The paradox of choice - Netflix
Some disclaimers: I don’t know anyone working at Netflix. The contents of this article based on public information and my intuition. Hopefully, some of it is interesting. Inevitably, some of the assumptions might be wrong.
I think I started using Netflix around four years ago. I didn’t have an account my own, and I was using one of my friend’s accounts. I could have taken an account, but I was a subscription-nazi. If my calculation is correct, I have spent almost 24 hours during these four years just scrolling through Netflix and deciding what to watch.
I opened my own Netflix account with my credit card around a month ago because I was bored due to corona. I genuinely had a lot of free time, which I wanted to spend lavishly binge-watching shows. I watched a lot of shows, and I added countless shows to my watch-list. And I decided to cancel my subscription at the end of the month.
I wanted to cancel the subscription because I realized that none of the shows makes me happy anymore.
Why did I cancel the subscription? Because Netflix’s suggestion system is not that great, or my preferences are pretty complicated for Netflix’s algorithm and failed to suggest matching shows.
Netflix is curating shows for me by analyzing my watching preference. I’m pretty sure that Netflix has a personality profile for me. But what kind of a person I am, when I watched “The Spy,” “Never have I ever,” “Russian Doll,” “The Half of it,” “Mindhunter,” and “Jerry Seinfeld: 23 Hours To Kill — NETFLIX COMEDY SPECIAL.” in a week.
What does this tell about myself? If I were to build a profile for myself for the shows I have watched, it would be like this — an easy-going person with a criminal mentality who likes pop-music and teenage shows. That’s so fucked up. As you can see, this sentence itself is a bit hard to digest, how Netflix’s algorithm can understand it?. It won’t, that’s why Netflix suggests a ton of irrelevant shows.
All of this comes to the paradox of choice or hick’s Law.
Hick’s Law predicts that the time and the effort it takes to make a decision increases with the number of options. The more choices, the more time users choose to make their decisions.
Netflix says, “Netflix has an extensive library of feature films, documentaries, TV shows, anime, award-winning Netflix originals, and more. Watch as much as you want, anytime you want.”
Effect of the paradox of choice and Netflix’s extensive catalog makes me scroll indefinitely instead of watching something.
Why can’t Netflix build a proper profile for me? Then I started to think more. Netflix has a personality graph based on your already watched shows, watch list, liked shows, skipped content while watching shows, and re-watched shows.
How Netflix build up this profile? To walk you through this, let me show you common assumptions we can make after watching some shows.
|The half of it||More than 5% likely to support LGBTQ community. Body affected by circadian rhythms.|
|Never have I ever||Chill. Easygoing. Likes Indian stuff. Likes brown color people. Likes teenage stuff.|
|Russian Doll||More than 5% likely to support Jewish people. Likes Dramedy. Most likely to watch cynical shows.|
|Descendants of the sun||Interested in South Korean people & culture.|
|Jerry Seinfeld: 23 Hours To Kill||Likes Jerry Seinfeld. Interested in Jerry’s other stand-up comedy shows, and stand-up comedy in general.|
Does this sound true to you? To me, yes, there is a 90% chance that all of this is true about myself.
Okay, now that you got an idea about how Netflix might be making assumptions about ourselves, let’s look into a detailed one.
|Mindhunter 1 & 2||Highly interested in violence & crime-solving. Casual with watching contents on prision, murder, and police investigation. Natural interest on dark content. Okay with dark, enigmatic background music.|
Each of these points has a specific score associated with it. This score depends upon the intensity of the content. Since Mindhunter is a show with most of the storyline revolves around crime-solving and criminal activities, we can most likely assign a score of 3 to Alvin instead of just 1.
If Netflix is were to build a public profile of our personality characteristics corresponding to our watch history, it might look like this. This score can be a combined score accumulated during a span of maybe one year by analyzing more than 50 shows and movies.
|Sexual Orientation||1 [ 0: Male, 1: Female ]|
|Okay watching LGBTQ content||3.0|
|Drug & abusive||0.5|
Let’s dive into some of the scenarios and try to define more characteristics. Let’s say that the user skipped a particular part while watching a show. Netflix can identify what that part was about and add an assumption based on that to the user’s personality profile.
|Intimate gay make out||Uncomfortable watching gay content|
|Raping||Uncomfortable watching violence|
|Romantic||Uncomfortable watching romantic/sexy content|
|Drugs, Killing, Violence||Uncomfortable watching violence|
|Politics, Court||Not interested in political content|
Let’s take a look at some other type of scenarios.
|Watched trailer, but didn’t watch the show||Didn’t like the show’s storyline or casting. Add this show’s storyline-keywords into user’s profile, and give less importance to those keywords while curating future shows.|
|Watched dubbed shows for sometime, and didn’t continue watching it||Not happy with dubbed shows, prefer native lip syncing|
Do you think Netflix knows your current emotion? Let’s take a look at how the Netflix can understand user’s current emotions.
