Sequential and Temporal Analysis of Human-generated Data
Machine learning and data mining have the potential to provide meaningful solutions to many real-world problems that could impact society. This dissertation explores machine learning methods that utilize the sequential or temporal aspect of data to solve socially-relevant problems in two domains: one in education, the other in medicine.
In the domain of education and learning analytics, I present two different methods that help better understand students' online learning behaviors based on their clickstream data. First, statistical change detection is used to detect when and how students change their behavior during a course relative to the student population as a whole. I group the students depending on the type of changes and examine how the changes can be related to the course outcomes. The second method I explore is the use of probabilistic clustering to allow different types of temporal behavioral patterns to emerge from clickstream data. The resulting patterns are analyzed in the context of procrastination and time-management, demonstrating that the procrastinating behavior and course outcomes are highly correlated.
Secondly, in the medical domain, different materials and methods are discussed to test the feasibility of using machine learning models to obtain structured information from patient-physician conversations in primary care visits, as a means to potentially address the physician burnout problem. Several hundred dialog transcripts of doctor-patient conversations are used to predict topic labels for talk-turns. Different machine learning models are trained to operate on single or multiple talk-turns (logistic classifiers, support vector machines, gated recurrent units [GRUs]) as well as sequential models that integrate information across talk-turn sequences (conditional random fields, hidden Markov models, and hierarchical GRUs). From the results, I show that incorporating sequential information across talk-turns improves the accuracy of topic prediction in dialogs. Moreover, I examine the degree to which topic classification accuracy drops by adding an automatic speech recognition (ASR) system for transcription to the pipeline. A systematic evaluation is carried out by measuring the performance of ASR (with word error rate) and the downstream classification accuracy.
In the second part of my work in the medical domain, I investigate the use of machine learning methods for predicting emotional valence of both doctors and patients at the utterance level, based on transcripts of doctor-patient visits. This is one of the first large-scale investigations of emotional valence prediction at the utterance level in long medical dialogs. Using a variety of evaluation metrics, I show that current machine learning methods can achieve accuracies for this task that are close to that of human performance.