A number of companies, Youtube, Netflix, Facebook to name a few, have used recommendations to achieve 100+X growth in the last decade. In this post, we will look at different entry points for recommendations that you could explore to propel the growth of your platform.
Recommendations in Search
When you think of recommendations you will probably first think of a “home page” and yes, we will come to it. However, I wanted to start with search since often search is typically the most common lever you will consider to help users quickly find what they are looking for.
Broad queries in search
If someone is searching a broad query, like say “t-shirt” in an e-commerce platform, the user is probably looking for inspiration. You can complement the topical results with deeper explorations.
The image above shows how to encourage deeper explorations.
Providing recommendations of what else to search can also be very useful in connecting users to what they are really looking for.
Diversity boosting in search results
Elaborated in the post Diversity of recommendations, if the user has not selected the top result and is moving down to the second result then you could guess that the user is looking for something else other than what the top result captures. Hence it would be good to boost the dissimilarity of the second result from the first and the third from the first two and so an so forth. It is an ML problem to understand how best to diversify, since the user could be looking for a different interpretation of the searched text, a different genre, a different publisher/creator or a more nuanced difference. However, diversity boosting, combined with deeper explorations as mentioned below in the “What-Next” section often leads to greater engagement.
What-Next
A great entry point for recommendations is when a user has selected a result. Think of “WatchNext”, videos recommended to a user while they are watching a video, or at the end of watching a video.
“Not this something else” journey: A key design principle to use is to not have any dead ends in a user journey. A user who is looking for something, and has not found it so far, should always have new ways to continue their exploration. Hopefully these ways provided by your platform build on the information the user has already provided and do not seem repetitive.
“After this” journey: Another user journey facilitated by What-Next recommendations is sequential consumption. Having watched the first episode of a season, a good recommendation is the second episode. Note that if the What-Next unit can be ordered dynamically, then early on in the consumption of the video, we can use it for Not-this-something-else and if we detect the user is satisfied with the item they are currently on the page of, the What-Next unit can switch to After-this user journey.
Think of how you might build an automatic personalized playlist on Apple Music for instance. When a user listens to a song and immediately skips it, you might want to play a different item. However, if the user listens to the entire song, the next song you play should probably one that is similar to the one you just played.
Recommendations on the home page
While much of what we discussed about search applies to recommendations on the home page, the home page has a broader purpose.
It helps to orient new users as to what the platform is about.
Think of the experience of a new user on Amazon. It shows the most popular user journeys Amazon aspires to fulfill.Even for repeat users, it is a way to recommend them to try journeys they might not have tried for a while. Think of Linkedin home feed asking you to try out events or jobs you may interested in or courses to learn in Linkedin Learning. These items on the home page serve not just to help users find what they are looking for now, but also to create future demand of journeys the user does not currently come to the platform for.
Explaining why
Early on, when your users are feeling out your recommender system and are building trust and familiarity with it, it can be very useful to provide explanations of why something was recommended.
Quoting from this paper by Balog and Radlinksi (Google Research), these are some of the major goals of explanations:
Effectiveness: Help users make good decisions
Efficiency: Help users make decisions faster
Persuasiveness: Convince users to try or buy
Satisfaction: Increase the ease of use or enjoyment
Scrutability: Allow users to tell the system it is wrong
Transparency: Explain how the system works
Trust: Increase users’ confidence in the system
Note that item, 5, Scrutability is something we will cover in the section “Giving users control” below. Apart from that, explanations for recommendations can be used to drive the other goals.
Giving users control
Recommender systems improve by incorporating user feedback. Explicit in-app feedback can be built to help the user provide both positive and negative feedback. One example of feedback is the 👍 action on feeds like Facebook
A more powerful and actionable type of feedback is explicit negative feedback. The figure below shows an example of how Linkedin encourages negative feedback.
If negative feedback is implemented well, it encourages:
Agency: your users will engage more with the recommender systems and start feeling like it is their creation.
Trust: users will trust something more if they feel they can control it in some part
Allowance for mistakes: All search/recommender systems will make mistakes, either in terms of results that are inappropriate or quite off topic. These actions provide users a way to help the platform correct it and reduce the perceived egregiousness of the error.
Sample complexity / efficiency: Allowing users to say things like not interested in a certain topic / genre / publisher / style can be a very quick way for the recommender system to stop making a habitual mistake. Most recommender systems take only a few samples to learn your tastes from what you click but many more samples of what you did not click to learn to what you don’t like.
Early indicators of drop off: Without explicit feedback the platform is much more likely to lose the user. Explicit feedback is an early indicator of drop off and also encourages users to give the platform a chance to learn from previous mistakes.
Recommendations for sellers’ journeys
So far we have been talking about the demand side. However if you are building a two sided marketplace then adding recommendations to the seller / creator journey can be immensely useful as well.
How to improve the reach of an item they have created
NLP techniques to suggest better titles
Suggest categories and tags to improve visibility of the item, especially to the right users.
Whether to write content ( description / blog ) to engage their audience and build demand and appreciation for their unique offering.
Pricing / promotions
Recommend pricing changes (to optimize sales, reduce inventory build up)
Share of voice
Provide sellers with an estimate of the share of voice (via search & recs) of the topics and keywords they most associate to.
Provide feedback on what parts of that narrative o those searches they are missing out on. (Think of a creator getting ideas of a subcategory of their genre they should focus on for growth)
In a future post, I will expand on applications of machine learning to sellers’ journeys. I believe this has a high return on investment for companies since supply is quickly becoming the differentiator among platforms today.
Disclaimer: These are my personal opinions only. Any assumptions, opinions stated here are mine and not representative of my current or any prior employer(s).