A Survey on Spatial-Keyword Search

Transcription

1 A Survey on Spatial-Keyword Search (COMP 6311C Advanced Data Management) Nikolaos Armenatzoglou 06/03/2012

2 Outline Problem Definition and Motivation Query Types Query Processing Techniques Indices Algorithms Strengths and Weaknesses Conclusion 2

3 A Survey on Spatial-Keyword Search PROBLEM DEFINITION AND MOTIVATION 3

4 Problem Definition and Motivation Definition: Spatio-Textual Object (STO) A STO o is an object that consist of a textual description (o.text) and a spatial location (o.loc). Example o 7 ( Shoes ) o 9 ( Cinema ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) o 2 ( Italian Restaurant ) o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) Hong Kong Times Square 4

5 Problem Definition and Motivation Definition: Spatial-Keyword Search (SKQ) Given a spatial constraint Q.loc and a set of keywords Q.text, a SKQ Q returns a list (ranked or not) of the STOs whose location satisfies Q.loc and whose textual description is relevant to Q.text. o 7 ( Shoes ) o 9 ( Cinema ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) o 2 ( Italian Restaurant ) o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) Hong Kong Times Square 5

6 Problem Definition and Motivation Definition: Spatial-Keyword Search (SKQ) Given a spatial constraint Q.loc and a set of keywords Q.text, a SKQ Q returns a list (ranked or not) of the STOs whose location satisfies Q.loc and whose textual description is relevant to Q.text. o 7 ( Shoes ) o 9 ( Cinema ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) q o 2 ( Italian Restaurant ) Current Location: (x, y) Keywords: Restaurant 1-NN o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) Hong Kong Times Square 6

7 A Survey on Spatial-Keyword Search QUERY TYPES 7

8 Query Types Basic Spatial-Keyword Range Query (SKR) Spatial-Keyword Nearest Neighbor Query (SKNN) Top-k Spatial Keyword Query (Top-k SK) Variations Approximation Spatial-Keyword Range Query (ASK Range) Top-k Spatial Boolean query (ksb) Spatial-Keyword Auto-Completion (SK AC) 8

9 Query Types (Basic) Spatial-Keyword Range Query (SKR) Given a region Q.loc and a set of keywords Q.text, a SKR query Q returns all STOs within Q.loc that contain all keywords in Q.text o 7 ( Shoes ) o 9 ( Cinema ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) o 2 ( Italian Restaurant ) 100 o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) Example Q.loc = [(200, 0), (500, 300)] and Q.text = { Chinese, Restaurant }. Ans(Q) = {o 1, o 6 } 9

10 Query Types (Basic) Spatial-Keyword Range Query (SKR) Given a region Q.loc and a set of keywords Q.text, a SKR query Q returns all STOs within Q.loc that contain all keywords in Q.text o 7 ( Shoes ) o 9 ( Cinema ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) o 2 ( Italian Restaurant ) 100 o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) Example Q.loc = [(200, 0), (500, 300)] and Q.text = { Chinese, Restaurant }. Ans(Q) = {o 1, o 6 } 10

11 Query Types (Basic) Spatial-Keyword Nearest Neighbor Query (SKNN) Given a query point Q.loc, a set of keywords Q.text and a positive integer k, a SKNN query Q returns the k nearest STOs to Q.loc, whose textual description contains all keywords in Q.text o 7 ( Shoes ) o 9 ( Cinema ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) o 2 ( Italian Restaurant ) 100 o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) Example Q.loc = (500, 200), Q.text = { Restaurant, Chinese }, k = 1. Ans(Q): {o 1 }. 11

12 Query Types (Basic) Spatial-Keyword Nearest Neighbor Query (SKNN) Given a query point Q.loc, a set of keywords Q.text and a positive integer k, a SKNN query Q returns the k nearest STOs to Q.loc, whose textual description contains all keywords in Q.text o 7 ( Shoes ) o 9 ( Cinema ) o 10 ( bookstore ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) Q.loc o 3 ( Restaurante de France ) o 2 ( Italian Restaurant ) ( Chinese Restaurant ) o 6 o 4 ( Italian Ristorante ) Example Q.loc = (500, 200), Q.text = { Restaurant, Chinese }, k = 1. Ans(Q): {o 1 }. 12

