UMass Machine Learning and Friends Lunch | Main / Domain Adaptation Using Manifold Alignment

In many situations, we want to adapt the knowledge learned in a source domain for use in a target domain, where the source and target domains share some latent but unknown underlying structure. This problem arises in a variety of applications in information retrieval, e-commerce, computer vision, and many other areas, where we have a few labeled instances in the target domain, but plenty of labeled instances in the source domain.

In this talk, I will introduce a manifold alignment based approach for domain adaptation, which automatically results in mapping functions that project the data instances from source and target domains to a new latent space simultaneously preserving the topology of each domain and minimizing the distance between corresponding instances. This proposed approach differs from the state of the art approaches in that it addresses the problem of domain adaptation when the source and target domains do not share any feature. It is also data-driven, and does not require advanced domain knowledge. Real-world applications including cross-lingual information retrieval and cross domain ranking will be presented, providing results showing useful knowledge transfer from one domain to another.