Recent Improvements of the Recommender System

A recently written article by Netflix Inc. developers, cited sixty-one times to date, aims to explore improvement opportunities for the utilized recommender system. The approach is member retention and medium-term engagement, through A/B testing experimentation using historical member engagement data and new subscribers. A/B testing shows two or more variants of the home page shown randomly to users; then statistical analysis is used to determine which variation performs better according to the set goal.It discusses innovation areas based on their algorithms. Some technical gaps to provide more accurate predictions have appeared and can be bridged. Although the Netflix recommender system has worked and achieved their goals, even surpassing them concerning subscriptions a year, they have identified opportunities to attune the algorithms.

According to the authors (Gomez et. al. 2016), revenue is proportional to the number of members, and three processes directly affect this number: the acquisition rate of new members, member cancellation rates, and the rate at which former members rejoin. If Netflix creates a more compelling service by offering enhanced personalized recommendations, they improve retention hence, decrease churn. Also, users with an improved experience can become influencers to new subscribers with the word-of-mouth effect.

The idea behind continuous improvement is regularly test data with their large samples. Ultimately, impacting in a better user experience and actively support the complicated decision-making process. The algorithms are based on mathematical and statistical models such as classification, regression and dimensionality reduction through clustering or compression. The algorithms that define the Netflix experience in a matrix-like layout are:

1. Personal video ranker (PVR) and top-ranked, for each profile it orders the entire catalog of videos or subsets of the list selected by genre, likability filters and head of ranking.
2. Trending now, shorter-term temporal trends related to news or select dates (e.g., climate change or valentine’s day). These are powerful predictors of videos that members will watch when combined with personal data.
3. Continue watching, orders the videos based on signals of the time elapsed since viewing and the point of abandonment.
4. Video-Video Similarity, because You Watched (BYW), it anchors its recommendations to a single video watched by the member.
5. Page Generation uses the output of the previous algorithms as well as the diversity of the page. It can select and order rows from a large pool of candidates to create relevance and offer variety.
6. Evidence for the displayed items. For example, indicate that a particular movie won an Oscar or instead demonstrate to the user that the film is similar to another recently watched video.
7. Search: twenty percent of the users use this option triggered by their input; videos, actors, or genres in the catalog; retrieval leverages the information.

The study concludes that recommender systems will continue to play a critical role in using large amounts of data to drive easier choices, guiding people to the best few options, resulting in better decisions. Furthermore, recommender systems can democratize access to long-tail products, services, and information, because machines have a much better ability to learn from vastly bigger data pools than expert humans, thus can make useful predictions for areas in which human capacity simply is not adequate to generalize usefully at the tail.