13 Query Types (Basic) Top-k Spatial Keyword Query (Top-k SK) Given a query point Q.loc, a set of keywords Q.text and a positive integer k, a Top-k SK query Q returns a list of k STOs ranked according to a ranking function f based on their spatial proximity to Q.loc and on their textual similarity to Q.text o 7 ( Shoes ) o 9 ( Cinema ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) o 2 ( Italian Restaurant ) 100 o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) Example f = a f IR + (1-a) f SP, where a (0, 1), f IR : text similarity and f SP : spatial proximity. Q.loc = (500, 200), Q.text = { restaurant }, k = 2 and a = 0.5. Ans(Q) = {o 2, o 1 }. 13

14 Query Types (Basic) Top-k Spatial Keyword Query (Top-k SK) Given a query point Q.loc, a set of keywords Q.text and a positive integer k, a Top-k SK query Q returns a list of k STOs ranked according to a ranking function f based on their spatial proximity to Q.loc and on their textual similarity to Q.text o 7 ( Shoes ) o 9 ( Cinema ) o 10 ( bookstore ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) Q.loc o 3 ( Restaurante de France ) o 2 ( Italian Restaurant ) ( Chinese Restaurant ) o 6 o 4 ( Italian Ristorante ) Example f = a f IR + (1-a) f SP, where a (0, 1), f IR : text similarity and f SP : spatial proximity. Q.loc = (500, 200), Q.text = { restaurant }, k = 2 and a = 0.5. Ans(Q) = {o 2, o 1 }. 14

15 Query Types (Variations) Approximate Spatial-Keyword Range Query (ASK Range) Given a region Q.loc, a set of keywords Q.text and a positive integer t, an ASK Range query Q returns the STOs in Q.loc that contain similar strings to all keywords in Q.text with edit-distance threshold t o 7 ( Shoes ) o 9 ( Cinema ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) o 2 ( Italian Restaurant ) 100 o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) Example Q.loc = [(200, 0), (500, 300)], Q.text = { Restaurant } and t = 1. Ans(Q) = {o 1, o 3, o 6 } 15

16 Query Types (Variations) Approximate Spatial-Keyword Range Query (ASK Range) Given a region Q.loc, a set of keywords Q.text and a positive integer t, an ASK Range query Q returns the STOs in Q.loc that contain similar strings to all keywords in Q.text with edit-distance threshold t o 7 ( Shoes ) o 9 ( Cinema ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) o 2 ( Italian Restaurant ) 100 o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) Example Q.loc = [(200, 0), (500, 300)], Q.text = { Restaurant } and t = 1. Ans(Q) = {o 1, o 3, o 6 } 16

17 Query Types (Variations) Top-k Spatial Boolean query (ksb) Given a query point Q.loc, a Boolean expression e on some query keywords and a positive integer k, a ksb query Q returns the k nearest STOs whose textual description satisfies e o 7 ( Shoes ) o 9 ( Cinema ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) o 2 ( Italian Restaurant ) 100 o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) Example Q.loc = (400, 300), e = restaurant Chinese and k = 2. Ans(Q): {o 5, o 2 }. 17

18 Query Types (Variations) Top-k Spatial Boolean query (ksb) Given a query point Q.loc, a Boolean expression e on some query keywords and a positive integer k, a ksb query Q returns the k nearest STOs whose textual description satisfies e o 7 ( Shoes ) o 9 ( Cinema ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) Q.loc o 1 ( Chinese Restaurant ) o 2 ( Italian Restaurant ) 100 o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) Example Q.loc = (400, 300), e = restaurant Chinese and k = 2. Ans(Q): {o 5, o 2 }. 18

