The benefits of content personalization for digital publishing
Artificial Intelligence (AI) and Machine Learning (ML) make it easy to create personalized user experiences. If you are a digital publisher, you can use AI/ML to display and recommend content that is tailored to the interests of an individual user. Personalization can improve customer satisfaction, increase user engagement, and reduce the workload of editorial and content creation staff. This blog post discusses how you can use Amazon Personalize, an AI/ML service from Amazon Web Services (AWS), to customize your publications.
The benefits of personalization
The value of personalization is proven across multiple industries. A 2021 report from McKinsey & Company found that 71% of customers expect companies to deliver personalized interactions online and preferred companies that recommended content and products based on their preferences. Companies that implemented personalization saw a 10 to 15% revenue lift from those activities.
Digital publishers achieve similar benefits from personalization. A large news site based in central Europe, for example, implemented a personalization service created by Ring Publishing, an Amazon Web Services Partner. Ring’s service builds homepages tailored to each user’s interests. The news publisher saw a 30% increase in user engagement as well as a 400% increase in diversity of content consumed. In other words, users were spending more time on the site and interacting with a much wider range of content. Editors also spent 50% less time managing homepages, enabling them to focus on other tasks, such as creating new content.
Using Amazon Personalize
Publishers traditionally personalize user experiences by creating complex rules: if the user matched criteria A, B, and C, then show content X, Y, and Z. However, rule-based systems tend to be brittle and hard to maintain because they do not learn from what the user actually does. Instead, a programmer needs to manually update the rules to handle changing circumstances.
AI/ML services like Amazon Personalize offer an easier, more scalable way to make recommendations. Rather than write special code, you run an AI/ML algorithm, or recipe, on a set of data to train it to make recommendations. Amazon Personalize offers multiple recipes that handle different recommendation scenarios. To personalize content shown to users, for example, you can use the “User Recommendations” recipe.
Figure 1. Typical Interactions with Amazon Personalize
To use the recipe, you start by importing data about your users, content items, and past user interactions with your content. You then train the recipe to create a model by having it read the data, usually from an Amazon Simple Storage Service (Amazon S3) bucket. At a minimum, you need to import at least 1,000 interactions and 25 users, each with at least two interactions. As of this writing, Amazon Personalize will consider up to 100 million users and 3 billion interactions for training purposes. The more training data, the better the recommendations will be. However, you will be charged for training time, so you need to balance your budget against your training goals. Once it is trained, you deploy the model as part of a campaign and can start requesting recommendations. More information on training, deployment, recommendation requests, and pricing is available in the Amazon Personalize documentation.
What happens when a new content item is published, such as a breaking news story or a movie review? Since the item is new, the model doesn’t have any interactions data to know which users are interested in the content. The User Recommendations recipe solves this problem in two ways. First, the recipe automatically includes a percentage of newly published items in recommendations so that it can collect data on how users react to the content. You can control how many new items are included in recommendations when you are setting up your campaign. The second way the system recommends new items is by examining any metadata provided when you loaded your content data, such as descriptions, release dates, or keywords. The recipe uses the metadata to determine how similar the new items are to previously published content. It can then make recommendations based on this similarity. Handling new or so-called cold items is discussed in greater detail here.
Improving recommendations with context and metadata
The metadata you provide about your users, items, and interactions are very important for making relevant recommendations. I’ve already discussed how metadata like descriptions and keywords help Amazon Personalize recommend new content. User metadata also helps the recipe understand what interests your users. For example, if you know that readers in different regions of the country are interested in specific topics, then you should include location information in the user dataset. When Amazon Personalize makes recommendations for the user, it will take their location into consideration.
