Machine Learning in Advertising Technology

“To Bid or Not To Bid: Machine Learning in Ad Tech” “Practical machine learning at scale for real-time video distribution” Presented by: Justin Fortier, Principal Data Scientist, ViralGains Outline for Presentation (subject to minor changes): I) What is AdTech? II) Who is ViralGains? III) Examples of the most common business problems we are trying to solve IV) Data used as model inputs V) ML algorithms considered / used VI) Productionizing our model to train daily, at scale VII) Business impact (e.g. Conversion; Cost per View savings) VIII) Next Steps IX) Q&A Abstract: If the digital advertising industry is an $83 billion ocean liner, advertising technology (“Ad Tech”) is the $20 billion engine that ensures that its precious cargo reaches its proper destination in the most efficient, most profitable way possible. Since it was founded in 2011, ViralGains has been a rapidly rising player in the ad tech space, serving 80 of the Fortune 500, and we specialize in video ad distribution across all device types in the United States (expanding soon). We are a technology platform developed and scaled by software engineers, and our decisions are powered by machine learning and artificial intelligence. Today, Justin Fortier, Principal Data Scientist and Michael Lubavin, Lead Software Engineer, will tell you more about how these two functions have been used to optimize performance for ViralGains’ clients across multiple industries (e.g. Auto; Consumer Goods; Finance; Retail; Pharma) for the last several years. ViralGains’ clients include both advertising agencies and enterprise brands, and they depend on us to find good opportunities to distribute their video advertisements online, win bids for ad space in a Real-Time Bidding (RTB) auction, and deliver an engagement experience (e.g. video; survey; product selector) that their target audience will feel compelled to engage with – all in milliseconds. This requires building robust machine learning models, re-training them on a frequent basis, and leveraging them to optimize business outcomes for our clients. This has to be done more than 2 billion times every day, so building a smart, scalable big data infrastructure is critical. To do this, ViralGains stores approximately 275 Terabytes of data per year about every interaction a user has with our platform. These data include a) demographic data about the user who’s just opened a browser, b) the various videos we could choose to show this user, c) the environment (e.g. device; geography; time of day), d) the domain where the ad space is being sold. Four primary decisions must be made lightning fast, each time a bid request comes in from an exchange (immediately after a user opens a browser): 1) Should we bid on this placement? 2) Which video ads’ targeting criteria match the placement? 3) Of those, which would be the most effective to show to this particular consumer? 4) How much are we willing to pay for the placement? To answer these questions, we use machine learning. To answer them fast, and at scale, we use software engineering. For machine learning, we train multiple different classification algorithms (e.g. Random Forest; Logistic Regression; XGBoost) until we find the one that works best in each particular situation (depending on which target variable we are trying to optimize, and how we measure success – e.g. k-fold cross-validation, Area Under the Curve, etc.). We also develop deep learning (e.g. Artificial Neural Networks) and artificial intelligence (e.g. Reinforcement Learning / Q Learning) algorithms, and we optimize the hyperparameters in each and use regularization to avoid overfitting. Now, Michael Lubavin is going to tell us about how we make these algorithms work fast, at scale, using software engineering. According to Lubavin, “It all starts with the data. While our Real Time Bidding server is reviewing millions of bid requests coming in from ad exchanges per minute, and bidding on a portion of those, winning on a portion of the bids, and finally some of the ads playing long enough in a consumer's browser or app to count as a "view", we are recoding millions of these real time events and storing them in structured JSON data files in Amazon S3. We then have a daily EMR job, which is Amazon's managed Apache Spark service, that goes through all the events for the past day, cleans the data and combines the related events into a single row for each bid that we made. After this cleaned data set is generated for the last day, we run another EMR job that uses it to actually train (and cross-validate) our machine learning models. Our system uses Amazon's apis to spin up EMR clusters on demand, and we only pay for the compute hours we actually use to do the data joining and model training.” How do we know if our machine learning / artificial intelligence algorithms are working? We measure success in two primary ways: A) From our clients’ perspective, are we exceeding their expectations in terms of a performance/price ratio on their desired KPI’s (e.g. engagement; completed views; clicks; online purchases)? For their budget, are we delivering them more clicks per engagement, etc., than they were getting without using our platform? B) Internally, by bidding smarter, are we delivering more engaged viewers – at a cheaper cost per view – with machine learning than we could do without it? We measure both metrics every day for every campaign we are supporting, and pivot quickly, as necessary, to optimize performance. This year, we will continue to focus on our clients, first, and we will continue to build and productionize models to drive their business performance higher than ever before.

Accomplished artificial intelligence, machine learning, and data science executive with multi-industry experience developing and communicating model-based profit-driving recommendations to executives and boards of directors. Recognized for having led Subway to more than 20% annual sales growth, having driven Staples to the highest customer satisfaction scores in its history, and most recently for having built dozens of machine learning algorithms for Constant Contact, Thermo Fisher Scientific, and ViralGains -- producing millions of dollars of incremental profit. Also known for having built world-class data science teams, and as a highly sought-after expert speaker at data science conferences.