Welcome to the second part of our Recommender Systems series. In our previous post, we outlined the history and basic components of recommendation systems, but here, we'll explore the technology of recommender systems and how they can specifically benefit publishers.
Exploration and Exploitation
A central challenge for any recommendation system is the paradox between exploitation and exploration. Exploitation refers to providing recommendations that are similar to suggestions that worked well in the past. But in contrast to exploitation, exploration provides recommendations that are less likely to succeed but have a high payoff if they do succeed, as the system expands its knowledge and models to drive future performance. In other words, exploration lets the algorithm learn about users’ interests that it didn’t know before, while exploitation lets the system drive performance consistently, but at the risk of boring people or trapping them in a filter bubble.
TikTok is notable for placing a relatively high emphasis on exploration compared to other platforms like YouTube. Every TikTok video seems to guarantee a certain minimum number of views. If the video performs well in that initial test, it will be served to larger batches of users. And from a user perspective, the algorithm keeps trying new topics even after it has found a set of topics that the user is interested in and will reliably watch, whereas YouTube focuses more on channels you subscribe to. TikTok’s algorithm treats each video more or less independently to assess its “viral” potential, caring relatively little about how many followers the creator has.
The new frontier – AI and Deep Learning
In recent years, deep learning algorithms have been the topic of conversation at the annual Recommender Systems (RecSys) Conference. AI and deep learning models now offer companies the ability to reverse engineer the way they calculate recommendations for a given user. The new models are learning machines, capable of detecting and learning features themselves that matter for users' behavior.
AI and deep learning models now offer companies the ability to reverse engineer the way they calculate recommendations for a given user.
Instead of manually adjusting parameters, now algorithm builders can essentially feed all that data into new algorithms, asking them to learn and uncover hidden patterns in the way users consume content. This is a huge leap from the “mass personalization” of current systems.
In a world where we have hundreds of data points on a user’s demographics, tons of metadata to understand the content they’ve consumed, and dozens of other feedback metrics such as clicks, likes, and shares, it’s not surprising that these approaches are on the rise. A number of neural networks and deep-learning techniques have been adopted in recent years. And what broadly unites them is their ability to work backward to identify performance directly from the data for each individual person.
This is genuine personalization, and it sounds easy, right? Just let the machine do the job.
In deep learning, the emphasis shifts to data acquisition and costs. Because these models essentially rely on as much data as possible, they can be hard to implement for publishers that don’t have sophisticated user-tracking or logged-in users. They can also be very expensive to run at scale, especially if you have casual users with a low revenue per page.
So, to answer the billion-dollar question: What does this mean for digital publishing?
The bar has been raised by users expecting better tools to find the right content and receive an amazing, seamless user experience. The addictive personalized content feeds of apps like Instagram are an innovation that many publishers are still struggling to adapt to. With that being said, it’s still worth the effort to start making the transition. Not only will prioritizing user experience benefit your users, but it’ll benefit you as well with more engagement and converting casual users to regular readers through better habit formation.
As an EX.CO client, you’ll automatically have access to a sophisticated content recommendation engine, a result of last year’s acquisition of Bibblio.
This includes contextual recommendations by default, ensuring every user sees relevant video recommendations in the player, corresponding to the article they are currently reading. This has shown a 49% increase in dwell time with the player in large-scale testing and doesn’t rely on sophisticated user tracking in order to deliver results.
Furthermore, EX.CO offers additional advanced recommendation types that are designed to optimize the yield of content shown in the player, or personalized recommendations based on our deep learning algorithm.
This represents a significant step forward for publishers looking to serve their users with highly relevant content, maximizing both engagement and yield from page real estate.
And the best part is: the recommendation engine comes as a free and integrated service when you choose EX.CO as your partner for video.
Need recommendations for your video strategy?