Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Trace-driven Context-Aware Mobile Networks: Towards Mobile ...


Published on

  • Be the first to comment

  • Be the first to like this

Trace-driven Context-Aware Mobile Networks: Towards Mobile ...

  1. 1. Trace-driven Context-aware Mobile Networks Towards Mobile Social Networks Ahmed Helmy Computer and Information Science and Engineering (CISE) Dept Mobile Networking Lab (NOMADS group) University of Florida [email_address]
  2. 2. Birds-Eye View : Mobile & Networking Lab Architecture & Protocol Design Methodology & Tools Test Synthesis ( STRESS ) Mobility Modeling ( IMPORTANT, TVC ) Protocol Block Analysis ( BRICS ) Query Resolution in Wireless Networks ( ACQUIRE & Contacts ) Robust Geographic Wireless Services (Geo-Routing, Geocast, Rendezvous) Gradient Routing ( RUGGED ) Worms, Traceback in Mobile Networks Behavioral Analysis in Wireless Networks ( IMPACT & MobiLib ) Multicast-based Mobility ( M&M ) Mobility-Assisted Protocols ( MAID ) Socially-Aware Networks Network Simulator ( NS-2 ) Protocol Independent Multicast ( PIM )
  3. 3. <ul><li>Future network devices are mobile & ‘personal’ </li></ul><ul><ul><li>Very tight coupling between devices & humans </li></ul></ul><ul><li>Network performance significantly affected by (and affects) users behavior: </li></ul><ul><ul><li>Movement, grouping, on-line activity, trust, cooperation… </li></ul></ul><ul><li>How do users behave in mobile societies? </li></ul><ul><li>What kinds of protocols/networks survive & perform well in highly mobile societies? </li></ul>Introduction & Problem Scope
  4. 4. Paradigm Shift in Protocol Design <ul><ul><li>May end up with suboptimal performance or failures due to lack of context in the design </li></ul></ul>Used to: Propose to: Modify to improve performance and failures for specific context Evaluate using models (random mobility, traffic, …) Design general purpose protocols Utilize insights from context analysis to fine-tune protocol parameters Design ‘application class’-specific parameterized protocols Analyze, model deployment context
  5. 5. Problem Statement <ul><li>How to gain insight into deployment context? </li></ul><ul><li>How to utilize insight to design future services? </li></ul><ul><li>Approach </li></ul><ul><li>Extensive trace-based analysis to identify dominant trends & characteristics </li></ul><ul><li>Analyze user behavioral patterns </li></ul><ul><ul><li>Individual user behavior and mobility </li></ul></ul><ul><ul><li>Collective user behavior: grouping, encounters </li></ul></ul><ul><li>Integrate findings in modeling and protocol design </li></ul><ul><ul><li>I. User mobility modeling – II. Behavioral grouping </li></ul></ul><ul><ul><li>III. Information dissemination in mobile societies, profile-cast </li></ul></ul>
  6. 6. The TRACE framework E mploy (Modeling & Protocol Design) MobiLib T race A nalyze C haracterize ( C luster) R epresent
  7. 7. Vision: Community-wide Wireless/Mobility Library <ul><li>Library of </li></ul><ul><ul><li>Measurements from Universities, vehicular networks </li></ul></ul><ul><ul><li>Realistic models of behavior (mobility, traffic, friendship, encounters) </li></ul></ul><ul><ul><li>Benchmarks for simulation and evaluation </li></ul></ul><ul><ul><li>Tools for trace data mining </li></ul></ul><ul><li>Use insights to design future context-aware protocols? </li></ul><ul><li> </li></ul>T race
  8. 8. Libraries of Wireless Traces <ul><li>Multi-campus (community-wide) traces: </li></ul><ul><ul><li>MobiLib (USC (04-06), now @ UFL) </li></ul></ul><ul><ul><ul><li> </li></ul></ul></ul><ul><ul><ul><li>15+ Traces from: USC, Dartmouth, MIT, UCSD, UCSB, UNC, UMass, GATech, Cambridge, UFL, … </li></ul></ul></ul><ul><ul><ul><li>Tools for mobility modeling (IMPORTANT, TVC), data mining </li></ul></ul></ul><ul><ul><li>CRAWDAD (Dartmouth) </li></ul></ul><ul><li>Types of traces: </li></ul><ul><ul><li>University Campus (mainly WLANs) </li></ul></ul><ul><ul><li>Conference AP and encounter traces </li></ul></ul><ul><ul><li>Municipal (off-campus) wireless </li></ul></ul><ul><ul><li>Bus & vehicular wireless networks </li></ul></ul><ul><ul><li>Others … (on going) </li></ul></ul>T race
