Description
This capstone project seeks to explore the usefulness of interactive visualizations for machine learning. Specifically, we look at online search advertisements, which represent one of the biggest use cases of the recent Big Data boom. We develop a tool aimed to help Ad Operations teams at search engine companies tune parameters for their auction process in an effort to balance tradeoffs between profit, advertiser satisfaction, and user satisfaction. Often times, business decision makers treat machine learning algorithms as a black box, as it is difficult to see what is going on underneath the hood. This project seeks to better inform the user of what is happening by shortening the feedback loop, thus allowing Ad Operations teams to quickly tune models and deploy changes. We accomplish this using a GPU-accelerated machine learning library called BIDMach for click rate prediction, and visualizations developed with a high performance JavaScript library called D3.js. This paper discusses the challenges in using Sparse Factor Analysis for click rate prediction, and how we turned to a Latent Dirichlet Allocation model to achieve better results. We also discuss the system architecture and technological choices for the visualization, and the challenges we faced in connecting it with the backend auction simulation.