Targeting audience personas with programmatic geographic segments using unsupervised methods

Abstract

The digital advertising industry, where clients advertise to existing and potential customers through digital channels, is going through a rapid transformation. This is necessitated by evolving privacy laws such as General Data Protection Regulation (GDPR) and California Consumer Privacy Act (CCPA), heading towards a cookie-less future by decreasing reliance on Personally identifiable information (PII). While approaches like Data clean rooms, Federated learning have started gaining prominence, the current AdTech space is highly fragmented regarding channels, standards, platforms, inventory, cookieless targeting approaches, and compatibility.

For marketers, it is crucial to have generalized and interoperable but privacy-compliant approaches to reach their audience. While it is safer not to limit to a specific advertising stack, it must not come at a substantial performance cost. When a brand starts an advertising campaign, they usually have objectives and a target audience in mind. While their objectives are aligned with business goals and performance metrics, the desired audience personas are defined by either market research or new/ historically well-performing consumer profiles suited to their products and services. This paper presents ways to convert multi-characteristic personas into geographic targeting without relying on cookie-based data. These geographic segments have broad compatibility across marketing platforms.

We sourced data from 5 privacy-compliant datasets containing ∼9000 variables aggregated at Forward Sortation Area (FSA) level by a leading Canadian data provider. These variables span a wide range of characteristics such as demographic, econometric, lifestyle and media choices, brand affinities, purchasing behaviors, and spending. The persona-related variables are optionally indexed, after which dimensionality reduction techniques, such as Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP), were applied. Unsupervised learning methods, such as KMeans, KMeans++ and Gaussian Mixture Models (GMMs), were then used to build an optimal number of FSA clusters. These clusters were then analyzed to identify those beneficial to targeting, which helps reduce the number of FSAs to target. The final output for marketing consumption is in the form of FSA segments for targeting aligned with our desired profiles. We analyze campaign metrics against different combinations of dimensional reduction and clustering techniques to assess what works well for advertising.