  9. 9. Wireless Networks and Mobility Measurements <ul><li>In our case studies we use WLAN traces </li></ul><ul><ul><li>From University campuses & corporate networks (4 universities, 1 corporate network) </li></ul></ul><ul><ul><li>The largest data sets about wireless network users available to date (# users / lengths) </li></ul></ul><ul><ul><li>No bias: not “special-purpose”, data from all users in the network </li></ul></ul><ul><li>We also analyze </li></ul><ul><ul><li>Vehicular movement trace (Cab-spotting) </li></ul></ul><ul><ul><li>Human encounter trace (at Infocom Conf) </li></ul></ul>T race
  10. 10. Case study I – Individual mobility
  11. 11. Case Study I: Goal <ul><li>To understand the mobility/usage pattern of individual wireless network users </li></ul><ul><li>To observe how environments/user type/trace-collection techniques impact the observations </li></ul><ul><li>To propose a realistic mobility model based on empirical observations </li></ul><ul><ul><li>That is mathematically tractable </li></ul></ul><ul><ul><li>That is capable of characterizing multiple classes of mobility scenarios </li></ul></ul>
  12. 12. IMPACT : Investigation of Mobile-user Patterns Across University Campuses using WLAN Trace Analysis* * W. Hsu, A. Helmy, “ IMPACT : Investigation of Mobile-user Patterns Across University Campuses using WLAN Trace Analysis”, two papers at IEEE Wireless Networks Measurements (WiNMee) , April 2006 - 4 major campuses – 30 day traces studied from 2+ years of traces - Total users > 12,000 users - Total Access Points > 1,300 <ul><li>Understand changes of user association behavior w.r.t. </li></ul><ul><ul><li>Time - Environment - Device type - Trace collection method </li></ul></ul>04/20/05-05/19/05 Event-based (Bldg) University campus Generic 4/20/05 – 3/31/06 USC 09/22/02- 10/21/02 Polling University campus PDA only 9/22/02 – 12/8/02 UCSD July ’03 April ’04 Event-based University campus Generic w/ subgroup 4/01/01 – 6/30/04 Dartmouth Whole trace Polling 3 corporate buildings Generic 7/20/02 – 8/17/02 MIT Analyzed part Collection method Environment User type Trace duration Trace source
  13. 13. Metrics for Individual Mobility Analysis <ul><li>What kind of spatial preference do users exhibit? </li></ul><ul><ul><li>The percentile of time spent at the most frequently visited locations </li></ul></ul><ul><li>What kind of temporal repetition do users exhibit? </li></ul><ul><ul><li>The probability of re-appearance </li></ul></ul><ul><li>How often are the nodes present? </li></ul><ul><ul><li>Percentage of “online” time </li></ul></ul>R epresent
  14. 14. <ul><li>Individual users access only a very small portion of APs in the network. </li></ul><ul><li>On average a user spends more than 95% of time at its top 5 most visited APs. </li></ul><ul><li>Long-term mobility is highly skewed in terms of time associated with each AP. </li></ul><ul><li>Users exhibit “on”/”off” behavior that needs to be modeled. </li></ul>Observations: Visited Access Points (APs) Prob.(coverage > x) Fraction of online time associated with the AP CCDF of coverage of users [percentage of visited APs] Average fraction of time a MN associates with APs
  15. 15. <ul><li>Clear repetitive patterns of association in wireless network users. </li></ul><ul><li>Typically, user association patterns show the strongest repetitive pattern at time gap of one day/one week. </li></ul>Repetitive Behavior
  16. 16. Mobility Characteristics from WLANs <ul><li>Simple existing models are very different from the characteristics in WLAN </li></ul>C haracterize Prob.(online time fraction > x) On/off activity pattern Skewed location preference Periodic re-appearance
  17. 17. Mobility Models <ul><li>Mobility models are of crucial importance for the evaluation of wireless mobile networks [IMP03] </li></ul><ul><li>Requirements for mobility models </li></ul><ul><ul><li>Realism (detailed behavior from traces) </li></ul></ul><ul><ul><li>Parameterized, tunable behavior </li></ul></ul><ul><ul><li>Mathematical tractability </li></ul></ul><ul><li>Related work on mobility modeling </li></ul><ul><ul><li>Random models (Random walk/waypoint): inadequate for human mobility </li></ul></ul><ul><ul><li>Improved synthetic models (pathway model, RPGM, WWP, FWY, MH) – more realistic, difficult to analyze </li></ul></ul><ul><ul><li>Trace-based model (T/T++): trace-specific, not general </li></ul></ul>
