Finding latent relationships between different data sets

Object matching is an important task for finding correspondences between objects in different domains. It can be used for matching words in different languages, matching images and texts, and matching users in different databases. Existing object matching methods require paired data or similarity measures. However, these information might not be available because of cost or the need to preserve privacy. We propose an unsupervised object matching method that can find many-to-many matching without paired data and similarity measures by embedding all objects in a shared latent space. The proposed method enables us to analyze different data sets simultaneously, and it leads to discovery of new hidden relations and knowledge.

Please click the thumbnail image to open the full-size PDF file.