Once we fall on the couch and click on the Lean-Back-Mode, then the selection of the right streaming content is really hard. What do you really want to do today? Luckily, there is a colorful mix of suggested videos, categories and your own watchlist that will help us now. Is there really only one more question: On which rules is this recommendation algorithm based?
The official version
The Netflix Help Center states, “The more you use Netflix, the better we can tailor our suggestions to your interests.” This statement initially suggests that the algorithm is trained solely by our vision and the ratings left to us to recommend more and more of the same. But this is not the case.
To some extent, individual films, series and documentaries do not really play such a big role – but the genres to which they are assigned. And Netflix does not just know simple genre titles like “sci-fi”. The keywording is so fragmented in sub-genres that even between “alien sci-fi by book template” and “criticized sci-fi from the 80s” is distinguished. All in all, the streaming service will boast an impressive 30,000 categories, which will be closely watched. Here you can even rummage through the codes for all the sub-genres.
Cluster analysis instead of true personalization
Of course, Netflix also knows that no one day by day sees through just one of these keywords. Tastes are finally diverse! And so the algorithm now begins with the actual work: With the help of a cluster analysis, it is looked at which groups of people prefer which genre mix. The data of all 150 million Netflix users are actually used for this – but evaluated only anonymously. After all, pattern recognition is no longer about the taste of the individual. Even if Netflix still acts as if every now and then:
Each user ends up in a crowd of stakeholders for whom content can be recommended based on identified streaming patterns. However, the work is not done yet with the display of a single movie or series within this genre mix. Now the algorithm tries to identify the best thumbnails with A-B tests. For example, groups that prefer to cast films with women in leading roles are more likely to be provided with motifs by Rachel, Monica and Phoebe in the series “Friends”. It does not matter if this leads to a false expectation – because it is not asked at all.
Data driven production decisions
Also, the development of new originals and the decision on their continuation is no coincidence. Because Netflix is just as data-driven approach. Producing what is (presumably) seen. The metrics collected are, for example, the amount of retrievals from which groups, the time intervals between the individual episodes or parts, as well as the preferred time of day, and which platform is used for viewing.
This detail obsession is not surprising, because successful originals determine whether a user will remain as a paying customer for a long time and thus on the success of the company. Correspondingly large is the tendency to mix originals among all recommendations. This approach is probably Netflix’s biggest challenge.
The end of chance
Because as we have already noted in our changes to the streaming landscape in 2019, the existing license agreements between Netflix and the major film studios are gradually phasing out. And once it’s done, Netflix will only have their own exclusive titles. But despite all the euphoria about successful series such as “Stranger Things” and films such as “Bird Box”: Are just under 320 titles (see list of all Netflix in-house productions on Wikipedia) sufficient in the end to cover an average-diverse taste?
According to Kelly Luegenbiehl (Netflix’s vice president of Creative International Originals), users browse through about 40 titles each time (!) Before deciding content. After eight streaming evenings but should not be over with the selection. Because then even the most sophisticated algorithm for recommendations does not help any further.