In our previous blog, we gave you an overview of recommender systems. Now let’s look at what makes a good recommender system and how some of the biggest media and eCommerce giants are making use of these systems to stay ahead of their rivals.
Metrics for Recommender System
We described how recommender systems are essentially information filtering systems that aim to provide the most accurate product suggestions for users to choose from, that will result in improved sales and customer engagement. But being accurate is not enough for a good recommender system. It must take several other metrics into consideration.
The user must trust the recommender system first. Trust is built into the system by providing the user with clear explanations of how it generates recommendations and why it suggests an item.
Recommender systems usually tend to offer products based on analysis of past user interactions. They will recommend products that are like the items that the user has previously selected or shown a preference for. Now users can become bored or annoyed when they see the same suggestions and may switch off. To avoid this and improve user satisfaction, recommender systems can offer serendipitous recommendations – items that are not only relevant to the target user but are also different from the items that the user has rated, and hence serendipitous or surprising. This is particularly complex as the concept of serendipity is subjective and difficult to put into practice.
Privacy is a big concern with recommender systems as they build user profiles using sensitive information and collaborative filtering. GDPR rules enforce data privacy and the datasets used by businesses remain a growing area of concern. Even anonymized datasets can be analyzed to reveal user identities, which becomes problematic as it encroaches upon data privacy.
Labeling recommendations is important as user satisfaction has been shown to be influenced by it. Recent studies indicate that recommendations labeled as ‘sponsored’ see a lower click-through rate than others.
Now that you have an idea of some of the metrics that are crucial while developing a recommender system, here’s a look at some of the best use cases in the industry.
How Amazon uses Recommender Systems
Amazon’s recommender systems have been studied ever since the e-commerce giant began offering customers personalized recommendations years ago. In a widely-cited research paper, data scientists Greg Linden, Brent Smith, and Jeremy York explained how it changed the game with its algorithms.
Algorithms till then, were user-based and recommended items based on profiles of users with similar interests and purchase patterns. Linden, Smith, and York explained how Amazon developed an algorithm that evaluated items that were like each other. It looked at the user’s previously purchased or rated items and paired them with similar items using a series of metrics and then compiled a list of recommendations. This type of item-to-item collaborative filtering scales to very large data sets and produces high-quality recommendations.
The other factor that distinguishes Amazon’s use of algorithms is how it integrates recommender systems into almost every part of the shopping process, from product discovery to checkout. Multiple product recommendation panels offer the customer a spectrum of choices. It then follows with a sophisticated and highly effective recommendation system that sends more product suggestions to customers via email which also have a high conversion rate.
How Netflix uses Recommender Systems
Netflix’s complex recommender systems are among the most widely-discussed topics in tech circles. With a presence in more than 160 countries, it faced several challenges for its recommender algorithms to work equally well across the globe.
Broadly speaking, it provides recommendations based on factors including user interactions, ratings, communities of similar users, the time of day users watch, the devices that users are watching Netflix on, and how long users watch.
One of the ways that it generates recommendations is by grouping communities of members with similar show preferences. It will then suggest shows under the panel ‘Customers also watched’, based on popular choices within the community.
Data scientist Carlos Gomez-Uribe, who led Product Innovation at Netflix explains that this allows Netflix to cater to a global community without limiting their choices to the shows that are popular within a single geographical region. He cites an example of the Anime genre in shows – their data reveals that anime watchers are not restricted to Japan. In fact, only 10 percent of Netflix’s Anime community is in Japan, the rest are from all over the world. In this case, he writes, pooling data from all over the world helps Netflix improve recommendations.
Companies are constantly working on creating better recommender systems that will help deliver personalized suggestions for every customer. There will continue to be opportunities to add more personalization to anticipate the customer’s needs, which will unlock tremendous business value for organizations in the long run.