19 Query Types (Variations) Spatial-Keyword Auto-Completion (SK AC) Given a point Q.loc, a string Q.text and a positive integer k, a SK AC query Q returns a list of k STOs, where Q.text is a prefix of their textual description, ranked according to a ranking function f that is based on their spatial proximity to p and several other static scores like popularity and ratings o 7 ( Shoes ) o 9 ( Cinema ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) o 2 ( Italian Restaurant ) 100 o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) Example Static scores are 1 for all STOs. Q.loc = (300, 0), Q.text = rest, k = 3. Ans(Q): {o 6, o 3, o 1 } 19

20 Query Types (Variations) Spatial-Keyword Auto-Completion (SK AC) Given a point Q.loc, a string Q.text and a positive integer k, a SK AC query Q returns a list of k STOs, where Q.text is a prefix of their textual description, ranked according to a ranking function f that is based on their spatial proximity to p and several other static scores like popularity and ratings o 7 ( Shoes ) o 9 ( Cinema ) o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) o 2 ( Italian Restaurant ) 100 o 10 ( bookstore ) o 3 ( Restaurante de France ) Example 0 ( Chinese Restaurant ) o 6 Q.loc o 4 ( Italian Ristorante ) Static scores are 1 for all STOs. Q.loc = (300, 0), Q.text = rest, k = 3. Ans(Q): {o 6, o 3, o 1 } 20

21 Query Types Query Type Spatial Categorization IR Categorization ASK Range Proximity Query Range SKR SKNN Contains all keywords Joint SKNN ASK knn Proximity Query ksb knn Boolean Query Model SK AC Auto-completion Top-k SK LkPT MkSK Moving knn Ranked Query Model RSTkNN Reverse knn mck GSK min-sum GSK min-max Aggregate Contains at least one keyword 21

22 A Survey on Spatial-Keyword Search QUERY PROCESSING TECHNIQUES 22

23 Indices Indexing Approach Query Processing Techniques On Answering SKR Query Separate indices for spatial and textual attributes SI: Separate Indices Approach Word Occurrences [Martins et al. GIR, 2005] o 7 ( Shoes ) n 4 o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) n 0 Athletic o 8 Bookstore o 10 Chinese o 1, o 6 Cinema o 9 France o 3 Greek o 5 n 3 o 9 ( Cinema ) n 1 o 1 ( Chinese Restaurant ) n 5 o 2 ( Italian Restaurant ) Italian o 2, o 4 Restaurant o 1, o 2, o 5, o 6 Restaurante o 3 Ristorante o 4 o 10 ( bookstore ) ( Chinese Restaurant ) o 6 n 6 o 3 ( Restaurante de France ) n 2 o 4 ( Italian Ristorante ) Shoes o 7, o 8 n 0 n 1 n 2 n 3 n 4 n 5 n 6 o 7 o 9 o 10 o 5 o 8 o 4 o 6 o 1 o 2 o 3 23

24 Query Processing Techniques On Answering SKR Query SI: Separate Indices Approach Query Processing 1 st Approach [Martins et al. GIR, 2005] Ans SP (Q.loc) = Answer to spatial constrains. Ans IR (Q.text) = Answer to textual constrains. Ans(Q) = Ans SP (Q.loc) Ans IR (Q.text) 2 nd Approach [Chen et al. SIGMOD, 2006] Ans SP (Q.loc) = Answer to spatial constrains. For each o Ans SP (Q.loc) check if o satisfies the textual constraints. Pros it is easy to maintain the two indices, it supports both textual, spatial and SKQ queries and updates and optimizations are performed in each index independently. Cons the pruning power of query processing algorithms is poor and leads to high I/O cost. 24

