only when its visibility is minuscule compared to its other stages, and the highest number of diggs accrues when the social-network effect is nonexistent. We therefore do not consider this feature (otherwise deemed important) a main contributor from a prediction point of view in terms of total popularity count.

conclusion

In this article we have presented our method for predicting the long-term popularity of online content based on early measurements of user access. Using two very popular content-shar-ing portals—Digg and YouTube—we showed that by modeling the accrual of votes on and views of content offered by these services we are able to predict the dynamics of individual submissions from initial data. In Digg, measuring access to given stories during the first two hours after posting allowed us to forecast their popularity 30 days ahead with a remarkable relative error of 10%, while downloads of YouTube videos had to be followed for 10 days to achieve the same relative error. The differing time scales of the predictions are due to differences in how content is consumed on the two portals; Digg stories quickly become outdated, while YouTube videos are still found long after they are submitted to the portal. Predictions are therefore more accurate for submissions for which attention fades quickly, whereas predictions for content with a longer life cycle are prone to larger statistical error.

We performed experiments showing that once content is exposed to a wide audience, the social network provided by the service does not affect which users will tend to look at the content, and social networks are thus not effective promoting downloads on a large scale. However, they are important in the stages when content exposure is constrained to a small number of users.

On a technical level, a strong linear correlation exists between the logarithmically transformed popularity of content at early and later times, with the residual noise on this transformed scale being normally distributed. Based on our understanding of this correlation, we presented a model to be used to predict future popularity, comparing its performance to the data we collected.

in the presence
of a large user base,
predictions can
be based on
observed early
time series, while
semantic analysis
of content is more
useful when no
early click-through
information is
available.

We thus based our predictions of future popularity only on values measurable at the time we did the study and did not consider the semantics of popularity and why some submissions become more popular than others; however, this semantics of popularity may be used to predict click-through rates in the absence of early-access data. 8 In the presence of a large user base, predictions can be based on observed early time series, while semantic analysis of content is more useful when no early click-through information is available.

However, we could not explore several related areas here. For example, it would be interesting to extend the analysis by focusing on different sections of the portals (such as how the YouTube “news & politics” section differs from the YouTube “ entertainment” section). We would also like to learn whether it is possible to forecast a Digg submission’s popularity when the diggs come from only a small number of users whose voting history is known, as it is for stories in Digg’s “ upcoming” section.

 

References

1. alexa Web Information service; http://www.alexa.com

2. Cha, M., Kwak, H., rodriguez, P., ahn, y.-y., and Moon, s. I tube, you tube, everybody tubes: analyzing the world’s largest user-generated content video system. In Proceedings of the Seventh ACM SIGCOMM Conference on Internet Measurement (san diego, oct. 24–26). aCM Press, new york, 2007, 1–14.

3. Cheng, x., dale, C., and liu, J. statistics and social network of youtube videos. In Proceedings of the 16th International Workshop on Quality of Service (enschede, the netherlands, June 2–4, 2008), 229–238.

4. digg aPI; http://digg.com/api/docs/overview

5. Feller. W. An Introduction to Probability Theory and Its Applications, Vol. 1. John Wiley & sons, Inc., new york, 1968.

6. Gill, P., arlitt, M., li, Z., and Mahanti, a. youtube traffic characterization: a view from the edge. In Proceedings of the Seventh ACM SIGCOMM Conference on Internet Measurement (san diego, oct. 24–26). aCM Press, new york, 2007, 15–28.

7. lerman, K. social information processing in news aggregation. IEEE Internet Computing (Special Issue on Social Search) 11, 6 (nov. 2007), 16–28.

8. richardson, M., dominowska, e., and ragno, r. Predicting clicks: estimating the click-through rate for new ads. In Proceedings of the 16th International Conference on the World Wide Web (banff, alberta, Canada, May 8–12). aCM Press, new york, 2007, 521–530.

9. Wu, F. and Huberman, b.a. novelty and collective attention. Proceedings of the National Academy of Sciences 104, 45 (nov. 2007).

10. youtube aPI; http://code.google.com/apis/youtube/ overview.html

 

Gabor Szabo ( [e-mail]) is a research scientist in the social Computing lab at Hewlett-Packard labs, Palo alto, Ca.

Bernardo A. huberman ( [e-mail]) is an HP senior Fellow and director of the social Computing lab at Hewlett-Packard labs, Palo alto, Ca.