1 of 9

    • A world of data.
    • Recommender Systems are an effective filtering tool.
    • Built on the centralization of users’ data.
    • FL solutions emerged to solve this.

1

2 of 9

Server

Hundreds of millions of clients

2

  • Fault tolerance issues.

Federated Recommender Systems : challenges

3 of 9

Server

Hundreds of millions of clients

2

  • Fault tolerance issues.
  • Scalability concerns.

Federated Recommender Systems : challenges

4 of 9

Server

Hundreds of millions of clients

2

  • Fault tolerance issues.
  • Scalability concerns.
  • Lack of personalization.

Federated Recommender Systems : challenges

5 of 9

Empowering User-Centric Recommender Systems over Gossip Learning

Yacine Belal

Supervised by : Sonia Ben Mokhtar & Vlad Nitu

16th EuroSys Doctoral Workshop (EuroDW 2022)

3

6 of 9

4

A user-centric gossip learning approach

  • A gossip approach.

​

7 of 9

4

A user-centric gossip learning approach

  • A gossip approach.

​

  • An aggregation that maximizes the local performance.

​

8 of 9

  • A gossip approach.

​

  • An aggregation that maximizes the local performance.

​

  • Better neighborhoods.

​

4

A user-centric gossip learning approach

9 of 9

yacine.belal@insa-lyon.fr

  • A gossip approach.

​

  • An aggregation that maximizes the local performance.

​

  • Better neighborhoods.

​

4

A user-centric gossip learning approach

  • An improved average and tail performance.