spatial-temporal models in location prediction

22
Spatial-Temporal Models in Location Prediction Jingjing Wang 03/29/12

Upload: marlis

Post on 23-Feb-2016

30 views

Category:

Documents


0 download

DESCRIPTION

Spatial-Temporal Models in Location Prediction. Jingjing Wang 03/29/12. Problem Definition. Given a sequence of movement: What is the location at a future time tF ? tF can be just the next timestamp to tN Any future timestamp. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Spatial-Temporal Models in  Location Prediction

Spatial-Temporal Models in Location Prediction

Jingjing Wang03/29/12

Page 2: Spatial-Temporal Models in  Location Prediction

Problem Definition

• Given a sequence of movement: – <(loc1, t1), (loc2, t2), …, (locN, tN) >

• What is the location at a future time tF?– tF can be• just the next timestamp to tN• Any future timestamp

Page 3: Spatial-Temporal Models in  Location Prediction

Traditional Methods• The order-k Markov Predictor• Sequential Pattern Mining + Association Rules

– Mine frequent patterns followed by the users.– Generating rules– Based on the similarity of the current movement sequence and the

premise of the rules, make prediction• Motion Function Prediction

– Estimate the object’s future location by motion functions– Linear model:

– Non-linear model• More complicated mathematical fomula (e.g. Recursive Motion Function)

Page 4: Spatial-Temporal Models in  Location Prediction

Combining Spatial and Temporal Information

• A Hybrid Prediction Model for Moving Objects (ICDE08)– Hoyoung Jeung and Qing Liu and Heng Tao Shen and Xiaofang Zhou

• NextPlace: A Spatio-temporal Prediction Framework for Pervasive Systems (Pervasive Computing 11)– Salvatore Scellato and Mirco Musolesi and Cecilia Mascolo and Vito

Latora and Andrew T. Campbell

• Exploring Spatial-Temporal Trajectory Model for Location Prediction (MDM11)– Po-Ruey Lei and Tsu-Jou Shen and Wen-Chih Peng and Ing-Jiunn Su

Page 5: Spatial-Temporal Models in  Location Prediction

A Hybrid Prediction Model for Moving Objects

• Main Problem to solve:– When query time is far away from current time,

motion function will fail• Methodology (combine motion function and

periodic patterns)– Detection of Frequent Regions– Discovery of Trajectory Patterns– The Hybrid Prediction Algorithm• Distance time query / non-distance time query

Page 6: Spatial-Temporal Models in  Location Prediction

Detection of Frequent Regions

• Decomposing a whole trajectory into sub-trajectories

• Grouping positions at time offset k in each sub-traj.• Applying a clustering method

/ 2

x

y

time

Tt +T

t

t +2T

T

G 0 G 1

G 2

G 3

G 4 G 5

R 0

t

y

x

0

R 1 0

R 3 0 R 4

0R 5

0

Page 7: Spatial-Temporal Models in  Location Prediction

Discovery of Trajectory Patterns

• Note : )...(,...,, 212

2

1

1 nmj

tcj

tj

tj

t ttttRRRR n

n

m

m

City

0 1 2

0

Shoppingcenter

Workplace

Beach

Home

R 2 1

R 0 0

t

P 3 R 2 1R 1

1: 0.4^R 0 0

P 20.5

^ R 2 0R 1

0R 0 0:

R 1 0

R 1 1

R 2 0 P 1 R 1

1: 0.8R 0 0

P 00.9

R 1 0R 0

0:

Page 8: Spatial-Temporal Models in  Location Prediction

The Hybrid Prediction Algorithm

• If no pattern at tq?

• If patterns at tq? – Distance time query / non-distance time query

Trajectory PatternsSpatio-temporal datasets

-----

-----

HPM

Discovery Existing motion function

Query time Object’s recent movements

Query answer

a motion function

computing top-k

Page 9: Spatial-Temporal Models in  Location Prediction

• Forward algorithm ( When )– An object is likely to follow the trend of recent movements

• Candidate filtering

• Ranking candidates– Sim(the current movements, the premise of each pattern) x confidence– Current movements are more important

n

n

m

m

jk

cjk

jk

jk RRRR ,...,, 2

2

1

1

Query Processing : Near Future

dtt cq

kqtct

Page 10: Spatial-Temporal Models in  Location Prediction

kqtct

Query Processing : Distant Future

• Backward algorithm ( When )– The recent movements are not so important for prediction

• Candidate filtering

• Ranking candidates– ( Sim(premise) x penalty + Sim(consequence) ) x confidence– As the query time , the importance of recent movements

dtt cq

n

n

m

m

jk

cjk

jk

jk RRRR ,...,, 2

2

1

1

Page 11: Spatial-Temporal Models in  Location Prediction

NextPlace: A Spatio-temporal Prediction Framework for Pervasive Systems

• The main idea: human behavior is strongly determined by daily patterns

• User location prediction based on nonlinear time series analysis of visits that users pay to their most significant locations

• Estimates the time of the future visits and expected residence time in those locations, based on this , prediction locations at certain time

Page 12: Spatial-Temporal Models in  Location Prediction

Significant Places Extraction

• Weighted Gaussian distribution

Page 13: Spatial-Temporal Models in  Location Prediction

Predicting User Behavior

• For each location, try to predict when the next visits will take place and for how long they will last

• The predictions obtained for different locations are analyzed, in order to produce a unique prediction of where the user will be at a given future instant of time

Page 14: Spatial-Temporal Models in  Location Prediction

Predict next timestamp and residence time

• User history:– (start time; duration)– ((t1; d1);(t2; d2); ... ;(tn; dn))

• Search in the time series C sequences of m consecutive values that are closely similar to the last m values

• The next value of time series C is estimated by averaging all the values that follow each found sequence

• The next value of d is computed similiarly

Page 15: Spatial-Temporal Models in  Location Prediction

Predict next location

• For each location, predict the sequence of the next k visits (starting with k = 1) . Then a global sequence of all predicted visits (loc1; t1; d1); … ;(locn; tn; dn) is created

• If there is a prediction (loci; ti; di) which satisfies ti < T + threshold< ti + di , then loci is returned as predicted location

• If no prediction satisfies the above, – If it is because the time range is not long enough, extend the

prediction length– Else the user is not at any significant place

Page 16: Spatial-Temporal Models in  Location Prediction

Exploring Spatial-Temporal Trajectory Model for Location Prediction

Tricky part

Page 17: Spatial-Temporal Models in  Location Prediction

Solution

• Frequent Region Discovery and Trajectory Transformation

• Location Prediction Using a suffix tree-based model with both spatial and temporal information (spatial-temporal trajectory model- STT model)

Page 18: Spatial-Temporal Models in  Location Prediction

Framework

Page 19: Spatial-Temporal Models in  Location Prediction

trajectories

Page 20: Spatial-Temporal Models in  Location Prediction

STT model

Page 21: Spatial-Temporal Models in  Location Prediction

Prediction

• When a new sequence comes, the similarity is based on both spatial and temporal information.

Page 22: Spatial-Temporal Models in  Location Prediction

Thanks!Questions?