mixed-drove spatio-temporal co-occurrence pattern mining: a summary of results
DESCRIPTION
Mete Celik, Shashi Shekhar, James P. Rogers, James A. Shine, Jin Soung Yoo Presented by: Mark Dietz, Jesse Vig. Mixed-Drove Spatio-Temporal Co-occurrence Pattern Mining: A Summary of Results. Background: Co-location Pattern Discovery. Extension of association rule mining to spatial domain - PowerPoint PPT PresentationTRANSCRIPT
Mixed-Drove Spatio-Temporal Co-occurrence Pattern Mining: A Summary of Results
Mete Celik, Shashi Shekhar, James P. Rogers, James A. Shine,
Jin Soung Yoo
Presented by: Mark Dietz, Jesse Vig
Background: Co-location Pattern Discovery
Extension of association rule mining to spatial domainTransaction replaced by neighborhoodFind object types that are associated by spatial
proximityFootball example:
3 object types: Wide receiver (WR), Cornerback (CB), and Quarterback (Q)
Particular instances of each type are indexed numerically, e.g. WR.1, WR.2
{WR, CB} forms a co-location pattern
Background: Participation Index
Measures “prevalence” of a co-location pattern Given a subset P of object types,
Participation ratio for each object type in P is the proportion of instances of that type that are co-located with instances of the other object types in P
Participation index is the minimum participation ratio of all object types in P
A co-location pattern is prevalent if the participation index of that pattern is above a threshold θp
Participation ratio of WR in {WR, CB} = 2 / 2 = 1Participation ratio of CB in {WR, CB} = 2 / 2 = 1Participation index of {WR, CB} = min(1,1) = 1
MDCOP: Intuition
Adds element of time to co-location patternsSee football example belowParticipation index only works for individual time slots
Participation index of {WR, CB} is 1 for t=0,2 but 0 for t=1,3
MDCOP: co-location patterns that persist over time MDCOP: Mixed-drove spatio-temporal co-occurrence
patterns
MDCOP: Formal definitions
Time prevalence: Fraction of time slots in which a given pattern occurs.
Mixed-drove prevalence: Composition of spatial prevalence (participation index) and time
prevalence. Assume a spatial prevalence threshold θp
Mixed-drove prevalence is the fraction of time slots with participation index ≥ θp
Example (below): If θp = .5, what is the mixed-drove prevalence of {WR, CB}?
Given a time prevalence threshold θtime ,an MDCOP is a mixed-drove prevalent pattern if mixed-drove prevalence ≥ θtime
Example: If If θp = .5 and θtime =.7, is the {WR, CB} mixed-drove prevalent?
Problem Statement
Given: A set P of Boolean spatio-temporal object-types A neighbor relation R over locations A spatial prevalence threshold θp
A time prevalence threshold θtime
Find: {Pi | Pi is a subset of P and Pi is prevalent MDCOP}
Objective: Minimize computation cost
Constraints: Solution set must be correct
i.e. all identified patterns are prevalent MDCOPs
Solution set must be complete i.e. finds all prevalent MDCOPs
Why is this important?
MilitaryIdentify patterns of attack
EcologyTracking predator-prey relationships
Homeland defenseSpotting suspicious behavior
Transportation Road and network planning
Why is this challenging?
Number of possible patterns grows exponentially with the number of different object types# of possible patterns = 2n, n = # of object
typesChallenge for pattern discovery in general
Interest measures are computationally expensive
Spatio-temporal datasets are huge
Limitations of Related Work
Mining uniform groups of moving objects, i.e. flock patterns
Doesn't apply to mixed object types
Mining mixed groups of moving objects, i.e. mixed droves
Only looks for patterns in consecutive time slots
Only looks for patterns between specific objects rather than object-types.
Problem Solution: MDCOP Miner
Finds MDCOP's relatively efficientlyUses apriori algorithm
Builds larger candidate patterns from smaller ones: see figure below (important)
MDCOP measure is monotonically non-increasing
Extension of co-location miner
Validation Analytical results
Mixed-drove prevalence measure is monotonically non-increasing
MDCOP Miner is correct and complete Total cost of MDCOP Miner is no worse than naïve
approach. Experimental results
Compared run-time of MDCOP Miner to naive approach Sample results :
Contributions & Assumptions
ContributionsMDCOP framework
Independent of time order Operates on object types rather than objects
MDCOP Miner Validated analytically and experimentally
AssumptionsAbsolute number of co-occurrences
irrelevantRelative proportion of object types
irrelevant
Suggestions for Re-write
Introduce a spatial “support” measureReflects absolute number of co-occurrencesUsed as an additional filter, could aid
performance of MDCOP-MinerEvaluate performance against more data
sets, including very large ones.Are the MDCOPs found meaningful?
Questions?