Netflix is an entertainment company providing streaming media and online video-on-demand. Recently it has also branched into other businesses, such as film and television production.
As of 2016, Netflix had 75 million subscribers in more than 190 countries.
The Netflix recommender system presents users with choices regarding movies, documentaries, television programs and other items from its list of available content. The recommender system’s success is vital to Netflix’s business success, as it leads to greater viewership, higher retention of subscribing customers, and lower advertising costs.
The recommender system is used by paying customers who are searching or browsing for movies, documentaries and television programs in the list of such items available from Netflix.
The users’ goal is to watch programs that appeal to their tastes and preferences, and which they would presumably find sufficiently entertaining to justify the subscription costs. Netflix’s goals are to get more people to watch their content, reduce the time spent searching and browsing, and also presumably to promote newer content and the Netflix-produced content.
The recommendation engine in use by Netflix is one of its key strategic assets and is being constantly improved by its researchers and engineers. Some features of the recommendation algorithm, that I could discern from the references, are as follows.
Data used by Netflix to derive its recommendation results can be loosely classified into two categories. Explicit data is generated by users taking specific actions to indicate their preferences, such as assigning ratings to movies. Implicit data is obtained from behavioral data, such as how often a user watched certain types of content. The majority of useful data is implicit.
The recommendation engine tries to avoid only recommending content that is exactly identical to a user’s known (or computed) preferences, since doing so would result in restricting of the available choices. For instance, if a user only watches crime movies, the recommendation system will occasionally recommend a comedy feature as well, so as to indicate the broader choice of genres available, beyond what the user typically watches. In the final analysis, however, recommendations are governed by users’ preferences.
Before incorporating new features into its user interface, Netflix quietly runs “a couple hundred” product tests with 300,000 users each year, and learns from the feedback before broadly deploying those features (or choosing to not deploy them at all).
One of Netflix’s challenges, as it attempts to expand its worldwide reach, is to apply its machine learning algorithms even where there is absence of data pertaining to the preferences of the local population. Netflix attempts to address this problem by eliminating region-based data on preferences, in favor of global data sets on which its algorithms are trained. This is an attempt to solve the “cold-start” problem arising from lack of data related to users in a new region or territory. I think that this is an area of machine learning that is relatively new, and offers opportunities for innovation and differentation. It seems that making its recommendation engine successfully incorporate and generalize over linguistic, cultural and regional variations across the globe would make for a very broadly powerful and capable system.
My main recommendation, in view of the point raised previously, would be to develop and fine-tune more powerful algorithms capable of dealing with differences across the globe, and generalizing across data from different geographies.
Another recommendation would be to test, and possibly incorporate, a “movie trailer” type of feature, in the manner used by traditional movie theaters, where users could just watch a few minutes of carefully-produced short introductory clips of movies and other content.
http://www.wired.co.uk/article/how-do-netflixs-algorithms-work-machine-learning-helps-to-predict-what-viewers-will-like https://www.quora.com/What-are-the-key-differences-between-the-recommendation-engines-of-Amazon-Prime-and-Netflix http://www.businessinsider.com/how-the-netflix-recommendation-algorithm-works-2016-2