Nowadays, automatic data acquisition is frequently deployed in manufacturing and service applications. Naturally, increasingly emerging data-generating processes have been activated. However, the data collected from lately activated target processes are usually of high variability and low volume, which challenges the applicability of traditional statistical methodologies. Fortunately, abundant data is usually available in existing ancillary processes that have operated for a long time, some of which share similar statistical structure with the target processes. One potential bridge between these processes is statistical transfer learning, which can handle related target tasks by exploiting previously acquired knowledge from ancillary tasks. This thesis is devoted to extending statis...[ Read more ]

Nowadays, automatic data acquisition is frequently deployed in manufacturing and service applications. Naturally, increasingly emerging data-generating processes have been activated. However, the data collected from lately activated target processes are usually of high variability and low volume, which challenges the applicability of traditional statistical methodologies. Fortunately, abundant data is usually available in existing ancillary processes that have operated for a long time, some of which share similar statistical structure with the target processes. One potential bridge between these processes is statistical transfer learning, which can handle related target tasks by exploiting previously acquired knowledge from ancillary tasks. This thesis is devoted to extending statistical methodologies in the transfer learning framework. In the first essay, we establish state space models for several closely related stations simultaneously. The aim is to predict the inflow passengers volume in a certain time lead. In the second essay, different sensors are connected via common Bayesian prior distributions on the auto-regressive coefficients to implement knowledge transfer. Meanwhile, ordered lasso penalty is imposed to incorporate the sparsity patterns within the AR models. The final essay concerns the model selection consistency of penalized likelihood methods in statistical transfer learning, where the bound control of variable selection is provided for the target estimators. The three essays above propose novel and efficient procedures to solve statistical problems involved in processes of high variability and insufficient data, the superiority of which is verified in extensive Monte Carlo simulations and in real case studies.