Introduction to Network Mathematics

By Bruce Hoppe

If your browser does not support mathematical symbols, you will not see this document correctly. If "⊆" looks like a box in quotes instead of a lazy U, or this math-symbol-test-page looks like a bunch of boxes, then you need to find a browser that supports mathematical symbols. In our experience, Firefox works and Internet Explorer does not.

Introduction to Network Mathematics provides college students with basic graph theory to better understand the Internet. Many passages are edited from Wikipedia, a few are from PlanetMath, and others are original writing by Bruce Hoppe, who teaches CS-103: "Introduction to Internet Technologies and Web Programming" at Boston University. This is a work in progress, first created in the spring semester of 2007, and now being used in the spring semester 2009.

Copyright © 2007, 2008, 2009 by Bruce Hoppe. Permission is granted to copy, distribute and/or modify these documents under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation; with no Invariant Sections, with no Front-Cover Texts, and with no Back-Cover Texts. A copy of the license is included in the section entitled "GNU Free Documentation License". If you believe a portion of this site infringes on your or anyone's copyright, please send us a note.

Contents


Sets

A set is a collection of objects. These objects are called the elements or members of the set. Objects can be anything: numbers, people, other sets, etc. For instance, 4 is a member of the set of all even integers. Clearly, the set of even integers is infinitely large; there is no requirement that a set be finite.

If x is a member of A, then we write x \in A. We may also say that x "belongs to" A, or that x "is in" A.

If x is not a member of A, then we write x \notin A.

Two sets A and B are defined to be equal when they have precisely the same elements, that is, if every element of A is an element of B and every element of B is an element of A. Thus a set is completely determined by its elements; the description is immaterial. For example, the set with elements 2, 3, and 5 is equal to the set of all prime numbers less than 6.

If the sets A and B are equal, then we write A = B.

We also allow for an empty set, a set without any members at all. We write the empty set as { }.

Explicit and Implicit Set Notation

Explicit Set Notation: The simplest way to describe a set is to list its elements between curly braces, known as defining a set explicitly. Thus {1,2} denotes the set whose only elements are 1 and 2. Note the following points:

Warning: This notation can be informally abused by saying something like {dogs} to indicate the set of all dogs, but this example would be read by mathematicians as "the set containing the single element dogs".

The simplest possible example of explicit set notation is { }, which denotes the empty set.

Implicit Set Notation: We can also use the notation {x : P(x)} to denote the set containing all objects for which the condition P holds, known as defining a set implicitly. For example, {x : x is a real number} denotes the set of real numbers, {x : x has blonde hair} denotes the set of everything with blonde hair, and {x : x is a dog} denotes the set of all dogs.

Cardinality

Given a finite set A, the cardinality of A is simply the number of elements in A. We write the cardinality of A as |A|, which we may read as "cardinality of A" or "size of A." Note that this vertical-bar notation is the same as used to denote absolute value, but the meaning of cardinality is quite different from absolute value. In particular, absolute value operates on numbers (e.g., |-3| = 3) while cardinality operates on sets (e.g., |{-3}| = 1). Other examples of cardinality are:

We can also consider the cardinality of an infinite set A, but that discussion is beyond the scope of this introduction.

Subsets

Given two sets A and B we say that A is a subset of B if every element of A is also an element of B. Notice that in particular, B is a subset of itself; a subset of B that isn't equal to B is called a proper subset.

If A is a subset of B, then one can also say that B is a superset of A, that A is contained in B, or that B contains A. In symbols, AB means that A is a subset of B, and BA means that B is a superset of A. For clarity, we use "⊆" and "⊇" for subsets (which allow set equality) and we use "⊂" and "⊃" are reserved for proper subsets (which exclude equality).

For example, let A be the set of American citizens, B be the set of registered students at Boston University, and C be the set of registered students in CS-103. Then C is a proper subset of B and B contains C, which we can write as C ⊂ B andBC. If we allow the possibility that every BU student might be registered for CS-103, then we would write C ⊆ B or BC to allow for this unlikely situation. Note that not all sets are comparable in this way. For example, it is not the case either that A is a subset of B nor that B is a subset of A.

Unions, intersections, and relative complements

Given two sets A and B, the union of A and B is the set consisting of all objects which are elements of A or of B or of both. It is written as AB.

The intersection of A and B is the set of all objects which are both in A and in B. It is written as AB.

Finally, the relative complement of B relative to A, also known as the set theoretic difference of A and B, is the set of all objects that belong to A but not to B. It is written as A \ B. Symbolically, these are respectively

A ∪ B := {x : (x \in A) or (x \in B)}
AB := {x : (x \in A) and (x \in B)}
A \ B := {x : (x \in A) and not (x \in B) }

Notice that B doesn't have to be a subset of A for A \ B to make sense.Venn Diagram

To illustrate these ideas, let A be the set of left-handed people, and let B be the set of people with blond hair. Then AB is the set of all left-handed blond-haired people, while AB is the set of all people who are left-handed or blond-haired or both. A \ B, on the other hand, is the set of all people that are left-handed but not blond-haired, while B \ A is the set of all people that have blond hair but aren't left-handed.

Venn diagrams are illustrations that show these relationships pictorally. The above example can be drawn as the Venn diagram at right:

Disjoint sets: Let E be the set of all human beings, and let F be the set of all living things over 1000 years old. What is EF in this case? No human being is over 1000 years old, so EF must be the empty set { }. When the intersection of any two sets is empty, we say those two sets are disjoint.

Similarity of sets: Jaccard index

We measure the similarity of two non-empty sets A and B with the Jaccard Index: J(A,B) = |A ∩ B| / |A ∪ B|.

If A and B are disjoint then J(A,B)=0. If A=B then J(A,B)=1. (Assuming A and B are non-empty.)

