Author: Raphael Schönenberger, Data Scientist@Adtrac

 

Welcome to the Data Science blog at Adtrac. Since this also marks my very first blog post ever, I want to take the opportunity to say hello and hope you will enjoy it and the blog in general. But let’s jump directly into the topic of this first post.  This is all about audience prediction.

Selling impressions over playouts in DooH

Therefore, imagine a DooH-media owner responsible for numerous screens placed on different locations from train stations to shopping malls. Typically, she is interested in making money out of her inventory by selling it to advertisement agencies. To do so, she sells a number of playouts to ad agencies which can then fill them with their content. But ad agencies do not care about how many times the content is played but rather prefer buying impressions/views for their advertisement. They even prefer buying just a small portion of the population, namely the one identified as the target group for a specific product. In addition, advertisers want these views  guaranteed before the campaign begins.
Unfortunately, selling views seems to be much harder in the real world than it is in the online one. So, our media owner has a tough time designing such a product, which is based on guaranteed views instead of playouts. But because she expects an increase in her clients satisfaction, she pushes for it. This is exactly where audience prediction starts to shine. 

As audience prediction we understand the process of forecasting, which demographic viewers will appear in front of a digital signage at a given time in the future.

Getting to know the Audience

Deciding for a smart product based on impressions raises questions about how it can be actually designed and what prerequisites are necessary.

Obviously, impressions are created by people crossing by the digital signage. Hence, our imaginary media owner has to know how the audience in front of her inventory looks like at any given point in time. Therefore, she must be able to answer questions like, how many people walked past the screens previous Monday morning. 

These are difficult questions to answer. Then, compared to the online world, where numerous tools help categorize each page view and user, it’s incomparably harder to do so in the offline world, say in front of a screen or a poster. This basically comes from the fact that the offline world lacks browser cookies or similar mechanism, which helps identifying and retracing each person passing by. Therefore, getting to know the audience requires some effort and usually the installation of additional equipment.
There are basically two types of systems, which helps overcoming these difficulties:

  • Panel solutions
  • Full-screening solutions

Both types collect helpful insights about the audience from the past until the presence. This data can then be used for forecasting the audience to future points in time.

Panel solutions

These solutions require a pool of people, called the panel. Each member of such a panel agrees on being monitored by a data collector. Moreover, the data collector ensures that the panel represents the population across a variety of features including demographic attributes as well as interests. Additionally to the panel it’s necessary to install beacons at each digital signage of interest. These beacons track the appearance of the panel members. The data collector or anyone else having access to this data is now able to estimate the audience in front of the monitored screens by extrapolating the passing members in relation to their fraction in the population.

One disadvantage of this type of solution is its need for active management in order to keep the panel in sync with the population. Otherwise it’s impossible to make any statement about the audience.
A further drawback might be that such panel solutions are only suitable for locations with many people passing by like a public train station or a shopping mall.
Nevertheless, these solutions are capable of measuring not just demographic features but also interests, which is a huge advantage over full-screening solutions. Moreover, panels also allow calculating the net reach of a digital signage network due to the possibility of retracing their members.

At Adtrac we trust in the solution provided by intervista (click here for further information).

Full-screening solutions

Full-screening solutions, unlike their counterparts, do not require any panel. Instead, they use special devices installed close to the digital signage to monitor the audience in real-time.
Such solutions range from simple counting systems to sophisticated computer vision tools capable of analyzing video feeds. One of these smart solutions is built by Advertima, which is our main data provider (click here for more information).
Advertima provides an artificial intelligence, which is not only capable of measuring how many persons are passing by its sensors, but also extracts features like age/gender as well as the attention towards the digital signage.

Increasing planning reliability

Installing one of the above types of solutions, our media owner now has access to huge data about the audience ranging from past to presence. But this is still not enough. Then, imagine an ad agency testing this new product of our media owner, where the agency wants to buy 100k impressions of females. Unfortunately, she is not capable of telling how long such a campaign needs to run, unless she starts thinking about forecasting the future audiences. This is exactly what audience prediction means.
Such predictive models can range from very elementary solutions like taking the average over the past to very sophisticated approaches including artificial intelligence (AI). For delivering the most accurate predictions for our own customer, we at Adtrac use an AI-model, which consists of multiple deep layers fed by hundreds of thousands of data points about past audiences. This allows us to ensure that we reach the specified campaign targets for impressions, while at the same time we have a precise idea of how the remaining audience for the future looks like.

Conclusion

We have seen that a paradigm switch from playouts towards impressions satisfies a need in DooH-marketing. But setting up such a solution requires media owners to collect data about the audiences in front of their screens. Moreover, in order to offer planning reliability it’s necessary to forecast future audiences, too. This is what we call audience prediction.

In a next blog post I will focus on how we collect and store the audience data and how our machine learning setup looks like. So stay tuned…