Speaker: Ping-Lin Chang 2009/04/12

 Introduction  ROAD Framework  Operation Designed  Empirical Results  Conclusions 2Fast Object Search on Road Networks

 Introduction  ROAD Framework  Operation Designed  Empirical Results  Conclusions Fast Object Search on Road Networks3

 Location-based services (LBSs) ◦ Blooming nowadays because of  the thriving development of mobile devices  the ubiquitous wireless communication networks  high precision geo-positioning technology  The core application of LBSs ◦ To answer user queries with respect to user-specified location Fast Object Search on Road Networks4

 The technological trend of LBSs ◦ Dynamic combination of content and map services  Content providers ◦ Stores, average users, etc.  Map service providers ◦ Google Maps, MapQuest, MS Virtual Earth, etc. Fast Object Search on Road Networks5

 Location-dependent spatial queries (LDSQs) ◦ A fundamental data access operations in LBSs ◦ Query the spatial objects (location dependent info.)  k-nearest neighbor (kNN) search ◦ Find the nearest bus station to the conference venue  Range search ◦ Find hotels within 10-minutes walk from the conference venue Fast Object Search on Road Networks6

 For an efficient LDSQ processing ◦ Flexibly and efficiently accommodate diverse objects ◦ Efficiently support various LDSQs ◦ Effectively support different distance metrics  However the prior works did not perform well on such an issue Fast Object Search on Road Networks7

 Review the deficiency of prior works ◦ Network expansion based approaches  inefficient due to an almost blind scan over entire search space  slow node-by-node expansion towards all directions ◦ Euclidean distance bound approaches  inefficient when paths are not in straight lines  not applicable to other network distance metrics, such as travel time or cost ◦ Solution based approaches  completely impractical due to extremely high preprocessing and storage costs  adapting poorly to other query types, and to object and network changes Fast Object Search on Road Networks8

 The proposed system framework ◦ Route Overlay and Association Directory (ROAD)  Two basic operations in processing LDSQs ◦ Network traversal (RO) ◦ Object lookup (AD)  Principle concepts ◦ Rnets, shortcuts, and object abstracts Fast Object Search on Road Networks9

 Introduction  ROAD Framework  Operation Designed  Empirical Results  Conclusions Fast Object Search on Road Networks10

 Preliminaries ◦ Φ = (N,E)  A road network can be modeled as a weighted graph Φ consisting of a set of nodes N and edges E ◦ A node n ∈ N  represents a road intersection ◦ An edge (n, n’) ∈ E  represents a road segment connecting nodes n and n’ ◦ |n, n’|  denotes the edge distance, which can represent the travel distance, trip time or toll of the corresponding road segment  the value is positive Fast Object Search on Road Networks11

 Preliminaries (cont.) ◦ A path P(u, v)  stands for a set of edges connecting nodes u and v and its distance |P(u, v)| = Σ (n, n’) ∈ P(u, v) |n, n’| ◦ The shortest path SP(u, v)  among all possible paths connecting node u and node v, the one with the shortest distance is referred to as the shortest path ◦ The network distance ||u, v|| between u and v  is the distance of their shortest path SP(u, v)  ||u, v|| = |SP(u, v)| Fast Object Search on Road Networks12

 Preliminaries (cont.) ◦ Assume that objects reside on edges (road segments) in a network  objects at nodes (i.e., road intersections) can be treated as they are located at the end of the edges ◦ O(n, n’)  represents a set of objects on edge (n, n’) ◦ δ(o, n) and δ(o, n’)  represents the distance from an object o ∈ O(n, n’) to the nodes n and n’ Fast Object Search on Road Networks13

 Basic idea ◦ Search space pruning  to skip some search subspaces that do not contain objects of interest ◦ We need  a hint about whether or what objects are on the path  an artifact at n 1 connecting n 5 ◦ A shortcut between two ending nodes is the shortest path between them Fast Object Search on Road Networks14

 Closed paths are usually short in road networks ◦ The performance gained by bypassing closed paths is limited ◦ Regional sub-networks (Rnets) is introduced  each Rnet encloses a subset of edges and is bounded by a set of border nodes Fast Object Search on Road Networks15

 Definition 1. Rnet ◦ In a network N = (N, E), an Rnet R = (N R, E R, B R ) represents a search subspace, where N R, E R and B R stand for nodes, edges and border nodes in R, and  (1) E R ⊆ E  (2) N R = { n | (n, n’) ∈ E R ∨ (n’, n) ∈ E R }  (3) B R = N R ∩ { n | (n, n’) ∈ E’ ∨ (n’, n) ∈ E’ }, where E’ = E − E R Fast Object Search on Road Networks16

 Definition 2. Object Abstract ◦ The object abstract of an Rnet R, O(R), represents all the objects residing on edges in E R  O(R) = ∪ e ∈ E R O(e)  Definition 3. Shortcut ◦ The shortcut, S(b, b’), between border nodes b and b’ ( ∈ B R ) of an Rnet R bears the shortest path SP(b, b’) and its distance ||b, b’|| ◦ It is noteworthy that the edges that contribute to SP(b, b’) might not necessarily be included in E R Fast Object Search on Road Networks17

