Netlix Recommendation System
Catherine Cho
November 5, 2021
Perform a Scenario Design analysis as described below. Consider whether it makes sense for your selected recommender system to perform scenario design twice, once for the organization (e.g. Amazon.com) and once for the organization’s customers.
- Who are the target users of Netflix? Adults in the United States as of 2020 has been the largest market with more than 70 million subscriptions making up more than 50% of the US adult population. However, the market is expanding internationally as well with more than 200 million total paid subscribers accross 190 countries.
- What are the user’s key goals? The key goal is to have access to a wide range of tv shows and movies of different genres, all in one streaming platform at a competitive monthly rate. Additionally, the user seeks to have access to content that may be original and critically acclaimed.
- How can Netflix help the user accomplish this goal? Netflix will help the user accomplish these goals by making their content diverse in genre, offer new releases frequently, and offer original content that cannot be found on other streaming platforms.
Attempt to reverse engineer what you can about the site, from the site interface and any available information that you can find on the Internet or elsewhere.
Selection Layout:
Figure 1: Netflix Page Layout: Ranking of Rows
Based on my experience navigating through Netflix, the recommendation system shows titles in row column format. Each row represents the genre of type of content (TV show or Movie) and the movie is ranked from highest to lowest recommendation (left to right). Then each column is displayed from highest to lowest (top to bottom) of what most likely appeals to the viewer. I would imagine that this is intentional based on what humans find most aesthetically pleasing as well as for data collecting purposes of the recommendation system. Aesthetically pleasing because I recently learned that anthropologists can trace the human’s preference for rectangles and rows to prehistoric times. Prehistoric humans have surveilled their environment at eye level from left to right or vice versa and have associated this type of gathering of information to safety and refuge by being aware of their surroundings. Other than appeal, this type of organization is also useful for Netflix to gather data on what the viewer is interested in and not based on how the user scrolls through the rows and columns.
The following summarizes a few ranking algorithms that the Netflix recommendation system uses. - The first is the Personalized Video Ranking (PVR) and this is a “general-purpose” ranker that filters through categories of genre or type of content. This will provide the top level categories to output recommendations. - Then there is the Top-N Video Ranker, which will provide the “top picks” of each category mentioned in the PVR. Then this “top picks” list will be presented in its own row. - The third, is the Trending Now Ranker, which will look at the temporary trends, outside of the user preferences. - The fourth is the Continue Watching Ranker which determines the probability that a user will continue watching an unfinished movie or show. The variables considered is how much of the video the user completed, how long it has been since the user “abandoned” the video, and what device was used. For instance, seeing if a pattern exists between the types of device used and how often this user completes a video can reveal some insight to the probability of the user coming back to complete that video.
Each of these types of recommendation algorithms will undergo some kind of ranking process to generate machine learning results. The data collected from the way the user interacts with the different rows is useful for both online and offline processing to update the ranking of each title. This type of processing is best done simultaneously such that more complex computations can be performed offline and applied after each event, while quick and efficient computations can be performed and update the recommendations live. This type of machine learning is incremental and can build an architecture that is stable and reliable.
Include specific recommendations about how to improve the site’s recommendation capabilities going forward.
Another algorithm that could improve the recommender system would be categorize the artists, composers, directors, creators of each title and implement this information into the ranking system. Oftentimes, I find myself gravitating towards films with similar composers or directors. Perhaps this can be in its own row even but implementing into the bread and butter of the pool of titles considered could reveal more information of the user’s preferences.
Another soft recommendation would be implement the time viewed of trailers or previews to the ranking system. Since the trailers are so readily available meaning, a user just needs to hover their cursor over the title to play the trailer, this is information that could be extracted as potential interests. However, I’m sure this is something that is already part of their system.