Parkinson's Disease (PD) is a neurodegenerative disorder that affects approximately one million Americans. Medications exist to manage the symptoms, but doctors must periodically adjust dosage level and frequency as a patient's disease progresses. These adjustments are typically based on observations made during short clinic visits, which provide an incomplete picture of a patient's daily quality of life. To provide a more accurate assessment of a patient's at-home experience, this thesis explores the use of wearable accelerometers for monitoring PD in the wild. A central challenge of this application is human variability, which makes it difficult for algorithms to generalize to \unseen" subjects who were not in the training set. This thesis investigates two major approaches to address this issue: (1) stratied algorithms for learning from in-the-wild data, and (2) personalization algorithms to actively adjust for human variability. A major challenge with in-the-wild data collection is obtaining labels of tremor symptoms. We developed a data collection protocol in which subjects used a cellphone app to log the approximate ([0-33%], [33-66%], or [66-100%]) amount of tremor experienced within short time segments and collected in-the wild data from six PD patients. We call these approximate percentages stratied labels, and we present a novel stratied weakly supervised learning technique that benefits from labels with this structure. Experiments demonstrated that this new technique leads to higher performance on this dataset compared to standard weakly supervised learning algorithms. Furthermore, stratied weakly supervised algorithms performed better on in-the-wild data than fully supervised algorithms trained on accurately labeled laboratory data. These findings suggest that collecting stratied labels from each subject in the wild is more effective than collecting accurate labels from other subjects in the laboratory. In summary, stratied learning provides a practical and efficient method for collecting weak labels in the- wild that greatly improves accuracy of person-specific PD tremor detection. This thesis also explores several unsupervised personalization algorithms, which either reweight the training data to align its distribution with the test data or learn a regression from distributions to classiers. We extended these personalization algorithms to the multiclass scenario. In our dataset, the personalization algorithms demonstrated modest improvement over non-personalized algorithms. We further analyzed the behavior of these algorithms with synthetic data, gaining insights into their limitations when learning appropriate weights. We addressed this issue by developing new constraints to the cost function, which improved performance on synthetic data, and experimented with global optimization.