Ranking in DB Laks V.S. Lakshmanan Depf. of CS UBC
10/16/20152 Why ranking in query answering? 1/3 Mutimedia data – fuzzy querying: e.g., “find top 2 red objects with a soft texture”. ObjScore D0.85 B0.80 A0.75 E0.65 C0.60 ObjScore A0.9 D0.8 C0.4 B0.3 E0.1 Combine scores Overall score
10/16/20153 Why ranking? 2/3 IR: “find top 5 documents relevant to `computational’, `neuroscience’ and `brain theory’. –IR systems maintain full text indexes; inverted lists of docs w.r.t. each keyword. –Same Q/A paradigm as before. Buying a home: several criteria – price, location, area, #BRs, school district. ORDER BY query in SQL. Finding hotels while traveling.
10/16/20154 Why ranking? 3/3 Data stream, e.g., of network flow data: “find 10 users with the max. BW consumption and max. #packets communicated”. – score may be complex aggregation of these two measures. In a social net, find 5 items tagged as most relevant to “lawn mowing” and blonging to users socially close to the seeker. And now, find top-k recs (recommender systems). etc. Fagin et al. – pioneering papers PODS’96, 01, JCSS Burgeoned into a field now. Focus on middleware algorithm, which given a score combo. function, computes top-k answers by probing diff. subsystems (or ranked lists).
10/16/20155 Computational model Naïve method. How to compute top-K efficiently? Access methods: –Sorted access (sequential access) [SA]. –Random access [RA]. Diff. optimization metrics: –Overall running time of algorithm. –SA < RA: minimize RAs. –RA not possible # : avoid RAs. –Combined optimization. Has led to a variety of algorithms. Memory vs. disk model. For the most part, assume score agg. is a monotone function; use SUM in examples. #: typical in IR systems.
10/16/20156 Fagin’s Algorithm (FA) m lists sorted by descending scores. Access (SA) all lists in parallel. –For each new object seen, fetch scores from other lists by RA. Overall score t(x) = t(x1, …, xm). Store (obj, score) in set Y. –Remember each object seen (under SA) in all lists in set H. Repeat until |H| >= K. Sort Y in descending order of scores, breaking ties arbitrarily, and output top K.
10/16/20157 Example of FA L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ Answers seen in >=1 list, i.e., Y unsorted. B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) Answers seen (under SA) in all 4 lists, i.e., H.
10/16/20158 Example of FA L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) Answers seen in >=1 list, i.e., Y unsorted. Answers seen (under SA) in all 4 lists, i.e., H.
10/16/20159 Example of FA L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ 3.30 B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) Answers seen in >=1 list, i.e., Y unsorted. Answers seen (under SA) in all 4 lists, i.e., H.
10/16/ Example of FA L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ 3.30 B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) Answers seen in >=1 list, i.e., Y unsorted. Answers seen (under SA) in all 4 lists, i.e., H. 2.65
10/16/ Example of FA L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ 3.30 B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) Answers seen in >=1 list, i.e., Y unsorted. Answers seen (under SA) in all 4 lists, i.e., H
10/16/ Example of FA L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) Answers seen in >=1 list, i.e., Y unsorted. Answers seen (under SA) in all 4 lists, i.e., H
10/16/ Example of FA L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) Answers seen in >=1 list, i.e., Y unsorted. Answers seen (under SA) in all 4 lists, i.e., H
10/16/ Example of FA L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) Answers seen in >=1 list, i.e., Y unsorted. Answers seen (under SA) in all 4 lists, i.e., H H
10/16/ Example of FA L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) Answers seen in >=1 list, i.e., Y unsorted. Answers seen (under SA) in all 4 lists, i.e., H H, G
10/16/ Example of FA L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) Answers seen in >=1 list, i.e., Y unsorted. Answers seen (under SA) in all 4 lists, i.e., H H, G, B, C 2.05 |H| = 4.
10/16/ FA Example concluded A, F – not seen in any list. Yet, we are sure they can’t make it to top-4. Why? Based on where the cursors are now, what’s the max. possible score for A, F? What assumptions are being made about t()? FA is shown to be optimal with very high probability [Fagin: PODS 1996]. But can be beaten by other algorithms on specific inputs. What about buffer size?
10/16/ Threshold Algorithm Do parallel SA on all m lists. For each object x seen under SA in a list, fetch its scores from other lists by RA and compute overall score. If |Buffer| < K add x to Buffer; Else if score(x) <= k-th score in buffer, toss; Else replace bottom of buffer with (x, score(x)) & resort. Stop when threshold <= k-th score in buffer. Threshold := t(worst score seen on L1, …, worst score seen on Lm). Output the top-K objects & scores (in buffer).
