Expedia Group Tech — Data
Enhancing Expedia Group’s Intraday Marketing Decisions
Using Marketing Attributed Transaction Events tool to improve strategy at Expedia Group
What is MATE?
In 2023 we released the Marketing Attributed Transaction Events (MATE) data product. It is the first cross-domain data product built using newly built data domains, the first single source for measuring marketing commercial performance, and is the first native signal that provides intraday marketing performance insights.
The focus of this article is to share some of the key lessons that were learnt along the way and to share a sneak preview into some of the capabilities MATE now allows us to unlock. To begin the journey, a little context is needed from back in 2019.
Why it was needed
In 2019 Expedia Group™ operated quite differently. There wasn’t a single source of truth for measuring or comparing the commercial performance of Expedia Group’s marketing mix for its various operating units. Each unit had the autonomy to define their marketing strategies and how success was measured.
When rolled up into a single picture, there was inconsistency, extensive manual intervention to try to align, and little opportunity for repeatability given a forever growing and changing business. It was truly square pegs into round holes. This led to multiple versions of the truth and maintenance overheads applied in several places.
In the same year, Expedia Group began its journey to become a single group; one team with one strategy and one platform, built on single definitions. No domain was left untouched, and in the face of increasing difficulty to operate as one group, due to the challenges of disparate sources, a clear vision to provide a single source of truth to power Expedia Group’s billion-dollar marketing machine was born.
What has the reaction been?
From launch in 2023, MATE has already been used;
- To measure VRBO platform migration to ensure marketing operations remained intact,
- To unlock new go-to market strategies for marketing channels through democratizing the Customer Lifetime Value signal, and
- As a source for Data Science and Machine Learning teams to find new ways to measure marketing channel performance. The ultimate measure of success for MATE is the deprecation of the previous incumbent products that were used prior to its introduction.
Architecting the solution for the future
Any person who has worked in or used data will know the industry has gone through multiple models over the last decade and a half. The industry has for a long time used Kimball Data Modelling, then Data Mesh was released as white paper in 2019. With an emphasis on thinner ETL layers for data truer to source and less data stitching, Data Mesh allows faster to-market capabilities that can scale and change much faster.
However, the business needs of cross-stitch domains to power key decision making remained, and a gap to leverage the new capabilities through a flexible architecture was identified, enter MATE.
For cross domain table architecture, single value field design can bring issues around scaling, maintenance, SLAs, data discoverability and eventually query performance over its lifetime.
Why does this happen?
Query engines must often go to a more memory intensive I/O for tables with over 500 fields. Though this is partially solved by using columnar file formats, the need to store metadata on each field means it’s intensive to render and maintain as the files grow over time. Additionally, this table design fails to recognize contextually similar attributes and metrics viewed through different lenses, like time-series which results in multiple fields, involving suffix/prefixes, which leads to ballooning the table over time. Marketing operations follow this pattern and therefore single-field design does not offer the ability to scale, add, or remove very easily and will lead to deprecated fields over time impacting usability of the table.
Data discoverability user experience is also poor when trying to search and collate all domain specific attributes and/or metrics due to table evolution, e.g. when appending new fields to a table. This forces the query UI application to group them which complicates the understanding of the data due to constant need to refer to documentation for field context.
So how did we architect to avoid these pitfalls?
The answer is by using more complex datatypes that Hive table format and file formats supports. When thinking about domains such as ‘customer’, there are a number of attributes and metrics associated with it which are often realized into single value fields. However, when talking cross-domain stitching, these can conflict when fields are similarly named with same context, like ‘country’ which would apply to the customer and the hotel they booked. So, where you consider say a property domain instead of a field for each attribute/metric such as ‘name’ and ‘country’ these become elements of single field instead of multiple single value fields, and can be stored as same context alongside a customer’s field which will also have name and country elements.
For table evolution, the field can easily scale by adding additional elements which will not cause as much maintenance and processing issues as single field design over its lifetime as highlighted previously. Data discoverability user experience is much improved as many query UI will present all elements of the domain with values together in query results. Additionally, users will immediately understand the lineage, and therefore context, of the elements they are viewing.
