Understanding Online Collection Growth Over Time: A Case Study of Pinterest.

A common feature of content discovery applications is the ability of users to save and organize digital items into collections, e.g., images into photo albums or songs into playlists. Understanding how these collections grow over time is important for user retention and collection as well as item recommendation. Here we study factors that affect collection growth over a long period of time. We conduct a large-scale longitudinal analysis of over 2.6 million collections, known as boards, on Pinterest, over a period of three years. We study the inter-event time distribution of pins saved to boards and find that it can be accurately described by a two-component lognormal mixture model. The mixture components reveal that board growth can be characterized by short-term fast-paced sprees of activity, and longer breaks between these sprees. Commonalities emerge in spree behavior; for example, sprees have consistent temporal dynamics and the content saved within the same spree is more focused compared to between sprees. Surprisingly, we observe that boards with longer initial sprees are less likely to have long-term growth. On the other hand, boards with more frequently occurring sprees continue growing for a longer time, and tend to have a larger size. Finally, we synthesize our findings into a series of predictive models which show that initial board evolution is a strong signal for long-term board growth in terms of size and lifespan. Overall, our research has important implications for the design of online content discovery applications and has immediate applications in user modeling and recommendation systems.