|Age: Male, Sex: In twenties, Location: India. Time: Afternoon||Had lunch. Mostly tired. Needs to rest, circadian rhythm in effect.|
|Age: Teenager, Sex: Female, Location: New York. Time: 09:00 AM||Might be in her period. Might have lost someone due to corona outbreak. Most possibly sad.|
|Age: In thirties, Sex: Male, Location: San Fransico. Time: 08:30 PM.||If there is only one Netflix user profile, then most likely unmarried. Must be having a drink. Possibly a stressed day, needs to calm down.|
|Age: In thirties, Sex: Female, Location: London. Time: 11:30 PM.||Multiple user profiles including kid profiles. Might be watching the show with the partner in bed. Most possibly just had sex, looking for something easygoing and relaxing. Or didn’t have sex, looking for something cheesy and romantic. Or had an argument with the partner, just want to relax. So ideally ‘relax’ is the key component here.|
These are some typical example scenarios. These kinds of conditions can place easily and curate recommendations accordingly.
1. User Personalization
In the early versions of Netflix, there was a “Friends” feature. A user can link his Facebook account and see what his friends are watching. And the user can watch shows on Friends recommendations.
Logically speaking, this is correct; there is nothing wrong with getting suggestions & recommendations friends on what to watch.
But what you watch genuinely shapes who you are, then it’s better to give preference to the user’s intuition rather than friends’ suggestions.
That’s why Netflix removed this feature and gave more importance to user personalization.
2. Artwork Personalization
Have you ever noticed that Netflix has shown you different cover images at different times?. Yes, Netflix is personalizing the cover image in real-time.
I have noticed that Netflix changed cover images of “Russian Doll” while I was casually scrolling around. Maybe Netflix saw that I was scrolling around on a specific genre [ For example, Comedy ], and decided to show me a different Russian Doll cover image that has a fun personality. I am not sure if this is the right way to solve it, but this is ideal too.
To show proper cover images, Netflix has to learn & analyze users watching history and make models on top of that using machine learning. And this is a time-consuming process, and Netflix has to make changes available in production so that the user can see the new cover images. This process takes time while the user is waiting. To reduce the user’s waiting period, Netflix introduced dynamic machine learning or called it ‘online machine learning.’
For artwork personalization, Netflix uses a framework called contextual bandits. Rather than waiting to collect a full batch of data, waiting to learn a model, and then waiting for an A/B test to conclude, contextual bandits rapidly figure out the optimal personalized artwork selection for a title for each member and context. For personalization, the member is the context as we expect different members to respond differently to the images.
The finalized artworks by the contextual bandit framework then added to the A/B pipeline, which also works dynamically and known as ‘online A/B testing.’
By doing this, Netflix can figure out the optimal artwork and analyze the performance by calculating the number of user interactions on it. If it has more interactions, then ship it out to more regions with localization.
The image & and text we see on the artwork are also dynamically generated. Everything is dynamic. Netflix’s computer vision algorithms scan through the shows to find suitable images that have good focal points. Netflix says it has around 2 million photos; that’s a lot for a bunch of tv-shows.
Broadly, we know that if you don’t capture a member’s attention within 90 seconds, that member will likely lose interest and move onto another activity. Knowing we have such a short time to capture user’s interest, images become the most efficient and compelling way to help members discover the perfect title as quickly as possible. After all, the human brain can process images in as little as 13 milliseconds.
3. Multi user profiles
Netflix knows you where your relationship going 😱. If you are using a Netflix account that has more than 1 user profile, I am pretty sure that Netflix has a relationship score for these profiles.
For example, both kids and parents watch horror & violent content all the time on Netflix. What does this tell about the family in general?. I believe you got the point.
Let’s take a different example. There are two user profiles. Alice & John. Let’s consider possible relationship types between Alice & John.
- Brother & Sister
- Husband & Wife
- Boyfriend & Girlfriend
- Friend & Friend
Now, let’s consider some of the shows Alice & John watched in two weeks. And try to analyze each person’s emotions and try to come up with appropriate assumptions.
Scenario 1: Brother & Sister
|Brother & Sister||Shows|
|Alice||Never have I ever. Gossip girls. The half of it. Sex Education. Friends. How I met your mother. Young Sheldon.|
|John||Narcos. Black Mirror. Mindhunter. Ozark. Daredevil.|
Alice: A teengage girl who likes casual and cute stuff.
John: A teenage boy who likes horror & violence.
Assumption: Alice is open and extrovert. There is a chance that John is an introvert. Both fight a lot inside the house, because a John carries a lot of aggression even though he doesn’t know it.
Scenario 2: Husband & Wife
|Brother & Sister||Shows|
|Alice||Working Moms. Gossip Girl. You Me Her. Feel Good. Easy. Frieds From College.|
|John||House of cards. Desinated Survivor. Bodyguard. Messiah.|
Alice: Likes Fashion, and most likely to cheat in the relationshop. Mostly sexy & horny. Open to change sexual orientation [ lesbian ], don’t know yet. Working Mom. Fun.
