Learning Significant Locations and Predicting User Movement with GPS Daniel Ashbrook and Thad Starner

1


Learning Significant Locations and Predicting
User Movement with GPSDaniel Ashbrook and Thad
Starner

• Avinash Parnandi

2
Learning Significant Locations and Predicting
User Movement with GPS

• Learning Significant Locations Significant
  Locations are those locations which we commonly
  visit spend time, over a period of time(say
  one semester).
• e.g. Home, Work, KOH, Leavey, DRB-Lab,
  Smart-Final etc.
• Predicting User Movement After collecting the
  GPS data and analyzing it, it uses Markov Model
  to predict user movement and answer queries
  like right now
  the user is at home, what is the likely place
  shell go next?
• -- how likely is he to stop at the grocery store
  on way from school to home.
• Location is one of the most commonly used
  context. Given this other information can
  inferred i.e. what are you doing(Stadium-Match,
  Classroom-Studying, Movie Hall- Movie etc.)

3
Applications Single User

• Associate a To-Do list or Reminders with
  location. So the reminders will pop up
  depending on your location e.g. Buy
  Vegetables when you are near a grocery store.
• This is where Prediction spices up the
  system
• Early reminders --- Suppose youve a library
  book to return the phone predicts that youll
  be taking the library route today so itll
  prompt a reminder before you leave your home
  instead of reminding you near the library.
• Location prediction abilities could allow a
  wearable computer to optimize its transmissions
  based on availability of service in various
  locations and the knowledge of how its user moves
  throughout the day.

4
Applications Multi User

• Simplest example Will I see Bob today?
• Another application is scheduling a meeting
  of several people depending on their
  calendar.
• Serendipitous meeting.
• Intelligent Interruption Location models can
  be used to make an intelligent guess about
  whether the user is interruptible or not.
  Suppose your location says that you are in a
  lecture or at your workplace and cannot be
  disturbed so itll automatically take your phone
  into silent mode.
• Requires sharing of information. Various
  mechanisms are available like central server,
  sharing only with trusted associates etc. Here
  privacy issues needs to be taken care of.

5
Implementation

• Hardware Collects the data using a GPS
  receiver data logger.
• Software Processes the data, then by using a
  Markov Model makes prediction about user
  movements based upon this data.
• Duration of experiment 4 months
• Data logger recorded the output from the GPS
  receiver at an interval of one second, but only
  if the receiver was moving at one mile an hour
  or greater. This helps pre-process the data.
• Limitation of GPS Accuracy of the GPS
  receiver was 15 meters this means that the same
  physical location will have a different GPS
  coordinate from day to day.

6
Significant Locations

• We are interested in the places where the user
  spends time and by looking at the time gaps in
  the GPS data we can find these significant
  locations.
• Time gaps will occur when the data logger is
  not logging which happens when either you are
  stationary(less than 1mph) in an open place or
  you are in a building with no GPS signal. Both
  these cases mean that you spend your time here.
• The value of time gap is an important
  parameter. Whenever a point is found that has
  more than a certain time t between it and the
  previous point, we conclude that the point marks
  a significant location.
• So now weve the significant locations

7
Determining places Clustering

• Because GPS measurements taken in the same
  physical location can vary
  by as much as 15 meters, the logger may
  record different point
  for a location even if the user stops at
  precisely the same point .
• Hence create clusters of some radius r .
  All GPS locations inside a cluster are now
  recognized by the cluster ID. So original
  hundreds of thousands of GPS coordinates are now
  just a few significant locations.
• Clustering done by a variant of K-mean
  clustering Algorithm.

8
K Mean Clustering

1. Take one place point and a radius.
   All the points within this radius of
   the place are marked, and the mean of these
   points is found.
2. The mean is then taken as the new centre
   point, and the process is repeated.
3. This continues until the mean stops
   changing.
4. When the mean no longer moves, all points
   within its radius are placed in its
   cluster and removed from consideration

9
To determine the Radius
knee
10
Radius of a ClusterIdentifying Sub-locations

• Problem with radius of the cluster--
• Large R small transits like SSL-RTH wont be
  predicted. All itll say is campus, which is
  a cluster.
• Small RMany locations will be individual
  clusters which is exactly what we want to avoid
  by clustering. Also broader trips will not be
  predicted.
• Variable R Sub-locations inside a cluster,
  Campus can be a cluster RTH, SSL, Lyon can be
  sub- locations.

11
So now we have our locations, what next??
Prediction!!!!
12
Prediction

• A Markov Model is created for each location
  in the map with transition to other locations.
• What is a Markov Model?
• Markov Models are State Transition models
  
  with the
  nodes being the states with
  corresponding state
  transition
  probabilities between the nodes.
• It follows the Markov Rule i.e. future
  state depends on the current
  state and observational
  data independent of past states .
• A simple random walk is an example of Markov
  Chain.

13
How does it work?

• This model predicts user movement. Given
  current location, it reads out transition
  probabilities for all possible locations from
  this location and the one with the highest
  probability is accepted and taken as the next
  move.
• Probability here is the relative
  frequency of transitions.
• If no transition ever occurs between
  two nodes then the transition
  probability between those
  nodes is zero.

14
Future Work Limitations

• Future Work
• This project was for single user, so the next
  obvious step is Multi-User collaboration .
• Design a Markov Model to support time based
  prediction. Can predict where someone will go
  next but not when.
• Limitations
• Changes in schedule may take a long time to
  be reflected in the model.
• Does not update the user models in real time.

15
Related Work

• Multi-User Learning locations Prediction
• Work done by the same group.
• Mostly similar work, few changes to the
  approach.
• The data collected was implemented for multi
  user collaborative applications.
• Still does not support time prediction.
• Adaptive mobility prediction for location
  management using mobile positioning
• Location management scheme for mobile wireless
  networks is presented
• The mobile periodically compares its
  current location (GPS) with the predicted locatio
  n and sends an autonomous location update
  whenever the prediction error exceeds a certain
  threshold.

16
Related Work

• Learning Significant Locations from GPS Data
  with Time Window (Tang Jian , Meng Lingkui
  2006)
• They identify significant locations but no
  prediction.
• Distinguishes the location of the same place
  where the user went at different time i.e. OHE
  at 11 AM is different from OHE at 330 PM.
• They make use of the fact that every GPS point
  is acquired with a time stamp.
• Can be the basis for time based prediction .

17
References

• Learning Significant Locations and Predicting
  User Movement with GPS
• Using GPS to learn significant locations and
  predict movement across multiple users Daniel
  Ashbrook and Thad Starner
• Learning Significant Locations from GPS Data with
  Time Window Tang Jian, Meng Lingkui.
• Adaptive mobility prediction for location manageme
  nt using mobile positioning Zhang Rui ,
  Bassiouni Mostafa A .
• http//intoverflow.wordpress.com/2008/05/27/what-i
  s-a-hidden-markov-model/

18
Thanks!!!
19

• Questions ?????