Suppose A = {1,2,3,4,5,6}. Below we use J(A,B) to measure the similarity of set A with 5 different example sets B:

Set B Venn diagram of sets A and B Jaccard index J(A,B)
B = {7,8,9,10,11,12}
jaccard-1

J(A,B) =

|{1,2,3,4,5,6} ∩ {7,8,9,10,11,12}|

|{1,2,3,4,5,6} ∪ {7,8,9,10,11,12}|

= 0/12;

Not at all similar

B = {4,5,6,7,8,9}
j-2

J(A,B) =

|{1,2,3,4,5,6} ∩ {4,5,6,7,8,9}|


|{1,2,3,4,5,6} ∪ {4,5,6,7,8,9}|

= 3/9;

Somewhat similar

B = {2,3,4,5,6,7}
j-3

J(A,B) = 5/7

Quite similar

B = {4}
j-4

J(A,B) = 1/6

B = {4,5,6,7, ... 25} j-5 J(A,B) = 3/25

Ordered Pairs and Ordered Lists

Intuitively, an ordered pair is simply a collection of two objects such that one can be distinguished as the first element and the other as the second element, and having the fundamental property that two ordered pairs are equal if and only if their first elements are equal and their second elements are equal.

We write an ordered pair using parentheses--not curly braces, which are used for sets. For example, the ordered pair (a,b) consists of first element a and second element b.

Two ordered pairs (a,b) and (c,d) are equal if and only if a = c and b = d.

In contrast, a set of two elements is an unordered pair which we write using curly braces and where there is no distinction between the first element and the second element. It follows that
We can also write ordered lists having any finite number of elements with an intuitive extension of the above notation. For example:

Graphs

Mathematicians have developed Six-Node Graph graph theory in order to study all kinds of networks. A graph is a set of objects connected by lines.

The objects in a graph are usually called nodes or vertices. The lines connecting the objects are usually called links or edges.

More formally, we define a graph G as an ordered pair G = (V,E) where

Example: The picture above represents the following graph:

Loops: An edge with identical endpoints is a loop: loop Unless explicitly stated otherwise, we disallow loops.

Trivial and empty graphs: The graph with only one vertex and no edges is called the trivial graph. The graph with no vertices and no edges is called the empty graph.

Undirected and Directed Graphs

Undirected graph: The edges of a graph are assumed to be unordered pairs of vertices. Sometimes we say undirected graph to emphasize this point. In an undirected graph, we write edges using curly braces to denote unordered pairs. For example, an undirected edge {2,3} from vertex 2 to vertex 3 is the same thing as an undirected edge {3,2} from vertex 3 to vertex 2.

Directed graph: In a directed graph, the two directions are counted as being distinct directed edges. In an directed graph, we write edges using parentheses to denote ordered pairs. For example, edge (2,3) is directed from 2 to 3 , which is different than the directed edge (3,2) from 3 to 2. Directed graphs are drawn with arrowheads on the links, as shown below:
Directed and Undirected

Adjacent and Incident

Two vertices are called adjacent if they share a common edge, in which case the common edge is said to join the two vertices. An edge and a vertex on that edge are called incident.

See the 6-node graph below right for examples of adjacent and incident:

Neighborhood and Degree

The neighborhood of a vertex v in a graph G is the set of vertices adjacent to v. The neighborhood is denoted N(v). The neighborhood does not include v itself. For example, in the graph below N(5) = {4,2,1} and N(6) = {4}.

The degree of a vertex is the total number of vertices adjacent to the vertex. The degree of a vertex v is denoted deg(v). We can equivalently define the degree of a vertex as the cardinality of its neighborhood and say that for any vertex v, deg(v) = |N(v)|.Six Nodes

The undirected graph at right has the following degrees:

Vertex Degree
1 2
2 3
3 2
4 3
5 3
6 1

In a directed graph, we define degree exactly the same as above (and note that "adjacent" does not imply any direction or lack of direction). It is also important to define indegree and outdegree. Recall that any directed edge has two distinct ends: a head (the end with an arrowhead) and a tail. Each end is counted separately. The sum of head endpoints count toward the indegree of a vertex and the sum of tail endpoints count toward the outdegree of a vertex. The directed graph at right has the following degrees, indegrees, and outdegrees:Directed Graph

Vertex Degree Indegree Outdegree
1 2 0 2
2 2 2 0
3 3 2 2
4 1 1 1

Pay particular attention to nodes 3 and 4 in the above table. If they seem confusing, try re-drawing the above graph using the "easier way to draw" illustrated previously.

Density and Average Degree

The density of a graph G = (V,E) measures how many edges are in set E compared to the maximum possible number of edges between vertices in set V. Density is calculated as follows:

The average degree of a graph G is another measure of how many edges are in set E compared to number of vertices in set V. Because each edge is incident to two vertices and counts in the degree of both vertices, the average degree of a graph is 2*|E|/|V|

Paths and Walks

A path P = (v1,v2,v3,...,vk) is an ordered list of vertices such that

  1. From each of its vertices there is an edge to the next vertex in the list
  2. No vertex occurs in the list more than once.

The first vertex of a path is called the origin and the last vertex is called the destination. Both origin and destination are called endpoints of the path.

Example: In the graph below, there are many paths from node 6 to node 1. One such path is P1 = (6, 4, 3, 2, 5, 1); another path from 6 to 1 is P2 = (6, 4, 5, 1).

Any node all by itself makes a trivial path. For example, (3) is the path that begins and ends at node 3. Another way to see (3) is as an ordered list with exactly one element: the node 3.

Six Nodes

We say walk to denote an ordered list of vertices that follows the first requirement of a path but not the second. In other words, a walk may visit the same vertex more than once, but a path must never do so. This implies that every path is a walk, but some walks are not paths (i.e., any walk that visits a vertex more than once). A cycle is a walk such that the origin and destination are the same; a simple cycle is a cycle that does not repeat any vertices other than the origin and destination.