  18. 18. <ul><li>Skewed location visiting preference </li></ul><ul><ul><li>Create “communities” to be the preferred area of movement </li></ul></ul><ul><ul><li>Each node can have its own community </li></ul></ul><ul><li>Node moves with two different epoch types – Local or roaming </li></ul><ul><ul><li>Each epoch is a random-direction, straight-line movement </li></ul></ul><ul><ul><li>Local epochs in the community </li></ul></ul><ul><ul><li>Roaming epochs around the whole simulation area </li></ul></ul>Time-variant Community (TVC) Model (W. Hsu, Thyro, K. Psounis, A. Helmy, “Modeling Time-variant User Mobility in Wireless Mobile Networks”, IEEE INFOCOM, 2007, Trans. on Networking) 25% 75% E mploy
  19. 19. Tiered Time-variant Community (TVC) Model <ul><li>Periodical re-appearance </li></ul><ul><ul><li>Create structure in time – Periods </li></ul></ul><ul><ul><li>Node moves with different parameters in periods to capture time-dependent mobility </li></ul></ul><ul><ul><li>Repetitive structure </li></ul></ul><ul><li>Finer granularity in space & time </li></ul><ul><ul><li>Multi-tier communities </li></ul></ul><ul><ul><li>Multiple time periods </li></ul></ul>E mploy
  20. 20. Using the TVC Model – Reproducing Mobility Characteristics <ul><li>(STEP1) Identify the popular locations; assign communities </li></ul><ul><li>(STEP2) Assign parameters to the communities according to stats </li></ul><ul><li>(STEP3) Add user on-off patterns (e.g., in WLAN, users are usually off when moving) </li></ul>
  21. 21. Using the TVC Model – Reproducing Mobility Characteristics <ul><li>WLAN trace (example: MIT trace) </li></ul>Skewed location visiting preference Periodic re-appearance * Model-simplified: single community per node. Model-complex: multiple communities ** Similar matches achieved for USC and Dartmouth traces
  22. 22. Using the TVC Model – Reproducing Mobility Characteristics <ul><li>Vehicular trace (Cab-spotting) </li></ul>
  23. 23. Using the TVC Model – Reproducing Mobility Characteristics <ul><li>Human encounter trace at a conference </li></ul>Inter-meeting time Encounter duration time A encounters B Encounter duration Inter-meeting time
  24. 24. Case study II – Groups in WLAN
  25. 25. Case Study II: Goal <ul><li>Identify similar users (in terms of long run mobility preferences) from the diverse WLAN user population </li></ul><ul><ul><li>Understand the constituents of the population </li></ul></ul><ul><ul><li>Identify potential groups for group-aware service </li></ul></ul><ul><li>In this case study we classify users based on their mobility trends (or location-visiting preferences) </li></ul><ul><ul><li>We consider semester-long USC trace (spring 2006, 94days) and quarter-long Dartmouth trace (spring 2004, 61 days) </li></ul></ul>
  26. 26. Representation of User Association Patterns <ul><li>We choose to represent summary of user association in each day by a single vector </li></ul><ul><ul><li>a = { a j : fraction of online time user i spends at AP j on day d } </li></ul></ul><ul><li>Summarize the long-run mobility in an “association matrix” </li></ul><ul><li>Office, 10AM -12PM </li></ul><ul><li>Library, 3PM – 4PM -Class, 6PM – 8PM </li></ul>R epresent Association vector: (library, office, class) =(0.2, 0.4, 0.4)
  27. 27. Eigen-behavior <ul><li>Eigen-behaviors: The vectors that describe the maximum remaining power in the association matrix (obtained through Singular Value Decompostion) with quantifiable importance </li></ul><ul><li>Eigen-behavior Distance calculates similarity of users by weighted inner products of eigen-behaviors. </li></ul><ul><li>Assoc. patterns can be re-constructed with low rank & error </li></ul><ul><li>Benefits: Reduced computation and noise </li></ul>
  28. 28. Similarity-based User Classification <ul><li>With the distance between users U and V defined as 1-Sim(U,V), we use hierarchical clustering to find similar user groups. </li></ul>USC Dartmouth * AMVD = Average Minimum Vector Distance
