Loading in 5 sec....

Finding Probabilistic Nearest Neighbors for Query Objects with Imprecise LocationsPowerPoint Presentation

Finding Probabilistic Nearest Neighbors for Query Objects with Imprecise Locations

- By
**romeo** - Follow User

- 100 Views
- Uploaded on

Download Presentation
## PowerPoint Slideshow about ' Finding Probabilistic Nearest Neighbors for Query Objects with Imprecise Locations' - romeo

**An Image/Link below is provided (as is) to download presentation**

Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.

- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -

Presentation Transcript

### Finding Probabilistic Nearest Neighbors for Query Objects with Imprecise Locations

Outline with Imprecise Locations

### Finding Probabilistic Nearest Neighbors for Query Objects with Imprecise Locations

Yuichi Iijima and Yoshiharu IshikawaNagoya University, Japan

Outline with Imprecise Locations

- Background and Problem Formulation
- Related Work
- Query Processing Strategies
- Experimental Results
- Conclusions

Imprecise with Imprecise LocationsLocation Information

- Sensor Environments
- Measurement Noise
- Frequent updates may not be possible
- GPS-based positioning consumes batteries

- Robotics
- Localization using sensing andmovement histories
- Probabilistic approach hasvagueness

- Privacy
- Location Anonymity

Nearest Neighbor Queries with Imprecise Locations

- Nearest Neighbor Queries
- Example: Find the closest bus stop
- Traditional problem in spatial databases
- Efficient query processing using spatial indices
- Extensible to multi-dimensional cases (e.g., image retrieval)

- What happen if the location of query object is uncertain?

NN of q

q

Example Scenario (1) with Imprecise Locations

- Query: Find the nearest gas station

Location estimation based on noisy GPS data

0%

35%

Nearest gas station

depends on the possiblecar locations

10%

55%

NN objects are definedin a probabilistic way

Example Scenario (2) with Imprecise Locations

- Mobile Robotics
- Location of the robot is estimatedbased on movement histories andsensor data
- Measurements are noisy
- Localization based on probabilisticmodeling
- Kalman filter, particle filter, etc.

- Estimated location is given as aprobabilistic density function (PDF)
- PDF changes on each estimation

Probabilistic Nearest Neighbor Query (1) with Imprecise Locations

pq(x)

- PNNQ for short
- Assumptions
- Location of query object qis specified as a Gaussian distribution
- Target data: static points

- Gaussian Distribution
- Σ: Covariance matrix

Probabilistic Nearest Neighbor Query (2) with Imprecise Locations

- Definition
- Find objects which satisfy the condition
- The probability that the object is thenearest neighbor of q is greater than or equal to θ

q

Outline with Imprecise Locations

- Background and Problem Formulation
- Related Work
- Query Processing Strategies
- Experimental Results
- Conclusions

Related Work with Imprecise Locations

- Query processing methods for uncertain (location) data
- Cheng, Prabhakar, et al. (SIGMOD’03, VLDB’04, …)
- Tao et al. (VLDB’05, TODS’07)
- Consider arbitrary PDFs or uniform PDFs
- Most of the case, target objects are imprecise

- Research related to Gaussian distribution
- Gauss-tree [Böhm et al., ICDE’06]
- Target objects are based on Gaussian distributions

- Gauss-tree [Böhm et al., ICDE’06]
- Our former work
- Ishikawa, Iijima, Yu (ICDE’09): Probabilistic range queries

Outline with Imprecise Locations

- Background and Problem Formulation
- Related Work
- Query Processing Strategies
- Experimental Results
- Conclusions

Naïve Approach (1) with Imprecise Locations

- Use of Voronoi Diagram
- Well-known method for standard (non-imprecise) nearest neighbor queries

Voronoiregion Ve

Naïve Approach (2) with Imprecise Locations

- PrNN(q, o): Nearest neighbor probability
- Probability that object o is the nearest neighbor of query object q
- It can be computed by integrating the probability density function pq(x) over Voronoi region Vo

- Problem
- Need to consider all targetobjects
- Numerical integration (MonteCarlo method) is quite costly!

pq(x)

pq(x)

Ve

Our Approach with Imprecise Locations

- Outline of processing
- Filtering
- Prune non-candidate objects whose PrNN are obviously smaller than the threshold
- Low-cost filtering conditions

- Numerical integration for the remaining candidate objects

- Filtering
- Two strategies
- -region-based Approach
- SES-based Approach
- SES: Smallest Enclosing Sphere

Strategy 1: with Imprecise Locations-Region-Based Approach (1)

- -region
- Similar concepts are often used in query processing for uncertain spatial databases
- Definition: Ellipsoidal region for which the result of the integration becomes 1 – 2 :The ellipsoidal regionis the -region

- Example: is specified as 1%,we consider 98% -region

q

-region

Strategy 1: with Imprecise Locations-Region-Based Approach (2)