Examples: In the graph above, there are many walks from node 6 to node 1. Our previous examples P1 = (6, 4, 3, 2, 5, 1) and P2 = (6, 4, 5, 1) are also walks from node 6 to node 1. In addition, W1 = (6,4,3,4,3,5,4,3,2,5,1) and W2 = (6,4,5,1,5,1) are examples of walks from node 6 to to node 1 that are not paths.

The definitions of paths and walks can easily be extended to apply to directed graphs, in which case we use the terms directed path, directed walk, and directed cycle.

Length, Distance, Diameter

The length of a path or walk is the number of edges that it uses, counting multiple edges multiple times. In the graph below, (5, 2, 1) is a path of length 2 and (1, 2, 5, 1, 2, 3) is a walk of length 5. Also, the trivial path (3) has length 0.

Six Nodes

The distance between two vertices x and y is written d(x,y) and defined as the length of the shortest path from x to y. The distance between any node and itself is 0. If there is no path from x to y (i.e., if x and y are in different connected components), then d(x,y) is infinity.

The average distance of a graph G is the average distance between all distinct pairs of vertices in G.

The diameter of a graph G is the maximum distance between any pair of vertices in G. If G is not connected, then the diameter of G is infinity.

Example 1: consider the graph G=(V,E) drawn below.

5-node ring

With a graph like the above example, diameter is similar to our geometric notion -- i.e., how far to get from one side of the circle to the other. In a graph context, we want to find the pair of nodes x and y such that the distance between x and y is not exceeded by the distance separating any other pair of nodes in the graph. One way to think about this is by filling in the table below, which shows the distance between each pair of nodes in G.

Nodes 1 2 3 4 5
1 0 1 2 2 1
2   0 1 2 2
3     0 1 2
4       0 1
5         0

Because G is an undirected graph, we know that the distance from x to y is the same as the distance from y to x, and so we can ignore the shaded cells in the above table for the purposes of computing diameter and average distance.

After computing the distance between every possible pair of nodes, we see that the maximum distance in the above graph is 2. Therefore the diameter of the graph is 2.

The average distance of the graph is the average of all the non-zero cells in the above table. There are 10 such cells and the sum of the numbers in those cells is 15, making the average distance of G = 15/10 = 1.5.

Example 2: consider the graph G=(V,E) drawn below.

diamter

The table below shows the distance between each pair of nodes in G.

Nodes 1 2 3 4 5 6 7 8 9 10
1 0 2 3 3 3 2 4 1 2 1
2   0 3 3 3 2 4 3 2 1
3     0 2 4 1 3 4 3 2
4       0 2 1 1 4 3 2
5         0 3 1 4 1 2
6           0 2 3 2 1
7             0 5 2 3
8               0 3 2
9                 0 1
10                   0

The average distance of G is the average of all the non-zero cells in the above table. There are 45 such cells and the sum of the numbers in those cells is 164, making the average distance of G = 164/45 = 3.6444.

The diameter of G is the maximum value in any single cell in the above table. One cell stores value 5--because the distance from node 7 to node 8 is 5. There is no value larger than 5 in any other cell--no pair of nodes in G has distance more than 5. Therefore the diameter of G = 5.

Induced Subgraphs

It is often useful to consider only part of a graph. Induced subgraphs are one particularly convenient way to define a specific sub-part of a larger graph.

Induced Subgraph
Graph G = (V,E)
Induced Subgraph
Subgraph induced by subset of nodes {1,2,3,4}

Given a graph G=(V,E), an induced subgraph is defined by a subset of vertices S ⊆V. Then the nodes of the subgraph induced by S are simply the nodes in S, and the edges of the subgraph induced by S are all edges with both endpoints in S.

The upper graph pictured at right shows G=(V,E) with V={x: x is an integer from 1 to 8} and edges as drawn.

The lower picture shows the subgraph induced by the subset of nodes S={1, 2, 3, 4}. Notice that this induced subgraph consists of two connected components, even though the graph as a whole is a single connected component.

Another Example: In lecture we have discussed the Facebook friends network. We can write this as a graph using implicit notation as follows: G = (V,E), where V = {x: x is a person on Facebook} and E = {{x,y}: x \in V and y \in V and (x and y are friends on Facebook)}. The Facebook friends network is enormous, and so graph G is impossible to draw by hand; however, we can choose a small subset S of nodes in V and draw the graph induced by that subset S. This is exactly what the Touchgraph Facebook browser does: It lets you choose set S to be some number of your "Top Friends" on Facebook, possibly with the addition of yourself. Given this set S, Touchgraph displays a subgraph of the entire Facebook friends network: specifically, the subgraph induced by your chosen number of "Top Friends" (possibly with the addition of yourself).

Below is the subgraph of the Facebook friends network induced by the subset S={Patti, Jenny, Bill, Deborah}. This set S is exactly Bruce Hoppe's "Top 4 Friends" (which does not include Bruce Hoppe). The edges are hard to see but they are {{Patti, Jenny}, {Jenny, Bill}, {Bill, Patti}, {Bill, Deborah}}. There are no edges joining Deborah and Patti, or Deborah and Jenny, because these people are not Facebook friends.

Subgraph by Touchgraph

Connected vertices, graphs, and components

Given an undirected graph, two vertices are connected if there is at least one path that joins them. For example, in the graph below, nodes 2 and 4 are connected. Nodes 1 and 6 are not connected.

Connected vs. Adjacent: Any two nodes that are adjacent are also connected (e.g., nodes 2 and 5 below). Two nodes that are connected may not be adjacent (e.g., nodes 2 and 4 below).

Furthermore, any node is connected to itself and is not adjacent to itself. For example, node 2 is connected to node 2 by the path (2)--i.e., the zero-length path that consists of nothing but the node 2. Node 2 is not adjacent to node 2 because there is no edge {2,2} (i.e., no edge {2}); and our prohibition against loops in graphs implies that no node can ever be adjacent to itself.

