Maximize revenue with a deep learning recommendation system

Today, product recommendations are an essential requirement for any eCommerce business looking to increase engagement, purchases and loyalty. However, a consistent challenge for marketers and merchandisers has been determining which products among a massive catalog of items to serve users with various preferences and levels of intent.

This need for greater accuracy is why many in the industry are moving towards advanced deep learning recommendation systems for delivering the next best products. In fact, in two-thirds of the use cases Mckinsey & Company studied, building recommendation algorithms based on neural networks (more on that below) improved performance beyond that provided by any other analytic technique.

Most product recommendation strategies in the market today are either global in nature, meaning they are based on popularity and trends, or contextual, tailored according to specific product attributes or the user’s traffic source, device, local weather, and so on. And while these can be highly effective in certain situations, ultimately, they are not truly tailored to the individual.

Even personalized strategies, which take into consideration a user’s affinity and recent activity to serve additional products of interest, are either unable to predict which items should be served next or fail to do so in real-time.

But as consumers have come to expect a high level of personalization in online retail interactions, merchandising teams need to better anticipate their needs and dynamically recommend products in the moment, and refine them over time.

Designed for this very purpose, Dynamic Yield’s deep learning algorithm instantly identifies customer intent, even from the first session. Enabling an understanding of consumers like never before, brands can now automatically showcase products predicted to drive action in the moment and over time.

As a neural network recommender system, the model driving deep learning recommendations at Dynamic Yield is inspired by the human brain, which is made up of multiple learning units which connect together like a web, each receiving, processing, and outputting information to nearby units. Unlike the vast majority of traditional machine learning applications, the architecture of our deep learning system allows for it to be rapidly trained (and with less data), adapt more freely based on learnings, and mine meaningful insights from complex information.

For example, Gmail popularly uses Natural Language Processing (NLP) to learn word associations and help their users write emails faster by suggesting complete sentences as they type. Similarly, Dynamic Yield learns the products in a user’s browsing history, in-session activity as well as trends seen across the site to recommend products they are predicted to engage with as they shop. This is done through item2vec, the learning model derived directly from its NLP counterpart, word2vec.

Our deep learning recommendation model self-learns quickly, frequently, and off of a huge amount of behavioral and product data, which is why it is able to instantly identify customer intent, even from the first session. And better yet, as new information comes in, the results are continuously refined.

The right set of parameters are automatically determined with our deep learning algorithm based on each user’s distinct behavior, location in the customer journey, as well as trends seen across the site. So say goodbye to applying custom filter rules and let the algorithm do its work.

Personalization is all about consistency, which is why the same advanced deep learning technology for recommending products predicted to drive engagement extends beyond the web to mobile apps as well as email, with results tailored at the time of email open.

Home to in-house brands as well as over 60 designer names, GlassesUSA understood the challenge of mapping the perfect eyewear to each shopper among thousands of styles available in its catalog. And after years of optimizing different experiences geared towards product discovery, the eCommerce team was ready to put a more sophisticated machine learning algorithm to the test.

Representing the very top of the funnel, GlassesUSA set up an experiment in an area just below the fold on its homepage to compare Dynamic Yield’s deep learning strategy against collaborative filtering for all desktop traffic.

Because the deep learning model is automatically configured per site, product feed, and individual, the team made but a few minor tweaks to the strategy before quickly seeing impressive results, most notably a 45% increase in add-to-cart rate, a 68.1% increase in purchases, and an 87.6% uplift in revenue. And after running a similar test on mobile, the advanced algorithm proved yet again to be the strongest performer when compared to the control, with the team at GlassesUSA making deep learning the sole strategy for its popular homepage widget on this channel.

Read more about how GlassesUSA.com deploys a deep learning algorithm to adapt its recommendations to each shopper.

Today, company’s must be willing to move beyond recommending similar or complementary items to those that are truly tailored to each user. And while affinity- and collaborative filtering-based algorithms are both considered personalized, deep learning combines each of their best qualities in that it works in real-time, off any feed size and is able to continuously learn and adapt to incoming data.

With Dynamic Yield, teams have access to this state-of-the-art technology out-of-the-box, allowing them to break the ceiling of what was originally thought possible without undergoing complex integrations with third-party providers or requiring in-house development.

And to top it off, our deep learning recommendation model is part of Dynamic Yield’s AdaptML™ system, which shares and applies all of the data and learnings collected from across our deep learning, ranking and predictive applications for enhanced decisioning power.

If you are a Dynamic Yield customer interested in implementing deep learning recommendations, please contact your Customer Success Manager. And click here to learn more about the prerequisites and best practices.