1. CIKM’2008 Presentation
Oct. 27, 2008
Napa, California
Mining Social Networks Using Heat Diffusion Processes for Marketing Candidates Selection
Hao Ma, Haixuan Yang, Michael R. Lyu and Irwin King
Dept. of Computer Science & Engineering
The Chinese University of Hong Kong

2. Background
According to , advertisement spending on worldwide social networking sites
2006, $445 millions
2007, $1.12 billions
2010, expected $2.28 billions

3. Growth of Social Networking Sites

4. Background Why has become so popular?
“Google's utility and ease of use have made it one of the world's best known brands almost entirely through word of mouth from satisfied users”
Other successful examples like Hotmail and MySpace

5. Background
Massive quantities of data are available on online social network sites
Blogs, Opinions
Knowledge sharing sites
Collaborative filtering systems
Newsgroups
systems, etc.
All of these social networks provide valuable information for decision-making in marketing campaigns

6. Objective Given a social network of consumers,
Try to convince a subset of influential individuals to adopt a new product.
The goal is to trigger a large cascade of further adoptions (coverage) of the product.
The research problem: which set of individuals should we select?

7. Chanllenges
Social Network is very complicated (clustering, power-law, etc.)
Marketing strategies are time-dependent
All kinds of information flows on social networks (positive, negative)

8. Our Solutions
Model the diffusion of innovations as the process of heat diffusion (3 diffusion models)
Undirected social networks
Directed social networks
Directed social networks with prior knowledge
Three candidates selection algorithms
Top-K Algorithm
K-Step Greedy Algorithm
Enhanced K-Step Greedy Algorithm

9. Undirected Social Network
Heat Diffusion Models
Undirected Social Network

10. Undirected Social Network
Heat Diffusion Models
Undirected Social Network
Thermal conductivity
Degree of node i
eth: diffusion kernel
Heat value of node i at time t
Vector of the initial heat distribution
Vector of the heat distribution at time t

11. An Example of Heat Diffusion on an Undirected Graph
At time 0, suppose node 1 is given 3 units
of heat, and not 2 is given 2 units of heat.
Set
A small undirected
social network graph
Curve of heat change
with time

12. Directed Social Network
Heat Diffusion Models
Directed Social Network
Thermal conductivity
Outdegree of node i
Vector of the initial heat distribution
Vector of the heat distribution at time t
Equal to 1 if node i has outlinks, else equal to 0

13. Directed Social Network
Heat Diffusion Models
Thermal conductivity
Directed Social Network
with Prior Knowledge
0.1
0.8
0.2
0.4
0.7
0.3
0.9
0.5
0.6
Outdegree of node i
The personality score of user j
Weight between node j and node i
Vector of the initial heat distribution
Vector of the heat distribution at time 1
Equal to 1 if node i has outlinks, else equal to 0

14. Candidates Selection Algorithms
Top-K Algorithm
Is(t) Influence set of s: the set of nodes influenced by the candidate s (at time t)

15. Candidates Selection Algorithms
Top-K Algorithm is very naive since it ignores the potential overlaps of top-k influence sets
Maximizing the coverage among top-k influence sets:
It is a set coverage problem and proved to be NP-hard

16. Candidates Selection Algorithms
K-Step Greedy Algorithm

17. Candidates Selection Algorithms
K-step Greedy Algorithm
First computes the influence set of each individual in turn
Then chooses the k sets with the maximum coverage
In reality, several diffusion sources diffuse information at the same time, not just from one single source

18. Candidates Selection Algorithms
Enhanced K-Step Greedy Algorithm
First select the Top-1 influential user
Then each time diffuse information using two users: the Top-1 user and one of other users. Choose the Top-2 influential users. ……
Each time diffuse information using K users, the Top-(k-1) users and one of other users. Select the Top-k influential users as the final results

19. Candidates Selection Algorithms
Enhanced K-Step Greedy Algorithm

20. Experiments
Dataset: Epinions.com (75,888 users, 508,960 relationships)
Every member of Epinions maintains a “trust” list and a “block (distrust)” list
We consider only the “trust” relationship between members of Epinions in the first experiment.

23. The social network graph of the Epinions dataset
20 users selected by the Enhanced K-Step Greedy Algorithm are set to red color

24. Negative Information Diffusion
Previous work does not consider the problem of passing negative information
In reality, some people may dislike the products and actively tell their friends
Our diffusion models can naturally simulate the diffusion of negative information by setting the initial value as negative

25. Negative Information Diffusion
Negative diffusion sources will significantly affect the adoptions of products
We propose a heuristic algorithm to defense against negative information diffusion when selecting the marketing candidates
Suppose we are given 10 product samples to market to consumers
For negative information case, suppose Top-1 user is the negative source

26. Negative Information Diffusion
In the Defense Algorithm
We select two users (Top-2 and Top-3) to have the most overlaps with the Top-1 user to diffuse the positive opinions
It generates better coverage than the Enhanced Greedy Algorithm with Negative Comments

27. Conclusions
Proposed three diffusion models, and three candidates selection algorithms
Considered social network clustering property and time-dependent marketing strategies
To the best of our knowledge, this is the first work which proposes how to defend against diffusion of negative information

28. Thanks!
Q & A