1. News and Notes, 1/12 Please give your completed handout from Tue to Jenn now Reminder: Mandatory out-of-class experiments 1/24 and 1/25 –likely time: either 5-7PM or 6-8 PM –both sessions are required –if you are registered and cannot make one or both sessions, send Prof Kearns email ASAP (including time constraints) please use “Experiments” as subject line

2. News and Notes, 1/17 You should be reading “The Tipping Point” Reminder: Mandatory out-of-class experiments 1/24 and 1/25 –likely time: either 5-7PM or 6-8 PM –both sessions (Tuesday and Wednesday) are required –if you are registered and cannot make one or both sessions, send Prof Kearns email ASAP (including time constraints) please use “Experiments” as subject line –confirmation of your attendance will be sent out later this week

3. News and Notes, 1/19 You should be reading “The Tipping Point” Two new assigned articles on the web pageweb page UPDATE ON EXPERIMENTS: –Tue 1/24, Wed 1/25 –Both sessions 6 – 8 PM; end time is approximate only –Location: 207 Moore –Need to arrive promptly and be present for entire session –Confirmation of your expected attendance(s) will be emailed to you –LAST CALL FOR CONFLICTS! –You must be present at Tuesday’s class to participate in either session

4. The Networked Nature of Society Networked Life CSE 112 Spring 2006 Prof. Michael Kearns

5. What is a Network? A collection of individual or atomic entities Referred to as nodes or vertices (the “dots” or “points”) Collection of links or edges between vertices (the “lines”) Links can represent any pairwise relationship Links can be directed or undirected Network: entire collection of nodes and links For us, a network is an abstract object (list of pairs) and is separate from its visual layout –that is, we will be interested in properties that are layout-invariant Extremely general, but not everything: –e.g. menage a trois –may lose information by pairwise representation We will be interested in properties of networks –often structural properties –often statistical properties of families of networks

6. Some Definitions Network size: total number of vertices (denoted N) Maximum number of edges: N(N-1)/2 ~ N^2/2 Distance between vertices u and v: –number of edges on the shortest path from u to v –can consider directed or undirected cases –infinite if there is no path from u to v Diameter of a network: –worst-case diameter: largest distance between a pair –average-case diameter: average distance If the distance between all pairs is finite, we say the network is connected; else it has multiple components Degree of vertex v: number of edges connected to v

7. Types of Networks

8. “Real World” Social Networks Example: acquaintanceship networks –vertices: people in the world –links: have met in person and know last names –hard to measure –let’s examine the results of our own last-names exercise Example: scientific collaboration –vertices: math and computer science researchers –links: between coauthors on a published paper –Erdos numbers : distance to Paul ErdosErdos numbers –Erdos was definitely a hub or connector; had 507 coauthors –MK’s Erdos number is 3, via Kearns  Mansour  Alon  Erdos –how do we navigate in such networks?

9. Online Social Networks Now outdated and discredited example: Friendster –vertices: subscribers to www.friendster.comwww.friendster.com –links: created via deliberate invitation –Here’s an interesting visualization by one uservisualization More recent and interesting: thefacebookthefacebook Older example: social interaction in LambdaMOO –LambdaMOO: chat environment with “emotes” or verbs –vertices: LambdaMOO usersLambdaMOO users –links: defined by chat and verb exchange –could also examine “friend” and “foe” sub-networks

10. MK’s Friendster NW, 1/19/04 Number of friends (direct links): 8 NW size (<= 4 hops): 29,901 13^4 ~ 29,000 But let’s look at the degree distributiondegree distribution So a random connectivity pattern is not a good fit What is??? Another interesting online social NW: –AOL IM BuddyzooBuddyzoo

11. Content Networks Example: document similarity –vertices: documents on the web –links: defined by document similarity (e.g. Google’s related search)related –here’s a very nice visualizationvisualization –not the web graph, but an overlay content network Of course, every good scandal needs a networkscandal –vertices: CEOs, spies, stock brokers, other shifty characters –links: co-occurrence in the same article Then there are conceptual networks –a thesaurus defines a networkthesaurus –so do the interactions in a mailing listmailing list

12. Business and Economic Networks Example: eBay bidding –vertices: eBay users –links: represent bidder-seller or buyer-seller –fraud detection: bidding rings Example: corporate boardscorporate boards –vertices: corporations –links: between companies that share a board member Example: corporate partnershipscorporate partnerships –vertices: corporations –links: represent formal joint ventures Example: goods exchange networksgoods exchange networks –vertices: buyers and sellers of commodities –links: represent “permissible” transactions

