6. es un poco
diferente.
(por ejemplo, en google, cuando investigas algo, necesitas saber
exactamente lo que quieres buscar…) un sistema de recomendación
es menos concreto…
25. ¿qué es el root-mean-squared error?
• El RMSE es la raíz cuadrada del
promedio de errores cuadrados.
• El efecto de cada error en el
RMSE es proporcional al
tamaño del error cuadrado;
• Por lo tanto, los errores
mayores tienen un efecto
desproporcionadamente grande
en el RMSE.
• Por lo tanto, el RMSE es
sensible a los valores atípicos.
26. the performance was measured in terms of
what is called root-mean-squared error.
27. Collaborative Filtering es
un método para hacer
predicciones automáticas
(filtrar) sobre los
intereses de un usuario
mediante la recopilación
de preferencias o
información sobre el
gusto de muchos
usuarios (colaboración)
Data Science for Business - Foster Provost Tom Fawcett (2013)
You see, in Google,
when you research something, you need to know exactly what you want to look for.
Google search is helping me find answer to the question.
When I know precisely what question I'm trying to and get answer.
Recommendation system is less concrete.
It's personalized depending on the person who is searching for the answer.
But it is necessity.
Take for example of YouTube.
You know how much content that's from YouTube generates?
Every minute, here's the statistics I came up with.
Every minute, 300 hours of video is uploaded on YouTube.
So let's hypothetically assume that you and
I are going to 100% of our time continue watching YouTube.
That it will take a million lifetime for
us to watch content generated in one lifetime.
In a sense, recommendation system is essential to discover the content that we
care about on YouTube otherwise we would not be able to deal with it
No, no, we're just scratching the surface, it's tip of an iceberg.
https://blog.gotinder.com/powering-tinder-r-the-method-behind-our-matching/
https://www.hackerearth.com/blog/algorithms/elo-rating-system-common-link-facemash-chess/
>> Wow, so it looks like recommendation systems really are everywhere.
Any other aspects that we should discuss?
>> Actually, it might be fun to discuss that Tinder example again.
Remember, we are Tinder veterans, without real Tinder accounts.
Recall the movie Social Network a la Facebook, Zuckerberg.
Remember the game of Hot or Not?
So, in an interactive world,
where individuals are available to engage fully, one way to
think of a recommendation system is like playing a game of 20 questions.
I want to know what you are thinking, and
I want to know that by asking as few questions as possible.
>> Absolutely,
it's like playing the game Guess Who against the Gary Kasperov of Guess Who?
He guesses right in a record number of questions.
>> Bingo, this is exactly how a recommendation can work
in an interactive situation.
Later in this module we will discuss a case study related to selecting beers and
gathering opinion of people using comparison like questions and answers.
https://blog.gotinder.com/powering-tinder-r-the-method-behind-our-matching/
https://www.hackerearth.com/blog/algorithms/elo-rating-system-common-link-facemash-chess/
>> Wow, so it looks like recommendation systems really are everywhere.
Any other aspects that we should discuss?
>> Actually, it might be fun to discuss that Tinder example again.
Remember, we are Tinder veterans, without real Tinder accounts.
Recall the movie Social Network a la Facebook, Zuckerberg.
Remember the game of Hot or Not?
So, in an interactive world,
where individuals are available to engage fully, one way to
think of a recommendation system is like playing a game of 20 questions.
I want to know what you are thinking, and
I want to know that by asking as few questions as possible.
>> Absolutely,
it's like playing the game Guess Who against the Gary Kasperov of Guess Who?
He guesses right in a record number of questions.
>> Bingo, this is exactly how a recommendation can work
in an interactive situation.
Later in this module we will discuss a case study related to selecting beers and
gathering opinion of people using comparison like questions and answers.
In 2006, the state of the art Netflix system
had a score of 0.95 and the winning team
dropped this number down to 0.85 after three years of hard work.