1. Mining Long, Sharable Patterns in Trajectories of Moving Objects
Győző Gidófalvi
Geomatic ApS
Center for Geoinformatik
and
Torben Bach Pedersen
Aalborg University
STDBM 2006, Seoul, South Korea

2. STDBM 2006, Seoul, South Korea
Outline
Why mine trajectories of moving objects?
From spatio-temporal data to spatio-temporal baskets: pivoting
From trajectories to transactions in 3 steps
Method for mining long, sharable patterns
Observations and associated steps
Experiments
Conclusions and future directions
STDBM 2006, Seoul, South Korea

3. STDBM 2006, Seoul, South Korea
Outline
Why mine trajectories of moving objects?
From spatio-temporal data to spatio-temporal baskets: pivoting
From trajectories to transactions in 3 steps
Method for mining long, sharable patterns
Observations and associated steps
Experiments
Conclusions and future directions
STDBM 2006, Seoul, South Korea

4. STDBM 2006, Seoul, South Korea
Outline
Why mine trajectories of moving objects?
From spatio-temporal data to spatio-temporal baskets: pivoting
From trajectories to transactions in 3 steps
Method for mining long, sharable patterns
Observations and associated steps
Experiments
Conclusions and future directions
STDBM 2006, Seoul, South Korea

5. STDBM 2006, Seoul, South Korea
Outline
Why mine trajectories of moving objects?
From spatio-temporal data to spatio-temporal baskets: pivoting
From trajectories to transactions in 3 steps
Method for mining long, sharable patterns
Observations and associated steps
Experiments
Conclusions and future directions
STDBM 2006, Seoul, South Korea

6. STDBM 2006, Seoul, South Korea
Outline
Why mine trajectories of moving objects?
From spatio-temporal data to spatio-temporal baskets: pivoting
From trajectories to transactions in 3 steps
Method for mining long, sharable patterns
Observations and associated steps
Experiments
Conclusions and future directions
STDBM 2006, Seoul, South Korea

7. STDBM 2006, Seoul, South Korea
Outline
Why mine trajectories of moving objects?
From spatio-temporal data to spatio-temporal baskets: pivoting
From trajectories to transactions in 3 steps
Method for mining long, sharable patterns
Observations and associated steps
Experiments
Conclusions and future directions
STDBM 2006, Seoul, South Korea

8. Why mine trajectories of moving objects?
Large amounts of location data from mobile devices
Locations data is interesting as a sequence
Extracting patterns in trajectories can:
aid the management, storage, and retrieval of trajectories
improve tracking of moving objects
be used to provide customized Location-Based Services (LBS) and Location-Based Advertising (LBA)
STDBM 2006, Seoul, South Korea

9. From spatio-temporal data to spatio-temporal baskets: pivoting
pivoting attribute
pivoted attributes
Pivoting is the process of grouping a set of records based on one or more attributes (pivoting attributes) and assigning the values of one of more different attributes (pivoted attributes) to groups or baskets.
Frequent itemset mining to discover relationships between items in the basket [AGR94]
pivoting
spatio-temporal item
STDBM 2006, Seoul, South Korea

10. From trajectories to transactions in 3 steps
STEP I: Identify trips
a trip ends when the total displacement in k consecutive GPS readings is less than δ
STEP II: Capture periodicity of patterns
Map date-time domain to time-of-day domain
STEP III: Eliminate the problem of noisy GPS readings
Substitute readings with spatio-temporal regions
STDBM 2006, Seoul, South Korea

11. Mining long, sharable patterns (LSP)
LBS example application: intelligent ridesharing
Find sharable 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
Unique challenges:
Patterns are extremely long
Interesting patterns have relatively low support
Not all sub-patterns are interesting
STDBM 2006, Seoul, South Korea

12. Sample trajectory database
Task: Find all patterns that have at least length 4, are parts of at least 4 trajectories, which belong to at least 2 distinct objects (shareable / common)
STDBM 2006, Seoul, South Korea

13. Filter infrequent itemsets
Observation: Infrequent items cannot be part of a pattern
DEF: An item is frequent if the number of transactions that contain that item >= MinSupp (for example 4), and the number of unique objects associated with those transactions >= n (for example 2).
STEP 1: Delete all infrequent items
infrequent items
STDBM 2006, Seoul, South Korea

14. Filter short transactions
Observation: A short transaction can never support a long pattern
DEF: A transaction is short if the number of items in it is less than or equal to MinLength (for example 4).
STEP 2: Delete all short transactions
short transactions
STDBM 2006, Seoul, South Korea

15. Project DB based on item with least n-support
Observation: All patterns that a given item i participates in can be found in the item-projected DB, T_i.
DEF: An item-projected DB, T_i, contains all the items from the transactions containing item i.
STEP 3: Project DB based on the item with the least n-support
Projecting item with least n-support
STDBM 2006, Seoul, South Korea

16. Find the single most frequent itemset in the projected DB
Observation: There is a single most frequent itemset in a projected DB, and its support is equal to the support of the projecting item
STEP 4: Identify single most frequent itemset
n-support in projected DB
Closed Frequent Itemset
projecting item
STDBM 2006, Seoul, South Korea

17. Delete unnecessary items from predecessor DB
Observation: Items that have the same support in the item projected DB as the projected DB, can be deleted from the later.
STEP 5: Delete unnecessary items from predecessor DB
support in projected DB
support in predecessor DB
Unnecessary items
STDBM 2006, Seoul, South Korea

18. Pseudo code for complete LCP mining
Recursive, projection-based LCP mining
Filter infrequent items
Filter short transactions
Project DB based on item with least n-support
Find the single most frequent itemset in the projected DB
Recursively mine the projected DB by further projecting on frequent items that are not part of the single most frequent itemset
Delete unnecessary items from predecessor DB
STDBM 2006, Seoul, South Korea

19. Extracted long sharable patterns
STDBM 2006, Seoul, South Korea

20. Experimental setup & performance results
INFATI data: 20 cars driving in Aalborg for 2 month, 1.8 million GPS readings, 3699 trips
LCP implemented on MS-SQL, running on 3.6 GHz, 2GB RAM
Minimum length experiments
Minimum support experiments
STDBM 2006, Seoul, South Korea

21. Visualization of extracted patterns
Large amount of spatio-temporal data
Some but not all patterns are visible, but not explicit
Difficult to analyse
28 explicit, at least 200-long patterns that are supported by at least 4 trips of at least 2 cars
Easy to analyse
STDBM 2006, Seoul, South Korea

22. Conclusions and future directions
Results:
Frequent itemset mining can be modified to extract long sharable patterns from trajectories
A RDBMS provides support for an easy but effective implementation
Challenges:
Some of the patterns are while different in terms of items, but they are not all that different in spatio-temporal space.
How can we further compress the patterns in a way that takes this spatial-temporal closeness into consideration?
Compression should take place during the mining process and should be exploited to further improve efficiency
STDBM 2006, Seoul, South Korea