Sequential Uncertainty Quantification with Contextual Tensors for Social Targeting,

Abstract

A common question that an online marketer asks is, given a social network, how one decides which users to target with their limited advertisement resources (such as coupon offers). The key technical task towards answering this question is the estimation of the user-user activation probability that quantifies the influence one user may have on another towards their purchasing decisions. In this paper, we propose a novel targeting strategy with sequential uncertainty quantification via probabilistic tensor regression. The proposed framework is designed to capture the heterogeneity in user preferences, product types, campaign strategies, etc. in the form of contextual tensor. For uncertainty quantification, we derive a closed-form online predictive distribution for the user response score, which is used in a bandit-style sequential decision making on who to receive marketing offers. We empirically confirm that the proposed algorithm achieves significant improvement in influence maximization tasks over benchmarks, which is attributable to its capability of capturing the user-product heterogeneity.