Predictive Optimization with Zero-Shot Domain Adaptation

Prediction in a new domain without any training samples, called zero-shot domain adaptation (ZSDA) (Yang and Hospedales, 2015a,b), is an important task in domain adaptation. To this end, an approach to utilize domain descriptions (Yang and Hospedales, 2015a,b), called domain attributes, has been developed. A goal of ZSDA is to obtain predictions in an unseen domain in which we did not observe any training samples. An application of ZSDA is the sales prediction of new products; regarding domains as products and given product attributes and sales data, we can use ZSDA to the sales prediction of a customer for a new product. Thanks to ZSDA, we can predict the response of input in an unseen domain; however, one potential aspect of ZSDA has been overlooked. We demonstrate another potential of ZSDA; by reversing the ZSDA prediction process, we can optimize domain attributes so that an evaluation metric of responses over customers is maximized, referred to as attribute optimization as shown in Figure 1. That is, instead of predicting responses given new domain attributes as in ZSDA, our task is to find new domain attributes given a prediction.