How Does Coveo ML Learn?

 

This first session of Coveo's three-part machine learning webinar series covers how to set up and optimize ML models effectively. Lipika Brahma, Customer Success Architect, and Jason Mlyniec, Director, Customer Onboarding, review the fundamentals of how Coveo's models learn and adapt.

 

Models like Automatic Relevance Tuning (ART), query suggestions, event and product recommendations, and Dynamic Navigation Experience (DNE) learn from user interactions around four key behaviors including:

• Search events
• Click events
• View events
• Purchase events

 

Each ML model is designed to enhance a specific aspect of the search experience, with the system continuously refining and personalizing search results as more user interactions are recorded. The models analyze connections between what users type into the search field and the content they engage with – this is how the most relevant results are surfaced based on user context.

 

Brahma notes that you don’t have to wait to benefit from Coveo’s out-of-the-box ML models, noting a few best practices for new users that include:

• Enabling machine learning from day one so the system can begin learning from user behavior immediately.
• Consider running an A/B test with 50% of traffic sent to a pipeline with ML enabled and 50% without. This lets you compare metrics and demonstrate ROI to stakeholders.
• Don't worry about waiting for a specific amount of traffic or data before enabling machine learning since there's no downside to turning it on right away – the models keep learning and adapting over time.

Tools To Optimize Your ML Models

 

Once your Coveo machine learning models are up and running, there are four "optimization levers" you can use to optimize ML performance: filter fields, custom context, custom model parameters, and default configurations.

 

Here’s a breakdown of each:

• Filter Fields - Filter fields allow you to create "sub-models" that learn from specific subsets of your data like language, search hub, or custom context. They’re a “check” that makes sure the right content is delivered to the right audience.
• Custom Context – Custom context allows you to pass additional information about a user or their interaction to the model (e.g., department, location, past purchases). The ML model uses this information to boost the ranking score of content that’s most appropriate to the user (e.g., an employee versus a customer) in the context of their search.
• Custom Model Parameters – This feature uses query pipeline parameters like thesaurus, ranking expressions, and ranking weights to fine-tune the behavior of your ML models. These parameters give you granular control over settings like the number of results boosted by ART or the minimum number of clicks needed for the model to learn.
• Default Configurations – Coveo's out-of-the box default ML configurations like "data period" and "refresh frequency" can have a big impact on performance. You can adjust these configurations based on your specific use case and traffic patterns.

 

Coveo’s optimization levers allow you to feed more more context to the ML models so that search results can be refined, personalized, and made as relevant as possible. There's no such thing as too much context — the more you provide, the better the ML models can learn and adapt.

 

Filter Fields and ML Sub-models

 

Brahma provides a comprehensive overview of yet another powerful ML optimization tool — filter fields. They allow users to create ML sub-models that learn from specific subsets of your data. Using dimensions like language, search hub, or tab, further refines how the ML model delivers content based on audience context and behavior.

 

For example, let's say you have a multilingual site with both a customer community and a technical knowledge base. You can use filter fields to create separate ML sub-models for each language and each content type. This way, when a user searches in the customer community in French, they'll get results that are relevant to that specific context.

 

Coveo's ML system creates a sub-model for every unique combination of filter field values. So, if you have two search hubs, four tabs, and one language, the system will create eight sub-models (2 x 4 x 1 = 8). Each sub-model learns from the events and interactions within its specific slice of the data.

 

Understanding Custom Context

 

Custom context is another powerful ML refinement tool that allows you to pass additional information about a user or their interaction to the ML model. This could include details like:

• The user's department or location
• The products they've purchased in the past
• The pages they've visited on your site

 

It’s a level of context that enables the ML model to personalize search results even more. For example, the model can boost the ranking of related accessories or complimentary items based on a shopper's past purchase history.

 

Customizing Your ML Model for Personalization

 

Throughout this webinar, Coveo's experts showcase the many tools and techniques you can use to optimize your Coveo ML models. Each lever is designed to help you fine-tune your ML models for more personalized, and highly relevant search experiences.

 

The key, as both speakers emphasize, is to leverage all these tools together. By segmenting your data with filter fields, providing rich user context, adjusting model parameters, and optimizing your default configurations, you can create an ML model that is perfectly tailored to your unique business needs and user behaviors.

 

Mlyniec drives this point home with a powerful example that demonstrates how to create a search experience that feels custom-made for each individual user. (You'll have to watch the webinar to learn more about that!) The result of ML model optimization is a search experience that serves up relevant results while anticipating — and adapting to — each user's unique needs and preferences.

 

To learn more about Coveo ML optimization and see real-world examples in action, be sure to watch the full webinar. You'll come away with a comprehensive understanding of how to harness the power of machine learning to create search experiences that delight your users and drive real business results.