Putting Contacts into Context: Mobility Modeling beyond Inter-Contact Times

Putting Contacts into Context Mobility Modeling beyond Inter-Contact Times Theus Hossmann ETH Zürich, Switzerland Thrasyvoulos Spyropoulos EURECOM, France Franck Legendre ETH Zürich, Switzerland

Mobility Modeling Mobile Ad Hoc Network (incl. DTN) research largely based on simulation Unrealistic mobility models can lead to wrong conclusions about protocol performance! [Bai et al Infocom `03] Many (many, many) good existing models Simple vs. Complex Location based vs. Social network based [email_address] RPGM SIMPS SLAW TVCM CMM HCMM SWIM GHOST

Known Mobility Properties MASTERED MASTERED [email_address] Individual Properties Diurnal & weekly periodicity [Henderson et al MobiCom `04] Location preference [Tuduce et al Infocom `05] Power law trip length [Lee et al Infocom `09] Pairwise Properties Heavy tailed aggregate inter-contact times (exponential cut -off) [Chaintreau et al Infocom `06] [Karagiannis et al MobiCom `07] [Cai et al MobiCom `07] Individual pairs with various distributions [Leguay et al Autonomics `07)]

Unexplored Mobility Properties What about correlations of more than two nodes? Community structure Hubs Social (Contact) Graph Quantify structure Protocols Simbet [Daly et al MobiHoc `07] BubbleRap [Hui et al MobiHoc `08] [email_address] ??? Structural Properties Community Structure [Hui et al MobiHoc `08] Community Connections ?? Do existing models correctly reflect structural properties ??

Methodology [email_address] Mobility Model ?? Synthetic Trace Contact Graph Contact Trace Contact Graph Community Structure? Modularity Community Connections? Bridges Structural Properties?

Mobility Traces [email_address] Self-reported “check-ins” (like Foursquare) ~ 440’000 users (October 2010) ~ 16.7 Mio check-ins to ~ 1.6 Mio spots 473 “power users” who check-in at least 5 out of 7 days

Mobility Models [email_address] TVCM (location based) [Spyropoulos et al ToN `09] HCMM (social network based) [Boldrini et al Comp. Comm. `10] SLAW (location based) [Lee et al Infocom `09]

The Contact Graph Represent contacts as Weighted Graph G(V,W) How to assess the tie strength? Contact frequency (many contacts -> short delay) Contact duration (long contacts -> high bandwidth) [email_address] time w 12 w 13 w 35 w 67 d f w (i,j) w ij Frequency f Duration d w ij (scalar weight) PCA

The Contact Graph [email_address]

Community Structure Louvain Community Detection Algorithm [Blondel et al `08] Heuristic to maximize modularity [Newman PNAS `06] MASTERED ? ? ? ? ? ? ? [email_address] Structural Properties Community Structure Modularity, heterogenous community sizes, etc. Community Connections

Community Connections Distribution of community connection among links and nodes Implications for networking? (Routing, Energy, Resilience) Different mobility processes? [email_address] Def: Bridging node u of community C i : Strong weights to many nodes of community C j Def: Bridging link between u of C i and v of C j : Strong weight but neither u nor v is bridging node

Node Spread / Edge Spread Example 2/5 3/5 MASTERED ? ? ? ? ? ? ? FAILED! [email_address] 3/5 TRACES MODELS Low spread (Bridging Links) High spread (Bridging Nodes) Structural Properties Community Structure Community Connections Bridging nodes, bridging links ?? Why ??

Context of Contacts Difference in mobility processes (speculation) Mobility Models: Nodes visit other communities Reality/Traces: Nodes of different communities meet outside the context and location of their communities Infer context of contacts in traces GOW: From spot category DART: From AP locations [email_address] Context INTRA- Community INTER- Community Academic 4.9% 32% Administration 1.4% 1.2% Library 0.12% 11% Residential 90% 45% Social 0.5% 3.5% Athletic 2.7% 6.5%

Location of Contacts Def: Location profile : Smallest set of locations which contains 90% of intra-community contacts [email_address] Confirmed DART Outside Home Locations “ At home” Speculation: Small spread edges happen outside community context and location

Synchronization of Contacts Do nodes visit the same “social” location synchronously? Overlap (Jaccard Index) of time spent at social locations Null model: Independent visits (same number, same durations) Result: many synchronized visits Do only pairs visit social locations or larger cliques? Detecting cliques of synchronized nodes [email_address] DART Geometric Distribution

Social Overlay Hypergraph G(N, E) Arbitrary number of nodes per Hyperedge Represent group behavior Calibration from measurements # Nodes per edge # Edges per node Adapted configuration model Drive different mobility models TVCM:SO HCMM:SO [email_address]

Evaluation Edge spread Original propreties [email_address] ✔✔✔✔✔ ✔✔✔✔✔ ✔✔✔✔✔ MODELS TRACES Small Spread

Conclusion Traces: Bridging links („narrow“ community connections) Models: Bridging nodes („broad“ community connections) Trace analysis shows Inter-community contacts happening outside the locations of communities cause bridging links Synchronized “social” meetings of two or more nodes Social Overlay TVCM:SO, HCMM:SO Create bridging links Maintain original model properties [email_address]

Putting Contacts into Context: Mobility Modeling beyond Inter-Contact Times

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Putting Contacts into Context: Mobility Modeling beyond Inter-Contact Times (20)

Recently uploaded (20)

Putting Contacts into Context: Mobility Modeling beyond Inter-Contact Times