When looking at marketing models though, they can have different results and share similar contextual attributes and metrics. In single field design these are realised as individual fields, which will suffer from field explosion. The challenge for data design is how these results can be consistently viewed through different lenses. The complex datatype map, a key-value pair where the value is primitive or complex, helps solve for this as it allows the flexibility to use the key to represent each marketing model/lense and the associated values. As the value can also be a complex type, the use of the appropriate data type can hold the attributes and metrics commonly shared between the models as a consistent view instead of being single fields.
It also offers easier scalability to add and remove lenses for each model over time, rather than having the maintenance of extending the table in the single field design with a field then later deprecating when falls out of use. Performance is similar between single value field design and complex data type design, where nested up to and including 3 layers.
How we did it
Whilst we overcame multiple hurdles, the challenges can generally be bucketed into 4 groups.
Influencers aren’t just on social media
Look beyond just the senior sponsors to the person who builds the forecast, the trading deck, the go-to data science oracle who can join everything together. These people will make or break your product launch. If you don’t find them they’ll find you, and without understanding why you’re moving from the comfort of single value fields to structs they’ll point out all these changes as problems. Their voice carries weight, no one wants to disrupt BAU so if they aren’t excited or engaged to adopt a product like MATE, you’re fighting an uphill battle. Win enough of them over, and your life just got a bit easier. For MATE, we built a focused beta group from iteration 1 through to full adoption which helped define requirements, grow the product documentation to build a strong onboarding artefact, and championed our product with the remaining user base.
Transparency is key
Moving to new domain data and single definitions meant senior leaders would see new trends. Existing trends are always status quo for stakeholders, but new BAU trends are looked at with greater scrutiny. A critical component to building confidence was the foresight to trend MATE vs the incumbent products week on week for key segments and KPIs, publish them and crucially have the evidence to explain the new trends to stakeholders. Demonstrating a desire to engage frequently with the ability to explain key trend changes of each increment built a far stronger base of confidence from stakeholders in the product instead of leaving it until the end and expecting adoption en masse. Clearly documented test cases with expected and actual results, as well as clearly linked defects, formed the basis of review with the beta groups and for senior leaders simple trending materials were published weekly. Office hours allowed stakeholders to expand upon questions and left with confidence.
Make adoption easy
In larger companies some products have stood the test of time, people are comfortable using them, with the technical changes of moving to a data mesh architecture getting people used to structs and new field names was a challenge, not everyone is excited to move their scripts when its already producing actionable insights.
We overcame this by;
- providing comprehensive onboarding materials including a mapping document of old attribute to new,
- building an FAQs based on the beta testers and feedback from the weekly trending,
- sharing easy to navigate testing and trending documentation with clear summaries,
- holding frequent multi-timezone office hours, and
- presenting often in key all-organisation, all-hands meetings to spread our message.
All of this coupled with building some influencers from the beta groups in each team allowed us to launch with huge success and led to a rapid deprecation of the existing products.
Believe what you’ve built is better
With such a broad stakeholder base, trying to replace long-standing incumbent products and moving people towards a cultural new world of single definitions through Data Mesh, you will not please everyone all the time. This happens even if you engage early and are transparent on differences. Despite this, you must believe that the benefits you’re bringing matter, and be willing to address the latent pain your stakeholders can’t see until it’s too late. Define your key benefits and extol them at every opportunity. MATE’s latency improvements have been a big win for each stakeholder team, now allowing us to hold weekly performance meetings a day earlier in the week.
To recap
MATE has met the challenge of the strategic direction of the company over the last few years being one of transformation, scalability, flexibility and greater speed to market for our customers and partners.
Its success lies in what it enables next, including new go to market strategies for marketing channels through democratizing the Customer Lifetime Value signal and is a source for Data Science and Machine Learning teams to test and learn their hypothesis for new ways to measure marketing channel performance.
Published by
Blake Whatley and Ravneet Mann
