The Effect of Correlation Coefficients on Communities of Recommenders

the effect of correlation coefficients oncommunities of recommenders

neal lathia, stephen hailes, licia capra

how do we model recommender systems?

a) machine learning

user ratings recommendationsmodel-based collaborative filtering

b) collaborative filtering

user ratings matrixrecommendationsmemory-based

collaborative filtering

how do we think about this?

collaborative filtering: a network of cooperating

usersexchanging opinions

nodes = userslinks = weighted according to similarity

community view of therecommender system:

0.75

-0.43

0.2

0.57

(a very small example)

or, put another way:

good

bad

good

good

(the relationships in the community)

the similarity values depend on how you derive

similarity

pearson:-0.50

weighted-pearson:-0.05

vector:0.76

= no agreement

ratings:[2,3,1,5,3]

ratings:[4,1,3,2,3]

pearson:bad

weighted-pearson:no similarity

vector:good

= no agreement

ratings:[2,3,1,5,3]

ratings:[4,1,3,2,3]

so what is the best way to build the recommender

system network?

like this?

good

bad

good

good

or like this?

bad

good

bad

good

or like this?

nosimilarity

good

good

bad

each way will change the distribution of values over

the network:

(let’s look at it on the movielens dataset)

pearson distribution:

other distributions:

a) accuracy: how well we can make predictions about

unknown items

what if we did this?

(random number)

(expect terrible results)

the results are far from terrible!

b) coverage: what proportion of items we can

find useful information about (to make predictions)

before:look for information from the top-k neighbours

(expect top-k to do quite well)

what if we did this?look for information from anyone who has rated the item

the results are terrible

(best coverage when all of community used)

why is all of this happening?

a) our error measures are not good enough?

N

rpMAE

iaia ,,

a) is there something wrong with the dataset?

…it does have the long-tail

c) is user-similarity not strong enough to

capture the best recommender relationships

in the network?

future: trust-based recommender systems

(neal’s phd)

the effect of correlation coefficients oncommunities of recommenders

neal lathia, stephen hailes, licia capra

all the details in the paper…