TRAJECTORY PATTERN MINING

Transcription

1 TRAJECTORY PATTERN MINING Fosca Giannotti, Micro Nanni, Dino Pedreschi, Martha Axiak Marco Muscat

2 Introduction 2 Nowadays data on the spatial and temporal location is objects is available. Gps, GSM towers, etc What can we do with this data?

3 Introduction 3 Nowadays data on the spatial and temporal location is objects is available. Gps, GSM towers, etc What can we do with this data? Data mine it for patterns!! (of course)

4 Trajectory Pattern Mining 4

5 Possible uses 5 Prediction of movement Aggregate movement behaviour Region of interests discovery Discovery of traffic flow & blockages

6 Structure of presentation 6 Introduction Mechanics Data structure Regions of Interest (ROI) T-Patterns using these ROI Results

7 Data structure 7 Raw Input List of (Lat, Long, timestamp) Multiple Moving Objects Convert Lat, Long, t covert x, y, t x, y R 2 T-Patterns T = s 0 α 1 s 1 α 2 α N s N

8 Example: Raw input 8

9 Example: T-Patterns 9

10 Example: T-Patterns 10

11 Support Threshold 11 Support: Percentage of trajectories which are contained within a pattern. Example: Support threshold of 0.2. A T-Pattern is kept only if 20% of the trajectories support it.

12 Curse of Dimensionality 12

13 Regions of Interest 13 Static pre-processed ROI ROI discovery List of candidate places Dynamic ROI Popular points detection ROI construction

14 Popular Points detection 14

15 ROI construction 15

16 T-Pattern Mining (1) 16 Two Approaches Static Dynamic Input: Trajectories and threshold parameters Neighbourhood and time threshold Support threshold Output: T-Patterns

17 T-Pattern Mining (2) 17 Step-wise Heuristic Any frequent T-Pattern of length n+1 is the extension of some frequent T-Pattern of length n.

18 Static discovery of T-Patterns 18

19 Dynamic Discovery of T-Patterns (1) 19 Y 4 trajectories. Aim is to find T- Patterns. First iteration prefix length is zero. X

20 Dynamic Discovery of T-Patterns (2) 20 Y A D E Detect Regions of Interest e.g. support(r) 0.5 B C X For each region find possible projections to other regions

21 Dynamic Discovery of T-Patterns (3) 21 Y A D E With respect to region A. Projections: B A B A C Add new configuration C (T, <A>) X

22 Dynamic Discovery of T-Patterns (4) 22 Y A B C D E With respect to region B. Projections: B C B D B E Add new configuration (T, <B>) X

23 Dynamic Discovery of T-Patterns (5) 23 Y A B D E With respect to region C. No Projections No new configuration C X

24 Dynamic Discovery of T-Patterns (6) 24 Y A D E With respect to region D. Projection D E B Add new configuration (T, <D>) C X

25 Dynamic Discovery of T-Patterns (7) 25 Y A D E With respect to region E. Projection E D B Add new configuration (T, <E>) C X

26 Dynamic Discovery of T-Patterns (8) 26 Y A D E Iteration 2 works on each configurations added in prior iteration. F (T, <A>) Recompute ROIs. X

27 Dynamic Discovery of T-Patterns (9) 27 Y A F D E Test projects from each of the regions. Results in configuration (T, <A F>) X

28 Dynamic Discovery of T-Patterns (11) 28 Y B D E (T, <B>) Recompute ROIs. Results in configuration (T, <B D>) (T, <B E>) X

29 Dynamic Discovery of T-Patterns (12) 29 Y D E (T, <D>) No possible projections. Same thing happens for region E. X

30 Dynamic Discovery of T-Patterns (11) 30 Y A F D E Iteration 3 (T, <A F>) Produces configurations (T, <A F D>) (T, <A F E>) X

31 Dynamic Discovery of T-Patterns (11) 31 Y Iteration 4 A D E No more paths to go therefore stops and outputs paths: B F C A F D A F E B D B E X

32 Results 32 Dynamic Approach Static Approach

33 Results 33

34 Results 34

35 To conclude 35 Lack of comparative results Idea of ROI is good for applications looking for simple and course results. Cited many times for ROI discovery algorithm Where Next: T-Patterns used to predict where a person will move next, given the current sequence. DAEDALUS: Extends SQL to enable TAS queries.

36 References 36 Giannotti, F., Nanni, M., Pinelli, F., and Pedreschi, D Trajectory pattern mining. F. Giannotti, M. Nanni, and D. Pedreschi. Efficient mining of sequences with temporal annotations pdf Monreale, A., Pinelli, F., Trasarti, R., and Giannotti, F WhereNext: a location predictor on trajectory pattern mining Ortale, R., Ritacco, E., Pelekis, N., Trasarti, R., Costa, G., Giannotti, F., Manco, G., Renso, C., and Theodoridis, Y The DAEDALUS framework: progressive querying and mining of movement data.

37 Thanks! 37