  29. 29. Validation of User Groups <ul><li>Significance of the groups – users in the same group are indeed much more similar to each other than randomly formed groups (0.93 v.s. 0.46 for USC, 0.91 v.s. 0.42 for Dartmouth) </li></ul><ul><li>Uniqueness of the groups – the most important group eigen-behavior is important for its own group but not other groups </li></ul>0.004 0.005 Other groups 0.727 0.779 Its own group Dartmouth USC Significance score of top eigen-behavior for
  30. 30. User Groups in WLAN - Observations <ul><li>Identified hundreds of distinct groups of similar users </li></ul><ul><li>Skewed group size distribution – the largest 10 groups account for more than 30% of population on campus. Power-law distributed group sizes. </li></ul><ul><li>Most groups can be described by a list of locations with a clear ordering of importance </li></ul><ul><li>We also observe groups visiting multiple locations with similar importance – taking the most important location for each user is not sufficient </li></ul>
  31. 31. Case study III – Encounter Patterns
  32. 32. Case Study III: Goal <ul><li>Understand inter-node encounter patterns from a global perspective </li></ul><ul><ul><li>How do we represent encounter patterns? </li></ul></ul><ul><ul><li>How do the encounter patterns influence network connectivity and communication protocols? </li></ul></ul><ul><li>Encounter definition: </li></ul><ul><ul><li>In WLAN: When two mobile nodes access the same AP at the same time they have an ‘encounter’ </li></ul></ul><ul><ul><li>In DTN: When two mobile nodes move within communication range they have an ‘encounter’ </li></ul></ul>
  33. 33. Prob. (unique encounter fraction > x) Prob. (total encounter events > x) CCDF of unique encounter count CCDF of total encounter count <ul><li>In all the traces, the MNs encounter a small fraction of the user population. </li></ul><ul><li>A user encounters 1 . 8% - 6% on average of the user population (except UCSD) </li></ul><ul><li>The number of total encounters for the users follows a BiPareto distribution. </li></ul>Observations: Encounters
  34. 34. Encounter-Relationship (ER) graph <ul><li>Draw a link to connect a pair of nodes if they ever encounter with each other … Analyze the graph properties? </li></ul>Group of good friends… Cliques with random links to join them R epresent
  35. 35. Av. Path Length Clustering Coefficient (CC) Small Worlds of Encounters Normalized CC and PL <ul><li>The encounter graph is a Small World graph (high CC , low PL ) </li></ul><ul><li>Even for short time period (1 day) its metrics ( CC , PL ) almost saturate </li></ul><ul><li>Encounter graph: nodes as vertices and edges link all vertices that encounter </li></ul>Random graph Regular graph Small World
  36. 36. Background: Delay Tolerant Networks ( DTN ) <ul><li>DTNs are mobile networks with sparse, intermittent nodal connectivity </li></ul><ul><li>Encounter events provide the communication opportunities among nodes </li></ul><ul><li>Messages are stored and moved across the network with nodal mobility </li></ul>A B C
  37. 37. Information Diffusion in DTNs via Encounters <ul><li>Epidemic routing (spatio-temporal broadcast) achieves almost complete delivery </li></ul>Robust to selfish nodes (up to ~40%) Robust to the removal of short encounters Unreachable ratio (Fig: USC) Trace duration = 15 days
  38. 38. <ul><li>Top-ranked friends form cliques and low-ranked friends are key to provide random links (short cuts) to reduce the degree of separation in encounter graph. </li></ul>Encounter-graphs using Friends <ul><li>Distribution for friendship index FI is exponential for all the traces </li></ul><ul><li>Friendship between MNs is highly asymmetric </li></ul><ul><li>Among all node pairs: < 5% with FI > 0.01, and <1% with FI > 0.4 </li></ul>
  39. 39. Profile-cast W. Hsu, D. Dutta, A. Helmy, Mobicom 2007 <ul><li>Sending messages to others with similar behavior, without knowing their identity </li></ul><ul><ul><li>Announcements to users with specific behavior V </li></ul></ul><ul><ul><li>Interest-based ads, similarity resource discovery </li></ul></ul><ul><li>Assuming DTN-like environment </li></ul>A B E C D Is B similar to V ? Is E similar to V ? ? Is C/D similar to V ?
