Recommender systems have long played indispensable roles in e-commerce by helping users find their preferred products efficiently and accurately. In recent years, some pioneering works have shown that users’ intrinsic characteristics, such as personality traits, can help improve the performance of recommender systems. However, they generally do not consider how to obtain users’ personality traits effectively without burdensome surveys and, therefore, cannot be used in new application domains without adequate personality information. Therefore, we propose a novel framework for building cross-domain personality-based recommender systems, especially for personality-scarce target domains. Specifically, we define the cross-domain personality trait classification problem and solve it in a semisupervised manner by leveraging the predictive text embedding method as the method for transfer learning from the source to the target domain. We then design a personality-boosted probabilistic matrix factorization method for personality-based recommendations. Extensive experiments conducted on five real-world datasets demonstrate that users’ personality traits can be recognized more precisely with cross-domain transfer learning, and recommendation performance is improved accordingly.