12/18 regular meeting paper

  • 857 views
Uploaded on

 

More in: Technology
  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
    Be the first to comment
    Be the first to like this
No Downloads

Views

Total Views
857
On Slideshare
0
From Embeds
0
Number of Embeds
4

Actions

Shares
Downloads
0
Comments
0
Likes
0

Embeds 0

No embeds

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
    No notes for slide

Transcript

  • 1. Mining Social Network for Targeted Advertising
    • Wan-Shiou Yang Jia-Ben Dia
    • Hung-Chi Cheng Hsing-Tzu Lin
    • Proceedings of the 39th Annual Hawaii Internetional
    • Conference on System Sciences(HICSS’06) Track 6
    • Repoter : Che-Min Liao
  • 2. Outline
    • Introduction
    • The problem
    • Mining cohesive subgroups from social network
    • Targeted advertising based on cohesive subgroups
    • Evaluation
    • Related work
    • Conclusion
  • 3. Introduction
    • The effectiveness of targeting a small portion of customers for advertising has long been recognized by businesses.
    • It is desirable to help customers wade through the information to find the product/service they want.
    • Understanding the needs of current and potential customers.
  • 4. Introduction
    • Traditional approach to targeted advertising (manually) :
      • Analyze a historical database of previous transactions.
      • Analyze the features associated with the customers.
      • Possibly with the help of some statistical tools.
    • Now, many recommender systems, whose basic idea is to advertise products according to users’ preferences :
      • Explicitly stated by the users.
      • Implicitly inferred from previous records.
  • 5. Introduction
    • The first type of recommendation technique was called the content-based approach :
      • Characterizes recommendable products by a set of content features.
      • Represents users’ interests by a similar feature set.
      • Selects target customers whose interests have a high degree of similarity to the product content profile.
    • Limitations :
      • It is inappropriate for products whose content is not electronically available.
      • It is inappropriate to advertise products to a new user and the possibility of providing surprising advertisements is low.
  • 6. Introduction
    • Another type of recommendation technique was called the collaborative approach : (social-based)
      • Looks for relevance among users by observing their ratings assigned to products in a training set of limited size.
      • The nearest-neighbor act as recommendation partners for the target user.
      • Advertises products to the target user that appear in the profiles of these recommendation partners but not in the target user’s own profile.
    • Limitations :
      • Failing to advertise newly introduced products that have yet to be rated by users.
      • Failing to advertise products to a new user who has yet to provide its rating data.
  • 7. Introduction
    • In order to remedy these problems, we propose a data mining framework that utilizes the concept of social network for the targeted advertising of products.
    • To apply data mining techniques to gathered social interaction data to cluster customers into a set of cohesive subgroups,
    • Based on the set of cohesive subgroups, we infer the probabilities of customer’s liking a product category from transaction records.
  • 8. The problem
    • A social network is a set of individuals connected through socially meaningful relationships, such as friendship or co-working.
    • Social network theory traditionally views social relationships in terms of nodes and links.
    • Nodes are the individual actors within the networks, and links are the relationships between the actors.
  • 9. The problem
    • The strength of a tie between two actors is much greater if two actors in question have another mutual acquaintance.
    • Actors with strong ties usually have some sort of common ground and thus often form a subgroup.
    • Because of the mutual influences, customers with strong ties are apt to exhibit similar purchase behavior.
  • 10. The problem
    • In this research, we investigate the problem of targeted advertisement in an environment with the following features :
    • A database that contains customer’s connectedness is available.
  • 11. The problem
    • A database that contains past customers’ transaction records is available
      • To facilitate data mining task, we group transaction of the same customer resulting in aggregated transactions.
  • 12. The problem
    • Out first goal is to identify a set of subgroups in which actors have relatively strong ties.
    • The second goal is to infer the probabilities of customer’s liking a product category from transaction records.
    • Therefore, given a product p, the problem of targeted advertising p is to identify a set of N potential customers who are most likely to purchase p based on the discovered subgroups.
  • 13. Mining cohesive subgroups from social network