10/16/ TA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30)
10/16/ TA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30)
10/16/ TA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) 3.30 Threshold Bar: x1 x2 x3 x
10/16/ TA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) 3.30 Threshold Bar: T = x1 x2 x3 x
10/16/ TA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) 3.30 Threshold Bar: T=3.60. x1 x2 x3 x X 3.05 X 3.15
10/16/ TA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) 3.30 Threshold Bar: T=3.30. x1 x2 x3 x X 3.05 X X
10/16/ TA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) 3.30 Threshold Bar: T=3.10. x1 x2 x3 x X 3.05 X X
10/16/ TA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) 3.30 Threshold Bar: T=2.90. ==> can stop! x1 x2 x3 x X 3.05 X X
10/16/ TA Remarks
TA is Instance Optimal 10/16/201528
TA IO Proof (contd.) 10/16/201529
Proof (contd.) 10/16/201530
Proof (contd.) 10/16/201531
Proof (contd.) 10/16/201532
Proof (concluded) 10/16/201533
10/16/ No Random Access Algorithm What if RA > SA or RA wasn’t allowed? Do SA on all lists in parallel. At depth d: –Maintain worst scores x1, …, xm. –x any object seen in lists {1, …, i}. Best(x) = t(x1, …, xi, xi+1, …, xm). Worst(x) = t(x1, …, xi, 0, …, 0). –TopK contains K objects with max worst scores at depth d. Break ties using Best. M = k-th Worst score in TopK. –Object y is viable if Best(y) > M. Stop when TopK contains >=K distinct objects and no object outside TopK is viable. Return TopK.
10/16/ NRA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) [0.95, 3.90] [1.00, 3.90] [0.95, 3.90] [1.00, 3.90]
10/16/ NRA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) [0.95, 3.65] [1.80, 3.65] [1.90, 3.75] [1.00, 3.65] [0.90, 3.60] [0.95, 3.60]
10/16/ NRA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) [1.85, 3.40] [1.80, 3.55] [1.90, 3.65] [1.85, 3.40] [0.90, 3.35] [1.80, 3.35] [0.70, 3.30]
10/16/ NRA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) [3.30, 3.30] [1.80, 3.45] [2.70, 3.55] [1.85, 3.30] [1.75, 3.20] [1.80, 3.25] [0.70, 3.15]
10/16/ NRA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) [3.30, 3.30] [1.80, 3.35] [2.70, 3.50] [2.60, 3.20] [1.75, 3.10] [3.15, 3.15] [1.50, 3.00]
10/16/ NRA Example L1L2L3L4 H(0.95) C(0.80 ABCDEFGHIJABCDEFGHIJ B(0.90) E(0.85) G(0.75) I(0.70) D(0.65) A(0.60) J(0.55) F(0.50) J(1.00) C(0.95) G(0.85) H(0.80) E(0.75) B(0.75) F(0.60) A(0.50) D(0.40) I(0.30) C(0.95) J(0.80) D(0.70) H(0.65) G(0.60) B(0.55) I(0.50) E(0.45) F(0.40) A(0.30) E(1.00) G(0.95) H(0.90) B(0.85) D(0.80) C(0.70) A(0.65) I(0.55) F(0.45) J(0.30) [3.30, 3.30] [1.80, 3.20] [3.40, 3.40] [2.60, 3.15] [3.05, 3.05] [3.15, 3.15] [1.50, 2.95] [0.70, 2.70]
10/16/ NRA Features What sort of t() do we need to assume, for NRA to work correctly? How large can the buffers get? How does the amount of bookkeeping compare with TA? NRA is instance optimal over algo’s not making RA (and of course, not making wild guesses).
10/16/ Combined optimization What if we are told cost(RA) = .cost(SA)? Can we find algo’s better than NRA and TA in this case? Combined algorithm = CA. (See Fagin et al.’s paper for details.)
10/16/ Worrying about I/O cost Based on Bast et al. VLDB Inverted lists of (itemID, score) entries in desc. score order, as usual, but on disk. Blocks sorted by itemID; across blocks still in desc. score order. Inverted Block Index (IBI) Algorithm. What is an IBI?
