Search by Algolia

Sorry, there is no results for this query

The anatomy of high-performance recommender systems – Part V
facebookfacebooklinkedinlinkedintwittertwittermailmail

Recommendations and User Actions

In the previous posts of this series, you learned what makes a high-performance recommender system, how you can feed it with the right data sources, how to engineer the right features from which it will learn, and what are the industrial-grade recommendation models and techniques to reach high-performance.

Now that you have a solid recommendation model in your hands, it’s time to get back to your users. Why do users even care about recommendations? It’s good to take some perspective here, to look at the increasingly complex decisions that users have to take when navigating online.

Quoting Netflix’s research: 

“Humans are facing an increasing number of choices in every aspect of their lives — certainly around media such as videos, music, and books, […], but more importantly, around areas such as health insurance plans and treatments and tests, job searches, education and learning, dating and finding life partners, and many other areas in which choice matters significantly.

“We are convinced that the field of recommender systems will continue to play a pivotal role in using the wealth of data now available to make these choices manageable, effectively guiding people to the truly best few options for them to be evaluated, resulting in better decisions.”

– Carlos A. Gomez-Uribe & Neil Hunt. 2015. The Netflix Recommender System: Algorithms, Business Value, and Innovation

We’ve reached a point where the amount of choice is paralyzing. Take YouTube: there are now 82 years of video content uploaded every day to their catalog. As a user, where do I even begin to explore this pile of content for the gem that I will value?

At the beginning of our series, we set out to “filter down users’ choices and provide them with the most suitable suggestions based on their requirements or preferences”.  Once you have a high-performance recommender system at hand, it’s time to make the most of the predictions it can generate, exposing them to your users at a strategic time and place to suggest the relevant actions to take.

What you can get from a high-performance recommender system

As we can see in the Netflix quote above, it’s too late for humans to be able to navigate alone the sea of options they face online. The rise of recommender systems is thus driven by business necessity: if humans have a hard time making choices to access your content, and if you have enough data to help them make the right choices, you’re losing business by not helping them.

What kind of recommendations could help your users make the most of your service? In Part IV of this series, we saw that different recommender models can provide different kinds of recommendations:

  • User-Based Collaborative Filtering: ”Users who clicked on Harry Potter might also enjoy Lord of the Ring
  • Item-Based Collaborative Filtering: ”If you rated Four Seasons Hotel Paris positively and now are looking at our ‘Week-ends in Berlin’ offers, you may enjoy the Movenpick Hotel Berlin”.
  • Content-Based Filtering: “Users who view the product page of an out of stock Loose Navy Blue Jeans might be interested in other Loose-fitting, Dark-colored pants”

From the user’s perspective, in each case they’re getting browsing shortcuts: instead of users randomly browsing for some time, hoping to eventually stumble on the right item, high-performance recommenders can directly point them to the destination of their journeys. 

More than that, users are also getting new openings: someone landing on this page might not know that some related items could interest them. These predictions allow you to widen your users’ exposure to the content you could offer, helping them make the most of an overwhelming number of options.

This is what enables businesses to offer users staggering amounts of new content, and yet to help them make sense of it all: if we go back to the YouTube example where 82 years of new content appear every day, the average session length is still quite long – 13m on average. As a result, YouTube users are highly engaged, with 2 billion Monthly Active Users of which 122 million are active daily users.

Likewise, TikTok is excellent at this: you start watching any kind of content, and your interactions inform the system of your interests. If you indicate interest in cooking videos, you might suddenly see a lot of these in your feed as the system learns to cater to your current tastes. 

Encouraging actions: showing how users can profit from strategic recommendations

In the previous section, we’ve seen why businesses which have a lot to offer have a crucial interest in exposing industrial-grade recommendations to help their users. However, it’s not just a matter of shoving recommendations in their face. Let’s take an extreme example, where an out-of-stock product page that displays your top 500 related items as alternatives: even with a high-performance system, the user is not likely to scroll through them all.

How do we go from getting good recommendations to showing strategic recommendations? 

Think of recommendations as a means rather than an end: recommendations are how you get your users to engage, but the action is the goal you’re directing them towards. Thus, a good recommendation needs to be:

  • Limited
  • Contextual
  • Effort-saving

Limited

Only show a small number of relevant items. 