    • We define a cohesion measure as Definition 3 to evaluate the cohesion extent of a set of actors :
  • 14. Mining cohesive subgroups from social network
    • In Figure 1, Cohesion(a 1 ,a 2 ) is equal to 1 since actors a 1 and a 2 have all neighbors in common while Cohesion(a 2 ,a 3 ) is 4/7 since only parts of their neighbors are common.
  • 15. Mining cohesive subgroups from social network
    • Our goal is to identify a set of subgroups, in which actors have certain extent of cohesion.
  • 16. Mining cohesive subgroups from social network
    • The cohesion measure decreases as the size of the subgroup increases. (i.e., Cohesion(a 1 ,a 2 ) = 1, Cohesion(a 1 ,a 2 ,a 3 ) = 4/7)
    • Similar to the support measure of the association rule discovery problem, the cohesion measure exhibits the downward closure property.
  • 17. Mining cohesive subgroups from social network
    • We adopted an iterative procedure similar to that in the Apriori algorithm.
  • 18. Mining cohesive subgroups from social network
    • In our algorithm, the nodes in a subgroup are kept sorted in their lexicographic order.
    • In order to improve the efficiency of the algorithm, two cohesive subgroups of size k-1 are joined if they differ only in one node.
    • The detailed algorithm for generating candidate subgroups is sketched as Figure 3.
  • 19. Mining cohesive subgroups from social network
  • 20. Mining cohesive subgroups from social network
    • Continuing with the previous example and assume the cohesion threshold is 0.5.
    • Calling the candidate generation function with L 1 gives 21 candidate subgroups in C 2 .
    • The procedure is repeated until the cohesive subgroups {a 1 ,a 2 ,a 3 ,a 7 } and {a 3 ,a 4 ,a 5 ,a 6 } are generated in L 4 .
  • 21. Targeted advertising based on cohesive subgroups
    • We then estimate the liking of each product category on each cohesive subgroup.
    • To calculate the likings for cohesive subgroup g j , we simple count the number of times each product category c i has appeared in the aggregated category transactions of the customers who belongs to g j .
    • We then normalize the liking Like(c i ,g j ) of each product category c i on cohesive group g j so that
  • 22. Targeted advertising based on cohesive subgroups
    • Suppose product p belong to category c i .The problem of selecting the set of targeted customers can be formulated as the following linear program :
  • 23. Targeted advertising based on cohesive subgroups
    • Obviously, the above linear programming problem is a special case of the fractional 0-1 knapsack problem.
  • 24. Targeted advertising based on cohesive subgroups
    • For a targeted new user, advertisements will be selected according to the likings of the cohesive subgroups that the targeted user belongs to.
    • Given a newly introduced product item, it will be advertised according to the category it belongs.
  • 25. Evaluation
    • We tested the effectiveness of our approach to the targeted advertising of library books in our university.
    • We collected the from/to email logs of all faculties between April 2005 and June 2005.
    • The library uses an online public access catalog (OPAC) system from INNOVATIVE.
    • The circulation data were also collected between April 2005 and June 2005.
  • 26. Evaluation
    • We adopted the first two levels of the Chinese Classification Scheme for the category of Chinese books.
    • We used five-way cross validation for conducting our experiments.
    • We randomly partitioned the set of books that appear in the transaction into five subsets of equal size.
  • 27. Evaluation
  • 28. Evaluation
  • 29. Evaluation
    • We then compare the following four targeted advertising approaches :
      • Group-based.
      • Single-based, which differs from group-based approach in regarding each user as a subgroup.
      • Neighbor-based, which differs from group-based approach in regarding each user and his direct neighbors as a subgroup.
      • Random, which randomly chooses users.
    • The hit rate for p is the ratio of the number of user in target set who had been issued with p to the total number of users who had been issued with p.
  • 30. Evaluation
  • 31. Related work
    • Referral Web system
    • The SNACK system
    • Viral marketing
    • Overtargeting problem
    • Multi-impressions and overselling.
    • Advertisement placement problem.
  • 32. Conclusion
    • The results from applying this approach to library-circulation data have demonstrated the effectiveness of the proposed targeted advertising algorithms.
    • It is essential to apply the proposed approach to a longer period library-circulation data.
    • The effectiveness of the proposed targeted advertising approach is influenced by the availability of a product category.
    • The adoption of a product may propagate through a social network. Integrating the effect of influence propagation to the targeted advertising approach is an interest direction.