Industry Media and Advertising, Hi-Tech
Specialization Or Business Function Media and Advertising (Clickrate Optimization), R&D (A/B Testing), Engineering and Design (Process Design/Support)
Technical Function Data Visualization (Dashboards & Scorecards, Statistical Graphics), Analytics (Predictive Modeling, Trend Analysis, Forecasting, Real-time Analytics, Machine Learning, Data Preparation, Simulation, Time Series Analysis, Deep Learning)
Technology & Tools Big Data and Cloud (Apache Spark, Google Cloud Platform), Data Analysis and AI Tools (Tableau, Pandas)
Taboola predicts RPM of ad impressions via a Deep Neural Network (NN). These predictions are used for ranking ad items in an auction. The predicted RPM which is the output of the NN contains bias, and requires calibration to the actual mean RPM achievable per each set of environmental parameters.
A PID controller performs the task of calibrating the output of the NN. Since the clicks are sparse, the noise in the measured actual RPM is significant and requires filtering. Moreover, each supply source has its own unique attributes - so a repeatable method of tuning controller parameters at scale to varying supply is required.
What we are looking for:
A consultation and hands-on practice - for tuning a PID controller for correcting bias in RPM predictions, and smoothing PV RPM measurement noise of said controller.
Our goal is to create a scalable PID controller which includes (a) an RPM measurement that is inherently resistant to noise, applicable to a wide range of supply and demand characteristics: input data rates, click through rates, CPC, other demand competing in the auction - i.e. highly varied RPM ranges and measurement window durations; and (b) the controller parameters should be tunable to correct for RPM bias over many supply placements and dynamic conditions.
A method should be devised for tuning the parameters of the controller and the noise filter for a wide range of such conditions.
The project will comprise an expert consultation for the tuning process, plus joint practical work with the team that is developing the controller. The deliverable will comprise several tuned instances of the controller, on choice placements (TBD) - and a simulation on real data - which will serve to prove the proposed concept for tuning the controller and for cleaning the noise. The result will be a well documented tuning procedure and noise filtering method to allow automation of the procedure by the team towards large-scale deployment.
The project has strong potential for follow-on projects and scope enlargement based on the success of the initial cooperation.
Your proposal should emphasize your background in control theory and practical implementations of control systems in software with very large scale data, as well as your background in Ad-technology and programmatic advertising or display advertising;
You should also outline which algorithms and approach you intend to use, and how you propose to measure the success of tuning and noise filtering.
We will share further information with the expert and jointly refine the approach and scope in detail with the relevant expert.
Note: the timezone of the customer is GMT+2, so there would be an advantage in communicating if near that timezone. But this is not a mandatory requirement.