Optimizing Advertising Performance with Advanced Machine Learning and Vector Database Technology

We integrated vector database technology into Criteo's retargeting recommendation engine to help marketers enhance their campaign performance.

These days, marketers are under unprecedented pressure. A changing and more expensive advertising landscape, along with new regulations and overall macroeconomic uncertainty, make them increasingly dependent on performance-based digital marketing activities. As consumers continue to heavily consider every purchase, high-performing bottom-funnel tactics play a crucial role in closing transactions and getting revenues in.

At Criteo, technology and innovation are at the heart of what we do, going back to our early days, when more than 15 years ago, we were running a movie recommendation service. AI (Artificial Intelligence) is part of our DNA, and every day we strive to enhance our AI-powered solutions and drive meaningful outcomes for our clients’ advertising campaigns. Today, we are thrilled to present DeepKNN, our latest addition to Criteo’s retargeting recommendation engine. DeepKNN is Criteo’s advanced vector database and deep learning engine that already powers major parts of our product line.

The impact of this approach goes beyond buzzwords — during our beta release of DeepKNN in our retargeting recommendation system, we’ve observed meaningful increases in all engagement metrics. While the performance increases will vary by client, we are driving improvements in both click-through rates and attributed revenue uplift that often go beyond 10%! Simply put, our advanced AI drives better performance for marketers, at scale.

How does DeepKNN manage to meaningfully drive better performance? In the next section we’ll give a brief explanation of what DeepKNN represents and how it helps us improve campaign performance.

Enhance Your Ad Perfomance

Discover the power of dynamic retargeting.

What is DeepKNN?

DeepKNN stands for “Deep Learning Representation for K-Nearest Neighbors Search.” DeepKNN combines a cutting-edge vector database and deep learning technology. This framework is now powering most of Criteo’s advertising campaigns, providing consumers with better product recommendations and improving consumer engagement metrics for advertisers.
DeepKNN represents the combination of two powerful technologies:

  • A Deep Neural Network that processes all our product and consumer activity information into a set of vectors (as shown in the processing pipeline pictured below in Figure 1).
  • A Vector Database (VectorDB) that stores the product and consumer vectors and allows us to find the most similar products and consumers to a query (as shown in the retrieval process pictured below in Figures 2a and 2b) in real time.
The process of indexing products in a VectorDB: Step 1) Ingest all the unstructured product information available (images, text, user activity). Step 2) Transform them into vectorial representations via deep learning. Step 3) Store them in the VectorDB.

The Power of Vectorial Representations

Why is it useful to associate vectors with all our products and consumers? Because vectors help us find the best-match consumers for a given product or service. By representing consumers and products as vectors, we are able to incorporate our notions of consumer-to-product affinities as distances between points. As shown on the right side of figure 2a below, the vectorial space contains a set of consumers and products placed in various positions on a 2D plane. We observe that by the simple act of placing them in a 2D space we can define implicit clusters of similar products, such as shoes, fashion, technology, and travel. Once the consumers and products are placed in the space, our vector database allows us to look very fast for the nearest neighbors, or closest consumer vector, for any given consumer or product. This approach is what powers at core our recommendation and audience products.

How Does DeepKNN Impact Marketers’ Campaigns?

DeepKNN helps advertisers improve their campaigns in several ways:

  • Personalized Recommendations: DeepKNN can provide personalized recommendations to consumers based on their past behavior and preferences. This can lead to higher conversion rates and better overall campaign performance. (See Figure 2b below).
  • Better Audiences: DeepKNN can help identify patterns and similarities in consumer behavior that may not be immediately apparent. By clustering consumers with similar behavior together, we can target our advertising more effectively, leading to better engagement and conversion rates. (See Figure 2a below).
  • Faster Data Retrieval: With DeepKNN, we can quickly search through large sets of data to find the most relevant results. This is essential for real-time decision-making in performance advertising, where every millisecond counts.
  • Improved Ad Creatives: By analyzing the content of our ads using DeepKNN, we can identify patterns and similarities that resonate with our target audience. This can help us create more effective ad creatives, that in turn drive better results.
Audience generation using search in a VectorDB: Starting from a product, we look up its vector in VectorDB, retrieve the closest user vectors, and return them as part of the relevant audience.

Improving Retargeting Product Recommendations with DeepKNN

Recommending the right products is the core of any retargeting campaign’s success. Our existing high-performing product recommendation solution for retargeting employs a two-stage architecture to ensure we can scale to thousands of requests per second. While in the first stage, we compute multiple types of similarities between different items along with popular and trending products, in the second stage we re-rank the most suitable products, based on the most recent consumer preference information, in order to generate the most relevant banners in real-time.

Integrating DeepKNN in our product recommendations allows us to now compute similarities of different items in a more sophisticated way: similarities between popular, rare, and new items, which were previously harder to discover, can be detected more easily – resulting in product recommendations which are even better aligned to consumers’ interests. Another big advantage of DeepKNN is that, by placing all our products and consumers in the same space, we can search hundreds of millions of “consumer-item pairs” in real-time and thus constantly update the lists of top similar products for every consumer.

Recommendation using search in a VectorDB: Starting from a consumer, we look up their vector in the VectorDB and retrieve the closest product vectors and return them for recommendation.

Performance Outlook with Vector Databases and Deep Learning

With the advent of programmatic access to AI capabilities such as OpenAI’s ChatGPT and embeddings APIs (that can turn any document into a vector), more and more companies are realizing the power of deep learning and vector search. We believe this trend highlights the importance of vector databases in modern digital marketing, as advertisers are starting to leverage the vast amounts of data available to them more effectively.

By using deep learning algorithms and advanced similarity search and clustering techniques, vector databases like DeepKNN are helping advertisers identify clusters of consumers with similar shopping preferences. This, in turn, can lead to more effective audience targeting, more relevant recommendations, and improved ad creatives, all done by a search system that is faster and more scalable than traditional serving architectures.

We’re proud to be at the forefront of this trend, and we believe that our new online decision engine represents a significant step forward in performance advertising. By leveraging the power of deep learning and vector search techniques, we’re able to offer our clients a powerful new tool for maximizing the effectiveness of their campaigns. We’re excited to see how our clients will utilize DeepKNN to drive better results and higher ROI, and we look forward to continuing to innovate in this space in the years to come.

To learn more about how DEEPKNN and machine learning can improve your campaign performance, talk to an expert.

Explore dynamic retargeting success stories

Discover why 19,000+ global clients trust Criteo as their preferred advertising partner.

+19%
Return on Ad Spend (ROAS)*
+217%
Revenue growth year over year (YoY)
Romain Lerallut

Romain Lerallut is VP Engineering at Criteo and head of the Criteo AI Lab, in charge of developing the uses of AI in digital advertising and commerce. Before the launch of the lab in 2018, he was a director in the engineering department, responsible for the development of large-scale machine ...

VP Engineering
Flavian Vasile

Flavian Vasile is a Principal ML Architect in the Criteo AI Lab, leveraging over 15 years of expertise in Machine Learning applications for Online Advertising. He specializes in developing Deep Learning techniques for Performance Advertising and actively explores the potential of cutting-edge AI ...

Subscribe
to our newsletter

Fresh sales trends and consumer insights to help
you plan and win.