25 Indexing Approach Query Processing Techniques On Answering SKR Query Inverted Files + Grids: words occurrences contains spatial information. TS: Textual - Spatial Approach [Vaid et al. SSTD, 2005] o 7 ( Shoes ) o 9 ( Cinema ) o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) o 2 ( Italian Restaurant ) Word Occurrences Athletic (c 3,4, {o 8 }) Bookstore (c 1,2, {o 10 }) Chinese (c 4,3, {o 1 }), (c 3,1, {o 6 }) Cinema (c 1,3, {o 9 }) France (c 4,2, {o 3 }) Greek (c 3,4, {o 5 }) Italian (c 6,3, {o 2 }), (c 5,1, {o 4 }) Restaurant (c 4,3, {o 1 }), (c 6,3, {o 2 }), (c 3,4, {o 5 }), (c 3,1, {o 6 }) Restaurante (c 4,2, {o 3 }) Ristorante (c 5,1, {o 4 }) Shoes (c 1,4, {o 7 }), (c 3,4, {o 8 }) Cell: c 5,1 25

26 Indexing Approach Query Processing Techniques On Answering SKR Query Inverted Files + Grids: words occurrences contains spatial information. TS: Textual - Spatial Approach [Vaid et al. SSTD, 2005] o 7 ( Shoes ) o 9 ( Cinema ) o 10 ( bookstore ) ( Chinese Restaurant ) o 6 o 8 ( Athletic Shoes ) o 5 ( Greek Restaurant ) o 1 ( Chinese Restaurant ) o 3 ( Restaurante de France ) o 4 ( Italian Ristorante ) o 2 ( Italian Restaurant ) Word Occurrences Athletic (c 3,4, {o 8 }) Bookstore (c 1,2, {o 10 }) Chinese (c 4,3, {o 1 }), (c 3,1, {o 6 }) Cinema (c 1,3, {o 9 }) France (c 4,2, {o 3 }) Greek (c 3,4, {o 5 }) Italian (c 6,3, {o 2 }), (c 5,1, {o 4 }) Restaurant (c 4,3, {o 1 }), (c 6,3, {o 2 }), (c 3,4, {o 5 }), (c 3,1, {o 6 }) Restaurante (c 4,2, {o 3 }) Ristorante (c 5,1, {o 4 }) Shoes (c 1,4, {o 7 }), (c 3,4, {o 8 }) Cell: c 5,1 26

27 Query Processing Techniques On Answering SKR Query TS: Textual - Spatial Approach Query Processing Approach A = {c cid R c cid Q.loc } For each keyword t Q.text get the STOs such that their cell id is in A. Finally, get the intersection. Example: Q.loc = [(200, 0), (400, 200)] and Q.text = { Chinese, Restaurant }. A = {c 3,1, c 3,2, c 4,1, c 4,2 }. Evaluation: Better performance than SI approach. Word Occurrences Athletic (c 3,4, {o 8 }) Bookstore (c 1,2, {o 10 }) Chinese (c 4,3, {o 1 }), (c 3,1, {o 6 }) Cinema (c 1,3, {o 9 }) France (c 4,2, {o 3 }) Greek (c 3,4, {o 5 }) Italian (c 6,3, {o 2 }), (c 5,1, {o 4 }) Restaurant (c 4,3, {o 1 }), (c 6,3, {o 2 }), (c 3,4, {o 5 }), (c 3,1, {o 6 }) Restaurante (c 4,2, {o 3 }) Ristorante (c 5,1, {o 4 }) Shoes (c 1,4, {o 7 }), (c 3,4, {o 8 }) 27

28 Indexing Approach Query Processing Techniques On Answering SKR Query Hybrid Index: Inverted File R*-Trees, R*-Tree Inverted Files [Zhou et al. CIKM, 2005] R*-Tree Inverted Files Inverted File R*-Trees n 0 Athletic..... Restaurant n 1 n 2 R*-Trees n 0 R*-Trees n 3 n 4 n 5 n 6 n 1 n 2 Bookstore Shoes Cinema {o 7 } {o 9 } {o 10 } Chinese Italian Restaurant Ristorante {o 6 } {o 4 } {o 6 } {o 4 } {o 1 } {o 2 } {o 5 } {o 6 } Athletic Greek Restaurant Shoes Chinese France Italian Restaurant Restaurante {o 8 } {o 5 } {o 8 } {o 5 } {o 3 } {o 2 } {o 1 } {o 1, o 2 } {o 3 } 28