6-node 2-component

A connected graph is one where each vertex can reach every other vertex by one or more paths. A connected graph is said to consist of a single connected component. If a graph is not connected, then it consists of two or more connected components. The graph above consists of two connected components. Below is another example.

Connected Components

Given an undirected graph G=(V,E), we define a connected component to be a subgraph induced by node set SV (i.e., GS = (S,ES)) with the following two properties:

  1. GS is connected
  2. There is no edge in E that joins a node in S to a node not in S

Six Degrees describes connected components metaphorically using buttons as nodes and threads as edges. In this metaphor, a connected component is what you get when you lift one button off the floor--up comes not only that one button but also all the threads tied to that button, all the other buttons tied to those threads, etc. The above 2 mathematical properties of a connected component translate into this metaphor as follows:

  1. GS is connected: The only buttons that rise off the floor do so because of the one button you are lifting and the threads that ultimately connect that button to other buttons (i.e., paths)
  2. There is no edge in E that joins a node in S to a node not in S: You have lifted the one button in your hand high enough so that no more buttons will come off the floor no matter how much higher you lift your hand.6-node 2-component

For example, the graph G=(V,E) drawn at right is not a connected component because it violates property #1: it is not connected.

The subgraph induced by S={3,4,5,6} (drawn with dark nodes and edges below as GS = ({3,4,5,6},{{3,4},{3,5},{3,6},{4,5},{4,6},{5,6}}) is not a connected component because it violates property #2. There are edges in E that join nodes in S to node 2, which is a node not in S.subgraph

 

 

Components, Cliques, and Clusters

Clusters are important features of the Web. To the extent that Web builders tend to link to sites similar to their own sites, clusters of interlinked Web pages represent collections of content about meaningful subtopics.

Clusters are hard to define mathematically. We will define clusters somewhat informally, by using precise definitions of connected components and cliques as stepping stones to a less precise definition of clusters.

In the preceding section we defined a connected component in an undirected graph G=(V,E) to be a subgraph induced by node set SV (i.e., GS = (S,ES)) with the following two properties:

  1. GS is connected
  2. There is no edge in E that joins a node in S to a node not in S

To that definition we add the notion of clique. In an undirected graph G=(V,E) a clique is a subgraph induced by node set SV (i.e., GS = (S,ES)) such that the density of GS is 1 (i.e., every pair of nodes in S is adjacent). Notice that the definition of clique does not depend on how many edges in E join nodes in S to nodes not in S.

Example: In the graph G drawn above, the subgraph induced by S={3,4,5,6} is a clique. Also, the subgraph induced by {2,6} is a (somewhat less interesting) clique. The subgraph induced by S={3,4,5,6} is the largest clique in G. Usually the largest clique in a graph is the most interesting clique in that graph.

The notion of cluster informally combines elements of our definitions of connected components and cliques. In an undirected graph G=(V,E) a cluster is a subgraph induced by node set SV (i.e., GS = (S,ES)) with the following two properties:

  1. The density of GS is "relatively high"; (a relaxed adaptation of clique-ness)
  2. There are "relatively few" edges in E that join a node in S to a node not in S; (a relaxed adaptation of connected component-ness)

Example: In the graph drawn below, the following subsets of nodes induce subgraphs that can fairly be called clusters (given our informal definition):clusters

Not all connected components are clusters. In the graph drawn at right, even the subgraph induced by S={9,10,11,12,13} is arguably not a cluster, because the density of that induced subgraph is relatively low.

Not all cliques are clusters--even relatively large cliques. The subgraph induced by S={1,2,3,4} is not a cluster because every single one of those nodes is adjacent to node 5; there are too many edges joining nodes in S to nodes not in S. Even the largest clique in the above graph--the subgraph induced by S={1,2,3,4,5}--is arguably still not a cluster because node 7 is adjacent to so many nodes in S.

Structural equivalence

Finding clusters is one way to discover meaningful collections of Web pages. Another way is based on the idea that Web pages x and y are similar to the extent that they have similar network neighborhoods -- even if x and y are not adjacent to each other. This is the idea behind structural equivalence.

Consider G=(V,E) drawn below. (Edge set E includes both black and grey edges; we will get to that distinction momentarily.) Suppose G is a subgraph of the Facebook friends network and you are logged in as node 26. Which node will Facebook suggest first as someone you may know? Facebook will not bother to suggest someone you already know; instead, it will look for someone who is not yet a Facebook friend of yours but who has many of the same Facebook friends that you have. For example, node 29 stands out as "someone you may know" if you are node 26. By comparison, node 27 appears to be someone that node 26 probably does not know, judging by the different sets of friends that nodes 26 and 27 keep. In the middle ground, nodes 28, 30, and 31 each have some neighbors in common with node 26 and other neighbors that are distinct from those of node 26. Structural equivalence is a metric that can be used to rank these nodes according to how similar their neighborhoods are to that of node 26.

se

Given two nodes x and y in a graph, the Jaccard index provides a simple and useful way to compare the similarity of their neighborhoods. Recall that the Jaccard index measures the similarity of two sets and is defined J(A,B) = |A ∩ B| / |A ∪ B|.

We define the structural equivalence of two nodes x and y as the similarity of their neighborhoods, as measured by the Jaccard index:

SE(x,y) = J(N(x),N(y)) = |N(x) ∩ N(y)| / |N(x) ∪ N(y)|.

This equals 1 when x and y have identical neighborhoods and 0 when the neighborhoods of x and y are disjoint.

Consider graph G drawn above and suppose x=26; then N(6) = {1,2,3,4,5,6}. Below we compute the structural equivalence of x with 5 example nodes y: 27, 28, 29, 30, and 31.

