By Sarath Avasarala | July 31, 2019

With rapidly expanding catalog sizes and time-strapped customers with few clicks to spare, recommendations have become essential — a sine qua non — for brands to help reduce decision making complexity and drive business value.

Viewing recommendations as a simple matching exercise between users and items will no longer cut it — brands need to build experiences around recommendations and guide customers on a path to self-discovery, while being sensitive to several pitfalls along this path.

We have a few ideas that we think would help — but first, some context.

In the year 2004, Chris Anderson of Wired wrote an article entitled “The Long Tail”. In his article, he speaks about how the traditional brick-and-mortar retailers — constrained by shelf space — favored the most popular products over niche (“long-tail”) products with low sales volumes. He argues that these long-tail products could potentially outsell the most popular products since they cater to individual tastes and retailers are no longer constrained by shelf spaces or the reach of distribution channels.

It’s 2019 and the “tail” is now longer than ever.

And we’re not just talking about retail. Look at the numbers yourself:

  • Spotify has over 50 million tracks in its catalog (Source: Spotify Newsroom)
  • Etsy has over 60 million items listed on its site and (Source: Etsy’s SEC Filing)
  • Netflix has anywhere between 2000 to 6000 titles based on your region (Source: Quartz Atlas)

and these counts will only keep getting bigger.

The problem now is not so much on the supply side: the marginal costs of distribution and the costs of inventory have gone down quite significantly.

The challenge is more on the demand (or the “customer-facing”) side, where there is a strong necessity to know the customer and personalize the product to cater to their needs. Therefore, it is important for managers to think about the various touchpoints in the customer journey where recommendations or product personalization can add value.

What follows is a list of key use-cases and goals that one needs to keep in mind to deliver personalized experiences.

Reduce decision-making complexity

The huge assortment of items in any catalog can sometimes be anxiety inducing — this ties to a well documented phenomenon known as the “paradox of choice” — where customers experience stress when presented with large collections of items without any guidance. That is where recommendation systems can be of great help.

When customers visit a website or use a mobile app, there are several implicit signals that can be captured and processed to help them with their decision-making process.

Let us take the example of Uber Eats: when a user opens the app for the very first time, there is little information about the user’s likes and dislikes.

However, the app still makes use of contextual signals such as (i) the location of the user to filter the number of restaurant recommendations and (ii) the time of day to trim down the list of options even further.

In addition to this, users visiting the app communicate intent through (i) search queries (ii) visits to specific cuisine pages or (iii) visits to restaurant pages. These seemingly simple, but powerful signals can be used to push recommendations that save a lot of time for the customer.

Source: Uber Eats

Notice how the app has several widgets stitched into the core user experience. What’s more interesting is the fact that each widget in the experience has a purpose and caters to a different user persona:

  • The “popularity” widget lets the users make a quick decision by surfacing the most popular content and also shows awareness of context by taking the user’s current location into account
  • The “freshness” widget surfaces new and possibly little-known places for users who like to try new places — in a marketplace, this ensures that new restaurants get sufficient visibility and that the “rich don’t get richer”.
  • The “recommended dishes” widget uses the signals captured about the user to jump from restaurant-level recommendations to item-level recommendations and
  • The “offers” widget surfaces attractive offers for a value-conscious customer

It is important to note that the personalization journey doesn’t end with implicit signals: there is a strong need to capture explicit signals about users' preferences, and this can be done through purchase data, item and category-level ratings, favorites, and reviews. These signals are fed back into the system to create a strong personalization engine that knows the customer very well and serves the most relevant recommendations.

Create diversity and serendipity

It is easy to fall into the trap of using implicit and explicit signals to only recommend items with a high probability of purchase.

While an e-commerce site may actually be doing this for certain items to increase repeat purchases, customers allow very little leeway for inaccurate recommendations and find them redundant and un-intelligent. 
Source: Twitter

However, with enough user data, this can create a “filter bubble”, where a user is repeatedly exposed to recommendations from a small set of categories.

This phenomenon is particularly pervasive on social media where a user watching certain category of videos is exposed to the same type of videos over and over again.

While this is a hard problem to solve, several brands have shown that this problem can be handled to a certain extent through editorial intervention (“featured”, “editor’s picks”), making recommendation diversity an explicit goal of the recommender system, or providing avenues for the customer to independently explore content through a “discover” space.

Spotify creates up to six “Daily Mixes” with personalized content
Source: Spotify

Spotify, for instance, uses an intelligent combination of personalized playlists (“Daily Mixes”) and curated content (“Editor’s Picks”), along with multiple categories of playlists (“Discover Weekly”, “New Music Friday”). This approach has reportedly led to a lift in listening diversity by close to 40% (Source: Spotify Insights)

Source: Twitter

Twitter lets the users choose between an algorithmically ranked feed or a reverse chronological feed. The same holds for trends, where the user can choose between personalized or non-personalized trends. These kinds of additions provide avenues for users to step outside of the filter bubble and help them discover new content.

Think outside of the core product experience

There are several occasions where users visiting an app or a website browse for content but are unable to complete the purchase flow (or perform a "success action") within a session.

One can observe drop-offs on category pages, product pages, or after a product is added to the cart; this presents a chance to take recommendations outside of the core product experience and into email  or notification campaigns, where a user can be gently nudged to finish an incomplete flow and be presented with similar product recommendations for purchase consideration.

On similar lines, there are occasions where users communicate intent through a search query, but the exact item is not available in the product catalog. On such instances, one can measure similarity between search query and the items in the catalog to surface similar items which are already in the catalog.

Source: Netflix Mobile
Source: Netflix Desktop

For instance, even when Netflix does not have a title related to your search query, it surfaces titles that are similar in some respect (the genre, actors, etc.) so that the user has alternative viewing options.

Evaluate downsides and create feedback mechanisms 

A critical part of recommender system design is to evaluate the cost of an inaccurate recommendation. This becomes all the more important in domains like healthcare, where an inaccurate recommendation can potentially cause a lot of harm.

The only way to solve this would be to have an open discussion involving a diverse group of people and build feedback mechanisms into the product to mitigate potential downsides.

Building simple feedback mechanisms to capture dislikes or offensive content goes a long way in improving the recommendation system.

Common implementations include (i) downvotes or thumbs-downs (ii) “see less often” options in social feeds (iii) close buttons in recommendation spaces to hide specific recommendations or (iv) full-fledged reporting modules which capture details about why a user didn’t like a certain recommendation.

Conclusion

Recommendation systems can span the whole gamut from popular to hyperpersonalized and context-unaware to context-aware.

A well-designed recommendation system can add a lot of business value in terms of increased frequency of product use to increased cart value and retention. On the customer side, it can reduce decision-making complexity and lead to moments of customer delight.

At the same time, understanding the limitations (bias, filter bubbles, cost of inaccurate recommendations) is extremely important and it is vital for system designers to ensure that the product has enough checks and feedback loops in place to protect the customer from potentially harmful or divisive content.

As we said before, providing a good recommendation is more than just matching a user with an item - it is about guiding the customer on a path to self-discovery!

About the Author

Sarath Avasarala - Product Manager @ YML Bengaluru

Sarath is a Product Manager at YML. With hands-on experience in design and a keen understanding of business and tech, Sarath loves to talk to customers, get his hands dirty with design, dive deep into data, and do whatever it takes to deliver customer delight.