29 Query Processing Techniques On Answering SKR Query Query Processing Example: Q.loc = [(200, 0), (500, 300)] and Q.text = { Chinese, Restaurant }. R*-Tree Inverted Files Inverted File R*-Trees n 0 Chinese..... Restaurant n 1 n 2 R*-Trees R*-Trees n 3 n 4 n 5 n 6 n 0 n 0 Bookstore Shoes Cinema Chinese Italian Restaurant Ristorante n 1 n 2 n 1 n 2 {o 7 } {o 9 } {o 10 } Athletic Greek Restaurant Shoes {o {o 4 } 6 } {o 6 } {o 4 } Chinese France Italian Restaurant Restaurante {o 1 } {o 6 } {o 1 } {o 2 } {o 5 } {o 6 } {o 8 } {o 5 } {o 8 } {o 5 } {o 3 } {o 2 } {o 1 } {o 1, o 2 } {o 3 } 29

30 Query Processing Techniques On Answering SKR Query R*-Tree Inverted Files & Inverted File R*-Trees Pros They reduce I/O cost and query processing. The Inverted File R*-Tree outperforms the R*-Tree Inverted File. Cons Inverted File R*-Tree: It does not exploit the spatial distribution of keywords. R*-Tree Inverted File: The pruning process is not powerful as we can get a lot of candidates from spatial restrictions. 30

31 Indexing Approach Query Processing Techniques On Answering SKR Query R*-Tree where the internal nodes are virtually augmented with textual information KR*-Tree: Keyword R*-Tree Approach [Hariharan et al. SSDBM, 2007] n 0 n 1 n 2 n 3 n 4 n 5 n 6 o 1 o 2 o 3 o 7 o 9 o 10 o 5 o 8 o 4 o 6 Word Nodes Athletic n 0, n 1, n 4 Bookstore n 0, n 1, n 3 Chinese n 0, n 2, n 5, n 6 Cinema n 0, n 1, n 3 France n 0, n 2, n 5 Greek n 0, n 1, n 4 Italian n 0, n 2, n 5, n 6 Restaurant n 0, n 1, n 2, n 4, n 5, n 6 Restaurante n 0, n 2, n 5 Ristorante n 0, n 2, n 6 Shoes n 0, n 1, n 3, n 4 31

32 Query Processing Techniques On Answering SKR Query KR*-Tree Approach Query Processing Example: Q.loc = [(200, 0), (500, 300)] and Q.text = { Chinese, Restaurant }. n 0 n 1 n 2 n 3 n 4 n 5 n 6 o 1 o 2 o 3 o 7 o 9 o 10 o 5 o 8 o 4 o 6 Word Nodes Athletic n 0, n 1, n 4 Bookstore n 0, n 1, n 3 Chinese n 0, n 2, n 5, n 6 Cinema n 0, n 1, n 3 France n 0, n 2, n 5 Greek n 0, n 1, n 4 Italian n 0, n 2, n 5, n 6 Restaurant n 0, n 1, n 2, n 4, n 5, n 6 Restaurante n 0, n 2, n 5 Ristorante n 0, n 2, n 6 Shoes n 0, n 1, n 3, n 4 Experimental evaluation: More efficient than the previous approaches. 32

33 Indexing Approach Query Processing Techniques On Answering SKNN Query R-Tree & Signature Files IR 2 -Tree: Information Retrieval R-Tree Approach [De Felipe et al. ICDE, 2008] Block1: Signature Files A text has many words Hash(text) = Hash(many)= Hash(words)= Signature (Block1) = False Positives: Block2: OR A text has many Signature (Block2) = words Block n 0 n 1 n 2 n 3 n 4 n 5 n IR 2 -Tree: o 7 o 9 o 10 o 5 o 8 o 1 o 2 o 3 o 4 o 6 33

