Recommender Systems
QuickTime™ and a
TIFF (Uncompressed) decompressor
are needed to see this picture.
>1,000,000,000
Finding Trusted Information
How many cows in Texas?
http://www.cowabduction.com/
Outline




What are Recommender Systems?
How do they work?
How can we integrate social information /
trust?
What are some applications?
Netflix
Amazon
How do they work?



Two main methods
Find things people similar to me like
Find things similar to the things I like
Example with People
St ar Wars Jaws
A lice
Bob
Chuck


4
5
3
The
Godf ather
3
3
2
5
2
2
Who is a better predictor for Alice?
Compute the correlation:



5
3
4
Wizard of
Oz
Bob 0.26
Chuck: 0.83
Recommend a rating of “Vertigo” for Alice



Bob rates it a 3
Chuck rates it a 5
(0.26 * 3 + 0.82 * 5) / (0.26+ 0.82) = 4.5
2001
1
1
2
Item similarity

Methods are more complex
Computed using features of items
E.g. genre, year, director, actors, etc.

Some sites use a very nuanced set of features


How good is a Recommender System?

Generally: error



Error = My rating - Recommended Rating
Do this for all items and take the average
Need alternative ways of evaluating systems


Serendipity over accuracy
Diversity
Trust in Recommender Systems




If we have a social network, can we use it to
build trusted recommender systems?
Where does the trust come from?
How can we compute trust?
Some example applications
In-Class Exercise
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
Your Favorite Movie
Your Least Favorite Movie


Some Mediocre Movie
Some Mediocre Movie
Some Mediocre Movie
Some Mediocre Movie
Some Mediocre Movie
Some Mediocre Movie
Some Mediocre Movie
Some Mediocre Movie








Sample Profile
Movie
Jaws
A Clockwork Orange
North by Northwest
Peeping Tom
The Godfather
The Matrix 2
Elf
Gone with the Wind
Madagascar
1408
Your Rating User 7's Rating Difference
10
1
10
1
7
4
8
7
4
3
4
7
4
7
6
5
7
8
3
4
Knowing this information, how much do you trust
User 7 about movies?
6
6
6
6
1
1
1
1
1
1
Factors Impacting Trust




Overall Similarity
Similarity on items with extreme ratings
Single largest difference
Subject’s propensity to trust
Propensity to Trust
160
Number of Ratings
140
120
100
80
60
40
20
0
1
2
3
4
5
6
Trust Value
7
8
9
10
Advogato

Peer certification of users




Master: principal author or hard-working co-author of
an "important" free software project
Journeyer: people who make free software happen
Apprentice: someone who has contributed in some
way to a free software project, but is still striving to
acquire the skills and standing in the community to
make more significant contributions
Advogato trust metric applied to determine
certification
Advogato Website




http://www.advogato.org/
Certifications are used to control permissions
Only certified users have permission to
comment
Combination of certifications and interest
ratings of users’ blogs are used to filter posts
MoleSkiing





http://www.moleskiing.it/ (note: in Italian)
Ski mountaineers provide information about
their trips
Users express their level of trust in other users
The system shows only trusted information to
every user
Uses MoleTrust algortihm
FilmTrust



Movie Recommender
Website has a social network where users rate
how much they trust their friends about movies
Movie recommendations are made using trust

Recommended Rating = Weighted average of all
ratings, where weight is the trust (direct or inferred)
that the user has for the rater
Challenges



Can these kinds of approaches create
problems?
Recommender Systems - recommending items
that are too similar
Trust-based recommendations - preventing
the user from seeing other perspectives
Conclusions




Recommender systems create personalized
suggestions to users
Social trust is another way of personalizing
content recommendations
Connect social relationships with online
content to highlight the most trustworthy
information
Still many challenges to doing this well