Senior Data Scientist, CtrlShift
Campaign Parameters Recommendation
Traders are overloaded with the menial tasks involved in setting up and running campaigns. This hampers their efficiency, potentially impacting the bottom line posing a possible speed bump to the business growth of the organization.
The Lab team were tasked with reviewing this problem statement as a potential product feature for The Hub. To help address the issue of scalability, The Lab wanted to introduce more support into the platform by letting machines handle some of the more basic and repetitive tasks associated with day-to-day trading operations.
The campaign parameters recommendation aims to provide parameters settings at the campaign planning stage. This tool uses numerous attributes of the Insertion Order and Media Plan. It also looks at the campaign settings.
The tool can then recommend a wide range of campaign parameters such as day of the week, time of day, ISP and preferred inventory source for clients and traders before they actually set up the campaign. One of the facets of the experiment is to be able to build a model that is extensible in its outputs and possibly in its inputs so as the feature evolves it is adapting to the changes in attributes
The Lessons Learnt:
We encountered a few challenges in building this tool. One challenge was due to the large data size and limited processing power. Changing how the machine-learning model approached incremental learning and digested data helped solved this. The other challenge was that the optimization algorithm initially took one minute to generate the result. Based on the requirements of The Hub’s product feature for campaign creation, 60 seconds was deemed too much latency for the API that provided the response or output of this model.
As such we had to work on improving the API response vastly so we can ensure a seamless user experience for the trader using The Hub to create a campaign. By refactoring the code to a low latency response, we are now able to reduce the processing time to 10 seconds.
This problem is more than a typical machine-learning problem from which you learn the relationships between the input and output. It also requires the optimization algorithm to make use of the machine-learning model and bring in real values to the traders.
We are now at Stage 1 of development for the parameters recommendation and will pre-generate recommendations for commonly used inputs and display on The Hub for users to leverage.
We have also compared the performance with a baseline campaign from Q4 2018. The model has managed to produce better performance in terms of CPM, CPC and CTR. We were initially aiming for higher CTR rate but managed to achieve better CPM and CPC rates as well.
As the amount of data that is required for pre-processing increases significantly, we have encountered very long processing time and out-of-memory issues. Thus we are optimizing the code to make it run smoothly and quickly to cater our daily refresh needs.
We are trying to include as many geographical territories as possible and put it into production, we currently include 234 territories. In tandem, work on trying to minimize the time it takes to generate parameters continues, to achieve a more efficient and more responsive API. We are also exploring the Google Cloud Platform to check if it can boost our efficiency.
Liked what you just read? Schedule a demo to find out more about The Hub, CtrlShift’s one-of-a-kind enterprise programmatic advertising platform for unified media planning, activation and insights. It aggregates platforms (social, search, RTB, native) to streamline digital media management and help brands and agencies achieve more.