One common question we hear from our customers is how much of the social network predictive effect can be accounted for by demographic attributes. Afterall, if birds of the same feather flock together, then perhaps the social network influence effect is nothing more than a simple artifact of demographic similarity. The implication of such an assertion runs deeper if you consider that most customers who have demographic data are already using this data for prediction. In this series of blog posts, we’ll discuss a recent 2009 paper that tried to tease out these two effects.
Let’s start with the dataset that was used for this analysis. Yahoo provided the instant messaging data for 27.4million users, the day-by-day adoption of the Yahoo! Go product over 5 months and the precise attribute and dynamic behavioural data on the users’ demographics. With such a rich set, it becomes possible to separate the demographic data from the social influence.
Results
First off let’s discuss the overall findings regarding social influence. The findings are so astounding that I’m going to quote them verbatim.
- “adopters have a 5-fold higher percentage of adopters in their local networks (t-stat=100.12, p<0.001; k.s. stat=0.06, p<0.001)”
- “adopters receive a 5-fold higher percentage of messages from adopters than nonadopters”
- “Both the number and percentage of one’s local network who have adopted are highly predictive of one’s propensity to adopt (Logistic: beta# = 0.153 p<0.001 beta% = 1.268 p<0.001)…
- and to adopt earlier (Hazard rate: beta# = 0.10 p<0.001 beta% = 0.003 p<0.001)
One interesting take away is that when analyzing social networks, you can construct the network using links or address lists, but the relevance of actual messages improves predictions further.
This graph above shows how much more likely you are to adopt if you had adopter friends as compared to no adopter friends. It means that if you have only 1 adopter friend in your local community, your chances of adopting doubles. If you have 4 adopter friends, your chances of adopting the product jumps to almost 4x, and so on.
So what?
Regardless of the reason for the adoption, it’s clear that customers whose social community has adopted a product is more likely to adopt the same product. This finding can be used by marketing in different ways: use this %-of-community as a predictor in your propensity modeling; make sure that you seed at least a few adopters in each social community etc. Your marketing starts to look different when you market to the “social community” and not to the individual customer!
In the next post, we’ll cover how they overlapped these results to show how much of it is due to demographics.
References
Sinan Aral, Lev Muchnik and Arun Sundararajan. Distinguishing influence-based contagion from homophily-driven diffusion in dynamic networks. Proceedings of National Academy of Science, Dec 22, 2009 vol 106 no 51 p. 21544-21549.
Tags: Marketing, predictive analysis, Social networks, viral marketing
Thanks, for bringing this study to attention, Nick.
Did the article cover influence-based contagion as a function of adopters/degree only or also adopters/(degree and connection between friends/neighbors). My thought is along the lines of triangular relationships creating more peer pressure than just dyadic – unrelated friend adopters.
What do you think?
Cheers, Doris
Hi Doris, very good question. The study does not explicitly address this question of connections between adopter friends. However, in part two of the series on this study, I discuss how they find that if you control for similarity in demographics and user behavior, the increase in contagion due to more adopter friends actually goes away. Contagion effect alone does not increase with more adopter friends. In the context of marketing application though, contagion or similarity matters less than whether you can actually influence behavior, ie. buying new product or staying with service.