The project focuses on access to information via systems such as Twitter and FaceBook. Our goal is to improve the efficiency of information access via OSN’s in terms of precision/recall. To that end, we will develop solutions for recommending both content items and potential contacts to users.
Such recommender systems will be constructed from the following primitives:
Reward mechanisms allowing users to give feedback on specific content items;
Inference mechanisms that exploit the former reward mechanisms for characterizing i) types of specific content and ii) tastes and expertise of individual users.
The resulting recommender systems should be light-weight, accurate, and resilient to users’ strategic behaviors.
The expected outcomes of the project are
A deeper understanding of the limits of collective processing and filtering of information in terms of precision / recall trade-offs.
Optimized schemes for identification of implicit communities of like-minded users and contact recommendation for helping users “rewire” the information network for better performance. In particular robust versions of spectral embedding and message-passing algorithms à la Belief Propagation will be elaborated in this respect. Limitations / relative merits of candidate schemes, their robustness to noise in the input data, will be investigated. Distributed mechanisms for evaluating users’ expertise on particular topics. Such “expertise” evaluation can be multi-dimensional.
Active learning strategies for quick categorization of content types. In particular we will design rules for adaptively selecting users from whom to obtain feedback on content, aiming to achieve accurate content categorization based on the smallest possible number of user solicitations.
Protection against selfishness of participating users: proposed reward mechanisms should induce desirable content editing and filtering by the users when they are selfishly trying to maximize their overall rewards. . We will also aim to protect the system against the attack whereby groups of users conspire and distribute among themselves rewards similar to FaceBook’s “likes” thus artificially boosting their individual expertise. A particular scheme that we will consider is an instance of the so-called “marginal utility” reward mechanism whereby a user is rewarded by the utility to others that would have been lost if this particular user had not actively participated to the system.
Validation on available datasets and experimental deployment of some of the proposed mechanisms.
LennartGulikersMicrosoft Research-Inria Joint Centre (PHD Student)
IanKashMicrosoft Research Cambridge
Anne-MarieKermarrecInria Rennes Atlantique
PeterKeyMicrosoft Research Cambridge
ArnaudLegoutInria Sophia Antipolis Méditerranée
MarcLelargeInria Paris Rocquencourt
MesrobOhanessianMicrosoft Research-Inria Joint Centre (Post Doctoral Student)
RémiVarlootMicrosoft Research-Inria Joint Centre (PHD Student)
LaurentViennotInria Paris - Rocquencourt
KuangXuMicrosoft Research-Inria Joint Centre (Post Doctoral Student)
Publications HAL de laurent,massoulie;george,giakkoupis;lennart,gulikers;ian,kash;anne-marie,kermarrec;peter,key;arnaud,legout;marc,lelarge;remi,varloot;laurent,viennot;kuang,xu;mesrob,ohanessian; du labo/EPI 65197