Spatio-temporal Rule Mining: Issues and Techniques

1
Spatio-temporal Rule Mining Issues and
Techniques

• Gyozo Gidófalvi
• Geomatic ApS
• Center for Geoinformatik
• and
• Torben Bach Pedersen
• Aalborg University

2
Outline

• Why mine spatio-temporal data?
• Frequent pattern mining background
• Frequent itemset mining / association rules
• Sequential patterns
• Taxonomy of spatio-temporal data
• Some examples STM, HUR, INFATI, DMI
• How to mine spatio-temporal rules?
• Pivoting -gt spatio-temporal baskets
• Some examples INFATI, HUR, DMI
• Mining long, common patterns in trajectories of
  moving objects
• Conclusions and future directions

3
Why Mine Spatio-temporal Data?

• Spatio-temporal data is being collected at
  enormous speeds (Tbyte/hour)
• Remote sensors on satellites
• Telescope scanning the skies
• Location data received from mobile devices
• Data needs to be analysed for various purposes
• Cataloging, classification, segmentation
• Scientific hypothesis formulation
• Study complex systems with autonomous mobile
  entities
• Aid the management, storage, and retrieval of
  spatio-temporal data
• Hidden information in data can be used to provide
  customized Location-Based Services (LBS) and
  Location-Based Advertising (LBA)

4
Frequent Pattern Mining

• First proposed by Agrawal and Sirkant for the
  analysis of customer purchase behaviour
• Frequent itemsets Discover items are bought
  together by customers frequently?
• bread, peanut butter, jelly
• Association rules Discover a possible causal
  relationship between the items in such a frequent
  itemset?
• bread, peanut butter -gt jelly (within trans.)
• Sequential patterns Discover sequences of items
  or itemsets that are frequent in sequences of
  transactions?
• Star Wars Episode I -gt Episode II -gt Episode
  III (between transactions)
• Episodes Discover periodic patterns in a long
  sequence?
• Patterns in other structures trees, graphs,
• How do we extend these to the spatio-temporal
  domain?
• EX Strøget,noon,businessman -gt cafe

5
Frequent Pattern Mining Cont

• Approaches to frequent itemset mining and
  association rules
• Apriori bottom-up, generate-and-test frequent
  itemsets
• BFS traversal of search space
• Pruning using support monotonicity of itemsets
• Projection-based (FP-growth) generate frequent
  itemset prefixes and extend the prefix by mining
  the prefix- projected database
• DFS traversal of search space
• Many other variants employing sophisticates
  in-memory data structures and representations of
  the data.
• Restrictions on frequent itemsets
• Closed frequent itemsets
• Maximal frequent itemsets

6
Taxonomy of Spatio-temporal Data

• Examples of spatio-temporal data
• Space Time Man (STM) activities performed by
  mobile users at particular times and locations
• HUR1 number of passengers getting on/off busses
  at particular times and locations
• HUR2 Personal chip cards recording travels of
  individuals
• DMI periodic atmospheric measurements like
  temperature, humidity, and pressure for 5 km grid
  cells
• INFATI daytoday movements of 20 private cars
  on the road network of Aalborg
• Criteria for categorization of spatio-temporal
  data
• Are the measured entities mobile or immobile?
• Are the attribute values of the measured entities
  static or dynamic?

7
How to Mine Spatio-temporal Rules?

• Knowledge extractable by association rules is
  about dependencies between items within baskets.
  -gt Need to construct spatio-temporal baskets.
• Pivoting is the process of grouping a set of
  records based on one or more attributes (pivoting
  attributes) and assigning the values of an
  another attribute (pivoted attribute) to groups
  or baskets.

• Spatio-temporal rules that can be mined from
  spatio-temporal baskets can be either implicit or
  explicit.

8
Illustration of Pivoting

• INFATI pivoting example pivoting attributes are
  Location and Time, pivoted attribute is
  CarID

9
Spatio-temporally Restricted vs. Unrestricted
10
Explicit Spatio-temporal Rule Mining 1
11
Explicit Spatio-temporal Rule Mining 2
12
DMI Dynamic Attributes of Immobile Entities
13
Mining Long, Common Patterns (LCP) in
Trajectories of Moving Objects

• Trajectories of moving objects contain
  regularities or patterns
• These patterns can be used in indexing, tracking,
  and LBS
• LBS example intelligent rideshare application
• Find common routes for a set of commuters and
  suggest rideshare possibilities to them
• Unique requirements
• Patterns should rather be long than frequent
• Patterns should be shareable, i.e. common
• Unique challenges
• Patterns are extremely long
• Interesting patterns have relatively low support
• Not all sub-patterns are interesting

14
Method to Mine LCP in Trajectories

• Pre-processing
• Identify trips, i.e. gaps
• Map date-time domain to time-of-day domain
• Substitute noisy GPS measurements with
  spatio-temporal regions
• Use / exploit unique requirements
• Prune search space if extractable patterns are
  doomed to be short
• Define unique support measure n-support of
  transactions satisfying an itemset if the number
  of distinct objects associated with those
  transactions gt n, 0 otherwise
• It can be shown that interesting patterns are
  closed frequent itemsets
• In current work, a projection-based FIM algorithm
  is being extended to meet and use and meet these
  requirements
• Illustrative example

15
Pre-processed Example Trajectory Database
16
Extracted Long Common Patterns
17
Conclusions and Future Directions

• Today
• Taxonomy of spatio-temporal data
• Pivoting to obtain spatio-temporal baskets
• Mining explicit and implicit spatio-temporal
  rules
• Spatio-temporally restricted vs. unrestricted
  mining
• Mining long, common patterns in trajectories
• Tomorrow
• Incorporate spatio-temporal indexes in
  spatio-temporal rule mining or vice versa
• Incorporate various spatio-temporal space
  partitioning methods into mining

18
Thank you for your attention! Questions?