10/16/ A Motivating Example List 1 List 2 List 3 Doc17 : 0.8 Doc25 : 0.7 Doc83 : 0.9 Doc78 : 0.2 Doc38 : 0.5 Doc17 : 0.7. Doc14 : 0.5 Doc61 : 0.3 · Doc83 : 0.5 · · · · · Doc17 : 0.2 · · · · Round 1 (SA on 1,2,3) Doc17 : [0.8, 2.4] Doc25 : [0.7, 2.4] Doc83 : [0.9, 2.4] unseen: ≤ 2.4
10/16/ A Motivating Example List 1 List 2 List 3 Doc17 : 0.8 Doc25 : 0.7 Doc83 : 0.9 Doc78 : 0.2 Doc38 : 0.5 Doc17 : 0.7. Doc14 : 0.5 Doc61 : 0.3 · Doc83 : 0.5 · · · · · Doc17 : 0.2 · · · · Round 1 (SA on 1,2,3) Doc17 : [0.8, 2.4] Doc25 : [0.7, 2.4] Doc83 : [0.9, 2.4] unseen: ≤ 2.4 Round 2 (SA on 1,2,3) Doc17 : [1.5, 2.0] Doc25 : [0.7, 1.6] Doc83 : [0.9, 1.6] unseen: ≤ 1.4
10/16/ A Motivating Example List 1 List 2 List 3 Doc17 : 0.8 Doc25 : 0.7 Doc83 : 0.9 Doc78 : 0.2 Doc38 : 0.5 Doc17 : 0.7. Doc14 : 0.5 Doc61 : 0.3 · Doc83 : 0.5 · · · · · Doc17 : 0.2 · · · · Round 1 (SA on 1,2,3) Doc17 : [0.8, 2.4] Doc25 : [0.7, 2.4] Doc83 : [0.9, 2.4] unseen: ≤ 2.4 Round 2 (SA on 1,2,3) Doc17 : [1.5, 2.0] Doc25 : [0.7, 1.6] Doc83 : [0.9, 1.6] unseen: ≤ 1.4 Round 3 (SA on 2,2,3!) Doc17 : [1.5, 2.0] Doc83 : [1.4, 1.6] unseen: ≤ 1.0
10/16/ A Motivating Example List 1 List 2 List 3 Doc17 : 0.8 Doc25 : 0.7 Doc83 : 0.9 Doc78 : 0.2 Doc38 : 0.5 Doc17 : 0.7. Doc14 : 0.5 Doc61 : 0.3 · Doc83 : 0.5 · · · · · Doc17 : 0.2 · · · · Round 1 (SA on 1,2,3) Doc17 : [0.8, 2.4] Doc25 : [0.7, 2.4] Doc83 : [0.9, 2.4] unseen: ≤ 2.4 Round 2 (SA on 1,2,3) Doc17 : [1.5, 2.0] Doc25 : [0.7, 1.6] Doc83 : [0.9, 1.6] unseen: ≤ 1.4 Round 3 (SA on 2,2,3!) Doc17 : [1.5, 2.0] Doc83 : [1.4, 1.6] unseen: ≤ 1.0 Round 4 (RA for Doc17) Doc17 : 1.7 all others < 1.7 done! Note deviation from round-robin.
10/16/ IBI Algorithm Same setting as NRA/CA, except use IBI. Maintain two lists: Top-K items (T = d1, …, dk) and StillHaveASHot (SHASH) (S = dk+1, …, dk+q) items. Pos_i = curr cursor position on list Li. high_i = score in Li at curr cursor position (upper bounds score of unseen items). For items d in S: –Which attr scores are known E(d). –Which attr scores are unknown E~(d). –Worst(d) = total score from E(d). –Best(d) = Worst(d) + {high_i(d) | i E~(d)}. (Exactly as Fagin.)
10/16/ IBI Algorithm (contd.) In each round, compute: –min-k = min{Worst(d) | d T}. –bestscore that any unseen doc can have = sum of all high_i’s. –For dj S: def_j = min-k – worst(d_j). [denotes deficit below qualification level for top-k.] T sorted in desc. Worst(); S sorted in desc. Best(). [sorting on (score, ItemID) for fast processing.] Invatiant: min-k >= max{Worst(d) | d S}. Termination: when min-k >= max{Best(d) | d S}. Can remove an obj from S whenever its Best <= min-k. stop when S = {}. Early termination AND minimal bookkeeping are BOTH important for performance.
10/16/ More on IBI Framework Instead of scheduling SAs using RR, use a differential approach for diff. lists based on expected score reductions at future cursor positions (Knapsack). Do SA*RA*. Order RAs based on estimated Prob[dj can get into top-k answers].