social search in small-world experiments

19
Social Search in “Small- World” Experiments Sharad Goel, Roby Muhamad and Duncan Watts 2009 Akhil Jain 2011CS50274

Upload: akhil-jain

Post on 19-Nov-2015

212 views

Category:

Documents


0 download

DESCRIPTION

A presentation on the paper by the title Social Search in Small World Experiments by Sharad Goel et alGoel, Sharad, Roby Muhamad, and Duncan Watts. "Social search in small-world experiments." Proceedings of the 18th international conference on World wide web. ACM, 2009.

TRANSCRIPT

  • Social Search in Small-

    World ExperimentsSharad Goel, Roby Muhamad and Duncan Watts

    2009

    Akhil Jain

    2011CS50274

  • Outline

    Algorithmic Small World Hypothesis

    Work done until now

    A new model of attrition

    Heterogeneity

    Unbiasedness

    Results

  • Algorithmic Small World Hypothesis

    Topological For a randomly chosen pair of individuals,

    there exists with high probability a short chain of

    intermediaries that connects them.

    Algorithmic Ordinary individuals can effectively

    navigate these short chains themselves with every

    individual having only local knowledge of the social

    network in question.

  • Topological Studies

    A majority of individuals in a given population are

    connected in a single giant component.

    The typical shortest path length connecting pairs of

    nodes within the giant component is of the order of the

    logarithm of the system size.

    By Daniel' (User:Dannie-walker) (Own work) [GFDL (http://www.gnu.org/copyleft/fdl.html) or CC-BY-SA-3.0-

    2.5-2.0-1.0 (http://creativecommons.org/licenses/by-sa/3.0)], via Wikimedia Commons

  • The problems with algorithmic studies

    Reliance on simulations hence not generalizable.

    Homogeneity assumption!

    Attrition Low rates of chain completion with length

    Lack of motivation

    Failure to receive the message

  • Stochastic modelling of attrition the

    story so far

    Parameter is the probability of termination at each step.

    Two estimators for true length in absence of attrition:

  • Problems with such attrition modelling

    Existence of multiple small worlds bowl of lumpy

    oatmeal!

    Search ability is not uniform but a function of social

    capital Homogeneity!

    Most source-target pairs live in separate populations that

    can only be reached via long or undiscoverable paths

    Unbiasedness.

  • A new model of attritionUnbiased and Heterogeneous estimators

  • Experiments

    The use of email to allow larger access at reduced cost

    1. Dec 2001 Aug 2003; 98865 people from 168 countries

    initiated 106295 chains directed at 18 targets in 13

    countries 491 (0.5%) reached the target.

    2. Aug 2003 Dec 2007; 85621 people from 163 countries

    initiated 56033 chains directed at 21 targets in 13

    countries 61 (0.1%) reached the target.

  • Dealing with heterogeneity

    Next step continuance probability

    Logistic multi-level regression

  • Dealing with heterogeneity

    9 categories 66 parameters

    Each category is modelled as a normal distribution with 0 mean

    Obtain regression coefficients for each attribute value in each category

  • A few interesting results

  • Dealing with Biasedness

    Space of paths that can be observed

    P() probability of choosing a path in

    Q() probability that path is observed

    A missing value is observed with probability ()

    Let denote the ith chain observed in the experiment

    Without attrition, expected true length is given by = () where f() is the length of path

  • Dealing with Biasedness

    Let m is the number of paths observed with non missing

    values, then an unbiased estimator for the man path

    length is given by

    This can also be used to find an unbiased estimate for pi the probability that path length is i

  • Estimating Chain Lengths

    Homogenous Attrition

    Uniform for all

    Different for the first link in the chain

  • Estimating Chain Lengths

    Heterogeneous attrition

  • Estimating Chain Lengths

    Randomised attrition

  • Results

  • Thank youQuestions?