How to harness ML to Personalize Products and Content

Author: Maryann Kongovi

Personalization is all around us. If you’ve ever received a relevant recommendation on a website, a notification from an app, or a promotion in your inbox, you’ve been delivered a personalized experience. Personalization systems are one of the most common applications of machine learning and have become essential for fueling growth in the modern economy. They’re also a must-have when consumers are watching their wallets, as these systems aim to identify the best ways to engage them.

In the case of email and text notifications, this can be particularly fraught for companies. Consumers will unsubscribe if too many irrelevant messages are sent. As every unsubscribe represents the loss of a valuable channel for revenue and engagement companies need to calibrate their strategies carefully. On the plus side, when notifications are done well, they should drive sales.

So how does personalization look to the user? Let’s take one of the most successful examples of recommendation and personalization in the industry: Spotify. Odds are, you have used Spotify playlists and have seen recommendations that line up with your interests. This is no coincidence. Personalization shows up in multiple ways on Spotify, the three most visible examples are:

1.  The homepage
   Personalized collection of playlists, artists, songs, podcasts
2.  Curated playlists
   Designed specifically for each user based on their listening history
3.  Recommended playlists
   Already made but matched to a listener’s likely tastes. 

So, what’s going on behind the scenes to power all of these highly customized experiences? In recent years Spotify has transitioned to a new machine-learning approach, Graph Neural Networks (GNNs),* which treats your data as a graph and learns the different relationships and interactions between the different entities. Spotify’s user data naturally resembles a three-dimensional graph (see image above), and by leveraging the structural properties of their data and uncovering the more complex and nuanced relationships, they were able to substantially improve their models and thus see significant gains in long-term user engagement and user experience. A recent example of how this approach delivered relevant and personalized recommendations is Spotify’s expansion into podcasts. Their GNN models were able to curate podcast playlists for each user – although they did not have historical data specific to podcasts, graph learning was able to leverage signal from other parts of the graph to predict what each user is likely to listen to.

Global fashion brands such as H&M and Zara also rely heavily on AI-driven personalization models to drive recommendations, both online and offline. These models enable brands to tailor the in-person retail experience by using analytics to identify the right apparel for each retail location given local purchasing patterns. In the case of H&M, this strategy was first employed in an attempt to reverse one of their biggest sales slumps. This saw them shifting from stocking all stores globally with the same merchandise to personalizing each store based on local preferences. They equipped their merchandising team, and a team of designers with insights to customize local offerings based. These were developed using AI-driven algorithms that analyzed store receipts, returns, loyalty data, and behavior. They also used their extensive data to predict overall demand across each category by store, which in turn meant they could optimize production, their supply chain, and design pipeline. The online experience is also heavily customized based on inputs such as search history, recent purchases, and different interactions that helped H&M identify new trends and behaviors. The end result is a unique homepage for every user that only shows the most relevant products, driving higher sales in the process.

In a similar vein, AirBnB uses personalization to match the right hosts with users looking for a place to stay, based on their preferences with the goal of increasing bookings. Even when a new host signs up, GNNs enable very effective recommendations for the newly registered host despite a lack of data. By understanding latent content on the new host as well as other connections on the platform, GNNs can quickly deduce information about the host based on other users on the platform.

Given the power of this approach, many companies with established data science teams have started adapting GNNs as the core of their recommender systems. What all of these companies have in common are large data science and engineering teams that can undertake the work to build these models. Many of them, including Spotify, also use open-source tools such as  PyTorch Geometric (PyG) to build GNNs to deliver on the promise of personalization. 

For organizations who want to accelerate the time to get predictions, while minimizing the burden on internal teams, Kumo can help. Whether you have a well-established data science function and ML pipelines already tackling personalization, or you are looking for new ways to build and develop an engine, contact us. We’d love to talk.