A dynamical measure of algorithmically infused visibility

TL;DR

Analysis of a large-scale, real-time dataset of topics of this work demonstrates that the proposed measure can explain a large share of the variability of the accumulated views of a topic.

Abstract

This work focuses on the nature of visibility in societies where the behaviours of humans and algorithms influence each other - termed algorithmically infused societies. We propose a quantitative measure of visibility, with implications and applications to an array of disciplines including communication studies, political science, marketing, technology design, and social media analytics. The measure captures the basic characteristics of the visibility of a given topic, in algorithm/AI-mediated communication/social media settings. Topics, when trending, are ranked against each other, and the proposed measure combines the following two attributes of a topic: (i) the amount of time a topic spends at different ranks, and (ii) the different ranks the topic attains. The proposed measure incorporates a tunable parameter, termed the discrimination level, whose value determines the relative weights of the two attributes that contribute to visibility. Analysis of a large-scale, real-time dataset of trending topics, from one of the largest social media platforms, demonstrates that the proposed measure can explain a large share of the variability of the accumulated views of a topic.

Figures (5)

• Figure 1: Trending trajectories. In the figure, we plot the trajectories (rank, $R$, versus time trending) of two different topics. One topic attains a higher maximum rank than the other, while the other topic trends for a longer period of time.
• Figure 2: Visibility versus discrimination level. In the figure, we plot visibility, $V(D)$, versus discrimination level, $D$, for two topics whose trajectories are given in Figure (1). The two topics spend different amounts of time trending: the values of their visibilities at $D=0$ indicate these different values (one topic trends for $363$ minutes, the other for $238$ minutes). For low discrimination levels, the topic with the longer trending time has a higher visibility. However, for sufficiently large $D$ ($D \gtrsim 0.37$), the other topic has a visibility that exceeds the first.
• Figure 3: Scatter plot showing $N_{reads}$ versus visibility. The figure contains a scatter plot of $\log_{10}( N_{reads} ))$ against $\log_{10}(V(D))$ for points associated with $23,993$ different topics. The figure also contains the best straight line through the data (red). A value of the discrimination level of $D=0.8$ was arbitrarily adopted, and the best straight line through the data yielded a value of $R^2$ (coefficient of determination) of $R^2 = 0.48$.
• Figure 4: Coefficient of determination versus discrimination level. We determined the best straight line of $\log_{10}(N_{reads})$ against $\log_{10}(V(D))$, for $23,993$ different topics, for a set of different values of the discrimination level, $D$. The plot contains the $R^2$ values of these lines against the corresponding values of $D$. We determined the value of $D$ that maximises $R^2$, and we write these quantities as $D_{max}$ and $R^2_{max}$, respectively.
• Figure :