Content Promotion for Online Content Platforms with Network Diffusion Effect

Abstract

Online content platform aims to maximize content adoption which is simultaneously driven by the platform promotion and network diffusion. We study the candidate generation and promotion (CGP) problem for online content with the diffusion effect in this paper. Motivated by real-world datasets from the industry partner, we propose a novel diffusion model for online content that captures time variant diffusion effect and the existence of promoting probability. Based on the diffusion model, we formulate the CGP problem as a two-stage optimization problem. We show that the problem is NP-hard and can be seen as a submodular maximization problem. By utilizing the two-stage structure and combining the greedy idea from submodularity, we propose an $(1-\frac{1}{e})(1-\epsilon)$-approximation algorithm. We also consider the online version of this problem where the platform is simultaneously estimating the diffusion parameters and deploying the content promotion decisions. We propose a state-dependent non-anticipatory online policy that achieves $\tilde{\mathcal{O}}(\sqrt{T})$ of $\alpha$-Cumulative Myopic Regret ($\alpha$-CMRegret). Numerical results from a real-world dataset show that the diffusion process for online content deviates drastically from the traditional diffusion models and our proposed model provides much better fit. We also demonstrate the efficiency of our offline and online algorithms with the dataset.