  40. 40. Profile-cast Use Cases <ul><li>Mobility-based profile-cast </li></ul><ul><ul><li>Targeting group of users who move in a particular pattern (lost-and-found, context-aware messages, moviegoers) </li></ul></ul><ul><ul><li>Approach: use “similarity metric” between users </li></ul></ul><ul><li>Mobility-independent profile-cast </li></ul><ul><ul><li>Targeting people with a certain characteristics independent of mobility (classic music lovers) </li></ul></ul><ul><ul><li>Approach: use “Small World” encounter patterns </li></ul></ul>
  41. 41. Mobility-based Profile-cast Scoped message spread in the mobility space Mobility space S D D S D N N N N Forward??
  42. 42. Profile-cast Operation <ul><li>Profiling user mobility </li></ul><ul><ul><li>The mobility of a node is represented by an association matrix </li></ul></ul><ul><ul><li>Singular value decomposition provides a summary of the matrix (A few eigen-behavior vectors are sufficient, e.g. for 99% of users at most 7 vectors describe 90% of power in the association matrix) </li></ul></ul>S N N N N 1. profiling Each row represents an association vector for a time slot An entry represents the percentage of online time during time slot i at location j Sum. vectors
  43. 43. Profile-cast Operation <ul><li>Determining user similarity </li></ul><ul><ul><li>S sends Eigen behaviors for the virtual profile to N </li></ul></ul><ul><ul><li>N evaluated the similarity by weighted inner products of Eigen-behaviors </li></ul></ul><ul><ul><li>Message forwarded if Sim(U,V) is high (the goal is to deliver messages to nodes with similar profile) </li></ul></ul><ul><ul><li>Privacy conserving: N and S do not send information about their own behavior </li></ul></ul>2. Forwarding decision S N N N N 1. profiling
  44. 44. Profile-cast Evaluation <ul><li>Epidemic: Near perfect delivery ratio, low delay, high overhead </li></ul><ul><li>Centralized: Near perfect delivery ratio, low overhead, a bit extra delay </li></ul><ul><li>Decentral: provides tradeoff between delivery & overhead </li></ul><ul><li>Random: poor delivery ratio </li></ul>Epidemic Decentral Decentral Decentral Random Random Random Random - Decentralized I-cast achieves: > 50% reduction in overhead of Epidemic >30% increase in delivery of Random * Results presented as the ratio to epidemic routing
  45. 45. Evaluation - Result <ul><li>Centralized: Excellent success rate with only 3% overhead. </li></ul><ul><li>Similarity-based: </li></ul><ul><li>(1) 61% success rate at low overhead, 92% success rate at 45% overhead </li></ul><ul><li>(2) A flexible success rate – overhead tradeoff </li></ul><ul><li>RTx with infinite TTL: Much more overhead undersimilar success rate </li></ul><ul><li>Short RTx with many copies: Good success rate/overhead, but delay is still long </li></ul>Success Rate Delay Overhead more overhead 92% 45%
  46. 46. Profile-cast Initial Results <ul><li>Adjustable overhead/delivery rate tradeoff </li></ul><ul><ul><li>61% delivery rate of flooding with 3% overhead </li></ul></ul><ul><ul><li>92% delivery rate with 45% overhead </li></ul></ul><ul><li>Better than single random walk in terms of delay, delivery rate </li></ul><ul><li>Multiple short random walks also work well in this case </li></ul>
  47. 47. Future Work <ul><li>Sending to a mobility profile specified by the sender </li></ul><ul><ul><li>Gradient ascend followed by similarity comparison (in the mobility space) </li></ul></ul><ul><li>Mobility independent profile-cast </li></ul><ul><ul><li>The encounter pattern provides a network in which most nodes are reachable </li></ul></ul><ul><ul><li>We don’t want to flood – How to leverage the Small World encounter pattern to reach the “neighborhood” of most nodes efficiently? </li></ul></ul>
  48. 48. Future Work <ul><ul><li>One-copy-per-clique in the “mobility space” </li></ul></ul><ul><ul><li>We expect this to work because similarity in mobility leads to frequent encounters </li></ul></ul>