 Rnet Hierarchy ◦ Large Rnets at the upper levels enclose smaller Rnets at lower levels ◦ At each layer, a network can be viewed as a layer of interconnected Rnets ◦ Original Rnet as the level-0 Rnet  does not have border node and partition it into p 1 ◦ At each subsequent level i  partition each Rnet into p i child Rnets ◦ As a result  at a level x ( ∈ [0, l]), the entire network is fully covered by x Π i=1 p i  for an Rnet hierarchy of l levels, there is l Σ h=0 ( h Π i=1 p i ) Fast Object Search on Road Networks18

Fast Object Search on Road Networks19

 Definition 4. Rnet partitioning ◦ Partitioning of an Rnet R = (N, E, B) where N, E, B are a set of nodes, edges and border nodes and B ⊆ N, forms p child Rnets, R 1, R 2, · · · R p where p > 1 and R i = (N i, E i, B i )  here, N = ∪ 1≤i≤p N i, E = ∪ 1≤i≤p E i, B = ∪ 1≤i≤p B i ◦ Also, the following three conditions must hold  edges of all child Rnets are disjointed  ∀ i ∀ j i ≠ j ⇒ E i ∩ E j = ∅  nodes in an Rnet are connected by edges in the same Rnet  ∀ i ∀ (n, n’) ∈ E i, n ∈ N i ∧ n’ ∈ N i  border nodes in an Rnet are common to its parent Rnet and some of its sibling Rnets  B i = N i ∩ [ B ∪ ( ∪ j ∈ ([1,p]−{i}) N j ) ] Fast Object Search on Road Networks20

 An ideal network partitioning generating ◦ Geometric approach  coarsely partitions a network into two with equal numbers of edges ◦ KL algorithm  fine tunes the two result Rnets by exchanging edges between them ◦ p i is set to be power of 2  recursively apply this binary partitioning until p i Rnets are formed Fast Object Search on Road Networks21

 Important property ◦ Object abstracts and shortcuts are constructed in a bottom-up fashion ◦ Shortcuts of a border node can be determined by adopting Dijkstra’s algorithm to explore paths for all other border nodes in the same Rnet ◦ Shortcuts in Rnets at level i can be calculated based on those in Rnets at level i+1 ◦ Explored shortcuts in Rnets can be used to determine other shortcuts of Rnets in the same level ◦ Some shortcuts that are composed of other shortcuts in the same Rnets can be safely ignored Fast Object Search on Road Networks22

Fast Object Search on Road Networks23

 Introduction  ROAD Framework  Operation Designed  Empirical Results  Conclusions Fast Object Search on Road Networks24

 Data structures ◦ Route Overlay (RO)  based on definition 4 that the border nodes in parent Rnets are always the border nodes in some of their child Rnets ◦ Association Directory (AD)  based on definition 2 that can examine Rnets quickly and determine whether bypass those Rnets or not Fast Object Search on Road Networks25

 Route Overlay Fast Object Search on Road Networks26

Fast Object Search on Road Networks27

 Association Directory Fast Object Search on Road Networks28

 Search algorithms Fast Object Search on Road Networks29

30

 Object update ◦ Simply changing the records in AD  Network update ◦ Only affects RO ◦ Filtering-and-refreshing approach is performed  in the filtering step, shortcuts that may be affected by the change are identified  in the refreshing step, the identified shortcuts are then updated Fast Object Search on Road Networks31

 Network update (cont.) ◦ Edge distance increased  only those shortcuts that cover (n, n’) might become invalid and need to be refreshed ◦ Edge distance decreased  may contribute to paths shorter than some expisting shortcuts Fast Object Search on Road Networks32

 Change of network structure ◦ Addition of a new edge (n, n’)  judging whether the nodes n and n’ are in the same Rnet ◦ Deletion of an existing edge (n, n’)  judging whether either n or n’ is border node ◦ Incorporating with the schema of network update Fast Object Search on Road Networks33

 Introduction  ROAD Framework  Operation Designed  Empirical Results  Conclusions Fast Object Search on Road Networks34

 Experimental environment ◦ Three real road network datasets, CA, NA, and SF ◦ Organize network nodes by CCAM ◦ Run on Linux servers with Intel Xeon 3.2GHz CPU ◦ All algorithms implemented in GNU C++ ◦ All indices are stored on disk  the page size is fixed at 4KB  memory cache of 50 pages Fast Object Search on Road Networks35

 Evaluation parameters Fast Object Search on Road Networks36

 Index construction time and index size Fast Object Search on Road Networks37

 Index construction time and index size Fast Object Search on Road Networks38

 Index update time Fast Object Search on Road Networks39

 Index update time Fast Object Search on Road Networks40

 Query performance - kNN query Fast Object Search on Road Networks41

 Query performance - range query Fast Object Search on Road Networks42

 Impact of Rnet hierarchy level ( l ) Fast Object Search on Road Networks43

 Introduction  ROAD Framework  Operation Designed  Empirical Results  Conclusions Fast Object Search on Road Networks44

 The proposed algorithm achieves clean separation between objects and network ◦ Better system flexibility and extensibility  The strategy of search space pruning ◦ Substantially speeds the object search  Range and kNN query for common LDSQs ◦ Shows high performance  Incremental framework maintenance techniques ◦ Update information in both efficiency and effectiveness Fast Object Search on Road Networks45

Fast Object Search on Road Networks46 Q & A