A Stochastic optimization framework for personalized location-based mobile advertising

Mobile location-based advertising has seen a lot of progress recently. We study the problem of optimal user targeting and monetization through advertising, from the point of view of the owner of a venue such as a shopping mall, an urban shopping district or an airport. The fundamental distinguishing characteristic of advertising in this setup is that the probability that the user will respond to an ad depends on timeliness of ad projection, hence it is important to target a mobile user with an appropriate ad or offer at the right time. A set of mobile users roam around the venue. Each user is profiled in terms of preferences based on prior visits. The system knows estimated instantaneous locations of users in the venue, e.g. through WiFi access point connectivity. A machine-learning model is used to derive a per-user time-varying probability of response to an ad, which depends on the relevance of the ad (store) to the user profile and on the time-varying physical proximity of the user to the store. Each store has a set of available ads, and each time the user responds to a projected ad, an amount is paid by the store to the venue owner. We use a stochastic-optimization framework based on Lyapunov optimization to address the problem of advertisement selection and allocation for maximizing the long-term average revenue of the venue owner subject to: (i) a constraint on maximum average ad projection rate per user for preventing user saturation, and (ii) a long-term average budget constraint for each store. We derive an algorithm that operates on a time slot basis by solving a simple assignment problem with instantaneous user locations while being agnostic to user mobility statistics. We test our algorithm with a real dataset of check-ins from Foursquare, complemented with data from user questionnaires. Our approach results in substantial improvement in revenue compared to approaches that are location- or relevance-agnostic.