Docstoc

Colocation

Document Sample
Colocation Powered By Docstoc
					Media Sharing based on Colocation Prediction in
               Urban Transport

                       Liam McNamara
                l.mcnamara @ cs.ucl.ac.uk


          Joint work with Cecilia Mascolo and Licia Capra
2
Talk Outline
           Challenges
           Approach


              - Selection Process
          - Temporal Mean
           Implementation


           Evaluation


              - Urban Transport Dataset
              - Content Generation
           Results


           Conclusions
Isn't this already possible?




   It is still manual, awkward and time consuming.

   We need intelligent automated wireless content sharing.


                                                             4
With Bluetooth, Isn't this easy?


   Finding others and sending a file is easy, but doing it
   successfully isn't.

   Its short range leads to very dynamic networks, with many short
   colocations.

   Who should be selected?
   Will the download complete?


                                                                5
6
Main Challenges


     Human proximity networks can have very high churn.

     Finding content that matches the user's interests.

     Most peers will be strangers, especially in cities.

     Required penetration would increase contention.


                                                            7
Approach: Colocation Prediction


  Human movement is seasonal, as are colocations.

  We estimate colocation using the mean past duration.

  Are colocations the same length through the day?

  We keep a mean for each time period throughout the day.

  A two-tier system of these profiles is employed.

                                                            8
Selection Process



                             Daniel
            Carol




                    Alice   Rock
                            Metal
            Bob             Blues


                                      9
                                      Pop
Selection Process                     Jazz
                                      Classical
  Rock
  Blues
  Country

                             Daniel
            Carol




                    Alice   Rock?
                            Metal?
   Pop      Bob             Blues?
   Rock

                                                  10
                                      Pop
Selection Process                     Jazz
                                      Classical
  Rock
  Blues
  Country

                             Daniel
            Carol




                    Alice   Rock?
                            Metal?
   Pop      Bob             Blues?
   Rock

                                                  11
Selection Process: Bob or Carol?

                  Pop                             Rock
                  Rock                            Blues
                                                  Country

   Bob                                                            Carol

 Carol may offer more appropriate content.

 Transfer completion is paramount.

 Therefore: choose host with longest remaining colocation time:

 remaining = predictedLength(host,time) -
 colocationStart(host)‫‏‬


                                                                          12
Approach: Temporal Mean

   A global anonymous profile is maintained.

   Recording all colocation lengths in the relevant time slot.
                                       Hour of day
             Overall   0-4      4-8      8-12    12-16    16-20   20-24

 Global        15.9    25.7     14.3      3.5        5     2.8     18


               12.4      -                23         10   13.8      -


                                                                          13
Approach: Temporal Mean


   If User1 is encountered frequently, it gains its own profile.


                                       Hour of day
             Overall    0-4     4-8      8-12    12-16   16-20     20-24

 Global        15.9    25.7     14.3      3.5        5     2.8      18


 User1         12        -        -       12         -     -         -


                                                                           14
Approach: Temporal Mean
   Colocations with User1 are then added to both
   the global and personal profiles.

   Personal timed > Personal overall > Global timed

                                    Hour of day
            Overall   0-4     4-8     8-12    12-16   16-20   20-24

 Global       15.9    25.7   14.3      3.5        5    2.8     18


 User1        12.4     -       -       13         7   13.8      -


                                                                      15
Talk Outline
             Challenges
             Approach
                    - Selection Process
                - Temporal Mean
             Implementation
             Evaluation
                - Urban Transport Dataset
                - Content Generation
             Results
             Conclusions
Implementation

 Nokia N70 Smartphones

 2GB MicroSD cards

 PyS60 Python Interpreter (<2KLOC)‫‏‬

 Experimental runs were performed on busy
 commuter trains.

 Achieved data rates went as low as
 100Kbps
                                            17
Urban Transport Dataset

  We used anonymous passenger movement traces from a large
  city's subway system.

   Recorded over 1 month
   1,000,000s of journeys


   100,000s of passengers




  Many passengers commute regular journeys, at the same time each
  day.



                                                                18
Dataset features




                   19
Dataset processing



  We needed to convert the format of the traceset.

  From journey:
    userid, start_station, start_time, end_station, end_time


  To colocation instance:
    userid1, userid2, start_time, end_time


                                                         20
Dataset processing
            Millbrae
                         SFO         San Bruno South SF    Colma Daly City   Balboa




       8.30 Alice in   8.45 Bob in                   9.30 Alice out 9.45 Bob out


  Would Alice and Bob have been on the same train?

  What time would they have shared the train?

  Journeys were then timed using official timetables, and intersected with
  all others.

      Alice            Bob           8.45                 9.30                        21
Content Modelling



  Social music website Last.fm was used to generate realistic music
  collections for users.

  Playlist information was recorded from over 500,000 users.

  User libraries are constructed to contain relative proportions of
  genres and the artists in those genres.



                                                                      22
Results: Selection Methods


   Random – Sources are chosen at random, and a transfer is
   always initiated.

   Prediction – Our approach, with both coarse and fine grained
   storage of colocation data.

   Oracle – Selects longest available neighbour, possessing future
   knowledge of every colocation.



                                                                     23
Results: Prediction success




                              24
Results: Efficiency




                      25
Results: Genre availability




                              26
Related Work

  Bluespots: "Bluetooth Content Distribution Stations on Public
  Transport”‫‏‬J.‫‏‬LeBrun‫‏‬and‫‏‬CN.‫‏‬Chuah,‫‏‬Mobishare‫.6002‏‬
  Uses infrastructure installed on buses to distribute data.

  "BlueTorrent: Cooperative Content Sharing for Bluetooth Users" S.
  Jung, U. Lee, A. Chang, DK. Cho, and M. Gerla, Percom 2007.
  Swarming is employed to overcome short colocations.

  “Wireless Ad Hoc Podcasting”‫‏‬V.‫‏‬Lenders,‫‏‬G.‫‏‬Karlsson‫‏‬and‫‏‏‬M.‫‏‬May,‫‏‬
  SECON 2007.
  Users subscribe to channels of content interest types.

                                                                      27
Conclusions


  Selection of download source has a large impact on the efficacy of the
  system.

  Not downloading can improve overall system performance.

  Movements of a subset of the strangers we meet in a city can be
  predicted.




                                                                     28
Future work



   Periodic colocation pattern identification and utilisation.

   Modeling of content attributes and finer grained user tastes.

   Larger deployment of our PyS60 implementation.




                                                                    29
          Questions?


http://www.cs.ucl.ac.uk/staff/l.mcnamara



                                           30

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:1
posted:11/6/2011
language:English
pages:30