  49. 49. Future Directions (Applications) <ul><li>Detect abnormal user behavior & access patterns based on previous profiles </li></ul><ul><li>Behavior aware push/caching services (targeted ads, events of interest, announcements) </li></ul><ul><li>Caching based on behavioral prediction </li></ul><ul><li>Can/should we extend this paradigm to include social aspects (trust, friendship, …)? </li></ul><ul><li>Privacy issues and mobile k-anonymity </li></ul>
  50. 50. On Mobility & Predictability of VoIP & WLAN Users J. Kim, Y. Du, M. Chen, A. Helmy, Crawdad 2007 Work in-progress Markov O(2) Predictor Accuracy VoIP User Prediction Accuracy <ul><li>VoIP users are highly mobile and exhibit dramatic difference in behavior than WLAN users </li></ul><ul><li>Prediction accuracy drops from ave ~62% for WLAN users to below 25% for VoIP users </li></ul><ul><li>Motivates </li></ul><ul><ul><li>Revisiting mobility modeling </li></ul></ul><ul><ul><li>Revisiting mobility prediction </li></ul></ul>
  51. 51. Gender-based feature analysis in Campus-wide WLANs U. Kumar, N. Yadav, A. Helmy, Mobicom 2007, Crawdad 2007 <ul><li>- Able to classify users by gender using knowledge of campus map </li></ul><ul><li>Users exhibit distinct on-line behavior, preference of device and mobility based on gender </li></ul><ul><li>On-going Work </li></ul><ul><ul><li>How much more can we know? </li></ul></ul><ul><ul><li>What is the “information-privacy trade-off” ? </li></ul></ul>visitors Males Females University Campus Fraternity Sorority traces traces
  52. 52. Instructor sensor-adhoc sensor-adhoc sensor-adhoc Embedded sensor network The Next Generation (Boundless) Classroom Students <ul><li>Integration of wired Internet, WLANs, Adhoc Mobile and Sensor Networks </li></ul><ul><li>Will this paradigm provide better learning experience for the students? </li></ul>Challenges Real world group experiments (structural health monitoring) sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor WLAN/adhoc WLAN/adhoc Multi-party conference Tele-collaboration tools WLAN/adhoc
  53. 53. Emerging Wireless & Multimedia Technologies Protocols, Applications, Services Human Behavior Mobility, Load Dynamics Future Directions: Technology-Human Interaction The Next Generation Classroom
  54. 54. Emerging Wireless & Multimedia Technologies Human Computer Interaction (HCI) & User Interface Educational/ Learning Experience Education Psycology Cognitive Sciences Mobility Models Traffic Models Protocol Design Context-aware Networking Engineering Application Development Service Provisioning Multi-Disciplinary Research Measurements How to Evaluate? How to Capture? How to Design? Social Sciences Human Behavior Mobility, Load Dynamics Protocols, Applications, Services
  55. 55. Disaster Relief (Self-Configuring) Networks sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor
  56. 56. On-going and Future Directions Utilizing mobility <ul><ul><li>Controlled mobility scenarios </li></ul></ul><ul><ul><ul><li>DakNet, Message Ferries, Info Station </li></ul></ul></ul><ul><ul><li>Mobility-Assisted protocols </li></ul></ul><ul><ul><ul><li>Mobility-assisted information diffusion: EASE, FRESH, DTN, $100 laptop </li></ul></ul></ul><ul><ul><li>Context-aware Networking </li></ul></ul><ul><ul><ul><li>Mobility-aware protocols: self-configuring, mobility-adaptive protocols </li></ul></ul></ul><ul><ul><ul><li>Socially-aware protocols: security, trust, friendship, associations, small worlds </li></ul></ul></ul><ul><ul><li>On-going Projects </li></ul></ul><ul><ul><ul><li>Next Generation (Boundless) Classroom </li></ul></ul></ul><ul><ul><ul><li>Disaster Relief Self-configuring Survivable Networks </li></ul></ul></ul>
  57. 57. Thank you! Ahmed Helmy [email_address] URL: MobiLib:
  58. 58. Emerging Wireless Communication <ul><li>Opportunities </li></ul><ul><li>Challenges </li></ul><ul><ul><li>Dynamic network structure </li></ul></ul><ul><ul><li>Decentralized service paradigm </li></ul></ul><ul><ul><li>Tight coupling between the devices and individuals </li></ul></ul>
  59. 59. Outline Complete case Detailed behavior analysis Future Work
