Leskovec et. al., ACM Conference on Electronic Commerce (EC), 2006.

From Cohen Courses
Revision as of 18:20, 1 February 2011 by Wcohen (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigationJump to search


The Dynamics of Viral Marketing by J. Leskovec, Lada Adamic, Bernardo Huberman. ACM Conference on Electronic Commerce (EC), 2006.

Online version

Jure Leskovec's website Arxiv [Extended version]


This paper presents a measurement and modeling approach to viral marketing. The authors study a product recommendation network of millions of users and half a million of different products in which users can recommend products that they bought to other people in the network. The data spans from June 2001 through May 2003.

The authors first present several statistics on the recommendation network. The network is formed such that nodes represent customers and edges represent recommendations of products. The edges also have time labels. They observe that the network consists of multiple small disconnected components. The largest connected component contained only ~2.5% of all the nodes in the network at the final time step.

They also study influence propagation and show that the cascade sizes follow power law distributions, with usually large exponents. This shows that many recommendation cascades are small and there are only a few cascades with large sizes. The authors next propose a simple propagation model which is shown to generate power law fittings to cascade sizes.

Next, the authors answer the following three questions: (1) Does receiving more recommendations increase the likelihood of buying? (2) Does sending more recommendations yield more purchases? and finally (3) Do multiple recommendations between two individuals weaken the impact of the bond on purchases?

Finally, they measure recommendation success by product category. To extend their understanding on the effects of recommendation success, they fit a linear regression model to the real data (around 50K products that had at least one purchase through a recommendation) after extracting several (network+non-network) features, such as the number of recommendations, the number of senders, the number recipients, price of product, average product rating etc. Their results show that the average product rating affects success the least compared to other features. In addition, senders and receivers have negative coefficients, showing that successfully recommended products are actually more likely to be not so widely popular. On the other hand, the price of product and the number of recommendations are the only features with positive coefficients. This suggests that more expensive and more recommended products have a higher overall success rate.

Related papers

There has been a lot of previous work in modeling links and text, including