COMP7801 Topic 3 Spatial Networks

作者: pseudoyu | 1544 字, 8 分钟 | 评论 | 2021-02-27 | 分类: Develop

comp7801, database, hku

翻译: EN, DE

Background

Network Distance

  • In many real applications accessibility of objects is restricted by a spatial network
    • Examples
      • Driver looking for nearest gas station
      • Mobile user looking for nearest restaurant
  • Shortest path distance used instead of Euclidean distance
  • SP(a,b) = path between a and b with the minimum accumulated length

Challenges

  • Euclidean distance is no longer relevant
    • R-tree may not be useful, when search is based on shortest path distance
  • Graph cannot be flattened to a one-dimensional space
    • Special storage and indexing techniques for graphs are required
  • Graph properties may vary
    • directed vs. undirected
    • length, time, etc. as edge weights

Modeling and Storing Spatial Networks

Modeling Spatial Networks

  • Adjacency matrix only appropriate for dense graphs
  • Spatial networks are sparse: use adjacency lists instead

Modeling_Spatial_Networks

Storing Large Spatial Networks

  • Problem: adjacency lists representation may not fit in memory if graph is large
  • Solution:
    • partition adjacency lists to disk blocks (based on proximity)
    • create B+-tree index on top of partitions (based on node-id)

Storing_Large_Spatial_Network

  • Given a graph G(V,E), and two nodes s,t in V, find the shortest path from s to t
  • A classic algorithmic problem
  • Studied extensively since the 1950’s
  • Several methods:
    • Dijkstra’s algorithm
    • A*-search
    • Bi-directional search
  • idea: incrementally explore the graph around s, visiting nodes in distance order to s until t is found (like NN)

Dijkstra_1

Dijkstra_2

Algorithm

Dijkstra_Algorithm

Example

Dijkstra_Example

Illustrating

  • Find the shortest path between a and b.
  • Worst-case performance O(|E| + |V|log|V| )

Description

  • Dijkstra’s search explores nodes around s without a specific search direction until t is found

  • Idea: improve Dijkstra’s algorithm by directing search towards t

  • Due to triangular inequality, Euclidean distance is a lower bound of network distance

  • Use Euclidean distance to lower bound network distance based on known information:

    • Nodes are visited in increasing SPD(s,v)+dist(v,t) order
      • SPD(s,v): shortest path distance from s to v (computed by Dijkstra)
      • dist(v,t): Euclidean distance between v and t
    • Original Dijkstra visits nodes in increasing SPD(s,v) order

A_Star_1

Example

A_Star_Example

Illustrating

  • Find the shortest path between s and t.
    • f(p) = Dijkstra_dist(s, p) + Euclidean_dist(p, t)
  • Dijkstra’s search explores nodes around s without a specific search direction until t is found
  • Idea: search can be performed concurrently from s and from t (backwards)
  • The shortest path tree of s and the (backward) shortest path tree of t are computed in concurrently
    • One queue Q_s for forward and one queue Q_t for backward search
    • Node visits are prioritized based on min(SPD(s,v), SPD(v,t))
    • If v already visited from s and v is in Qt, then candidate shortest path: p(s,v)+p(v,t) (if v already visited from t and v in Q_s symmetric)
    • If v is visited by both s and t terminate search; report best candidate shortest path

Example

Bi_Directional_Example

Discussions

  • A* and bi-directional search can be combined to powerful search techniques
  • A* can only be applied if lower distance bounds are available
  • All versions of Dijkstra’s search require non-negative edge weights
    • Bellman-Ford is an algorithm for arbitrary negative edges

Spatial queries over spatial networks

Introduction

Source/Destination on Edges

  • We have assumed that points s and t are nodes of the network
  • In practice s and t could be arbitrary points on edges
    • Mobile user locations
  • Solve problem by introducing 2 more nodes

Source_Destination_on_Edges

Spatial Queries over Spatial Networks

  • Data:
    • A (static) spatial network (e.g., city map)
    • A (dynamic) set of spatial objects
  • Spatial queries based on network distance:
    • Selections. Ex: find gas stations within 10km driving distance from here
    • Nearest neighbor search. Ex: find k nearest restaurants from present position
    • Joins. Ex: find pairs of restaurants and hotels at most 100m from each other

Spatial_Queries_over_Spatial_Networks

Methodology

  • Store (and index) the spatial network
    • Graph component (indexes connectivity information)
    • Spatial component (indexes coordinates of nodes, edges, etc.)
  • Store (and index) the sets of spatial objects
    • Ex., one spatial relation for restaurants, one spatial relation for hotels, one relation for mobile users, etc.
  • Given a spatial location p, use spatial component of network to find the network edge containing p
  • Given a network edge, use network component to traverse neighboring edges
  • Given a neighboring edge, use spatial indexes to find objects on them

Evaluation of Spatial Selections (1)

  • Query: find all objects in spatial relation R, within network distance ε from location q
  • Method:
    • Use spatial index of network (R-tree indexing network edges) to find edge n_1n_2, which includes q
    • Use adjacency index of network (graph component) and apply Dijkstra’s algorithm to progressively retrieve edges that are within network distance ε from location q
    • For all these edges apply a spatial selection on the R-tree that indexes R to find the results

Example

  • Example: Find restaurants at most distance 10 from q

  • Step 1: find network edge which contains q

Evaluation_of_Spatial_Selections_1

  • Step 2: traverse network to find all edges (or parts of them within distance 10 from q)

Evaluation_of_Spatial_Selections_2

  • Step 3: find restaurants that intersect the subnetwork computed at step 2

Evaluation_of_Spatial_Selections_3

Evaluation of Spatial Selections (2)

