Transfer Learning for Individualized Prediction

Transfer learning provides an effective strategy for improving prediction accuracy when functional responses need to be inferred from limited data. However, existing transfer learning mainly focuses on population-level learning in a process rather than a specific experimental unit. This talk presents two complementary frameworks for individual unit-level transfer learning from stochastic process and neural process perspectives. The first part introduces a transfer learning formulation for stochastic kriging, a Gaussian process–based model designed for functional responses with heterogeneous replications and non-i.i.d. noise. Conventional stochastic kriging focuses on estimating the mean functional relationship of a process, which may be insufficient when the interest lies in predicting the response curve of a particular replication supported by only limited observations. The proposed framework addresses this limitation by modeling both within-process and between-process correlations through a tailored covariance structure, enabling information from data-rich source processes to support prediction in a data-scarce process. The second part extends this idea to neural processes, which combine neural networks with stochastic process modeling to represent distributions over functions. A destination-centric architecture transfers information from source processes while integrating observations from the destination process, improving predictive accuracy and uncertainty quantification under severe data scarcity. Both methods are validated using numerical studies and real-world data from engineering processes.