John: Mostly interested in politics and religion. Sel-motivated. No direct interest or support on LGBTQ.
Assumption: Alice is open and extroverted. Sexy, cheesy, and easygoing. John is an introvert. John could be a professor. They fight less — a causal couple.
Similarly, we can come up with assumptions for Boyfriend/Girlfriend & Friend/Friend user profiles.
4. Popularity in personalization
What about popularity in personalization? Netflix doesn’t give much importance popularity instead focuses more on user’s behavior and watching preferences.
I would say no to popular shows. Currently, Netflix solves this by showing the top 10 popular shows in my country.
Netflix has a neat categorization of shows. These categories include “Casual Viewing,” “Cynical Shows”, “Dramedy”, and “Because you watched Mindhunter,” etc. Categories dynamically generated by adjusting user’s personality score by analyzing user’s watching preferences.
Also, Netflix suggests random categories like “International crisis” or “Life in China” to understand how you are interacting with these random categories.
If I select the “International crisis” category and watch a couple of documentaries on migrant-crisis, then Netflix can increase my score on my trait “Kind.”
Recently I watched a Korean tv-show called Descendants of sun. It was very long, each episode up to 60-minutes and around 26 episodes. When I finished the show and liked it, Netflix filled up my entire home screen with Asian shows. I was furious. I have no idea how Netflix could do this to me. These categories included “Indian shows,” “Chinese shows,” “Korean Romantic shows,” “Asian Action shows,” and “Top Asian shows,” etc.
Categorization should be calculated solely on user’s personalty score. Categorization is excellent, but when there are more, it’s overwhelming.
1. Skip - Sex, Violence, Drinking [ Warning alert ]
Netflix has a user’s personality score index. Let’s say that a user’s violence-tolerance score is 0.2. This score means that the user is not happy viewing violent content. If the user is watching a show and the show contains horrific violence at the 40th minute, Netflix can warn the user, saying that intense scenes are coming up on the 38th minute, and the user can skip this part if the user likes.
For example, The climax violent scenes in the movie Once upon a time in Hollywood.
2. Ask feedbacks
Currently, the users can only give feedback by clicking on the like or dislike button. But I might have more to say rather than clicking on these options. I would love to have a feedback loop, at least in Netflix’s mobile app. So that completing the show or movie or an episode, the user can write thoughts. This is a piece of great information, because these thoughts are specific to the user.
After watching an episode on Black Mirror, I might have so many things to say, this might include my likes, dislikes, and neutral content. For example, The episode was great, I didn’t particularly like the killing part, it was too much for me, I couldn’t take it. But I loved the mind-bending stuff.
If the Netflix team can run a pattern-matching and machine learning system on top of this, it’s easily understandable that the user is not happy to watch violence but interested in mind-bending casual shows.
Instead of keeping just like / dislike buttons bring in more smileys.
3. Treat unsubscribed member as guest user
I should be able to login to Netflix and see my watch-list and trailer of new & existing shows. But I won’t be able to watch any of these shows because I don’t have an active Netflix subscription.
There is a higher chance that the user will re-subscribe if we give the user a chance to scroll around. Psychology.
Since Netflix can’t fix the paradox of choice problem, I would like to have a blind-date option inside Netflix. Netflix will suggest 1-5 shows that have a matching percentage of more than 90%. And I must have to choose one of these shows from the list of 5. And I won’t be able to select other shows unless I change the settings.
5. Netflix wrapped
I love seeing my yearly listening history in Spotify. I know that I have watched a ton of shows and movies on Netflix. If Netflix can show me a list of my yearly watch list, it will be an excellent overview of my genre preferences and tastes.
And the user can click like / dislike button here if the user forgot to do that before.
6. Embedding Netflix
I usually watch Netflix trailer on YouTube. But I think Netflix should build it’s own embedding form so that other websites can directly embed Netflix’s form instead of YouTube’s form.
The idea behind doing is this that Netflix can add a button saying “Subscribe Now” with a 5% off on subscription plan. Users can directly click on the button and go to the website. But still, even without a discount price, Netflix gets unique attention.
7. We miss you
I really enjoyed Netflix’s WhatsApp reminder on newly arrived shows. I was able to dedicate time to watch these shows. But after unsubscribing from Netflix, I didn’t get any notification. I don’t know if this does sound correct, but I would love to get a text from Netflix, let’s say two or three times in a year saying that, “we miss you, please come back.” Make it sound more cute than formal.
These types of user behavior assumptions are typical, and the same logic applies to Spotify or youtube or Swiggy. Or any brand that has something to do the showing a bunch of stuff to the user to pick. Typically e-commerce or on-demand video streaming services.
Time moves forward
The user behavior attributes that Netflix has supplied to the online machine learning system won’t be applicable after three or four years. Humans are adopting new technologies and lifestyles so fast that businesses have to adapt new trends to understand the unique behavior of the users.
Would love to hear your thoughts and comments.