I’ve lately been thinking about how recommendation engines handle situations where a single account has multiple users with widely divergent preferences. For example, my family has a single Netflix account, but my husband and I have very different preferences. He likes “Back to the Future” and Alfred Hitchcock movies. I, on the other hand, like “Maid in Manhattan”, “Price & Prejudice” and “The Mummy”. Won’t this confuse recommendation engines?
An obvious solution is to base recommendations on categories. If someone has a preference for a romantic comedy such as ‘Maid in Manhattan”, show them another in the same category such as “My Best Friend’s Wedding”. This way, there are recommendations for the different users based on their categories of interest.
However, the disadvantage of only relying on recommendations within categories is that it does not leverage the richness you can get by having recommendations across categories. For example, people who prefer crime thrillers such as “The Bone Collector” may also like action movies such as “Terminator” and comedies such as “There’s something about Mary” Providing recommendations across categories provides many more choices and will delight users.
The problem is that basing these recommendations on data from multi-user accounts can confuse the recommendation engine. For example, based on preferences within my account, the engine could come up with “Maid In Manhattan” for someone who liked “Back to the Future”. But this would not work for a single user account owned by a woman who only liked “Maid In Manhattan”.
How to solve this problem?
Thumbs Down/ Thumbs Up: Letting users give the thumbs down or thumbs up to recommendations or categories of recommendations could allow the engine to give richer recommendations based on multiple categories while allowing some users to filter out irrelevant categories. In the above case, the woman with the single user account can give a Thumbs Down to “Back in the Future” (and its category) and never have to see similar recommendations again.
A better solution is:
Obtain multi-user information and tailor preferences: It would make sense for recommendation engines to ask if there is more than one user, and obtain the preferences of each user. This way, it can tailor recommendations for individual users. This way, for my husband, based on his interest in Alfred Hitchcock movies, it can show screwball comedy recommendations such as “There’s something about Mary”, while it would not do so for me. The UI will have to be adapted accordingly to support this.
In addition, it can also provide ‘linked’ recommendations by identifying patterns where users typically interested in romantic movies tend to have partners interested in action thrillers and sometimes tend to have children interested in Disney movies. Thus, it can make recommendations for one user based on their partner’s very different preferences.