34 Query Processing Techniques On Answering SKNN Query IR 2 -Tree Approach Query Processing n n 1 n n 3 n 4 n 5 n o 7 o 9 o 10 o 5 o 8 o 1 o 2 o 3 o 4 o 6 Example: Q.loc = (500, 200), Q.text = { restaurant, Chinese } and k = 1. Signature(Q.text): Priority Queue: N 0,

35 Query Processing Techniques On Answering SKNN Query IR 2 -Tree Approach Query Processing n n 1 n n 3 n 4 n 5 n o 7 o 9 o 10 o 5 o 8 o 1 o 2 o 3 o 4 o 6 Example: Q.loc = (500, 200), Q.text = { restaurant, Chinese } and k = 1. Signature(Q.text): Priority Queue: N 2,

36 Query Processing Techniques On Answering SKNN Query IR 2 -Tree Approach Query Processing n n 1 n n 3 n 4 n 5 n o 7 o 9 o 10 o 5 o 8 o 1 o 2 o 3 o 4 o 6 Example: Q.loc = (500, 200), Q.text = { restaurant, Chinese } and k = 1. Signature(Q.text): Priority Queue: N 5, 0.0 N 6,

37 Query Processing Techniques On Answering SKNN Query IR 2 -Tree Approach Query Processing n n 1 n n 3 n 4 n 5 n o 7 o 9 o 10 o 5 o 8 o 1 o 2 o 3 o 4 o 6 Example: Q.loc = (500, 200), Q.text = { restaurant, Chinese } and k = 1. Signature(Q.text): Priority Queue: N 6, o 1,

38 Query Processing Techniques On Answering SKNN Query IR 2 -Tree Approach Query Processing n n 1 n n 3 n 4 n 5 n o 7 o 9 o 10 o 5 o 8 o 1 o 2 o 3 o 4 o 6 Example: Q.loc = (500, 200), Q.text = { restaurant, Chinese } and k = 1. Signature(Q.text): Priority Queue: o 1, o 6,

39 Query Processing Techniques On Answering SKNN Query Multi-level IR 2 -Tree (MIR 2 -Tree) Uses different signature lengths for different levels. More complex update operations Fewer False Positives Experimental evaluation: MIR 2 -Tree is more efficient than IR 2 -Tree. False Positives. No comparison with other indices. 39

40 Indexing Approach Query Processing Techniques On Answering Top-k SK Query R-Tree where the internal nodes are augmented with textual & scoring information IR-Tree: Inverted file R-Tree Approach [Cong et al. PVLDB, 2009] Inverted file I 0 n 0 n 1 n 2 Inverted file I 1 Inverted file I 2 n 3 n 4 n 5 n 6 I 3 I 4 I 5 I 6 o 7 o 9 o 10 o 5 o 8 o 1 o 2 o 3 o 4 o 6 Inverted file I 3 Cinema: <o 9, 3> Bookstore: <o 10, 2> Shoes: <o 7, 2> Inverted file I 1 Cinema: <n 3, 3> Bookstore: <n 3, 2> Shoes: <n 3, 2>, <n 4, 3> Greek: <n 4, 7> Restaurant: <n 4, 2> 40

41 IR-Tree Approach Query Processing Techniques On Answering Top-k SK Query Query Processing Similar to the IR 2 -Tree query processing algorithm (Branch-and-Bound) They propose two metrics for scoring rectangles and objects. MIND ST (Q, N) D ST (Q, O) Both metrics are based on Euclidean distances and IR scores. Shortcoming n 1 Inverted file I 1 n 3 n 4 I 3 I 4 o 7 o 9 o 3 o 1 o 2 o 1 : restaurant, weight 5 o 2 : Chinese, weight 3 o 3 : restaurant, weight 4 o 3 : Chinese, weight 2 41

