How do you recommend products or content for a very specific interest at a moment in time?
In this post, we’ll show you how to build a world-class recommendation engine using Primal’s data service.
Why Recommendation Engines Fail
Current recommendation engines tend to fall into the following approaches:
- Recommendations based on social networks or collaborative filtering.
- Recommendations based on broad classifications of attributes or features.
- Recommendations based on behavioral analyses of past activities.
None of these approaches gets down to the heart of the problem. If your customers are focusing on a specific interest, give them recommendations that reflect that specific interest! Your recommendations will be more timely, direct, and productive for it.
Before: Recommendations that are not timely or specific
In the screen capture below, note the list of article recommendations at right, under the heading News For You. Whatever the reasons for recommending these articles, you can tell at a glance that they have little if anything to do with the subject matter represented in the article.
And there’s the problem: The reader is clearly interested in this content, but the article recommendations aren’t targeted to this specific interest.
After: Recommendations that are timely and specific
In the following screen, we’ve updated the article recommendations using Primal’s data service. Note that the recommendations are now highly targeted to the specific interest that the user is expressing at this moment in time.
We’ll now walk you through the solution presented above. The challenge here is that you have no preexisting information about your users’ individual interests. The opportunity is to use Primal’s data service to generate this interest data to power better recommendations.
As a signal of a user’s interests, we’ll use the news abstracts that they’re reviewing as highly targeted and specific indicators of their interests at the specific moment of time.
Note: This same approach can be applied to product recommendations simply by using product summaries.
Here are the steps:
Step 1: After reviewing a list of abstracts, our user selects an article to read. Each abstract (or alternatively, the URL) is used as inputs to Primal.
Step 2: Primal synthesizes a unique interest graph based on the abstract to power the recommendation engine.
Step 4\3: Primal filters content in real-time using each unique interest graph, generating highly targeted recommendations based on the interests.
Step 1: Capturing interests
We have no data representing our anonymous user, only the highlighted abstract that was selected by the user.
With only this focused indicator of interest, Primal can build a rich interest graph to provide recommendations.
Step 2: Create an interest graph using Primal
Primal uses the same objects of interest that people use as inputs (such as text, URLs, queries, and topics), making it simple and intuitive.
For this demo, only 5 terms were used as input. Using this unique collection of terms, Primal synthesized a unique interest graph comprising about 200 semantic terms, as an economical and concise representation of these specific interests.
Step 4: Filter content using interest graph
The interest graph created in Step 3 can be used to filter and rank any type of textual content, including real-time social media, news, products, ads, and multimedia sources.
This screen cap above is a visualization of a portion of the interest data synthesized by Primal’s AI.
This data is used to filter and recommend the following selection of news articles.
If you’d like to build your own world-class recommender, check out our Developer site to get started.