Utilizing a clean room for privacy-safe data ingestion and audience building

Challenge:

A leading financial products company needed an identity-enabled platform, with a fair-lending-compliant third-party data and modeling strategy for building look-alike audiences to bolster its marketing efforts.

To help the client develop these prospecting audiences, Merkle engaged Merkury, its proprietary, privacy-safe identity resolution and data platform. Merkury’s clean room combines disparate digital and terrestrial identities into a common, interchangeable, secure and anonymous, person-based Merkury ID. Merkle ingested the client’s customer and survey data into the clean room environment, with the addition of Merkury’s third-party data, after passing the data sets through the identity resolution process. After rigorous fair-lending scrutiny of all data sources, applicable third-party data elements were trained using machine learning algorithms to develop high propensity look-alike audiences.

Approach:

Merkle rebuilt the client’s second-party data audience within Merkury, with the help of a research partner (data collected from surveys). All the data was ingested into a clean room environment, and Merkle was able to build look-alike models by combining the client’s data, Merkury third-party data, and acquired second-party data from the client’s research partner, using a gradient-boosting algorithm. Since these data sets were only accessible through the clean room, all data points were anonymized for privacy-safe matching and analysis.

The finalized audience was built at a person-ID level and scaled to meet the client’s in-market reach goals. Merkle then engaged parent company dentsu’s media agencies for test setup and help to build an in-market activation plan. The testing plans included measuring the audiences against holdout and random samples for conversion and brand lift. The client audiences were activated on Verizon, DV360, and other integrated Merkury connections. Lastly, Merkle performed an ad-hoc analysis and reporting for model approval and validation, and finally conducted closed-loop reporting to present in-market performance.

Results

✓

Two separate audiences were developed for key client segments.

110%

The Merkury-built look-alike audiences delivered a 110 percent lift over the unexposed groups.

127 Million

unique impressions were processed, exposing six million unique individuals.

70%

Machine learning boosted the modeled audiences to ~70 percent AUC (area-under-curve), despite stringent fair lending restriction on available attribute classes.