A Physicist in a Sea of Data
Michael Prim has been the Data Science Team Leader at Adello since the beginning of the year. His day-to-day work shows remarkable parallels to his previous work in particle physics.
From Physics to Advertising
Michael Prim has been the Data Science Team Leader at Adello since the beginning of the year. Prior to that, he worked for a year as a data scientist at Adello. Advertising, sales, and image itself are not, in principle, areas that would spark Prim’s main interest. But that’s not necessary, as he explains: «Statistics and the procedures and methods used in particle physics are universally applicable. This form of data processing can be applied to all industries. As a data scientist at Adello, I only have to deal with advertising to the extent that I need to define the algorithms and prepare obvious knowledge for the machines.»
To explore topics in the industry, Prim joined the software company Blue Yonder five years ago. Big Data in the context of programmatic advertising has been his area of expertise for quite some time. Moreover, the two fields are closely related. However, there is one major difference between physics and data processing. And it is precisely this difference that drew Prim away from physics and into advertising. In physics, the typical question is: «Does the given theory still make sense based on the data we’ve analyzed so far, or can I disprove the theory?» So, while the same research and analytical steps are taken—with data analysis focused on evaluating interesting events based on the probability that the desired event will occur—«essentially a prediction of individual probabilities,» as Prim summarizes. In advertising, measurement can determine whether, for example, ad impressions actually correspond to the desired outcome, whether the ad is actually being displayed, and whether sales figures or brand image actually improve to the desired extent after the campaign. In short: An effect is immediately visible. This is what fascinates Prim about programmatic advertising: «Continuous learning from data analysis drives constant change. In physics, the laws of nature do not change. That means I can only use the data to test the theory and, if necessary, narrow it down further. Not a very big change.»
Michael Prim has always had this curiosity and desire for a challenge. As a child, he wanted to be an astronaut—«but that’s what all boys want,» he says. At the same time, Prim was exposed to computers at a very early age. He was already sitting at a computer regularly by the time he was about six years old, since his father had gotten into IT early on. «I’ve always found it interesting. After all, computer science isn’t just mechanical data processing; to a certain extent, it also involves a creative process. You build systems and develop algorithms—there’s a creative process behind it all.» But while studying physics, it became clear to Prim that he didn’t want to stay in research. «In research, there was too little change for my taste,» he explains. It’s completely different in programmatic advertising. «Users today show very different interests than they did yesterday, for example. So theories can change very quickly, and the effects can be measured immediately. I like that.» And so, a few years ago, the job posting for a data scientist immediately fascinated him—the ability to use algorithms to calculate the probability of an event occurring and then measure that probability in real time. «Of course, I can’t predict what a specific target group will be looking for in six months. But I can calculate with great precision how likely it is that a specific user—given their background and the campaign I’m showing them—will perform a specific action at a specific point in time. From this, I can make what’s called an expectation-based forecast, which makes it very easy to run the numbers and make the right decision.» Of course, the models have to be continuously adjusted if events don’t unfold as predicted. «But that’s exactly what makes my job so exciting,» Prim says enthusiastically, «things are constantly changing.»
On Data and Elementary Particles
Prim's data science team is responsible for developing and refining algorithms that analyze billions of ad impressions every day and display them on a patented real-time analytics platform. For each impression, the algorithms examine thousands of attributes to make real-time predictions about where the best conversion will occur. Every step of the customer journey is measured and fed back into the system. This results in a self-optimizing system that learns from every ad view, every interaction, every website visit, and every purchase. These systems have since led to various patents and award-winning mobile advertising products from Adello, which are managed by fifty employees around the globe. Just a few weeks ago, the «Silicon Valley Technology Innovation & Entrepreneurship Forum» (SVIEF) presented the Swiss mobile technology company in Santa Clara with the 2015 Top 30 Innovation Award. Adello received the award for its proprietary technology, which boosts the efficiency of mobile ads by using machine learning to identify the right customers for the highest ROI. As a result, users see fewer ads, and advertisers gain more customers while maintaining the same advertising budget.
