When one uses sites like Netflix or Amazon, they keep recommending things to you based on your past searches and purchases. Most of the time, one can see why they recommended the items. But sometimes, they are real head-scratchers where the recommendation seems wildly off or to make no sense whatsoever, and today I experienced one such case.
When Marcus Ranum today reviewed the acclaimed 1969 film Z by director Costa-Gavras, it reminded me that this was a film that I had long meant to see but had somehow forgotten about until he reminded me. This film was released before the days of videotapes and DVDs which have essentially enabled people to have a movie theater in their own homes and so allowed people like me to see films that I missed during their first theatrical release.
I immediately went to Netflix and queued it up but what struck me was that when I did so, the site said that this film had been “recommended based on your interest in The Office (UK) and Monty Python’s Life of Brian”!
My reaction was, of course, “What?!” How could those two British comedies possibly suggest to Netflix that I might be interested in a Greek film of political intrigue? I could see no similarities at all that would explain the choice.
My guess is that the amazonian algorithm heurist is British, and sees “Z” as zed, and not zee, and therefore you must prefer british films.
I’ve never designed anything like this but if you want an amateur hour guess at how they’re using the info behind the scenes it probably entails a whole slew of tags put on the flims. Some of these are shown to help users find stuff but there are probably a whole lot of hidden tags. Some tags probably have more weight in the algorithm than others. From there you have two very basic possibilities.
1. It’s matching because it sees a significant number of people who watched one film went on to watch another.
2. It’s matching because there are enough tags in common to merit a match even if the genre is different.
#2 is the easier way to go. #1 is more of a “big data” approach.
Fun follow-up question: If the recommendation is due to significant numbers of users viewing both films, is the recommendation there because users search out and watch both films? Or did it get there by a fluke but stay there because it’s recommended and the greater visibility causes more views after the other film is watched? Working around this chicken and egg problem is why you need a complicated algorithm rather than one that simply matches tags and counts views.
If our esteemed host made an explicit request for Z and Netflix told him they’d figured out his interest by their own arcane calculations, it would seem their error lies in attributing every find to their file of his previous requests and disregarding his own input.
Prof. Singham may find it entertaining to ask them for a variety of titles remote from his previous selections and see what self-promotions they prefer to proffer.
There are clear affinities between Z and Life of Brian. Hero wants a better society, ends up getting killed for his efforts, etc. The Office is a stretch, but “horrible human beings in charge” could be a link.
I believe net flix uses collaborative filtering- I.e. If you and I watch the same things then if I have watched something that you do not have then it can be recommended to you . It’s generally item based I.e. What do people who watch item 1 and 2 also watch…
I once had Netflix recommend I watch Terminator due to my interest in movies with a strong female lead.
Amateur two cents’ worth here…
In looking for patterns in Netflix’s recommendations, one thing I have noted is the inclusion based on a starring actor or actress in common with previously viewed films.
In the case of Amazon, in addition to your demographic and purchase history data, previous purchases tell them specific items you own and have shopped for previously. These data certainly tell them what related items you are likely to purchase.
Pierce R Butler has it right:
You searched for it and queued it up explicitly, and they took the credit for recommending it? WT actual F? Does Trump run Netflix now?
@deepak shetty: about fifteen years ago I had a billion-dollar idea -- taste-dating. The idea worked in stages.
Step 1: sign up, and fill in a survey describing your tastes, in music, films, television, art, whatever.
Step 2: algorithm compares your tastes with those of other members. Then it points out that you have a 90% taste-match with person X. They love the things you love and hate the things you hate… except about 10% of their “love” list isn’t on yours. First revenue stream: for a paltry sum ($2?) we’ll simply give the list -- chances are good there’s something on there you haven’t discovered that you’ll love too, so it’s worth having that list.
Step 3: Second revenue stream: we’ll SELL you the stuff on the list (this was in the days of DVDs -- remember those?).
Step 4: Third revenue stream -- if you’re both in the market for it, for a slightly higher sum we’ll put you in touch with your taste-match. Even if you’re not a romantic match, you’ll at least be guaranteed tastes in common.
Never followed it up, obvs. It does look like some recommendation algorithms use at least some of this, though -- “People who bought this also bought” etc.
Another tidbit: None of the three films are “American”-made. “Foreign” films were often seen as a niche interest in the States in the past, and I presume still are.
The weirdest one I’ve ever had was just a couple of weeks ago. I enjoyed the independent comedy Obvious Child that about unplanned pregnancy and abortion (the movie is very much pro-choice). For some reason, because I watched it, Netflix thought I might like Left Behind (the Nicolas Cage version).
I think that Netflix simply assumes that users go by their recommendations when they select a film and retrospectively finds possible reasons for the choice. Their algorithms may have thrown up a list of probable cases and they picked the highest ones even though the probability of all of them were low.
At least this is what I am guessing, not really knowing how it works.
Ooh! I remember Z, now that you mention it. Good movie!
For interested folks
https://en.wikipedia.org/wiki/Netflix_Prize
@sonofrojblake
Heh -- these algorithms have been around from a long time so no billion dollars for you!
Amazon claims a good percentage (10+ if memory is correct) of its revenue is from those people who bought X also bought Y or people who viewed X also viewed Y kind of recommendations.
Two relevant articles:
https://www.businessinsider.com.au/how-the-netflix-recommendation-algorithm-works-2016-2
https://www.howtogeek.com/295952/how-to-get-better-netflix-recommendations/
In a word, for American viewers, all three films are appreciated by “intellectual” types, such as us. So, compared with most American Netflix viewers, their analysis was correct.