Learning transferable representations by topic guided graph adversarial network

To reduce the need for costly labeled data in a target domain, it is desirable to learn transferable representations from a related source domain, where it is feasible to obtain labeled data. Facing distribution shift between different domains, one often aligns the marginal feature distributions by performing feature-level adversarial learning. Despite the recent success, existing approaches often ignore the consistency between feature and label, which is a challenging problem since there is no supervision in the target domain. Hence, we are motivated to transfer the discriminative information from the source to the target domain to realize better domain alignment. To this end, we propose a topic-guided graph adversarial network (TGAN), composed of a graph constructor, a graph feature extractor, a domain discriminator, and a classifier. Specifically, to learn a graph describing the relational structure among samples from different domains, we propose a semantic disentangled topic model to extract domain-shared and domain-specific topics, so that the graph edges can be defined by the sample similarities in a domain-shared semantic space. Then, TGAN aggregates the discriminative characteristics of source nodes and propagates them to the target nodes by attentive message passing through the graph, with the final node embeddings used for adversarial learning between source and target domains. TGAN has achieved the state-of-the-art performance on sentiment classification and clinical risk prediction tasks. Moreover, the discovered domain-invariant discriminative topics show interpretable meanings, which is beneficial to understanding the prediction results, especially for biomedical researches.