Node y N(y) = Neighborhood of y SE(x,y) = J(N(x),N(y))
27
{7,8,9,10,11,12}

J(N(x),N(y)) =

|{1,2,3,4,5,6} ∩ {7,8,9,10,11,12}|

|{1,2,3,4,5,6} ∪ {7,8,9,10,11,12}|

= 0/12; Not at all similar

28
{4,5,6,7,8,9}

J(N(x),N(y)) =

|{1,2,3,4,5,6} ∩ {4,5,6,7,8,9}|


|{1,2,3,4,5,6} ∪ {4,5,6,7,8,9}|

= 3/9; Somewhat similar

29
{2,3,4,5,6,7}

J(N(x),N(y)) = 5/7; Quite similar

30
{4}

J(N(x),N(y)) = 1/6

31
{4,5,6,7, ... , 25}
J(N(x),N(y)) = 3/25

To emphasize the underlying congruence between measuring the similarity of nodes using structural equivalence and the similarity of sets using the Jaccard index, we have arranged the above examples of structural equivalence to correspond exactly to the table of examples illustrating the Jaccard index.

Structural equivalence, social bookmarking, and collaborative filtering: By far the most important Web application of structural equivalence is its use by sites such as Delicious and Amazon. When you look up a specific tag on Delicious and it recommends related tags, that recommendation is based on structural equivalence. When you shop on Amazon and it recommends other items you might also like to buy, that recommendation is based on structural equivalence. Websites that work in this way are often said to use collaborative filtering.

Our above definition of structural equivalence works the same in both the Facebook "someone you may know" examples and the Delicious/Amazon "other things you may be interested in" collaborative filtering scenario. However, the underlying graph used in the Delicious/Amazon scenario is a special kind of graph, often called an affiliation network or a bipartite graph.

We define bipartite graphs formally in the next section. For now, we simply re-draw the above example graph without any of the grey edges. The new graph is G'=(V,E'):

se4

The graph above not only omits the grey edges of E, but also distinguishes nodes by two types. In a Delicious example, the red squares would be tags, the blue circles would be Web pages, and the links would represent bookmarks on Delicious that associate a tag with a Web page. In an Amazon example, the red squares would be books, the blue circles would be Amazon users, and the links would represent purchases.

Unlike Facebook, where any node can invite any other node to be friends, Delicious and Amazon do not much care (relatively speaking) about any edges other than those between red squares and blue circles. By eliminating edges that join blue circles to blue circles, and edges that join red squares to red squares, we get a graph that yields to much simpler algorithms while still modeling the connections that are most important to the task at hand (i.e., recommending new items of interest).


Network Structure

Bipartite (Affiliation) Graphs

A bipartite graph is as a graph whose nodes can each be colored red or blue such that there are no red-red edges and no blue-blue edges. The following graph, for example, is bipartite and so can be colored red and blue without creating red-red edges or blue-blue edges:

Red-Blue Net

Another way to think of a bipartite graph is by drawing the red nodes on one side of the picture and the blue nodes on the other side of the picture:

Red-Blue 2-Columns

Bipartite graphs are very commonly drawn in this way, with one subset of vertices on one side and the other subset on the opposite side. For the graph to be bipartite, edges may join vertices across columns but may never join vertices in the same column.

Bipartite graphs are useful for modelling matching problems. For example, which actors have worked on which movies? Which websites are bookmarked by which people? These are bipartite graphs.

Most graphs are non-bipartite. A graph is non-bipartite when there is:
  1. No way to color its vertices red and blue without creating red-red or blue-blue edges
  2. No way to draw its vertices in two columns so that edges only cross from column to column and never join vertices in the same column
Note that the above two condidtions are both Non-bipartite graphsaying the same thing. For example, the graph below is non-bipartite:

Suppose we color the first node red; then we must color the second node blue. But then there is no way to color the third node red or blue without creating either a red-red edge or a blue-blue edge.

Social bookmarking, bipartite graphs, and structural equivalence

One application of bipartite networks is the organization of user-created tags, such as the tags in del.icio.us. Together with the website URLs they are associated with, these tags form a bipartite network. For example, the yellow nodes below left are tags on del.icio.us, and the orange nodes below right are some of the URLs associated with these tags:

Bipartite Tags

Even a machine that doesn't understand the tag-words themselves can analyze how the tag-words relate to each other based on how others associate those tags with website URLs.

The simplest indicator of tag relatedness is the group interlock network, which is explained in Six Degrees. Unfortunately, the group interlock network is too crude for many purposes, because it provides only a yes/no indicator and does not indicate relative amount of relatedness.

A more useful measurement of tag (or node) relatedness is based on structural equivalence, which we defined previously.

The structural equivalence of two tags is 1 when they have identical website associations and 0 when two tags have not a single common website associated with them.

Examples based on the above yellow & orange graph:

The structural equivalence of YouTube and Video is

|N(YouTube) ∩ N(Video)| / |N(YouTube) ∪ N(Video)|

= |{A, B, C, F} ∩ {A, B, C, D, E, F}| / |{A, B, C, F} ∪ {A, B, C, D, E, F}|

= |{A, B, C, F}| / |{A, B, C, D, E, F}|

= 4 / 6

= 2/3

 

The structural equivalence of MySpace and Video is

|N(MySpace) ∩ N(Video)| / |N(MySpace) ∪ N(Video)|

=|{H, I, K} ∩ {A, B, C, D, E, F}| / |{H, I, K} ∪ {A, B, C, D, E, F}|

=|{ }| / |{A, B,C, D, E, F, H, I, K}|

= 0 / 9

= 0

[stop]

Trees 

A tree is a graph in which any two vertices are connected by exactly one path.

Here are three different trees made from the same set of nodes:
trees

Rooted (or Hierarchical) Trees

In computer science, we often use trees for hierarchical organization of data, in which case we use rooted trees.
rooted tree
In the above tree, node 3 is specially designated as the root of the tree. Every other node in the tree is implicitly ranked by its distance from the root.

