Recommender Systems
COMPSCI 753 Kaiqi Zhao
Overview
¡ì What is recommender system (RS)?
¡ì Recommendation algorithms ¡ì Content-based methods
¡ì Collaborative filtering ¡ì Latent factor model
¡ì Advanced topics in recommendations ¡ì Context-aware recommendations
¡ì Session-based recommendations
1
What is recommender system?
¡ì Recommender systems are a way of suggesting like or similar items for users.
¡ì Recommender systems try to automate the process of finding other people with similar tastes and then ask them to suggest new things.
¡ì Recommender systems can change the traditional way that a user obtain information
2
Search v.s. recommendation
¡ì Search
¡ì A passive way for information to be
retrieved
¡ì You know what you are looking for
¡ì Recommendation
¡ì Active way of information delivery
¡ì You can discover wonderful things that you even don¡¯t know them!
3
Why do we need it?
¡ì Information overload
¡ì Data gets bigger than ever
¡ì People are facing more data to consume. How to choose?
¡ì Some statistics
¡ì According to IBM – 2.5 quintillion bytes of data are produced every day ¡ì Statista (March 2019):
¡ì YouTube users watch 4.5 million videos every day ¡ì More than 390K apps downloaded every day
¡ì More than 347K user scrolling on Instagram
4
The value of recommendations
¡ì Recommender systems can largely increase the revenue of the online business
¡ì Netflix: 2/3 of the movies watched are recommended
¡ì Google News: 38% more clickthrough rate from recommendations ¡ì Amazon: 30% sales from recommendation
[Statistics from Xavier Amatriain]
5
Industrial recommender systems
Social & News
LBS Services
Entertainment
E-Commerce
6
A brief trace of recommender systems
1992 —
Goldberg proposed Tapestry, a personalized mail recommender system based on collaborative filtering
1994 —
Resnick proposed Grouplens, a collaborative filtering system for news.
1996 —
The concept of ¡°recommender system¡± was established
2006 —
Netflix Prize, a 1 million competition on recommendation algorithms
2007 —
The first ACM conference of recommender system (RecSys)
2016 —
Google proposed the Wide & deep system for scaling the recommender system to big data setting
7
The recommendation problem
¡ì Predict how a user likes an item based on
¡ì Similar items to the past consumed items
¡ì Items that are consumed by others with similar preference ¡ì Context, i.e., session, cart
¡ì A simple setting – rating prediction
I1
I2
I3
I4
U1
5
?
?
3
U2
?
4
?
?
U3
3
?
2
?
U4
?
3
?
2
U5
?
5
?
3
Explicit feedback
Ratings, Liked, Favorite
Implicit feedback
Click, Browse, Stay Time
8
Challenges in RS
¡ì Key problems ¡ì Sparsity:
¡ì The user-item interaction matrix is often extremely sparse
¡ì Millions of items, but a user can only interact with small number of items
¡ì Cold-start:
¡ì New items have no ratings yet
¡ì New users have no historical behavior
Item cold
U1
U2
U3
U4
U5
I1
5
?
3
?
?
I2
?
4
?
3
5
I3
?
?
2
?
?
I4
3
?
?
2
3
?
U6
?
?
?
?
?
User cold
I5
?
?
?
?
9