Personalization in Search and Recommendations

What is Personalized Relevance? When and Where should you use it?

By Nikhil Dandekar, Engineering Manager, Search at Foursquare

What is personalization? How much does it matter in a search engine or in a recommendation app? Is adding personalization a game-changer, that makes your product 100x more relevant? Or is it just another buzzword, which doesn’t really matter in real life?

What is Personalization?

Personalization, in the context of a relevance product, means adding user-specific signals on top of all the other signals to make your product much more relevant. Based on user’s demographics, their likes, their follows, their past behavior etc., you can build a fairly rich model of the user that you can then use to personalize your product. For example:

1. For a search engine, you can use the user’s past searches & result clicks to boost certain results and to demote others. My team at Bing built this for web search a few years ago (Link). And at Foursquare, we do an even richer use of past visits, likes, dislikes, tips etc. to personalize your local search & recommendations.
2. In a movie recommendation app like Netflix, you can promote movies similar to the ones you have watched in the past, while demoting others that you are unlikely to watch.
3. In Google News, you can show more news higher up on the page, on topics which you have shown an interest in, and from websites that you are more likely to click on. E.g. if you always scroll to the Sports section & click on the news from espn.com, promoting Sports articles from espn.com would make the app more relevant to you.

Personalization can be used in an implicit manner, i.e. to boost ranking of certain results. Or it can be explicit, i.e. when you change the UI or add/remove explicit elements in the app based on personal data. Often times, it’s both implicit & explicit, i.e. you use personal signals to both boost the ranking of a result and also explain the result in the UI. E.g. “You visited this page” text in the 1st screenshot above.

A broad definition of personalization also includes social signals i.e. behaviors of friends or followers/followees of the user, assuming those are available to you. For the sake of simplicity, I’m going to skip the nuances that come with social signals in this article.

Where does personalization help?

Personalization is often thought of as a hammer, which will magically make your product much better. This leads to people often over-investing in personalization. Personalization is useful in a lot of cases, but not very useful in others. You need to be careful where and how you use it and to what extent.

That said, the answer is quite simpler and intuitive. Here is the general principle that you need to know:

The importance of personalization is inversely proportional to how specific the user intent is.

Personalization is more important when the user intent is broad. When the user intent is specific, i.e. the user knows what he or she exactly wants, personalization doesn’t matter much. As we go from a broader intent to a more specific intent, personalization becomes less and less important.

Let’s look at this in context of different products.

1. Personalization matters a lot in most news feeds and query-less recommendation engines. E.g. Google News, Netflix, Medium.com homepage etc. In this case, the user intent is very broad (e.g. reading news or watching movies). If you are building the Netflix home screen, the potential list of movies you can recommend is huge. Personal signals can help you nail the top few movies that you think the user is most likely to watch.
2. In a search engine, for informational queries, the intent is fairly broad. If you search for “lunch” in the heart of San Francisco, the number of potential answers is huge and having personal signals helps to better pinpoint the few results that are right for you.
3. As we move along the spectrum from broad, informational queries to specific, navigational queries, the importance of personalization decreases. Navigational queries have a very specific intent and only have 1 right answer. E.g. If you search for “YouTube” on Google, there is only 1 right answer (youtube.com). Personal signals don’t matter at all in such cases.

Broadest (query-less recommendations) to the most specific (navigational query) intent in Foursquare

Another pattern emerges from these examples. The corollary to the personalization rule I stated above is:

The importance of personalization is directly proportional to the number of “right answers”.

If there is only 1 right answer, as in navigational queries, personalization doesn’t add any value to your system. But if the number of potential answers is huge, (e.g. news/movie/what-to-buy recommendations or informational queries), personalization signals are very useful in selecting those few results which really matter.

This second rule is more general in scope. It goes beyond search & recommendations and applies to every product where there is a list of potential results, and you need to figure out the top few results that are most relevant to the user.

The “No Intent” Scenario

Beyond traditional search & recommendations, there is a scenario that is coming more & more into prominence now, where personalization matters even more than in a query-less recommendation scenario. That’s the no intent scenario.

The most visible examples of the no intent scenario are the notifications that mobile apps push on your phone. In this case, the user never expresses any intent. Instead, based on the current context, the app decides that the user might have an intent at this point of time and pushes content to the user.

The quality bar for doing the right thing is much higher for mobile notifications. E.g. if Google Now decides to push me a notification about the current weather, it better be sure that I’m planning to head out and face the elements in the next 30 minutes. Otherwise, the notification becomes an annoyance.

To get the notifications right, you really, really need to know the user well. Personal signals matter a lot more in this scenario.

Going by the relevance and quality of notifications that I get on my phone, I feel we are in the very early stages of figuring out how to send highly relevant, well-targeted notifications. It’s not surprising that up to 60% of people opt out of push notifications in some app categories. But that’s a topic for another day.