1 / 94

Statistical Analysis of Network Configurations in a Communication Network

This chapter explores statistical analysis of network configurations in a communication network, using ERGM examples and random graphs.

fgunderson
Download Presentation

Statistical Analysis of Network Configurations in a Communication Network

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. MIS 644Social Newtork Analysis2017/2018 Spring Chapter 7 C Statistical Analysis of Networks: ERGM Examples

  2. Outline

  3. Random Graphs • Some configurations may occur by chance even for random graphs • ties are fromed independently of one another • no dependence • Fig 4.1a random graph • n=38, m=146 • Communication network Fig 4.1b • Selected network statistics Table 4.1

  4. Distributions of Graphs • whether there are significantly more network structures than expected by chance • simuate a large sample of random directed nets • on 38 edges and 146 arcs • sample of graphs from a graph distribution • U|L: uniform condition on L: # of arcs • uniform - each graphs of 146 arcs has equal probability • 1,000 simple random graphs Fig 4.2 • distribution of reciprocated arcs

  5. ranges 0 – 15 • on average 7.8 reciprocated arcs – by chance • unusual to see more than 15 • RT=44 • 1- the network is not a random graph • 2- there is dependence among network ties • 3- there are reciprocity processes active in the network • prob of observing 44 ties in a RG is alomst 0 • if :%5 less than 

  6. null hypthesis: graph is random U|L model • # of RA is significantly greater than what is expected from U|L • there is significant reciprocation in the network • Similar conculusion about transitive traids • from 1,000 graphs mean TT:55 with std: 7.8 • observed 212 significantly greater

  7. Two limitations • 1- comparison distribtion to a random graph • another distribution • 146 ties and 44 RA • # TT observed data may not be extream • 2- consider each effect one at a time • ignore nesting of configurations • There are significantly more TTs • do not know • increased number of stars or • triangulation process

  8. saying that signficant number of RAs and TTs • rejecting the null: observed graph from U|L • not saying anything about the alternative • wheter reciprocation or transitivity are significant processes

  9. ERGM tries to do • model the effects of interest • e.g., reciprocated ties or transitive traids • in relation to an observed netowrk • to find a distribution of graphs • the observed data – at the centered • fit the model – by esitmating parameters

  10. for the communication net – ERGM • 38 nodes, • average of 146 ties, 44 RA, 212 TT • 312 in-2 start, 283 out-2 strat for nested effects • which configurations are importnat • which effects – independent explanatory value • whch can be expalined by other effects

  11. ERGM – compair the observed net with other possible ways it could be arranged • saample space – tie can be arranged • model estimation - assign probabiilities • observed net is central not an exteam • in terms of the effects being modeled

  12. ERGM examine competing theories - formation of network ties – within single analysis • test one network theoretical concept against others • homophily and/or reciporcity

  13. Applied Example • communication in the Corporation • in entertainment industry • n=38 executives • communication network - directed • tie – response of an actor to a survay item about another executive with whome it was important to communicate to coplete a work • Fig 4.1b, Table 4.1 • understand structure of informal communication ties

  14. information abour executives • # of projects they involved • level of seniority • office membership • Information about advice relations within the organization – binary directed network • ties pointing executives form whom adive is recieved • Download data and PNet models from • http://www.sna.unimelb.edu.au/

  15. network self-organization – communication ties • Fig 5.1 - # traids • triadic structure • some regions densly clustered • nodes 24 and 34 – hubs • Fig 5.2 – mutual ties undirected reciprocated ties • communication - partially explained by reciprocity • actor attributes – network structure • Fig 5.3: node size  # projects completed • suggest – homophily effect (experience)

  16. Fig 5.4: seniority • Fig 5.5: branch office • may be homophily from seniority • brach office – not clear • communication ties – affected from adsvice relationships • Fig 5.6. ties align (entrained) between communbication and advice networks • 24 ties • whether by chance or entainment effect significant

  17. many possible compating explanations for network structure • if examine one process at a time • may overestimate its worth • ERGM infer • whether – independent tendencise for a configuration to occur • or whether – presence of condiguration more parsimoniously explained by combined presene of other feects

  18. ERGM model and interpretation • parameter estimates –strength and direction of patterns Table 5.1 • a positive (negative) more (less) of the configuration then expected – given the ogther effects in the model • magnitutes not be comparable – scaling is different

