"What AI Can Teach Us About Human Bias in Decision Making: A Brief History of Netflix Personalization"

"What AI Can Teach Us About Human Bias in Decision Making: A Brief History of Netflix Personalization"

The collective power of the weak is mightier than that of the strongest individual. If only things were that simple. We are all human. We have limited knowledge. We may understand others well if they have had similar life experiences to ours. Human psychology tricks us, too. It makes us want to be around people who are like us — people with whom we can sympathize. That is all very innocent until you consider the cost to society. 'The lack of women's participation in clinical trials is the result of a long-held belief that the male perspective is the norm' Gatekeepers, including doctors, surgeons, policymakers, and medical researchers, are predominantly male. Their viewpoints have become entrenched in the way we approach medicine. If you do not actively pursue diversity, the status quo will remain. That is just how decisions are made. The way we see the world is shaped by our personal experiences. Observable traits, such as race, ethnicity, gender, and age, are the starting point of diversity of thought.

However, the most compelling explanation can be found in the field of artificial intelligence and the algorithms associated with it. An approach called the ensemble method deserves mention here. It essentially harnesses cognitive diversity in the world of AI. The ensemble method works by dividing an existing dataset into small samples, which can then independently train different algorithms at the same time. In these smaller sample sets, these algorithms tend to be weaker. They make more errors in their predictions on their own. The wonderful thing is that you can aggregate their results. The collective insight, or total wisdom of these algorithms is far superior to that of a single algorithm. In fact, it is better than a single algorithm that is trained by all the data in one go. There are two key reasons for the power of this collective: firstly, algorithms tend to make different types of errors, and secondly, they learn from each other and improve as they go along. This is what computer scientists call co-evolution. The main idea is that a diverse set of algorithms will have superior performance. This is because of their complementary strengths and weaknesses. And they shine especially when making predictions in the new, unseen data.

Just as with machine learning, humans need to actively seek out diverse perspectives of people who are wired differently because of their alternative life experiences. Only then can we build a better, more accurate picture of the truth. Furthermore, we can make better decisions for society when we understand things better. Explore and exploit are not mutually exclusive choices. Companies must have both. Still, in industries that are turbulent and changing quickly, companies tend to explore more. 'If you do not actively pursue diversity, the status quo will remain. That is just how decisions are made.' In the second graph, we plot how different industries learn. The X-axis shows how oriented an industry is towards learning. Again, we see that pharmaceutical and technology industries are more learning-oriented than other industries. They are more likely than others to pay more attention to diversity.

Now, let's shift gears and take a look at the history of Netflix's personalization journey. In 2007, Netflix made a groundbreaking move by launching its streaming service. This marked the beginning of a new era in entertainment consumption. However, they soon realized that simply offering a wide range of movies and TV shows wasn't enough. They needed to personalize the user experience to keep customers engaged.

In 2007, Netflix introduced the "Netflix Prize," a competition that offered a $1 million reward to any team that could improve the predictive power of their collaborative filtering algorithm by 10%. This sparked a wave of innovation and exploration into the world of recommendation systems. Contestants discovered that recent ratings provided more predictive power than older ratings, highlighting the importance of timeliness in personalization algorithms.

The Next Big Netflix Prize came in 2009, with the goal of further improving the algorithm. However, when Netflix executed the new algorithm in a large-scale A/B test in 2010, there was no measurable retention difference. This was a disappointing result, but it taught Netflix an important lesson about the complexity of personalization and the need for continuous experimentation.

In 2011, Netflix introduced its Movie Genome Project, a new algorithm called "Category Interest." This algorithm allowed Netflix to suggest movies to users and provide context for why they might like them. It was a major step forward in personalization, as it went beyond simply recommending movies based on previous ratings.

Netflix's personalization approach has three components: a forced-rank list of titles for each member, an understanding of the most relevant filters for each member, and the ability to understand the most relevant rows for each member. This multi-layered approach ensures that Netflix can tailor its recommendations to individual tastes and preferences.

Over the years, Netflix has continued to refine its personalization algorithms and test new approaches. They have even ventured into original content, using their knowledge of member tastes to make data-driven decisions on what shows and movies to produce. This has given them a significant advantage in right-sizing their content spend and maximizing viewer engagement.

In 2016, Netflix conducted an A/B test of the five-star rating system against a thumbs up/down system. The result was surprising: the simpler thumbs system collected twice as many ratings. This highlighted the importance of simplicity in user interactions and the need to continuously iterate and optimize the personalization experience.

In 2017, Netflix made another significant change by replacing star ratings with a "percentage match" system. This shift was driven by the insight that star ratings do not necessarily reflect a user's enjoyment of a movie. The percentage match system focuses on predicting how much a user will enjoy a movie, regardless of its quality.

Looking ahead, Netflix has a long-term vision for personalization. They aim to eliminate both the "Play Something" button and their personalized merchandising system. Instead, they envision a future where the perfect movie for a user's mood will automatically begin playing. This ambitious vision is fueled by their commitment to understanding and predicting member preferences.

So, what can we learn from AI and Netflix's personalization journey? Here are three actionable pieces of advice:

1. Actively seek out diverse perspectives: Just as AI algorithms benefit from cognitive diversity, humans need to actively seek out diverse perspectives to make better decisions. Embrace diversity in your teams and encourage different viewpoints to build a more accurate picture of the truth.

2. Continuously experiment and iterate: Netflix's personalization journey is a testament to the power of continuous experimentation. Don't be afraid to try new approaches, learn from failures, and iterate on your existing algorithms or processes. Personalization is an evolving field, and staying stagnant is not an option.

3. Keep simplicity in mind: Netflix's shift from star ratings to a simpler thumbs up/down system highlights the importance of simplicity in user interactions. Aim for streamlined experiences that make it easy for users to engage with your personalization efforts. Complexity can be a barrier to adoption and engagement.

In conclusion, AI and Netflix's personalization journey offer valuable insights into human bias in decision-making and the power of diversity. By harnessing cognitive diversity and continuously refining algorithms, we can make better decisions and create personalized experiences that delight users. The future of personalization holds immense potential, and by embracing these lessons, we can unlock new possibilities for innovation and growth.