2. Introduction Viral marketing techniques rely on the social network among customers to spread product recommendations Key Assumption Peer-influence is both static, and independent of the product type
3. Motivation Ann Janet John Bob and Mary will definitely be interested. However, I think Ann is not much into movies Mary WOW… I’ll send it over to everyone Women’s fashion Store(Invite a friend and get 10% off your next purchase) MovieRental.com(Refer a friend and get $10 off your next rental) Bob
4. Motivation Peer-influence is dynamic Dependent on previous user interactions Viral marketing strategies have an implicit effect on the underlying social network Sometimes changing the structure of the underlying network altogether
5. Motivation Peer-influence is dependent on the type of product being spread Users have varying preferences for different products Both factors play a role in product adoption
6. Background Popular Diffusion Models Threshold Models (e.g. Linear threshold model) Cascade Models (e.g. Independent cascade model) Influence probabilities are assumed to be static, insensitive to the product type, and known in advance
7. Objectives Capture the diversity in user preferences for different products Model the change in influence probabilities across multiple campaigns Design a viral marketing strategy that takes these factors into account
8. Outline Case Study: Digg.com Differential Adaptive Diffusion Model Adaptive Viral Marketing Conclusion and Future Work
9. Case Study: Digg.com Social news website Users “submit” stories in differenttopics, which can then be “digged”by other users Users can “follow” other users to get their submissions and diggs on their homepage
10. Case Study: Digg.com Following links define the social network User submissions serve as proxy of user preferences for different topics User diggs are analogous to product adoptions
11. Dataset Social Network (user-user following links) 11,942 users 1.3M follow edges Digg Network (user-story digging links) 48,554 news stories 1.9M digg edges 6 months (Jul 2010 – Dec 2010)
12. Observation 1User Submissions vs. Diggs Smaller percentage of users who digg stories in topics that vary significantly from the topics they post in Most users adopt only stories of interest to them
13. Characterizing User Submissions Focused UsersHighly skewed preferences toward one or two topics Biased UsersLess skewed preferences toward a larger set of topics Balanced (Casual) UsersAlmost uniform preference across all topics
14. Observation 2Effect of Homophily on Adoption Peers with different topic preferences lose confidence in each other’s recommendations over time (followed diggs/adoption ) Peers with similar topic preferences gain confidence in each other’s recommendations over time (followed diggs/adoption )
15. Outline Case Study: Digg.com Differential Adaptive Diffusion Model Adaptive Viral Marketing Conclusion and Future Work
16.
17. Experimental Evaluation Evaluate the model performance in predicting future adoptions We use the first four months in Digg.com dataset for learning the influence probabilities, and the last two months for testing
18. Baselines We compare our approach with two baselines* that incorporate temporal dynamics in learning the influence probabilities Bernoulli Each product recommendation Bernoulli Trial Influence probabilities are estimated using MLE over a given contagion time for each user Bernoulli-PC Same Bernoulli representation Partial credit for each recommending peer within the contagion period *A. Goval, F. Bonchi, and L. Lakshmanan. “Learning influence probabilities in social networks.” In Proceedings of the Third ACM International Conference on Web Search and Data Mining (WSDM’10), 2010
19. Results The Adaptive model, taking both the diffusion dynamics and the users heterogeneity into account, yields better performance
20. Outline Case Study: Digg.com Differential Adaptive Diffusion Model Adaptive Viral Marketing Conclusion and Future Work
21. Adaptive Viral Marketing User recommendations are most effective when recommended to the right subset of friends Highly selective behavior Limited exposure Spamming lower confidence levels, limited returns What is the appropriate mechanism for maximizing both the product spread and adoption?
22. Adaptive Rewards Successful recommendations are awarded (αxr)units, while failed ones are penalized ((1-α) xr) units α conservation parameter Most existing viral marketing strategies assume α=1 (no reason for the user to be selective) The penalty term helps maintain the average overall confidence level between different peers
23. Experimental Setup Agent-based models to simulate the behavior of customers in different settings When an agent adopts the product, it makes a probabilistic decision to send a recommendation based on its knowledge about the peers’ preferences The objective of each agent is to maximize its expected reward according to the existing strategy
24. Experimental Setup Two sets of experiments Fully observable: The agents are allowed to directly observe the preferences of their peers Learning preferences: The agents have to learn the peer’s preferences based on their response to previous recommendations Simulate the diffusion of 500 campaigns for products from 5 different categories We use a linear kernel for adjusting the confidence levels between peers after each campaign
25. Fully Observable Intermediate values for α (e.g. α= 0.5) consistently maintains high adoption rates and high overall confidence over large number of marketing campaigns
26. Learning Preferences Allowing agents to learn the preferences accounts for both the product preference as well as the confidence level
27. Effect of Spammers To test the robustness of our proposed method, we inserted spamming agents in the network A spamming agent forwards all product recommendation for all its peers, regardless of their preferences We set (α = 0.5) for all the other agents, and vary the number of seeded spammers
28. Effect of Spammers The network adapts to the presence of spammers (dropping their confidence levels), and continues to maintain adoption levels through trusted links
29. Outline Case Study: Digg.com Differential Adaptive Diffusion Model Adaptive Viral Marketing Conclusion and Future Work
30. Conclusion Network dynamics and users heterogeneity have a considerable impact on user interactions The proposed adaptive diffusion model incorporates both aspects to better model the diffusion process Adaptive rewarding mechanism for viral marketing maintains higher confidence levels over time
31. Future Work Potential Applications Social Recommendation Collaborative Filtering Analyzing the impact of the proposed model on opinion leader identification Incorporating the time-variability aspect of user-product preferences in the model
----- Meeting Notes (7/18/11 15:48) -----encouraging her
Add a slide for describing digg.com / add key words (submissions vs. diggs)
One interpretation for the high divergence users that they are imitatorsWe use the topic distribution of the user postings as an influence-independent source for measuring preferencesFig1: KL-divergence between the topics of user submissions and diggsFig2: KL-divergence between user submissions (prefs) and uniform distribution of topics
* Reduce text- In order to capture both the heterogeneity in user preferences as well as the temporal dynamics, we split the influence probability into two components- The product preferences are either given or can be inferred from an influence-independent source (such as user submission in the case of Digg)
Mention that these baselines use the independent cascade model
Color the series if possible We compared against two models that take the changing dynamics of the influence probability into account, but doesn’t address the user preference / heterogeneity aspect
If a user is very selective and makes each recommendation to only a few friends, then the chances of success are slim due to limited network exposure. On the other hand, recommending a product to everyone may have limited returns as well, due to the effect of irrelevant recommendations on the confidence levels between peersThe natural question to ask now is
Typical viral marketing strategies reward users for successful recommendation, but don’t penalize them for failed ones 0 representing fully conservative behavior and 1 representing fully nonconservative behavior.
No text for alpha=1,0 – just arrowarrows for this and next
In more realistic settings, the user friends learn them from her responses to different recommendation*The settings where users are allowed to learn preferences are better than the settings where users directly observe preference (prev slide) as this one
Varying the percentage of spammers (individuals recommending any product to all their peers), and analyzing the over all effect (set alpha = 0.5)As percentage of spammers increases, the overall adoption and confidence level decreases. However, the network is able to adapt to their presence by assigning lower confidence levels to
*Add future work / implicationsUtilized in social recommendations (incorporate both “trust” in a friend and their preference combined with your preference to make better automated recommendations – Facebook, x like this page)