You can also add metadata to your interactions data. At a minimum, interactions must contain identifiers for your user and content item, as well as a timestamp for when the interaction occurred. Contextual metadata can also be extremely valuable. For example, users often change their reading habits depending on the device they are using. On cell phones, users may prefer shorter content items. When using a desktop or a tablet, they may prefer longer-form content. It is therefore a good idea to capture the device type in the interaction dataset so that Amazon Personalize can learn from this behavior. When you submit a recommendation request, you can include the user’s current device-type as a context parameter. Amazon Personalize will then factor that context into its response. Other contextual fields that might be important are the available bandwidth, the geographic location where the interaction is occurring, or perhaps even the weather. If you publish content about food and drink, for example, you might want Amazon Personalize to recommend more iced drinks and cool desserts when the weather is hot!
You will need to experiment with the metadata you provide in your datasets and work iteratively until you get the best recommendations for your users and content. A blog post that describes the importance of metadata and context can be found here. Track the metrics for how many content items your users view and how long they spend interacting with your content. You can then run A/B testing to compare different metadata options. If you see improvements in your engagement metrics, then you know you are providing the right context and metadata.
Should you always show users what interests them?
When personalizing content, you need to be mindful of creating “filter bubbles” or “echo chambers”. Since users only see what they are interested in, they lose the ability to explore and discover new items. Filter bubbles can happen in any industry that relies on personalization. One ML engineer at a food delivery service, for example, noted in a blog post that “building quality recommendations and personalization requires delicately balancing what is already known about users while recommending new things that they might like.”
Amazon Personalize recognizes the need for users to explore and can include so-called exploration items in its recommendations. The User Personalization recipe has
exploration_item_age_cut_off parameters that you can configure when setting up the system. If you set the weight to 0, then no exploration items are included in recommendations. If you set the weight closer to 1.0, then more items are included. The default value is 0.3. The age cutoff parameter controls the maximum age of exploration items, counted in number of days. For example, if you set the value to 7, then content over 7 days old will not be included as exploration items.
News and academic publishers are especially sensitive to how much personalized content should be displayed to users. These publishers rely on being an authoritative voice and often want to emphasize content selected by their subject matter experts. As the Chief Technology Officer of the New York Times explained in 2018: “Readers have mixed feelings about personalization in news. They want two different things—the world rendered with judgement and authority, but also a review that reflects their interests.”
The New York Times maintains editorial control by showing personalized content only in special “For You” sections on its homepage. AWS Partner Ring Publishing takes a similar approach to balancing personalization with editorially curated content. With Ring’s solution, editors can designate a specific block on the homepage that is entirely personalized, while other portions of the homepage might be selected by the editors or contain a mix of editorially and AI-selected content.
Amazon Personalize makes it possible for you to show recommendations that are a mix of personalized and manually selected items. The User Recommendations recipe supports this use case through Promotions. In your items dataset, you first identify editorially selected content with a metadata field. For example, you might have a field called “SELECTED” that you set to “True” or “False”. When requesting recommendations, you then specify the percentage of promoted items to include, plus a promotion filter that selects the metadata field(s) that you created. In this case, the filter expression might look like this:
INCLUDE ItemID where Items.SELECTED IN (“True”)
More information on promoting content items in recommendations can be found in the Amazon Personalize documentation here.
In this blog post, we discussed how Amazon Personalize can create better user experiences. If you are a digital publisher, personalizing content for users can:
- Increase user engagement
- Lower costs by making staff more productive
- Maximize your investment by diversifying content consumption
Personalization is not a one size-fits all solution. Every publisher needs to tailor the approach to meet business needs—how much new content versus older content should be recommended? What percentage of manually selected content should be recommended versus AI-chosen content? How much should you rely on content that aligns with user interests versus encouraging exploration of new content? What metadata and context fields are the most relevant? Publishers should treat personalization as an iterative process, which they refine and adjust over time to create the best experience for users.
As a publisher, you decide how personalization will increase customer satisfaction, lead to innovation, reduce costs, and increase your efficiency. To find out more about Amazon Personalize, visit the service documentation or speak to your AWS account representative about how it might benefit your business.