Kinesso’s marketing intelligence engine introduces a new tool AdShift etto diversify the audiences targeted by its advertising model.

Berkeley, CA, November 2020 — Kinesso announces its new technology to improve media and marketing performance. With AdShift, we aim to understand and expand advertisement audiences to optimize Kinesso’s targeting strategies. AdShift uses data analysis and metrics to quantify differences in distributions to account for bias in Kinesso’s current reached audiences in their advertisement. With this information on bias and marginalized groups that are not showing up in their reached audiences and originally accounted for in their target audience, Kinesso is able to make modeling decisions to incorporate such groups. 

Identifying inherent bias within reached target audience datasets will provide Kinesso with the tools and opportunity to reach audiences that currently fall beyond their targeting algorithm. This increase in diversity of target audience will not only increase revenue for the business, but also will benefit society and individuals as a whole. Currently, the algorithms used to award advertising space to the highest bidder are limited by brand marketing budgets, which leads to brands prioritizing cheaper ad space for more easily reachable audiences over more expensive ads for less represented people groups.  This, combined with assumptions in the data used as a foundation for targeting models, creates a negative feedback loop that perpetuates the issue of inequality.  In other words, layers of decisions that allocate impressions are subject to bias.  This is bad for data science, bad for business, and bad for the world. This project will debias the existing model and diversify audiences in the digital advertising ecosystem.  

Unfortunately, only about 1% of the audience that Kinesso and most other advertisement models actually target logs an impression. Due to the nature of the feedback loop, such models are more likely to prioritize the population behind that 1% of advertisements, narrowing the scope of the targeting algorithm more and more over time. This illustrates the large opportunity for identifying features that perpetuate bias and introducing functionality for explicitly trying to reach historically underrepresented people groups.

Github Repository: