Collaboration graph

Abstract: Social network-based Sybil defenses exploit the algorithmic properties of social graphs to infer the extent to which an arbitrary node in such a graph should be trusted. However, these systems do not consider the different amounts of trust represented by different graphs, and different levels of trust between nodes, though trust is being a crucial requirement in these systems. For instance, co-authors in an academic collaboration graph are trusted in a different manner than social friends. Furthermore, some social friends are more trusted than others. However, previous designs for social network-based Sybil defenses have not considered the inherent trust properties of the graphs they use. In this paper we introduce several designs to tune the performance of Sybil defenses by accounting for differential trust in social graphs and modeling these trust values by biasing random walks performed on these graphs. Surprisingly, we find that the cost function, the required length of random walks to accept all honest nodes with overwhelming probability, is much greater in graphs with high trust values, such as co-author graphs, than in graphs with low trust values such as online social networks. We show that this behavior is due to the community structure in high-trust graphs, requiring longer walk to traverse multiple communities. Furthermore, we show that our proposed designs to account for trust, while increase the cost function of graphs with low trust value, decrease the advantage of attacker.

Abstract: We consider the fully decentralized machine learning scenario where many users with personal datasets collaborate to learn models through local peer-to-peer exchanges, without a central coordinator. We propose to train personalized models that leverage a collaboration graph describing the relationships between user personal tasks, which we learn jointly with the models. Our fully decentralized optimization procedure alternates between training nonlinear models given the graph in a greedy boosting manner, and updating the collaboration graph (with controlled sparsity) given the models. Throughout the process, users exchange messages only with a small number of peers (their direct neighbors when updating the models, and a few random users when updating the graph), ensuring that the procedure naturally scales with the number of users. Overall, our approach is communication-efficient and avoids exchanging personal data. We provide an extensive analysis of the convergence rate, memory and communication complexity of our approach, and demonstrate its benefits compared to competing techniques on synthetic and real datasets.

Abstract: The entity disambiguation task partitions the records belonging to multiple persons with the objective that each decomposed partition is composed of records of a unique person. Existing solutions to this task use either biographical attributes, or auxiliary features that are collected from external sources, such as Wikipedia. However, for many scenarios, such auxiliary features are not available, or they are costly to obtain. Besides, the attempt of collecting biographical or external data sustains the risk of privacy violation. In this work, we propose a method for solving entity disambiguation task from link information obtained from a collaboration network. Our method is non-intrusive of privacy as it uses only the time-stamped graph topology of an anonymized network. Experimental results on two real-life academic collaboration networks show that the proposed method has satisfactory performance.

Abstract: In this work we study preference systems suitable for the Peer-to-Peer paradigm. Most of them fall in one of the three following categories: global, symmetric and complementary. All these systems share an acyclicity property. As a consequence, they admit a stable (or Pareto efficient) configuration, where no participant can collaborate with better partners than their current ones. We analyze the representation of such preference systems and show that any acyclic system can be represented with a symmetric mark matrix. This gives a method to merge acyclic preference systems while retaining the acyclicity property. We also consider properties of the corresponding collaboration graph, such as clustering coefficient and diameter. In particular, the study of the example of preferences based on real latency measurements shows that its stable configuration is a small-world graph.