‘Recommended for you’: where does it come from?

Online recommendations based on algorithms generate a series of suggestions for items (products) that users might be interested in, and are based on previous choices. Clicking and reading a science piece generates recommendations for other articles from the science genre, for example. This way, users extend their engagement with the platform’s content.

Product attributes or customer preferences

Many companies make recommendations in a similar way. Today’s recommender systems frequently adopt a hybrid approach that uses common attributes across products as well as common preferences across customers. However, companies are free to choose how to explain the algorithm. Some, like the New York Times, emphasise common attributes across products, for example the New York Times uses ‘More in Science’, and Spotify uses ‘Similar to [what you have listened to]’. Other companies focus on the overlap in customers’ preferences, for example Amazon uses ‘Customers who viewed this item also viewed… and Netflix uses ‘Customers also watched…’.

Researchers Phyliss Jia Gai and Dr Anne-Kathrin Klesse investigated which of the two ways of explaining recommendations is more effective at triggering clicks; item-based or user-based. Customers are told which information has been used to make recommendations, either matching the attributes of the product, or recommending something chosen by other customers who also chose the first product. Item-based framing matches products by their attributes, whereas user-based framing matches products by their consumers.

A guarantee of good taste

Importantly, user-based framing also tells users that the recommendation is based on matching their tastes with other similar users with shared interests, delivering a sort of ‘double guarantee’.

But does user-based framing outperform item-based framing? Does it produce more click-throughs? To find out, the researchers conducted two field studies in WeChat, China’s top social media app.

Many WeChat users subscribe to daily news articles from media accounts. The researchers collaborated with a media company that publishes popular science articles and summaries of academic research on WeChat. Each day they embedded a pair of recommendations at the end of the focus article; one recommendation from user-based framing – what other people like; and one recommendation from item-based framing – more of the same kind of product or similar product.

More click-throughs

In both studies, the researchers found that user-based framing (e.g., “People who like this also like”) increased the click-through rates of recommended articles compared to item-based framing (e.g., “Similar to this item”). When the WeChat subscribers were asked about their understanding of the two kinds of recommendations, they said both types suggest product matching as the basis for recommendations, but they recognised that user-based framing also signals taste matching. This confirms that user-based framing provides additional information to customers.

However, customers do not always see taste matching as successful. If consumers perceive the suggested taste match is inaccurate, then user-based framing has no advantage over item-based framing – it can even become disadvantageous.

Successful taste matching has a couple of critical contributory factors: how much experience customers have already accumulated with that type of product, and the presence of other users’ profiles.

Customers with more experience tend to develop more refined tastes, and see their own tastes as distinctive. As a result, they find it harder to believe their tastes can be accurately matched with those of others when based on a single item.

Personal profiles can backfire

Other users’ profiles are sometimes displayed alongside product recommendations, but this tactic backfires if customers perceive themselves as different to other users. The researchers show that dissimilarity cues like age and gender make people think their tastes are different to those of other users; it makes customers avoid the user-based recommendations.

The researchers’ findings suggest that not all explanations are equally effective. Using a user-based explanation rather than item-based one can make the same recommendation much more effective. It can even drastically increase customers’ tendency to click on the recommendation. Importantly, adapting this explanation costs almost nothing so it becomes an effective tool that can help companies maximise returns in recommender systems.

The researchers also highlight situations in which user-based framing is more effective than item-based framing – and when it becomes disadvantageous. By leveraging these findings, managers can tailor the framing of their recommendations for different customers and products and so boost click-through rates.