UC Irvine

There has been great interest and demand for individualized modeling and personalized prediction, with applications ranging from medicine to education programs and marketing. For example, it is well-known that subject’s heterogeneity exists in demographics, genetic characteristics, medical history, lifestyle, and even many unobserved attributes. Due to these heterogeneities, some subjects are expected to show more favorable treatment outcomes while others have adverse effects given the same treatment. This dissertation consists of two topics addressing heterogeneity on personalized modeling and predictions arising from recommender systems and individualized treatment regimes.

Recommender systems are information filtering tools that seek to match customers with products or services of interest. Most of the prevalent collaborative filtering recommender systems, such as matrix factorization and AutoRec, suffer from the “cold-start” problem, where they fail to provide meaningful recommendations for new users or new items due to informative-missing from the training data. To address this problem, we propose a weighted AutoEncoding model to leverage information from other users or items that share similar characteristics. The proposed method provides an effective strategy for borrowing strength from user or item-specific clustering structure as well as pairwise similarity in the training data, while achieving high computational efficiency and dimension reduction, and preserving nonlinear relationships between user preferences and item features. Numerical studies confirm the advantage in prediction accuracy of the proposed model compared to current state-of-the-art approaches.

Developing individualized treatment regimes is essential in precision medicine, yet still remains challenging since patients' responses to treatments could be heterogeneous and multifaceted. Existing studies mainly focus on optimizing treatment efficacy but ignore potential side effects and cost restrictions. Our target is to develop a treatment regime which maximizes efficacy while taking individualized risk control into consideration. We focus on a clinical study to design effective schizophrenia treatment (Stroup et al., 2003), and meanwhile control side-effect risk for each individual. In addition to the response heterogeneity and multifaceted consideration, this dataset contains multiple stages of treatments, which adds more challenges to the modeling of individualized risk control. Motivated by this application, in this paper, we consider a general setting to control non-negligible negative effects for each individual. Specifically, we propose a restricted outcome weighted learning method that optimizes efficacy outcomes while satisfying individual-level negative effect constraints. Our method is developed for multi-stage treatment decision problems that include single-stage decision as a special case. In particular, we propose an efficient learning algorithm that utilizes difference-of-convex algorithm and Lagrange multiplier to solve the nonconvex optimization with nonconvex risk constraints. We also establish theoretical properties including Fisher consistency and strong duality result for the proposed method. Extensive simulation studies and an application to an intervention effectiveness study for schizophrenia patients have demonstrated the superior performance of the proposed method over existing approaches.