Will the Winner Take All? Competing Influences in Social Networks Under Information Overload

TL;DR

This paper investigates the dynamics of influence competition in social networks, revealing that information overload prevents a single influence from dominating entirely, leading to coexistence of influences rather than a winner-takes-all outcome.

Contribution

It introduces a novel framework considering incomplete network observations and information overload, deriving conditions for influence dominance and coexistence.

Findings

Information overload acts as a boundary preventing total dominance.

Before overload, the 'winner takes all' phenomenon is likely.

After overload, influence shares stabilize based on initial conditions.

Abstract

Influence competition finds its significance in many applications, such as marketing, politics and public events like COVID-19. Existing work tends to believe that the stronger influence will always win and dominate nearly the whole network, i.e., "winner takes all". However, this finding somewhat contradicts with our common sense that many competing products are actually coexistent, e.g., Android vs. iOS. This contradiction naturally raises the question: will the winner take all? To answer this question, we make a comprehensive study into influence competition by identifying two factors frequently overlooked by prior art: (1) the incomplete observation of real diffusion networks; (2) the existence of information overload and its impact on user behaviors. To this end, we attempt to recover possible diffusion links based on user similarities, which are extracted by embedding users into a latent space. Following this, we further derive the condition under which users will be overloaded, and formulate the competing processes where users' behaviors differ before and after information overload. By establishing the explicit expressions of competing dynamics, we disclose that information overload acts as the critical "boundary line", before which the "winner takes all" phenomenon will definitively occur, whereas after information overload the share of influences gradually stabilizes and is jointly affected by their initial spreading conditions, influence powers and the advent of overload. Numerous experiments are conducted to validate our theoretical results where favorable agreement is found. Our work sheds light on the intrinsic driving forces behind real-world dynamics, thus providing useful insights into effective information engineering.