Objective
- Leveraging cross-domain dataset to overcome the cold-start and data sparsity problem, which is common in POI (point of interests) recommendation problems
- Protecting user privacy while preserving utility of data from the auxiliary domain
Challenges
- No existing privacy protection model for cross-domain recommender system
- Decoding obfuscated data to improve the accuracy of recommendation requires novel decoding algorithm designs
Solutions
- Generalized a powerful location-privacy preserving model, geo-indistinguishability, to restrict obfuscation within the same category of locations.
- Calculated a confidence score for each obfuscated interaction to reflect on what scale the collective matrix factorization (CMF) trusts each record.
Publication
Accepted by IMWUT 2019 as a full paper [link].