How the Knot Leverages Taxonomies and ML to Help Enable the Perfect Wedding
In a talk at Taxonomy Boot Camp 2018, Suzanne Carroll, product director, Data Intelligence, XO Group (The Knot), and Julianne Marzulla, associate taxonomist, XO Group (The Knot), covered the basics of machine learning, how taxonomies enhance it, and how to recognize machine learning opportunities.
Taxonomy Boot Camp 2018 is a part of a unique program of five co-located conferences that includes KMWorld 2018, Enterprise Search & Discovery, Text Analytics Forum ’18, and Office 365 Symposium.
The XO Group’s taxonomy team related machine learning techniques that everyone encounters in everyday apps to demonstrate how they’re integrated into products and applications. They shared the major role taxonomies play in The Knot’s personalization and recommendation algorithms and how they adjusted and adapted taxonomies and taxonomy management to enhance their machine learning solutions.
A key element to enable such personalization is preparing the data for machine learning. This involves identifying what data can (and should) be collected to improve the product in the future. Every interaction is a chance to learn and intentionally creating feedback loops. It is important to collect and maintain quality data, they emphasized.
Looking at the taxonomy development behind, particularly the recommendations and the automated venue classification, they said they use machine learning to suggest auto-classified tags for their venues and then based on those tags recommend those venues to couples and users, while making sure the taxonomy supports The Knot’s learning endeavors as best as it possibly can.
When they were redesigning their taxonomies, they sought to improve the overall quality of the taxonomy, especially as it relates to clarity and precision in vendor tagging. This in turn improved the precision and recall of their ML recommendations and classification.
For The Knot, it basically boiled down to two to methods: The first was logic-based, and, broadly, they said, they evaluated their taxonomy’s adherence to the well-established principles of mutual exclusivity and collective exhaustiveness, and were looking specifically with an eye for findability and specificity. Under that umbrella, they did three main things:
- Re-grouped terms, in which they changed the fundamental structure of their taxonomy by creating new hierarchies of similar items and dismantling others
- Promoted concepts to a higher level in the taxonomy
- Flattened their hierarchies
According to Carroll and Marzulla:
- Flatter hierarchies are desirable, but they shouldn’t come before specificity
- “Findability” is key
- Utilize data associated with taxonomy concepts (like synonyms) to improve ML capabilities
- Learn from and test with user data and subject matter experts
Many Taxonomy Boot Camp presentations, including this one, have already been made available online at www.taxonomybootcamp.com/2018/Presentations.aspx and others will become available after the presentations are given.
KM World 2019 will be held November 5-7, 2019 at the JW Marriott in Washington, DC, with pre-conference workshops on November 4.