  60. 60. Trace Sets <ul><li>Available information from WLAN traces </li></ul><ul><ul><li>MAC addresses of the devices as identifiers </li></ul></ul><ul><ul><li>Location/Time of users (our main focus) </li></ul></ul>2197745 4433 2230200 13320 2257917 643 2285119 1017 2297134 7153 2304287 6744 Node: e0_12_29_fc_ba_8c Association Start time Location_ID Duration T race
  61. 61. Summary (Case Study I) <ul><li>We observe some omni-present mobility characteristics from WLANs. </li></ul><ul><li>These characteristics are not captured by existing synthetic mobility models (i.e., hence the models are not realistic) </li></ul><ul><li>We propose the Time-variant Community (TVC) model, which is realistic, theoretically tractable, and flexible </li></ul>
  62. 62. Theoretical Tractability <ul><li>For the TVC model, we can derive </li></ul><ul><ul><li>Nodal spatial distribution – the demographic profile of the mobility model </li></ul></ul><ul><ul><li>Average node degree – important for cluster maintenance and geographic routing </li></ul></ul><ul><ul><li>Hitting time/ Meeting time – important for routing performance analysis </li></ul></ul><ul><li>With low error when the communication range is small compared to the community sizes (communication disk < 25% of community) </li></ul>
  63. 63. Theoretical Tractability
  64. 64. Theory Derivation – Hitting Time <ul><li>Hitting time – the time for a node to move into the communication range of a randomly chosen target coordinate, starting from the stationary distribution </li></ul>(hit)
  65. 65. Theory Derivation – Hitting Time <ul><li>Weighted average conditioned on the relative location of the ‘target’ </li></ul><ul><li>Calculate the unit-time hitting probability for each scenario </li></ul><ul><li>Calculate hitting probability for the whole time period </li></ul><ul><li>Calculate the conditional hitting time </li></ul>
  66. 66. Application II : Trace-based Mobility Modeling <ul><li>Skewed location preference </li></ul><ul><ul><li>Nodes spend 95% of time at top 5 preferred locations. </li></ul></ul><ul><ul><li>Heavily visited “preferred spots” </li></ul></ul><ul><li>Repetitive behavior </li></ul><ul><ul><li>Nodes show up repeatedly at the same location after integer multiples of days. </li></ul></ul><ul><ul><li>Periodical “daily/weekly schedules” </li></ul></ul>
  67. 67. Similarity-based User Classification: Association-based Representation* <ul><li>For a given day d , user assoc. vector is defined by n -element vector </li></ul><ul><ul><li>a = { a j : fraction of online time user i spends at AP j on day d } </li></ul></ul><ul><ul><li>‘ n ’ is the number of APs </li></ul></ul><ul><ul><li>Use zero vector for off-line users </li></ul></ul><ul><li>Vector elements quantify relative attraction of AP to user </li></ul><ul><li>User Association Consistency </li></ul><ul><ul><li>User i is ‘consistent’ if daily assoc. vectors can be grouped into few clusters </li></ul></ul><ul><ul><li>Use clustering with Manhattan distance measure </li></ul></ul>Association vector: (AP1, AP2, AP3) =(0.2, 0.4, 0.4) * W. Hsu, D. Dutta, A. Helmy, Mobicom 2007 (AP2:library, 1:30PM-2:30PM) (AP1:office, 10AM-12PM) (AP3: class, 6PM-8PM)
  68. 68. Summarizing user associations <ul><li>Association matrix: concatenate user association vectors for all days into a matrix. </li></ul><ul><li>To summarize, perform SVD and store the top-k eigen values/vectors. </li></ul><ul><li>What value of k we have to use for a good representation of the matrix? </li></ul><ul><ul><li>Captured matrix power = </li></ul></ul><ul><li>How much is the reconstruction error? </li></ul><ul><ul><li>Matrix norms ||X-Xk|| p /||X|| p where </li></ul></ul>Daily association vector
  69. 69. Assoc. patterns can be re-constructed with low rank and low error * Matrix reconstruction error < 5% Summarizing user associations 90% 4 94% Dartmouth (Bldg) 90% 6 92% Dartmouth (AP) 90% 6 95% USC (Bldg) power # vectors (rank) % users Trace
