Network-Based Audience Influence Analysis Using Multiple Regression Models

Abstract

The increase of diffusion speed by online media makes information publication and audience shaping in heterogeneous communication network more complex. Thus, a key challenge in network analysis and media-analytics research is to investigate how network structure impacts the size of audience reach. In the present work, we constructed a directed network among media outlets based on data from multiple sources on their interaction, and in which nodes represent media outlets and edges encode diverse interactions between media outlets with associated weights. Structural network properties, such as in-degree, out-degree, weighted degree measures and ranks (PageRank), betweenness centrality and closeness centrality were calculated to measure the positional importance of every node in the network.

To analyses how audience size could be related to these properties of the network Ridge Regression with leave-one-out cross-validation and Ordinary Least Squared (OLS) Multiple Linear Regressions were applied. The results show that the network-based features can convey relevant explanatory power towards predicting audience size, which emphasizes the involvement of centrality and connectivity in medium influence. The Performance of the Regression The regression performance analysis reveals good prediction capacity, 0.72 and 0.74 for Ridge and OLS models' R2, respectively, as well as low error measures which ensure that the proposed framework is robust. Furthermore, influence diagnostics indicate the existence of a small subset of highly influential nodes, thus highlighting the necessity to address network heterogeneity in regression type models. In conclusion, the method introduced in this paper provides a systematic and interpretable tool to study audience influence through network-aware regression model that has applications outside traditional areas of media studies and beyond, including information diffusion analysis and digital communication research.