13. Physical Networks Example: the Internet –vertices: Internet routersInternet routers –links: physical connections –vertices: Autonomous Systems (e.g. ISPs)Autonomous Systems –links: represent peering agreements –latter example is both physical and business network Compare to more traditional data networkstraditional data networks Example: the U.S. power gridU.S. power grid –vertices: control stations on the power grid –links: high-voltage transmission lines –August 2003 blackout: classic example of interdependenceinterdependence

14. Biological Networks Example: the human brain –vertices: neuronal cells –links: axons connecting cells –links carry action potentials –computation: threshold behavior –N ~ 100 billion –typical degree ~ sqrt(N) –we’ll return to this in a moment…

15. Network Statics Emphasize purely structural properties –size, diameter, connectivity, degree distribution, etc. –may examine statistics across many networks –will also use the term topology to refer to structure Structure can reveal: –community –“important” vertices, centrality, etc. –robustness and vulnerabilities –can also impose constraints on dynamics Less emphasis on what actually occurs on network –web pages are linked… but people surf the web –buyers and sellers exchange goods and cash –friends are connected… but have specific interactions

16. Network Dynamics Emphasis on what happens on networks Examples: –mapping spread of disease in a social network –mapping spread of a fad –computation in the brain Statics and dynamics often closely linked –rate of disease spread (dynamic) depends critically on network connectivity (static) –distribution of wealth depends on network topology Gladwell emphasizes dynamics –but often dynamics of transmission –what about dynamics involving deliberation, rationality, etc.?

17. Network Formation Why does a particular structure emerge? Plausible processes for network formation? Generally interested in processes that are –decentralized –distributed –limited to local communication and interaction –“organic” and growing –consistent with (some) measurement The Internet versus traditional telephony

18. Structure, Dynamics, Formation: Two Brief Case Studies

19. Case Study 1: A “Contagion” Model of Economic Exchange Imagine an(y) undirected, connected network of individuals –no model of network formation Start each individual off with some amount of currency At each time step: –each vertex divides their current cash equally among their neighbors –(or chooses a random neighbor to give it all to) –each vertex thus also receives some cash from its neighbors –repeat A transmission model of economic exchange --- no “rationality” Q: How does network structure influence outcome? A: As time goes to infinity: –vertex i will have fraction deg(i)/D of the wealth; D = sum of deg(i) –degree distribution entirely determines outcome! –“connectors” are the wealthiest –not obvious: consider two degree = 2 vertices… How does this outcome change when we consider more “realistic” dynamics? –e.g. we each have goods available for trade/sale, preferred goods, etc. What other processes have similar dynamics? –looking ahead: models for web surfing behavior

20. Case Study 2: Grandmother Cells, Associative Memory, and Random Networks A little more on the human brain: –(neo)cortex most recently evolved(neo)cortex –memory and “higher” brain function –closest to a crude “random network” all connections equally likely Hebbian learning of correlations: –cells learn to fire when highly correlated neighboring cells fire –entirely decentralized allocation process Problem of associative memory: –consider the phrase “Pelican Brief” –or “Networked Life” –under “localist” assumption, requires that neurons representing “pelican” and “brief” have a common neighbor

21. A Back-of-the Envelope Analysis Let’s try assuming: –all connections equally likely –independent with probability p So: –at some point have learned “pelican” and “brief” in separate cells –need to have cells connected to both to learn conjunction –but not too many such cells! In this model, p ~ 1/sqrt(N) results in any pair of cells having just a few common neighbors! Broadly consistent with biology

22. Remarks Network formation: –random connectivity –is this how the brain grows? Network structure: –common neighbors for arbitrary cell pairs –implications for degree distribution Network dynamics: –distributed, correlation-based learning There is much that is broken with this story But it shows how a set of plausible assumptions can lead to nontrivial constraints

23. Recap We chose a particular, statistical model of network generation –each edge appears independently and with probability p –why? broadly consistent with long-distance cortex connectivity –a statistical model allows us to study variation within certain constraints We were interested in the NW having a certain global property –any pair of vertices should have a small number of common neighbors –corresponds to controlled growth of learned conjunctions, in a model assuming distributed, correlated learning We asked whether our NW model and this property were consistent –yes, assuming that p ~ 1/sqrt(N) –this implies each neuron (vertex) will have about p*N ~ sqrt(N) neighbors –and this is roughly what one finds biologically (Note: this statement is not easy to prove)