For any edge in a rooted tree, one end of the edge must be closer to the root than the other. We call the node closer to the root the parent of the other node, and we call the node farther from the root a child of its parent node. For example, nodes 9 and 4 are children of node 3. Similarly, nodes 10, 6, and 5 are all children of node 4. (And node 2 is the only child of node 9.) Any node in a rooted tree may have any number of children (including zero). A node with no children is called a leaf (examples leaf nodes are 1, 10, 7, and 5). Every node except the root node must have exactly one parent. For example, the parent of node 7 is node 6. Note that the above rooted tree is undirected, but each edge in the rooted tree has an implicit direction based on the parent-child relationship represented by that edge.

Any node of a tree can be a valid root of that tree. For example, if we decided that node 6 was actually the central anchor of the above network, we could re-root the exact same tree at 6 instead of 3:

rerooted tree
Check and see that the two graphs above have exactly the same sets of nodes and edges. What happens to parent-child relationships when a tree is re-rooted?

In terms of its graph theoretic properties, a rooted tree is equivalent to the traditional hierarchical organizational chart:
Org Chart
The Division Officer at the top of the above hierarchy would probably argue that changing the root of the corresponding tree is a bad idea. However, if we are using trees to connect computers or concepts rather than people, we are less likely to encounter resistance when considering different alternatives for designating the root of a tree.

Important properties of trees

A tree (rooted or not) may also be defined by the following equivalent conditions:
  1. A connected graph with no cycles.
  2. A connected graph with |E| = |V| - 1
Definition #1 above gives us an easy way to eyeball a graph and see if it is a tree. Does the graph consist of one connected component or does it fall apart into several connected components? A tree must be connected. Does the graph have any cycles (intuitively, a path that ends where it started)? A tree cannot have any cycles.

Which of the following three graphs are trees or not trees?
Not Trees and tree

Definition #2 above gives us a precise relationship between the number of nodes and edges in a tree, which can be very useful when designing algorithms that involve trees.

It is somewhat remarkable that these three conditions are completely equivalent:
  1. G is a connected graph with no cycles.
  2. G is a connected graph with |E| = |V| - 1.
  3. G is a graph in which any two vertices are connected by exactly one path.
To understand better why these three are in fact equivalent, we will use mathematical induction.

Mathematical Induction

When you write a computer program, how do you know it will work? A mathematical proof is a rigorous explanation of such a claim. (Example claim: "MapQuest finds the shortest path between any two addresses in the United States.") A proof starts with statements that are accepted as true (called axioms) and uses formal logical arguments to show that the desired claim is a necessary consequence of the accepted statements.

Mathematical induction (or simply induction) is one of the most important proof techniques in all of computer science. Induction is a systematic way of building from small example cases of a claim to show that the claim is true in general for all possible examples.

Every proof by induction has two parts:
  1. Base case: Look at the smallest possible example case of the desired claim and show that the claim is true. The base case is often disarmingly trivial.
  2. Inductive step: Here we get to assume the inductive hypothesis: that the claim holds for all example case of size i or less. Given that assumption, we must prove that the claim also holds for all example cases of size i+1.
That's it! Here's an example proof by induction. We will prove something we discussed in class weeks ago:

Suppose we want to prove that the maximum possible number of edges in an undirected graph with n nodes is n*(n-1)/2. Let's prove this by induction:
  1. Base case: Consider a graph with one node. This graph can have no more than 0 edges. Checking our claimed formula when n=1, we verify 1*(1-1)/2 = 0.
  2. Inductive step: The inductive hypothesis lets us assume that any graph with i nodes can have i*(i-1)/2 edges but no more. Now we have a graph with n=i+1 nodes. What is the maximum possible number of edges in this graph? We hope to show that this is (i+1)*((i+1)-1)/2 = (i+1)*i/2.
We take advantage of our inductive hypothesis by aribitrarily numbering the nodes of our graph 1, 2, 3, ..., i, i+1. Then we consider the first i nodes as one part of our graph and node i+1 as the other part of our graph:
Max Edge Induction
To count the maximum possible edges in our graph, we consider
  1. The maximum possible number of edges in the red circle (nodes 1, 2, 3, ..., i)
  2. Plus the maximum possible number of edges outside the red circle--those edges incident to node i+1
The inductive hypothesis tells us the maximum possible number of edges in the red circle is i*(i-1)/2.

The maximum number of edges incident to node i+1 is simply i (which would happen if node i+1 is adjacent to every other node in the graph).

Adding the above two terms gives us the maximum number of edges in a graph with n=i+1 nodes equal to


i*(i-1)
------- + i
2

=
i2 - i + 2i
-----------
2

=
(i+1)*i
--------
2



Voila! Check for yourself that we have completed both the base case and the inductive step. We now have a complete proof by induction that the maximum number of edges in a graph with n nodes is n*(n-1)/2.

Now we return to our original goal, to understand better why these three conditions are equivalent:
  1. G is a connected graph with no cycles.
  2. G is a connected graph with |E| = |V| - 1.
  3. G is a graph in which any two vertices are connected by exactly one path.
We will use induction to prove that a connected graph with no cycles has |E| = |V| - 1.
  1. Base Case: Consider a graph with one node. A one-node graph is trivially connected and has no cycles. It cannot have any edges. We immediately get |E| = |V| - 1 = 0.
  2. Inductive Step: Consider any connected graph with no cycles and no more than i nodes. The inductive hypothesis lets us assume that the number of edges is one less than the number of nodes for any such graph. Now consider a connected graph with no cycles and i+1 nodes. We aim to show that any such graph has exactly i edges, or that |E| = |V| - 1 = (i+1)-1 = i.
We take advantage of the inductive hypothesis by aribitrarily numbering the nodes of our graph 1, 2, 3, ..., i, i+1. Then we consider the first i nodes as one part of the graph and node i+1 as the other part of the graph.

Because the graph is connected, there is at least one edge from node i+1 to the rest of the graph. There may be more than one edge from node i+1 to the rest of the graph. Below we draw the case where there are three such edges:
Edges Per Tree Induction
Here is the crux of the inductive step: If we remove node i+1 from the graph, then the rest of the graph breaks up into sub-trees. The number of sub-trees is exactly equal to the number of edges incident to node i+1.