42 Indexing Approach Query Processing Techniques On Answering Top-k SK Query Enhanced indexing: At construction time, take into account both the area parameter and the textual similarity between the indexed objects. DIR-Tree: Document Inverted file R-Tree Approach IR-Tree (similar to R-Tree) Area cost of inserting STO o into rectangle R: DIR-Tree [Cong et al. PVLDB, 2009] AreaCost(R) = area(r ) area(r) area(r ): area of R after inserting o SimAreaCost(R, o) = a AreaCost(R) maxarea + (1 a)(1 IRsim(R. text, o. text)) a (0, 1) Query Processing: Similar to IR-Tree. Experimental Evaluation: More efficient. 42

43 Indexing Approach Query Processing Techniques On Answering Top-k SK Query Enhanced indexing: Enhance IR-Tree with clustering. CIR-Tree: Clustering Inverted file R-Tree Approach [Cong et al. PVLDB, 2009] Goal: Tighter bounds for rectangles i.e., MIND ST. STO where word w o.text STO where word w o.text N (node of IR-Tree) Query Processing: Similar to IR-Tree. Experimental Evaluation: Most efficient. 43

44 Indexing Approach Query Processing Techniques On Answering Top-k SK Query Enhanced indexing: Enhance IR-Tree with clustering. CIR-Tree: Clustering Inverted file R-Tree Approach [Cong et al. PVLDB, 2009] Goal: Tighter bounds for rectangles i.e., MIND ST. C 1 C 2 STO where word w o.text STO where word w o.text N (node of IR-Tree) Query Processing: Similar to IR-Tree. Experimental Evaluation: Most efficient. 44

45 Indexing Approach Query Processing Techniques On Answering Top-k SK Query R-Tree where the internal nodes are augmented with textual & scoring information. IR-Tree2 d 0 n 0 [Li et al. TKDE, 2010] n 1 n 2 d 1 d 2 n 3 n 4 n 5 n 6 I 1 I 2 I 3 I 4 I 5 I 6 I 7 I 8 I 9 I 10 max tf Inverted file I 1 Shoes: o 7 d 1 < A 1 = [(40, 120), (280, 390)] D 1 = 5 M 1 = Cinema: 1, 1 Bookstore: 1, 1 Shoes: 2, 1 Greek: 1, 1 Restaurant: 1, 1 > df Cinema: 1, 1 45

46 Query Processing Techniques On Answering Top-k SK Query IR-Tree2 Query Processing! Q.loc = region, spatial proximity to the center of Q.loc Step 1: Compute idf = D Q.loc /D Q.loc,t, where D Q.loc is the number of STOs located in Q.loc and D Q.loc,t is the number of STOs in Q.loc s.t. t o.text. Step 2: Similar to IR-Tree. 46

47 Indexing Approach Query Processing Techniques On Answering Top-k SK Query Inverted File Aggregate Spatial Index S2I: Spatial Inverted Index Approach [Rocha-Junior et al. SSTD, 2011] Infrequent term Frequent term Athletic Restaurant o 10, o 10.loc, 9. o 1, o 1.loc, 4 n n 1 n n 3 n 4 n 5 n o 7 o 9 o 10 o 5 o 8 o 1 o 2 o 3 o 4 o 6 47

48 Query Processing Techniques On Answering Top-k SK Query S2I Approach Query Processing Threshold Algorithm Round Robin Strategy Q.text = 2, k = 1 Let S i be a source responsible for the i th query keyword. Given the Q.loc and a positive integer j, S i returns the STO o with the j th highest score. Iteration j S 1 S 2 UB 1 UB 2 UB unseen 1 o 1 o o 4 o o 3 o o 9 o Experimental Evaluation: More efficient than DIR-Tree. No comparison with CIR-Tree. 48

49 A Survey on Spatial-Keyword Search CONCLUSION 49

50 Conclusion Spatial-keyword queries. Indexing schemes and the corresponding algorithms for query processing. Pros and cons of each proposed framework. There are still much left to cover on this topic There is no experimental evaluation that compares the frameworks properly. Spatial-keyword search on road networks. Combination of two indices! 50

51 Questions 51