Network Intelligence and Networked Intelligence Deyi Li ziqinlipublic

Network Intelligence and Networked Intelligence 网络智能和网络化智能 Deyi Li ( 李 德 毅 ) ziqinli@public 2. bta. net. cn Aug. 1, 2006 1

Challenge to AI for Knowledge Representation 2

Study on Knowledge Representation l one-dimensional representation: predicate calculus, natural language understanding, etc. l two-dimensional representation: pattern recognition, neural network learning, etc. l attention on evolutional networks with uncertainty was less paid unfortunately. 3

Networks are present everywhere. All we need is an eye for them. 4

We are witnessing a revolution in the making as scientists from all different disciplines discover that complexity has a strict architecture. We have come to grasp the important knowledge of networks. 5

l Networks interact with one another and are recursive. l Getting such a diverse group to agree on a common core of knowledge representation about networks is a significant challenge to both Cognitive Science and Artificial Intelligence. 6

Networks Evolution and Growth drive the fundamental issue that forms our view of network representation and network intelligence. 7

ER pure random graph(1960) Paul. Erdos WS small world model (1998) Duncan Watts BA scale-free model(1999) Albert Barabasi 8 Alfred Renyi Steven Strogatz Reka Albert

l “Small worlds” and “power law distributions” are generic properties of networks in general. l There is a new knowledge representation out there that is the network representation. 9

It’s the fact that all of these real world networks can be explained and understood using the same concepts, and the same mathematics, that makes network representation so important in AI research in the information age. 10

Mining Typical Topologies from Real Complex Networks 11

typical topologies with randomness 12

An evolutional and growth network may be by and large characterized by an ideal typical model 13

Mining Typical Topology from Real World Networks at Multi-scale Expectation of topologies at different scales: l Small world network l Scale free network l Hub Network l Star Network 14

Extend more properties of networks l the mass of a node l physical distance between two nodes l the age of a node l betweenness of a link l betweenness of a node 15

With the extended properties of networks, we may map relational data into networked representation and propose a new direction that is networked data mining. 16

A detailed networked data 17

Mining typical topology with a middle granularity 18

Mining typical topology with a large granularity 19

Discover critical links and important communities from a real network 20

Many networks are inhomogeneous, consisting lot of an undifferentiated mass of nodes, but of distinct groups. 21

Mining communities Classification: The typical problem in networked data mining is that of dividing all the nodes of a network into some number of groups, while minimizing the number of links that run between nodes in different groups. l Clustering: Given a network structure, try to divide into communities in such a way that every node belongs only to one of the communities. l 22

v. Community model can capture the hierarchical feature of a Network. 23

A link removal method based on link betweenness Input：Initial network topology，the number of community Output：network communities Step 1. Calculate the betweenness for all links in the network. Step 2. Remove the link with the highest betweenness. Step 3. Re-calculate betweennesses for all links affected by the removal. Step 4. Repeat from step 2 until generating specified numbers of communities. 24

Mining Communities 25

Mining clusters in a complex network using data field method and finding virtual kernels l Given a traffic network, find virtual traffic centers 26

Node mass may represent its degree from data field point of view 27

Node mass may also represent its betweenness from data field point of view 28

Emergence Computation 29

A subtle urge to synchronize is pervasive in nature indeed l l l l l synchronized clapping fireflies flashing menstrual cycles of women adaptive path minimization by ants wasp and termite nest building army ant raiding fish schooling and bird flocking pattern formation in animal coats coordinated cooperation in slime molds 30

Nature Vol. 403, 24 Feb. 2000 31

The emergence of synchronized clapping is a delightful expression of self-organization on a human scale 32

emergence mechanism l For everybody in the audience there are 3 measurements: 1. time difference at the beginning of the applause (TDB) 2. interval time of a clap to the next one (IT, represented by △t) 3. the clapping strength (CS) 33

If there is no any interaction among audience, the distributions of everybody’s TDB, IT and CS, even the number of clap times all follow a kind of poisson curve like. l If there are interactions among audience, the influence to each other depends on the distance (say rij) between them. l 34

Assume: l all the clap strengths are the same. l “following the many” is fundamental mechanism and pervasive applicable. l Therefore the relationship of persons in the audience, that is the structure of the network, encoding how people influence each other is set up in formula 1 35

0 ０ １ ２ ３ ４ ５ ６ ７ 1 2 3 4 5 6 7 somebody’s just-happened clap moment is ti and the next IT (say △ti’) is based on his current IT (say △ti) and influenced by the distanced person who’s just -happened clap is measured by △tj and clapped moment tj l σ represents distance influence factor l c 1 and c 2 are coupling factors 36 l

The formula shows the fact that there is no an invisible control to all the audience, every body affects others and affected by others equally. 37

Single Clapping Single continuous clapping 38

general applause in a theatre 39

all the palms in theatre came together after a long time applause 40

The ‘up here down there’ applause in the square 41

An experimental platform of emergence computation Visualization of courtesy applause and synchronized applause 42

It is difficult to distinguish the virtual general ap l It is also difficult to distinguish the virtual synch l 43

Network is the key to representing the complex world around us. Small changes in the topology, affecting only a few of the nodes, can open up hidden doors, allowing new possibilities to emerge. 44

Sum up l Challenge to AI for Knowledge Representation l Mining Typical Topologies from Real Complex Networks l Discover Critical Links and Important Communities from a Real Network l Emergence Computation 45

To be studied in the future: 1. better measurements of network structure in network representation 2. better understanding of the relationship between the architecture of a network and its function 3. better modeling of very large networks 4. mining common concepts of a network across different scales 5. robustness and security of networks 6. networked data mining 7. virtual reality of emergence. 46

Thanks 李 德 毅 ziqinli@public 2. bta. net. cn 47