As a heterogeneous type of data becomes an extremely popular NoSQL data model, more and more researchers have focused on the collection and statistical analysis of key-value data with local differential privacy (LDP). However, it is difficult to achieve a good balance between privacy and utility, because the key-value data has strong correlation between keys and values. In this paper, we study the problem of frequency and mean estimation on key-value data by proposing a user-centric key-value data collection scheme called UEKV-GRR. To address the privacy preferences of different users, a Privacy Budget-based Sampling (PBS) algorithm and Privacy Budget-based Sampling for Encoding Perturbation (PBS-E) algorithm are designed to provide personalized privacy protection based on different users’ privacy preferences. By theoretically analyzing the estimation error of different sampling methods, we select sampling method adaptively. Thus, the accuracy and usability of frequency estimation and mean estimation can be improved on the server side. Experimental results show that our proposed scheme UEKV-GRR is superior to existing schemes in accuracy.