  19. Arc: - negagtive arc effect • like intercept in regression • baseline propensity fo the occurene of ties • Reciprocity: employies likely to reciprocat3ed communbication

  20. Popularity and actgivity (in and out degree effects): tendencies for centrilization in and out degree distrubution • in-degree param. (-)insignificant – no distinctgively poular employee - net of other effects • out-degree param. (-) significant • absence of centralization in network activity • people tend to be uniform in the number of choice of comunication patterns

  21. simple 2-pats: not significant • neither more nor less 3-paths than expected given other params • no evidence – people send more ties also receive them • if (+) and significant – • actors most popular also most active • positive correlation between in and out degree distributions

  22. Multiple 2-paths: depth of local connectivity between pairs of nodes • 2-path – center of the two path • explaining - connectivity between pairs of nodes at the end of the path • not significant – local connectivity • neither stronger not weaker than expected

  23. Transitivity: transitive path closure of multiple 2-paths • (+) significant • tendency for hierarchical path closure • not use single triangles – problematic • not coherent models - not converge • triangles tend to occur together in denser regions of the network

  24. multiple 2-paths – nested in multiple triangles • combination of the two param estimate – • when mulitple 2-paths occur – tend to be closed in transive form

  25. Cyclic closure: cyclic closure of multiple 2-paths • (-) negative • nonhierarchical (generalized exchange) network closure • associated with mulitpe 2-patghs - together • when multipe 2-paths occur • tend to be closed in transitive form • and not to be closed in cyclic form

  26. onces these two closure processes takein into accout together • no other evident tendencies for multple 2-paths • present or absent • local connectivity explained by • tendency for transitive closure • and tendency against cyclic closure

  27. Sender effects: degree to chich actors with a specific attribute send more ties compared to others • (-) not quite significant for binary attribute seniority • managers not senior (seniority = 0) - tendency to send ties • number of project – continuous variable • no sender effect – experience in projects effect tendency to communicate • for varible “office” no sender and receiver effects

  28. Receiver effefct: degree to chich actors with a specific attribute has the propensity to receive more ties compared to others • no significant effect for variales seniority or number of projects

  29. Homophily: (+) significant • for seniority and number of projects • tend to communicatre wtih others with the same seniority and similar experience • the param is – for number of projects • absolute differences • no significant effect for office

  30. Covariae advice network: • exogenous to the communication network • fixed in the model • advice may have an effect on communication • but communication has no effect on advice • advice ties may effect communication ties • (+) significant: • advice and communication ties occur together

  31. Multiple explanations for network structure • there are significant effects for • purly structural • actor relation • covaraite network effects • each subtype effect – independent explanatory capacity

  32. Illustrations: Simulations, Estimations and Goodness of Fit • how to simulate ERGMs • as a heuristic method for GOF • to see wheter a model can fit to many emprical features of networks • Short examples of simulation and GOF • Corporation Communication Network

  33. Goals • simulation – effects of parameters • how they can work together • model interpretation • suggestions for model specification • with a GOF example • show how GOF may guide model selection • so effects may be added to improve fit • can be replicated with PNet

  34. Simulation • simulations about – • parameters and their interpretations • once burn-in • take sample of graphs • mean and SD of statiscits • n=30, burn-in 100,000 • 1,000,000 interatioons, 1,000 samples at intervals of 1,000 • last graphs is visualization • fixed density of 0.1 (43 ties) – undirected nets

  35. when on effects other than density 0.1 • random graph distribution U|L=43 • Fig 13.1 – illustative exaple • baseline compariosn

  36. Triangulation • effect of triangulation parameters T(on network closure Fig 13.2 • left-column: with different values of Markov TP • right-column: AT parameters with =2 • MTP of 0.5 or 1.0 – not differnt from RG • from 1.0 to 1.5 – sudden jump of triangulation to a cliquelike structure • for fixed # of edge – clique – maximum # of triangles

  37. phase transmition – • why MRG are not good • sample of graphs from RG • mean # triangles = 3.6, SD = 1.8 • a weak MTP increases samohow • when MTP=1.0 # triangles=12.1 SD=6 • t test significantly different from RG • when MTP=1.5 # triangles=102.5, SD=3.9

More Related