To illustrate why this is so, we have colored the edges incident to node i+1 in the above example along with the sub-trees they connect to. Notice how each of the three "NO" scenarios violates one of our assumptions, either by creating a cycle (which either of the two red dashed edges would do) or by creating a disconnected graph (which the grey node would do).

Re-iterating: However many edges are incident to node i+1, then the nodes 1, 2, 3, ..., i break up into exactly that many sub-trees when we remove node i+1 and its edges.

Because each of the sub-trees has no more than i nodes, we can use the inductive hypothesis to tell us that each sub-tree has exactly one less edge than it has nodes.

In the above example, this means
# edges in blue sub-tree =
# edges in green sub-tree =
# edges in yellow sub-tree =
|Vblue| - 1
|Vgreen| - 1
|Vyellow| - 1

Although we don't know how many nodes are in each sub-tree, we do know the total number of nodes in all the sub-trees is i. So if we add the three equalities above, we get

# edges in all sub-trees =
i - 3
Now we add the three edges incident to node i+1 (three--as drawn in the example above) to get

|E| =
i - 3 + 3 = i

Three is a number we arbitrarily chose for the sake of this example--and notice how the 3's cancel each other out in the above sum. No matter whether there are 3, 2, 1, or whatever number of edges incident to node i+1, when we look at a connected graph with no cycles and i+1 nodes, we will end up with |E| = i, just as we aimed to prove. This completes our proof by induction that a connected graph with no cycles has |E| = |V| - 1.

Minimum cost trees

Suppose your new office building is being wired for Internet. For the service to be worth much, your network must be connected (a path between every pair of computers); however, you probably don't want to install direct cables connecting every possible pair of computers.

Trees provide a useful starting point for any network service provider facing this sort of problem. This is because a tree connects a network with the minimum possible number of edges. (Think about why this is true.)

Simply finding any tree to connect a network isn't enough though. Not all computers (or nodes) are equally costly to join directly. Consider the network below:
K1
There are numbers next to each edge. Each number represents the cost (or weight) of the edge. In the above example, the weight of each edge corresponds to how far apart its endpoints are.

Suppose we want to choose a subset of edges to connect the above 7 nodes in a tree. Any tree connecting 7 nodes must have exactly 6 edges, so finding a "minimum edge tree" is meaningless. What we want are the 6 edges that form a tree for the least total cost.

We call this problem the minimum cost tree problem. It is often called the minimum spanning tree problem--"spanning" being a term applying to any subgraph that includes every possible node of the original graph.

The minimum cost tree problem turns out to one that is both remarkably important in real life and also easy to compute. Below is an algorithm that starts with a graph G = (V,E) and computes a set of edges T that form a tree (V,T) that has minimum possible total cost:
  1. create a set T, initially equal to { }
  2. create a set F, initially equal to E
  3. if F = { } then go to step 7
  4. remove an edge {x,y} with minimum cost from F
  5. if x and y are already connected by a path of edges in T, then discard edge {x,y}; otherwise add edge {x,y} to T
  6. go to step 3
  7. done--the minimum cost tree is (V,T)
The above algorithm is called Kruskal's algorithm.

Example

We will apply Kruskal's algorithm to the graph above. Recall that set F initially includes every edge. Set T is initially empty. As we add edges to T in the example below, we color those edges green.

K1

This is our original graph. The numbers near the arcs indicate their weight. None of the arcs are highlighted.

K2

{A,D} and {C,E} are the shortest arcs, with length 5, and {A,D} has been arbitrarily chosen, so it is highlighted.

K3

However, {C,E} is now the shortest arc that does not form a cycle, with length 5, so it is highlighted as the second arc.

K4

The next arc, {D,F} with length 6, is highlighted using much the same method.

K5

The next-shortest arcs are {A,B} and {B,E}, both with length 7. {A,B} is chosen arbitrarily, and is highlighted. The arc {B,D} has been highlighted in red, because it would form a cycle (A,B,D,A) if it were chosen.

K6

The process continutes to highlight the next-smallest arc, {B,E} with length 7. Many more arcs are highlighted in red at this stage:

  • {B,C} because it would form the cycle (B,C,E,B)
  • {D,E} because it would form the cycle (D,E,B,A,D)
  • {F,E} because it would form cycle (F,E,B,A,D,F).

 

K7

Finally, the process finishes with the arc {E,G} of length 9, and the minimum spanning tree is found.


PageRank

PageRank is the algorithm behind Google. According to Tim Berners-Lee, PageRank and its kin "form the heart of the critical crawlers and rank-assessment algorithms behind Web search."

Google describes PageRank:

PageRank relies on the uniquely democratic nature of the web by using its vast link structure as an indicator of an individual page's value. In essence, Google interprets a link from page A to page B as a vote, by page A, for page B. But, Google looks at more than the sheer volume of votes, or links a page receives; it also analyzes the page that casts the vote. Votes cast by pages that are themselves "important" weigh more heavily and help to make other pages "important".

In other words, a PageRank results from a "ballot" among all the other pages on the World Wide Web about how important a page is. A hyperlink to a page counts as a vote of support. But, as mentioned above, not all votes count equally.

Centrality

The core network mathematical concept in PageRank is the notion of centrality. This is an intuitive notion that is surprisingly hard to pin down mathematically. For example, which node is the most central in the following graph?

Google Centrality

There is no one right answer to the question, because the question is too vague. There are many different ways to define centrality more precisely. Here are three important ones, all of which rely critically on the direction of links (something we have largely glossed over until now):

The PageRank formula is described and illustrated with examples and an interactive PageRank calculator at http://www.webworkshop.net/pagerank.html. You can also find a rigorous mathematical introduction in Wikipedia (not for the faint of heart).