- Query Processing
- -region for the query is computed at first
- -region can be derived using r-table:
- It is constructed for the normal Gaussian (S = I, q = 0): Given , it returns appropriate r

- Final -region can be obtained by transformation

- -region can be derived using r-table:
- Derive the bounding box of-region
- Objects whose Voronoiregions overlaps with the box are the candidates
- {a, c, d, f, g},in this example

- -region for the query is computed at first

a

c

d

g

f

θ-region

Strategy 1: with Imprecise Locations-Region-Based Approach (3)

- Derivation details
- First, we consider the normal Gaussian
- Using r-table, we can get the appropriate r for given

- Second, a spherical -region is derived based on transformation
- Transformation is performed by analyzing

- Third, the bonding box is calculatedwhere (Σ)iiis the (i, i) entry of Σ

xj

xi

wj

q

r

wj

q

q

wi

wi

Strategy 2: SES-Based Approach (1) with Imprecise Locations

- SES: Smallest Enclosing Sphere
- For each Voronoi region Vo, we compute its SES SESo beforehand

SESe: SES for Voronoi region of e

Strategy 2: SES-Based Approach (2) with Imprecise Locations

- Integration over SESo gives the upper-bound for PrNN(q, o)
- Integration over a sphere region is more easier to compute
- We use a table called U-catalog constructed beforehand

- Integration over a sphere region is more easier to compute

pq(x)

Strategy 2: SES-Based Approach (3) with Imprecise Locations

- What is U-catalog?
- Given two parameters and , it returns corresponding integral
- U-catalog is made for different (, ) pairs by computing the integral of normal Gaussian over sphere region R

Strategy 2: SES-Based Approach (4) with Imprecise Locations

- To use U-catalog, another approximation is required since it is only useful for normal Gaussian
- Use of upper-bounding function pq(x)
- pq(x) tightly bounds pq(x) and has sphericalisosurfaces
- For pq(x), we can easily derive its integral over SESo using U-catalog

- In summary, we use twoapproximations:

T

T

T

T

pq(x)

pq(x)

Strategy 2: SES-Based Approach (5) with Imprecise Locations

- Bounding Function
- Original Gaussian PDF
- Upper-Bounding Functionwhere

is satisfied

Outline with Imprecise Locations

- Background and Problem Formulation
- Related Work
- Query Processing Strategies
- Experimental Results
- and Conclusions

Setup of Experiments (1) with Imprecise Locations

- Target Data
- 2D point data (50K entries)
- Based on road line segments in Long Beach

- Computation of Voronoiregions and SESs
- LEDA package was used

- Comparison
- Strategies 1, 2, and their hybrid approach
- Evaluation metric: Response time

Setup of Experiments (2) with Imprecise Locations

- Default Parameters
- Covariance matrix
- For this matrix the shape of the isosurface of pq(x) is an ellipse titled at 30 degrees and the major-to-minor axis ratio is 3:1
- γ : Parameter for controlling vagueness (default: γ = 10)

- Probability threshold value: θ = 0.01
- No. of samples for Monte Carlo method: 1,000,000

- Covariance matrix

pq(x)

Candidate Objects in Strategy 1 with Imprecise Locations

Bounding box of the θ-region

Query center

Voronoi regions for candidate objects

Voronoi regions for answer objects

Candidate Objects in Strategy 2 with Imprecise Locations

Query center

Voronoi regions for candidate objects

Voronoi regions for answer objects

Candidate Objects in Hybrid Strategy with Imprecise Locations

Bounding box of the θ-region

Query Center

Voronoi regions for candidate objects

Voronoi regions for answer objects

Experimental Results with Imprecise Locations - Default Parameters

- Most of the processing time is spent in numerical integration
- A strategy that can prune more objects has better performance

Experimental Results with Imprecise Locations - Different γValues

γ = 1(almost exact)

γ = 50

(too vague)

γ = 10

Query is costly when impreciseness is high

Experimental Results with Imprecise Locations - Different Shapes

Circle

Default Ellipse

Narrow Ellipse

Narrow ellipse (correlated distribution) needs to consider additional candidates

Conclusions of Experimental Results with Imprecise Locations

- Most of the processing time is spent in numerical integration
- A strategy that can prune more objects has better performance

- Superiority or inferiority of two strategies depends on the given query and the specified parameters
- No apparent winner

- The hybrid strategy inherits the benefits of two strategies and shows the best performance

- Background and Problem Formulation
- Related Work
- Query Processing Strategies
- Experimental Results
- Conclusions

Conclusions with Imprecise Locations

- Nearest neighbor query processing methods for imprecise query objects
- Location of query object is represented by Gaussian distribution
- Two strategies and their combination
- Reduction of numerical integration is important
- Proposal of two pruning strategies and their evaluation

- Future work
- Evaluation for multi-dimensional cases (d 3)

Yuichi Iijima and Yoshiharu IshikawaNagoya University, Japan

Download Presentation

Connecting to Server..