Database Reference
In-Depth Information
4.1 Experimental Setup
We ran our experiments on the PeerSim simulator 1 . We used three different
datasets: MovieLens , Flickr and LastFM . MovieLens dataset is composed of users
that rated movies. Flickr dataset is composed of users that submitted or added
a picture to their favorites. Each user also associates tags to the pictures he/she
submits. Finally, LastFM dataset is composed of users who listen and associate
tags to artists. Each dataset has different features, in particular users are more or
less redundant if the number of items per user is more or less respectively. The
characteristics of the datasets are summarized in the following table.
dataset items #items#usersavg items/user
MovieLens Movies
3 , 900
6 , 040
166
Flickr
Pictures 2 , 029
2 , 000
3 . 7
LastFM Artists
23 , 346
2 , 000
98
The queries used in the experiments consist of: In MovieLens , for each user, a ran-
dom subset of movies are shared and the rest are used as the queries to submit.
In particular, the words in the title are used as separate keywords. In Flickr and
LastFM queries are computed as the random association of several tags submit-
ted by a given user on a given item. An experiment is composed of two parts.
First, all users gossip during 400 rounds until convergence. Then, every 20 gos-
sip rounds all users submit one of their queries. The experiment stops at 500 gos-
sip rounds. We measure the average recall results. The recall enables to compute
the fraction of items that has been successfully recommended as presented in Sec-
tion 2. On the MovieLens dataset, the recall value is 1 if the movie has been found
and 0 otherwise. On Flickr and LastFM , the recall is the proportion of pictures in
the whole dataset that contains all query's keywords that have been returned to
the user. On the Flickr and LastFM experiments, we have computed the variance
which enables to compute the variability of the recall and is computed as follows:
V ( X )=1 /N × i =1 ( x i − m ) 2 where m is the average recall.
In our experiments, we use the following relevance scores:
overlap ( u , v )= |I u ∩ I v |
over big ( u , v )= |I u ∩ I v | + |I v |
Jaccard ( u , v )= |I u ∩ I v | / |I u ∪ I v | cosine ( u , v )= Ir u × Ir v / ||Ir u || × ||Ir v ||
where I u and I v are the items shared by u and v , respectively, and where Ir u
and Ir v are the set of ratings u and v gave to the items they share. We have fixed
the U-Net 's size to 16 and TTL to 3. Other values have been tested and showed
similar results. The size of the random view (5 in our case) is not important as
it only modifies the convergence speed.
4.2 Experiments
Figure 2 presents the results of our experiments. More precisely, Figures 2a, 2b
and 2c compare the recall results of the used relevance scores with and without
including our usefulness score, while Figures 2d, 2e and 2f compare the recall
results of several diversification methods.
1 www.peersim.sourceforge.net
 
Search WWH ::




Custom Search