Description

  • Query: find all objects in spatial relation R, within network distance ε from location q
  • Alternative method based on Euclidean bounds:
    • Assumption: Euclidean distance is a lower-bound of network distance:
      • dist(v,u) ≤ SPD(v,u), for any v,u
    • Use R-tree on R to find set S of objects such that for each o in S: dist(q,o) ≤ ε
    • For each o in S:
      • find where o is located in the network (use Network R-tree)
      • compute SPD(q,o) (e.g. use A*)
      • If SPD(q,o) ≤ ε then output o

Example

  • Example: Find restaurants at most distance 10 from q

  • Step 1: find restaurants for which the Euclidean distance to q is at most 10: S={r1,r2,r3}

Evaluation_of_Spatial_Selections_Example_1

  • Step 2: for each restaurant in S, compute SPD to q and verify if it is indeed a correct result

Evaluation_of_Spatial_Selections_Example_2

Evaluation of NN search (1)

  • Query: find in spatial relation R the nearest object to a given location q
  • Method:
    • Use spatial index of network (R-tree indexing network edges) to find edge n_1n_2, which includes q
    • Use adjacency index of network (graph component) and apply Dijkstra’s algorithm to progressively retrieve edges in order of their distance to q
    • For each edge apply a spatial selection on the R-tree that indexes R to find any objects
    • Keep track of nearest object found so far; use its shortest path distance to terminate network browsing

Example

  • Example: Find nearest restaurant to q
  • Step: in ppt 31

Evaluation of NN search (2)

  • Query: find in spatial relation R the nearest object to a given location q
  • Alternative method based on Euclidean bounds:
    • Assumption: Euclidean distance lower-bounds network distance:
      • dist(v,u) ≤ SPD(v,u), for any v,u

Evaluation_of_NN_search

Spatial Join Queries

Description

  • Query: find pairs (r,s), such that r in relation R, s in relation S, and SPD(r,s)≤ε
  • Methods:
    • For each r in R, do an ε-distance selection queries for objects in S (Index Nested Loops)
    • For each pair (r,s), such that Euclidean dist(r,s)≤ε compute SPD(r,s) and verify SPD(r,s)≤ε

Notes on Query Evaluation based on Network Distance

  • For each query type, there are methods based on network browsing and methods based on Euclidean bounds
  • Network browsing methods are fast if network edges are densely populated with points of interest
    • A limited network traversal can find the result fast
  • Methods based on Euclidean bounds are good if the searched POIs are sparsely distributed in the network
    • Few verifications with exact SP searches are required
    • Directed SP search (e.g. using A*) avoids visiting empty parts of the network

Advanced indexing techniques for spatial networks

Shortest Path Materialization and Indexing in Large Graphs

  • Dijkstra’s algorithm and related methods could be very expensive on very large graphs
  • (Partial) materialization of shortest paths in static graphs can accelerate search

Shortest_Path_Materialization_and_Indexing_in_Large_Graphs.png

Hierarchical Path Materialization

  • Idea: Partition graph G into G_1,G_2,G_3,… based on connectivity and proximity of nodes
  • Every edge of G goes to exactly one G_i
  • Border nodes belong to more than one G_i’s
  • For each G_i compute and materialize SPs between every pair of nodes in G_i (matrix M_i)
    • Partitions are small enough for materialization space overhead to be low
  • Compute and materialize SPs between every pair of border nodes (matrix B)
    • If border nodes too many, hierarchically partition them into 2nd-level partitions

Hierarchical_Path_Materialization

algorithm

Hierarchical_Path_Materialization_algorithm

Illustrating

  • Good partitioning if:
    • small partitions
    • few combinations examined for SP search
  • Real road networks:
    • Non-highway nodes in local partitions
    • Highway nodes become border nodes

Hierarchical_Path_Materialization_Illustration

Compressing Materialized Paths

  • Distance matrix with successors has O(n_2) space cost
  • Motivation: reduce space by grouping targets based on common successors

Compressing_Materialized_Paths

algorithm

  • Create and encode one space partitioning defined by targets of the same successor
  • For each node s, index Is a set of <succ,R> pairs:
    • succ: a successor of s
    • R: a continuous region, such that for each t in R, the successor of s in SP(s,t) is succ

Compressing_Materialized_Paths_Algorithm

  • To compute SP(s,t) for a given s, t:
    1. SP=s
    2. Use spatial index Is to find <succ,R>, such that t in R
    3. SP = SP + (s,succ)
    4. If succ = t, report SP and terminate
    5. Otherwise s=succ; Goto step 2

Summary

  • Indexing and search of spatial networks is different than spatial indexing
    • Shortest path distance is used instead of Euclidean distance, to define range queries, nearest neighbor search, and spatial joins
  • Spatial networks could be too large to fit in memory
    • Disk-based index for adjacency lists is used
  • Several shortest path algorithms
  • Spatial queries can be evaluated using Euclidean bounds
  • Advanced indexing methods for shortest path search on large graphs

相关文章

2021-03-06
COMP7801 Topic 4 Top-k
2021-02-06
COMP7801 Topic 2 Spatial Data Management
2021-01-30
COMP7801 Topic 1b Database Indexing
2021-01-23
COMP7801 Topic 1a Relational Database
pseudoyu

作者

pseudoyu

后端 & 智能合约研发工程师,HKU ECICer。喜欢探索新技术,空闲时也折腾博客搭建、效率工具等。 在 GitHub 关注我。在我的 Telegram 频道了解更多。


Comments