Presented By O'Reilly and Cloudera
Make Data Work
March 13–14, 2017: Training
March 14–16, 2017: Tutorials & Conference
San Jose, CA

A contextual real-time bidding engine for search engine marketing

Mahesh Goud T (Ticketmaster)
4:20pm5:00pm Thursday, March 16, 2017
Business case studies, Strata Business Summit
Location: 210 D/H Level: Intermediate
Secondary topics:  Data Platform, ecommerce, Streaming, Text
Average rating: **...
(2.00, 1 rating)

Who is this presentation for?

  • Data scientists and big data engineers

Prerequisite knowledge

  • Familiarity with real-time pipelines (Storm, Kafka, HBase, and Spring Boot)
  • A basic understanding of online machine-learning algorithms

What you'll learn

  • Learn how Ticketmaster built a real-time contextual bandit platform for optimizing keyword bids


Ticketmaster has built a real-time online learning pipeline for optimizing keyword bids on Google AdWords both for effective spend and scalability. (AdWords-based marketing is the key revenue driver for Ticketmaster’s SEM team.) The previous process of manual bid adjustment for high-performing keywords was neither effective nor efficient, and Google’s automated bid management software is limited in terms of adopting aggressive bid strategies and enabling customization for meeting different business requirements, such as optimizing marketing for primary or resale seats or other metrics across life span of a live event.

Mahesh Goud offers an overview of Ticketmaster’s technology stack: The streaming pipeline assists in making real-time bid decisions across different keywords using contextual bandit algorithms; the streaming pipeline is built with Storm, with Kafka and HBase as persistence layers; the ELK stack is primarily used for real-time visualizations, insights, and monitoring; a Hadoop minicluster is used to run end-to-end automated integration tests for testing the distributed infrastructure; and continuous integration is done using Jenkins. Mahesh Goud then walks you through the current architecture in which messages are streamed through various modules in the following order:

  1. Various fetchers capture keyword, campaign, and event-related contexts.
  2. Aggregation logic maps multigranular contexts to keyword-specific contexts.
  3. Business rules are applied on the keyword-specific contexts. This module determines whether to turn off or activate bidding on the respective keywords.
  4. Contextual bandit predictions are run in parallel across different keywords to determine bids.
  5. Reward signal for previous bid decisions made by the engine is captured in real time.
  6. The learning module builds new models from the keyword context, bid decision, and reward signal and updates the prediction service with the new model, based on the model performance.

Mahesh Goud concludes by sharing some lessons learned while building the system, a comparison of results from Ticketmaster’s learning platform and Google’s Double Click Manager (an automated bid manager), and a demo of an interesting real-time visualization.

Photo of Mahesh Goud T

Mahesh Goud T


Mahesh Goud is a data scientist on Ticketmaster’s Data Science team, where he focuses on the automation and optimization of its paid customer acquisition systems and works on quantifying and modeling various data sources used by the Paid Acquisition Optimization Engine to increase the efficacy of marketing spend. He has helped develop and test the platform since its conception. Previously, he was a software engineer at Citigroup, where he was involved in the development of a real-time stock pricing engine. Mahesh holds a master’s degree in computer science specializing in data science from the University of Southern California and a bachelor’s degree with honors in computer science specializing in computer vision from the International Institute of Information Techonology, Hyderabad.