1. Probabilistic Contextual Skylines D. Sacharidis 1, A. Arvanitis 12, T. Sellis 12 1 Institute for the Management of Information Systems — “Athena” R.C., Greece 2 National Technical University of Athens, Greece

2. static vs. dynamic skyline (another) hotels example Price and Distance are Statically Preferred (SP) attributes with fixed preferences: lower is better Ignore Amenity (assume all amenities are equally preferable) h 4 and h 5 are in the static skyline Amenity is a Relatively Preferred (RP) attribute, preferences are defined per query h 3, h 4 and h 5 are in the dynamic skyline

3. contextual skyline just like dynamic skyline, but preferences are associated with some context what if no preferences are specified for the current context? two issues: can we extract them from previous situations? what does it mean to be in the skyline?

4. extract preferences key idea is to combine preferences from similar contexts to the current first assess the similarity between the current context C q and all past contexts C j : contexts may have conflicting preferences, and we model uncertainty with probabilities value u is better than v for the context C i with some probability probabilities can be extracted based on context similarities

5. probabilistic contextual skylines dominance relationships are uncertain (assuming independence among attributes) tuple t dominates t’ for the context C i with probability the skyline probability of a tuple is defined as probabilistic contextual skyline query, p-CSQ, returns all tuples with

6. example skyline probability (1) (2) (3)

7. non-indexed algorithms (1/2) for RP attributes (unlike standard and tuple-probabilistic skylines) no monotonic visit order exists transitivity is not preserved we have to apply BNL-like methods (not SFS, BBS, etc.) Basic Iterative Algorithm (BIA) for each tuple scan the database and compute skyline probability (abort when below threshold)

8. non-indexed algorithms (2/2) Candidate Selection Algorithm (CSA) it identifies candidates group tuples by their values on the RP attributes tuples that are dominated in an RP-group have 0 probability tuples that are in the skyline w.r.t. only the SP attributes have probability 1 CSA applies BIA only for the candidates (needs to check them against all tuples, though)

9. index-based algorithms (1/3) the algorithms only consider the candidates Basic Group Counting (BGC) idea: tuples in an RP group that dominate a candidate t contribute the same probability use COUNT aR-tree per RP group but, don’t just issue a range query per tree… we don’t care about the exact count if tuple’s probability is below threshold instead visit nodes from all trees in parallel and use a single priority queue the node from the tree which has the highest expected probability of dominating tuple t has the largest priority

10. index-based algorithms (2/3) Super Group Counting (SGC) there can be a lot of RP groups with only a few tuples to mitigate this, assign groups to super groups use a GROUP-COUNT aR-tree per super-group entry: where c[g j ] is the number of tuples beneath node e i that belong to the j-th group same algorithm as BGC… you only need to redefine the expected dominance probability to take into account multiple groups

11. index-based algorithms (3/3) Batch Counting Algorithm (BCA) all previous algorithms compute the skyline probability of one tuple at a time BCA examines candidates in batches (as many as fit in memory) extra bookkeeping with each heap entry to avoid double counting e 1 is deheaped e 1 + dominates t 1 but not t 2 entire e 1 contributes to t 1, but for t 2 we need to expand e 1 and enheap its children e 2, e 3 also remember e 1 + with e 2, e 3

12. Experiments Non-indexed BIA, CSA Index-based SGC, BCA

13. Total Time vs. Dataset Cardinality Non-indexed BIA, CSA Index-based SGC, BCA

14. Total Time vs. RP domain size Non-indexed CSA Index-based SGC, BCA

15. Total Time vs. Dimensionality Non-indexed CSA Index-based SGC, BCA

16. thank you!