It’s easy to be tempted to just display more options, thinking it can only provide more opportunities. However, this intuition is easily disproved, and there’s a body of research demonstrating that a small number of options (3-6) can convert up to 10x more than a high number (10-24). We’re here to help the user filter irrelevant options, not to bring back the Too Much Information syndrome that we had set out to solve in the first place!

Contextual

A recommended item should be useful here and now

The value of a recommendation derives from how much it serves the current user at their current time and place; this immediate relevance makes context-awareness an important aspect of providing great recommendations.

Here are some examples:

  • On a Product Landing Page, after a user has added an item to the cart, the user is wondering if they have all they need to purchase. This is the right time to recommend complementary, Frequently Bought Together products with a one-click add-to-cart button.
  • On the other hand, on a Product Landing Page which is out of stock, the user is looking for other options to address the same need; recommending alternative, Related Products with a link to see their details would surely bring more value.
  • After a video is played, recommending strongly similar videos is unlikely to be satisfying; it’s better to recommend content that was Frequently Watched Together, maximizing your chances to keep the user engaged.
  • On a category landing page, users are looking for inspiration or a new idea; thus, showing a carousel with the best recommendations across this category, for example, by displaying Related Products to top-sellers in this category, is likely to perform well at converting first-time visitors into customers.

Effort-saving

A recommendation is only valuable if it saves a user some time and helps them achieve their objectives. 

  • On the one hand, this means you need to be conservative: recommending wildly varying content, although increasing serendipity, brings the risk of user frustration. It’s good to put ourselves in the shoes of the users, browse the service with a beginner’s perspective, and wonder: “do these recommendations help me better reach my goals?”
  • On the other hand, this means you can be innovative: recommendations can be useful in many more ways than just as a carousel on your website. There is a lot of value to be unlocked by feeding such recommendations into other tools like marketing platforms and  your CMS. For example:
    • Email campaigns, e.g., SendGrid: you can customize your mailings based on relevant recommendations. This could turn generic weekly emails into a dynamic, highly relevant newsletter!
    • Marketing Automation, e.g., Shopify: you could empower apps on your Store to leverage the predictions from Algolia, combining these user/item insights with your existing setup to merchandise and promote content, thereby improving your current Shopify Apps’ usefulness.
    • CRM integration, e.g., HubSpot: you could augment the customer’s profile with segmented recommendations of potentially interesting services for them, to inform your sales force to bring up the right new feature during their next call.
    • Data platforms like Segment or BigQuery: from these you can use the generated recommendations to augment your own data. While these systems generate popular services or actions users take, their results can feed into other back-office systems to enhance automated or manual decision-making, to build useful features for your next projects/analyses/…

If you follow these guidelines, you can be confident that the solid recommendations you get from your system are put efficiently at the service of your users.

You might only need data and algorithms for a recommender system, but a recommender system doesn’t make satisfied users on its own. Deciding how to leverage its recommendations to maximize your users’ satisfaction by suggesting the right actions to the right person at the right time is the key to putting this high-performance recommender system in service of your user-base, and ultimately to make the most of recommendations for your business.

How do we measure the success of your recommendations?

In this blog post, we’ve learned what makes for good recommendations, which actions users can take based on them, and how they can best serve your users to guide their actions across your service. To learn about Algolia Recommend, get started here or check out our recommendations API docs.

But how do we know it was the right action to take? How can you measure the effectiveness of your recommendations against your business metrics? How does one avoid optimizing for short term performance (e.g., hooking users every day to a new series of videos until 2AM) at the expense of long term business goals (ensuring this user won’t cancel their membership after one-too-many late night binge-watching sessions?)

Stay tuned as the next part of this series, Results and Evaluation, dives into these critical aspects of leveraging a recommender system in your products!

If you want to continue the conversation, reach out to Paul-Louis on Twitter: https://twitter.com/PaulLouisNech

About the author
Paul-Louis Nech

Senior ML Engineer

githublinkedintwitter

Recommended Articles

Powered byAlgolia Algolia Recommend

From search to recommendations – framing your customer’s online retail journey
product

Peter Villani

Sr. Staff Writer

The anatomy of high-performance recommender systems - Part 1
ai

Ciprian Borodescu

AI Product Manager | On a mission to help people succeed through the use of AI

What is a product recommender (or product recommendation engine)?
product

Catherine Dee

Search and Discovery writer