Exploiting Diversification in Gossip-Based Recommendation - Data Management in Cloud, Grid and P2P Systems - page 35

Figures 2d, 2e and 2f show that all diversification methods enable to increase
the recall values compared to undiversified methods. Among them, usefulness
obtains the best gain in terms of recall closely followed by xQuad . Finally, MMR
shows the worst gain in terms of recall. Indeed, MMR chooses users that minimize
the maximum similarity between any two users in u 's U-Net . Therefore, it prefers
users that are a little bit similar with every user in u 's U-Net , and that do not
necessarily increase recall results.
5 Related Work
Distributed recommendation for web data based on collaborative filtering has
been recently proposed with promising results. In this section, we compare our
recommendation approach with state of the art solutions.
In [15], Loupasakis and Ntarmos propose a decentralized approach for social
networking with three goals in mind: privacy, scalability with profitability and
availability. They propose an architecture based on a DHT for keywords query
search. Since, DHTs are better suited for exact-match queries, the author pro-
pose to decompose each query into several single word exact-match queries. The
main drawback is that responses that have medium scores with respect to each
keyword but high scores with respect to all the keywords are likely to be missed.
P2PRec [3] is a gossip-based search and recommendation solution where the
profile of each user u is represented as a set of topics computed based u ”s items.
Then, using gossip protocols, similar users in terms of topics, are clustered to-
gether and used to guide recommendation as we do. However, since diversity is
not taken into account, users within each cluster can be redundant, thus limiting
recall results. In [6], Kermarrec et al. focus on recommendation and propose to
combine gossip algorithms and random walks. The users are clustered based on
relevance through gossip protocols. A user has knowledge of the items shared by
its neighbors. To compute the recommendation, each user runs locally a random
walk using a transition similarity matrix. However, the computational complex-
ity of the algorithm with respect to the size of the neighborhood and the number
of items. reduces the complexity of the approach. Moreover, Kermarrec et al. [7]
claim that, since users are heterogeneous, the similarity measure used to cluster
users should also be heterogeneous. Nevertheless, the concept of diversity is dif-
ferent from ours as it represents the usage of various relevance scores depending
on each user”s profile. As a consequence, each user”s cluster may still carry re-
dundant user profiles, because there is no explicit diversification. In [1], Bai et
al. propose a solution for personalized P2P top-k search in the context of col-
laborative tagging systems, called P4Q . In this solution, the users are clustered
based on relevance through gossip protocols. The users in each cluster are split
into two groups: 1) the c closest users from which u replicates all items metadata
( i.e. tagging actions) and 2) the n less similar users from which u knows only
the profile. Still, diversity is not taken into account and users within the clusters
are likely to be redundant.