithm in terms of interactivity because it produces Skyline points in a somewhat random order (based on the clustering of the data).

36. 36 Comparing Algorithm Variants Figure 7 shows the performance of four variants for a 5-dimensional Skyline with an anti- correlated database and 1 million points. measured the following variants: Laisser-faire Propagate Merge:laisser-faire and merge regions whenever a duplicate is found. Hybrid: propagate to the first 20 % entries of the to-do list.

37. 37 Comparing Algorithm Variants

38. 38 Conclusion Skyline queries are important for several database application. All algorithms have their particular virtues. The NN algorithm is the only algorithm that gives the user control over the process and allows the user to give preferences. The B-tree algorithm gives “extreme” points preference (i.e., points good in one dimension) and returns points which are good in many dimensions very late. The Bitmap algorithm scans the database and uses Bitmaps in order to detect whether a point is part of the Skyline.

39. 39 Conclusion(cont.) future work: First, improve the performance of the NN algorithm by exploiting new techniques for nearest neighbor search. Second, we would like to investigate special- purpose main-memory index structures Third, we would like to investigate how the NN algorithm can be combined with other algorithms.

40. 40 Thanks for your listening!!