But what exactly is behind data processing using algorithms from particle physics? Michael Prim makes the answer easy to understand—the approach to handling large amounts of data, borrowed from physics, is almost fascinating. «Particle physics deals with the question of what holds the world together at its core—the question of the fundamental forces of nature,» Prim begins. «To answer the questions in these fields of research, scientists build very large detectors, such as those found at CERN in Geneva. These machines are the size of houses and produce data rates per second that, if written to CDs over the course of a week, would stack up to the moon.» Adello also generates these volumes of data, albeit on a smaller scale. To be able to process and analyze these data volumes nonetheless, they must be filtered. And this must be done very quickly. That covers the context of individual particles. But this data also needs to be evaluated instantly. «So, for example, if I have 1,000 events per second, I have to discard 999 of them immediately; I can only save one of them for analysis,» says Prim. More precisely, the algorithms save approximately 1,000 out of 40 million events per second. «The decision, then, lies in which data I keep or discard,» he explains. Anything that is deleted is irretrievably lost. This requires the algorithms to evaluate the data very quickly and then either store or delete it. This is only possible if it is specified very precisely which data is valuable or unknown. After all, only this information is truly worthwhile for further research. On the other hand, the algorithms should capture information that has long been known. «Data that we know, that we have understood,» as Prim explains. The algorithms are capable of making this decision if they are programmed with the appropriate triggers. To illustrate the scale of data volumes in particle physics, Prim gives an example: «At CERN, a dozen data centers are linked together just to store the data and then process it further. That’s still far beyond what a single computer could handle.» The advantage here, however, is that the entire dataset is stored and can be searched using many interconnected computers to identify the 1,000 events of interest that are specifically related to the process under investigation. For his doctoral thesis, for example, Prim filtered 1,000 such events that had accumulated over more than 10 years of measurement—even though billions of events had been recorded. The analysis ultimately takes place on this relatively small subset of data. «Perhaps we’ll then find out whether certain theories are truly important—we can use the analysis to support an existing theory or, in the rarest of cases, even falsify it,» Prim summarizes. To get there, data processing takes center stage. Data processing is primarily about establishing a scale that does not yet exist in industry. In industry, however, the volume of data is a factor of 1,000 smaller than in particle physics. And this is where the data scientist comes in. Until now, business analysts have been sifting through such data sets, but the trend is shifting toward particle physics.
The same is true for Adello in the field of mobile advertising. «We have 10 billion ad impressions every day,» says Prim. These are distributed across 8 billion people, 1 billion smartphones, and several hours of mobile usage per day. «The industry’s traditional methods fall short when it comes to analysis,» explains Prim. Existing programmatic approaches cannot meet the challenge of rapidly and accurately predicting user behavior in conjunction with the enormous volumes of data. This is because these approaches are focused on individual computers, even though distributed data analysis is required to ultimately use multiple separate computers to filter out the individual ad impressions that are of interest to the customer. In his view, therefore, the problem of rapid, instantaneous data reduction and decision-making is best solved using scientifically proven methods from particle physics. However, customer satisfaction and growth demonstrate to Prim on a daily basis that, thanks to its chosen approach, Adello is now successfully holding its own internationally against the industry giants.
In a nutshell:
Michael Prim has been the Data Science Team Leader at Adello since January 2015. Prior to that, he worked for one year as a big data scientist at Adello. Before that, he worked as a Data Analyst at Blue Yonder for over three years, during which time he completed his Ph.D. at the Karlsruhe Institute of Technology. He also earned his master’s degree in physics from the Karlsruhe Institute of Technology.
– October 2004 – January 2010: Bachelor’s degree in physics. Karlsruhe Institute of Technology (KIT)
– November 2009 – December 2012: Data Analyst. Blue Yonder, Karlsruhe
– February 2010 – December 2013: Ph.D. in Physics, Doctor of Natural Sciences. Karlsruhe Institute of Technology (KIT)
– January 2014 – December 2014: Big Data Scientist. Adello, Karlsruhe.
– January 2015–present: Team Lead, Data Science. Adello, Karlsruhe.
Ursina mason
