Competition for Popularity in Bipartite Networks

We present a dynamical model for rewiring and attachment in bipartite networks in which edges are added between nodes that belong to catalogs that can either be fixed in size or growing in size. The model is motivated by an empirical study of data from the video rental service Netflix, which invites its users to give ratings to the videos available in its catalog. We find that the distribution of the number of ratings given by users and that of the number of ratings received by videos both follow a power law with an exponential cutoff. We also examine the activity patterns of Netflix users and find bursts of intense video-rating activity followed by long periods of inactivity. We derive ordinary differential equations to model the acquisition of edges by the nodes over time and obtain the corresponding time-dependent degree distributions. We then compare our results with the Netflix data and find good agreement. We conclude with a discussion of how catalog models can be used to study systems in which agents are forced to choose, rate, or prioritize their interactions from a very large set of options.

💡 Research Summary

The paper introduces a dynamic model for the evolution of bipartite networks in which one set of nodes represents agents (e.g., Netflix users) and the other set represents items (e.g., movies). Unlike static network analyses, the authors explicitly model how edges—ratings in the Netflix context—are created, rewired, and attached over time. Two distinct catalog scenarios are considered: (i) a fixed‑size catalog where the number of items remains constant, and (ii) a growing catalog where new items are continuously added. In both cases, edge formation is decomposed into a “rewiring” process (existing users rating additional items or existing items receiving new ratings) and an “attachment” process (new users entering the system or new items receiving their first rating).

Mathematically, the evolution of the degree distribution for users, (P_u(k,t)), and for items, (P_v(k,t)), is captured by a pair of ordinary differential equations (ODEs). The rate at which a node of degree (k) gains an additional edge consists of a preferential‑attachment term proportional to (k) (the classic “rich‑get‑richer” mechanism) and a random‑attachment term that is independent of degree. To reproduce the empirically observed cutoff in the tail of the distribution, the preferential term is multiplied by an exponential damping factor (e^{-\beta k}). Solving the ODEs yields time‑dependent degree distributions of the form

\