  70. 70. Clustering Users with Similar Behavior <ul><li>Exhaustive comparison of assoc. vectors: </li></ul><ul><ul><li>Find average of | a jd - a id | over all days d for all i,j pairs </li></ul></ul><ul><ul><li>Drawback: O ( nd 2 ) for each pair </li></ul></ul><ul><li>Compare similarity of eigen-vectors obtained from SVD </li></ul><ul><li>Use weighted inner products of eigen vectors U , V </li></ul><ul><ul><li>, </li></ul></ul><ul><ul><li>w ui = proportion of power of SV </li></ul></ul><ul><ul><li>D(U,V) = 1 - Sim(U,V) </li></ul></ul><ul><ul><li>Corr > 91% with exhaustive </li></ul></ul>Can achieve very good clustering efficiently using distributed computation A handful of eigen-vectors can capture most of the behavior power
  71. 71. Encounter Events <ul><li>Derived from simultaneous associations to the same locations </li></ul><ul><li>How many other nodes does a node encounter with? </li></ul>Prob. (unique encounter fraction > x) 0.5 On avg. only 2%~7% of population
  72. 72. Encounter-Relationship graph <ul><li>To our surprise, disconnected pairs of nodes are low!! </li></ul>Disconnected Ratio (%)
  73. 73. Summary (Case Study II) <ul><li>We use SVD to obtain eigen-behaviors of individual users. </li></ul><ul><li>We use the eigen-behavior distances and hierarchical clustering to classify WLAN users into similar groups. </li></ul><ul><li>This finding is useful for mobility modeling (identifying group sizes and their frequently visited locations), network management, abnormality detection, and group-aware protocol (i.e., profile-cast, our future work) </li></ul>
  74. 74. Summary (Case Study III) <ul><li>The distribution of encounters in real WLAN trace is very different from synthetic models </li></ul><ul><li>The encounter-relationship graph displays SmallWorld characteristics </li></ul><ul><li>Despite a low encounter ratio of the whole population, the encounter events lead to a robust, reachable network (with long delay). </li></ul>
  75. 75. Future Work – Profile-cast
  76. 76. Goal <ul><li>To send messages to a group of nodes within the general population </li></ul><ul><ul><li>The group is defined by the intrinsic behavior patterns of the nodes (CISE students, library visitors, moviegoers) </li></ul></ul><ul><ul><li>The sender does not know the network identities (addresses) of the destinations </li></ul></ul><ul><li>Different from multi-cast: No join/leave, no group maintenance </li></ul>
  77. 77. Largest number of female users is in social sciences and is much higher than the male WLAN users in those buildings. Female users are surprisingly high (vs males) in the first 2 samples. WLAN activity was down Feb 07 due to lower enrollment in Spring and potential changes in the network.
  78. 78. Females in social, economic, admin and comm/journalism generally have longer session durations than males in those majors. In Engineering, music and chemistry the opposite is true. Session durations are decreasing indicating potential increase in mobility.
  79. 79. Apple consistently more popular in females than males Intel (PCs) are more popular in males than females Increase in use of Apple and Intel in general, and degradation in other brands
  80. 80. Mobility Profile-cast (intra-group) Goal S S Flooding S Flood-sim S Single long random walk S Multiple short random walks
  81. 81. Mobility Profile-cast (inter-group) Goal Flooding Flooding_sim S T.P. S T.P. S T.P. Gradient-ascend S T.P. Single long random walk S T.P. Multiple short random walks S T.P.
  82. 82. Performance Comparison Gradient ascend helps to overcome the difficult case – when the source is far from T.P. Few long RW is better when S is far from T.P. but many short RW is better when S is close to T.P.
  83. 83. Performance Comparison Gradient ascend helps to overcome the difficult case – when the source is far from T.P. Few long RW is better when S is close toT.P. but many short RW is better when S is close to T.P. Gradient ascend has some extra delay comparing with flooding
  84. 84. Mobility Independent Profile-cast Goal Flooding SmallWorld-based Single long random walk Multiple short random walks S S S S S