Fig. 2.

Users are heterogeneous, mostly specialized, and predictable. (A) We calculate the frequency with which each user downloads files from each content type. We hierarchically cluster users according to these frequencies and identify 17 user profiles (see SI Appendix for alternative partitions of the users into groups, which support the conclusions of the manuscript). (B) For each group, we depict the average download frequencies, which provide a stylized profile of the users in the group. We label each profile according to the most prevalent content types in the profile. For instance, users with a Music profile download, on average, Small files (4% of the times), Music (70%), TV Shows (11%), Movies Low Definition (5%), Movies Standard Definition (3%), Movies High Definition (4%), and Large (2%). Users are often highly specialized in few content types. Indeed, for 8 of the 17 user profiles, one content type alone accounts for more than 50% of the downloads, and for 10 of the 17 two content types account for more than 70% of the downloads. We classify as generalists the users that download contents proportionally to their availability and as specialists the users that focus primarily on one or two content types. (C) The effective number of contents E is indicative of how the downloads of a user or a group of users are concentrated in a small number of content types (SI Appendix). We plot the effective number of contents as a function of the number of observed downloads, for specialists (red), generalists (blue), and a hypothetical average users that download files randomly chosen from all observed downloads (black; the gray region corresponds to the 95% confidence interval). (D) To evaluate the potential implications for privacy of user specialization, we use a simple model (SI Appendix) to infer the profile of users from their downloads alone. We find that specialists can be profiled quite easily with this simple model. Indeed, after having only five downloads we can correctly identify the profile of more than 50% of them. After 100 downloads, our accuracy goes up to 75%. In contrast, generalist users are more difficult to profile; around 50 downloads are necessary to achieve 50% accuracy. For comparison, random guessing of the user profile yields an accuracy of 6% (null model 2) and assigning all users the most frequent profile (Movies low) has 22% accuracy (null model 1).