Cross-media Age Regression with Textual Adaptation

In realistic scenarios, an age regression model learned from one social media (named the source media) generally performs rather poorly when it is tested on another social media (named the target media). In this paper, a textual adaptation approach is proposed to cross-media age regression which aims to improve the age regression performance by exploiting textual features in the labeled data from the source media and unlabeled data from the target media. The basic idea to achieve this lies in the fact that many textual features are shared by the data from both social media. Specifically, two views generated by random subspace generation (RSG) are leveraged to train two separate regressors in a co-training algorithm for adding automatically-labeled samples in the target media. Moreover, we tackle the confidence evaluation challenge in co-training by the query by committee (QBC) approach. Empirical studies demonstrate the effectiveness of the proposed approach to cross-media age regression.