WebWorkshop writes the PageRank formula in a manner that is compact but subtle. We restate the formula here and then introduce an easier, simplified forumulation. We begin by defining a damping factor d to be a fixed coefficient at least 0 but no more than 1. (Google uses d=0.85.) Then for any graph, we define the PageRank of each of its nodes as follows:

For any node x, let

  • {y1, y2, ..., yn} denote the set of all nodes with directed links to x,

Then

PageRank(x) =

(1-d) + d*(PageRank(y1)/Outdegree(y1)

+PageRank(y2)/Outdegree(y2)

...

+ PageRank(yn)/Outdegree(yn))

Computing PageRank from the above definition presents both major and minor difficulties:

Another problem with the above definition of PageRank is understanding the damping factor, which is rather obtuse. Here is the best way we know to explain the damping factor d:

One way to consider PageRank is as a kind of probability distribution. Suppose a Web surfer starts at a random page and clicks at random for a while (i.e., clicking on any link that happens to be on the current page, etc.). Occasionally the Web surfer types in a new URL (randomly chosen from all Web pages) instead of link-clicking. Eventually the surfer stops at his "destination" page. The PageRank of a Web page x is essentially the probability that x will be the destination page of that surfer's random walk of clicking and URL-typing. In this scenario, the damping factor d is essentially the probability that the "random Web surfer" clicks a link instead of typing in a new URL. So...

  • If d=1 then the resulting PageRank function assumes that the "random Web surfer" starts at a random page and then always clicks and never types in a new URL.
  • If d=0 then the resulting PageRank function assumes that the "random Web surfer" never clicks; he always types in another random URL. (Admittedly the d=0 model corresponds to an unrealistic and/or pathological version of Web surfing; hence Google uses d=0.85, which is closer to 1 than 0.)

Here we present HopRank as a simpler version of PageRank. The HopRank function takes two inputs: a Web page x and a positive integer i. It is then written HopRanki(x). There are three good ways to interpret HopRanki(x):

  1. HopRanki(x) is a a simple way to estimate PageRank. It is essentially equivalent to the PageRank formula referenced above with damping factor d=1 and with no dividing by node outdegrees.
  2. HopRanki(x) is a more sophisticated way to measure the indegree of x. HopRanki(x) takes into account not just how many incoming links are incident to x; it also accounts for the ranks of where those links are coming from. One incoming link from a high-ranking node can be more important than many incoming links from many low-ranking nodes.
  3. HopRanki(x) simplifies the "random Web surfer" model of PageRank in the following way: HopRanki(x) = the number of directed walks of length i that end at node x.

Perhaps the easiest way to understand HopRank is not by interpreting it (as above), but by showing how it is calculated, which we do below. The simplest case of HopRanki(x) is when i=1:

HopRank1(x) = InDegree(x)

The above equality is simply the definition of HopRank1(x). From that starting point, we will proceed iteratively and define HopRank2 based on HopRank1, then HopRank3 based on HopRank2, and so forth.

Given a directed graph G=(V,E), we lay the foundation of the HopRank calculation by defining the InNeighborhood of a node x ∈ V to be

InN(x) =
the set of all nodes that have directed edges into x
=
{y: (y,x) ∈ E}

Then HopRank1(x) = |InN(x)| = the indegree of x. Below we illustrate this using same the 6-node graph as above:

Node x
InN(x)
HopRank1(x)
1
{4,6}
2
2
{3}
1
3
{2}
1
4
{3,5,6}
3
5
{4}
1
6
{1,2,3,5}
4
Google Centrality

HopRank1 is a very local measure of node centrality. We can measure a slightly more system-wide sense of node centrality by defining HopRank2 as follows: For any node x, HopRank2 of x is the sum of the HopRank1's of all nodes in InN(x).

Node x
InN(x)
HopRank1(x)
HopRank2(x)
1
{4,6}
2
7 = 3+4
2
{3}
1
1 = 1
3
{2}
1
1 = 1
4
{3,5,6}
3
6 = 1+1+4
5
{4}
1
3 = 3
6
{1,2,3,5}
4
5 = 2+1+1+1

We can then repeat the above process to calculate HopRank3 based on HopRank2, then HopRank4 based on HopRank3, etc. Below is a table of five successive calculations of HopRanki, based on the same graph as above:

Node x
InN(x)
HopRank1(x)
HopRank2(x)
HopRank3(x)
HopRank4(x)
HopRank5(x)
1
{4,6}
2
7
11
21
38
2
{3}
1
1
1
1
1
3
{2}
1
1
1
1
1
4
{3,5,6}
3
6
9
19
29
5
{4}
1
3
6
9
19
6
{1,2,3,5}
4
5
12
19
32

Below we redraw the same table, highlighting exactly how the row for node 4 is calculated. Because InN(4) = {3,5,6}, in each iteration i the calculation of HopRanki(4) adds the HopRank(i-1) of nodes 3, 5, and 6:

inrank

The above tabular calculation leads us to our most concise complete definition of HopRanki(x). Again, we assume we are given a directed graph G=(V,E), a node x∈V, and a positive integer i:

if i = 1:
HopRank1(x) =
|InN(x)|
if i > 1:
HopRanki(x) =
sigma
y∈InN(x)
(HopRank(i-1)(y))

Looking back at the tabular calculation, notice that relative rankings of nodes change from column to column in the table. Sometimes node 6 is ranked highest; sometimes node 1 is ranked highest. After "enough" iterations, the rank order of relative nodes will stabilize. Even more importantly, the relative ratios of HopRanki for different nodes will stabilize.

In other words, when the value of i is "big enough," then the relative ratios of HopRanki for different nodes will not change in any significant way from iteration i to iteration i+1. According to http://www.webworkshop.net/pagerank.html, Google uses 40-50 iterations to calculate its version of PageRank. So we can informally say that HopRank50 will give us a good simplified version of PageRank, where the relative rankings of nodes are stable.