A key purpose of building a model from clinical data is to predict the outcomes of future individual patients. This work introduces a Bayesian patient-specific predictive framework for constructing predictive models from data that are optimized to predict well for a particular patient case. The construction of such <i>patient-specific models</i> is influenced by the particular history, symptoms, laboratory results, and other features of the patient case at hand. This approach is in contrast to the commonly used <i>population-wide models</i> that are constructed to perform well on average on all future cases. The new patient-specific method described in this research uses Bayesian network models, carries out Bayesian model averaging over a set of models to predict the outcome of interest for the patient case at hand, and employs a patient-specific heuristic to locate a set of suitable models to average over. Two versions of the method are developed that differ in the representation used for the conditional probability distributions in the Bayesian networks. One version uses a representation that captures only the so called <i>global structure</i> among the variables of a Bayesian network and the second representation captures additional <i>local structure</i> among the variables. The patient-specific methods were experimentally evaluated on one synthetic dataset, 21 UCI datasets and three medical datasets. Their performance was measured using five different performance measures and compared to that of several commonly used methods for constructing predictive models including naïve Bayes, C4.5 decision tree, logistic regression, neural networks, k-Nearest Neighbor and Lazy Bayesian Rules. Over all the datasets, both patient-specific methods performed better on average on all performance measures and against all the comparison algorithms. The <i>global structure</i> method that performs Bayesian model averaging in conjunction with the patient-specific search heuristic had better performance than either model selection with the patient-specific heuristic or non-patient-specific Bayesian model averaging. However, the additional learning of local structure by the <i>local structure</i> method did not lead to significant improvements over the use of global structure alone. The specific implementation limitations of the local structure method may have limited its performance.

Private states (mental and emotional states) are part of the information that is conveyed in many forms of discourse. News articles often report emotional responses to news stories; editorials, reviews, and weblogs convey opinions and beliefs. This dissertation investigates the manual and automatic identification of linguistic expressions of private states in a corpus of news documents from the world press. A term for the linguistic expression of private states is subjectivity. The conceptual representation of private states used in this dissertation is that of Wiebe et al. (2005). As part of this research, annotators are trained to identify expressions of private states and their properties, such as the source and the intensity of the private state. This dissertation then extends the conceptual representation of private states to better model the attitudes and targets of private states. The inter-annotator agreement studies conducted for this dissertation show that the various concepts in the original and extended representation of private states can be reliably annotated. Exploring the automatic recognition of various types of private states is also a large part of this dissertation. Experiments are conducted that focus on three types of fine-grained subjectivity analysis: recognizing the intensity of clauses and sentences, recognizing the contextual polarity of words and phrases, and recognizing the attribution levels where sentiment and arguing attitudes are expressed. Various supervised machine learning algorithms are used to train automatic systems to perform each of these tasks. These experiments result in automatic systems for performing fine-grained subjectivity analysis that significantly outperform baseline systems.

"Social Navigation" for the Web has been created as a response to the problem of disorientation in information space. It helps by visualizing traces of behavior of other users and adding social affordance to the information space. Despite the popularity of social navigation ideas, very few studies of social navigation systems can be found in the research literature. In this dissertation, I designed and carried out an experiment to explore the effect of several factors on social navigation support (SNS). The purpose of the experiment was to identify situations under which social navigation is most useful and to investigate the effect of personal factors, e.g., interpersonal trust, and gender on the likelihood of following social navigation cues. To gain a deeper insight into the effect of SNS on users' information seeking behavior, traditional evaluation methodologies were supplemented with eye tracking. The results of the study show that social navigation cues affect subjects' search behavior; specifically, while under time pressure subjects were more likely to use SNS. SNS was successful in guiding them to relevant documents and allowed them to achieve higher search performance. Reading abilities and interpersonal trust had a reliable effect on the SNS-following behavior and on subjects' subjective opinion about SNS. The effect of the gender was less pronounced than expected, contrary to the evidence in the literature.

Bayesian modeling of unknown causes of events is an important and pervasive problem. However, it has received relatively little research attention. In general, an intelligent agent (or system) has only limited causal knowledge of the world. Therefore, the agent may well be experiencing the influences of causes outside its model. For example, a clinician may be seeing a patient with a virus that is new to humans; the HIV virus was at one time such an example. It is important that clinicians be able to recognize that a patient is presenting with an unknown disease. In general, intelligent agents (or systems) need to recognize under uncertainty when they are likely to be experiencing influences outside their realm of knowledge. This dissertation investigates Bayesian modeling of unknown causes of events in the context of disease-outbreak detection. The dissertation introduces a Bayesian approach that models and detects (1) known diseases (e.g., influenza and anthrax) by using informative prior probabilities, (2) unknown diseases (e.g., a new, highly contagious respiratory virus that has never been seen before) by using relatively non-informative prior probabilities and (3) partially-known diseases (e.g., a disease that has characteristics of an influenza-like illness) by using semi-informative prior probabilities. I report the results of simulation experiments which support that this modeling method can improve the detection of new disease outbreaks in a population. A key contribution of this dissertation is that it introduces a Bayesian approach for jointly modeling both known and unknown causes of events. Such modeling has broad applicability in artificial intelligence in general and biomedical informatics applications in particular, where the space of known causes of outcomes of interest is seldom complete.

In this dissertation, I explore roles of visual working memory, global perception and eye-movement in complex visual problem solving. Four experiments were conducted and two models were built and tested. Experiment one and model one showed that global information plays an important role and there is an interaction between external representation and internal VWM on global information representation. Experiment two and model two showed that this interaction is achieved by encoding global information with eye-movements throughout the duration of solving a problem. A very regular eye-movement pattern is observed in experiment two. Experiment three further tested the hypothesis that this eye-movement pattern is a result of the individuals VWM limitation by measuring the correlation between individual differences in the quantitative features of the eye-movement pattern and VWM size. The second model assumes that global and local information share a unified VWM capacity limitation. In the fourth experiment, I tested this hypothesis along with several alternative hypotheses. Results of the fourth experiment support the unified capacity hypothesis best and thus make a complete story for the interaction between VWM, global information processing and eye-movements in complex visual problem solving. Even with such a limited amount of VWM capacity, human visual cognition is able to solve complex visual problems by keeping a balanced amount of global and local information in VWM. This balance is achieved by eye-movements that encode both types of information into a unified VWM. Thus, although VWM has such a limited capacity, through frequent eye-movements, visual cognition is able to encode complex visual information in a temporal manner. At each instance, the amount of information encoded is limited by the capacity limitation of VWM but the global information encoded can further guide eye-movements to acquire information that is needed to make the next decision.

A user simulation is a computer program which simulates human user behaviors. Recently, user simulations have been widely used in two spoken dialog system development tasks. One is to generate large simulated corpora for applying machine learning to learn new dialog strategies, and the other is to replace human users to test dialog system performance. Although previous studies have shown successful examples of applying user simulations in both tasks, it is not clear what type of user simulation is most appropriate for a specific task because few studies compare different user simulations in the same experimental setting. In this research, we investigate how to construct user simulations in a specific task for spoken dialog system development. Since most current user simulations generate user actions based on probabilistic models, we identify two main factors in constructing such user simulations: the choice of user simulation model and the approach to set up user action probabilities. We build different user simulation models which differ in their efforts in simulating realistic user behaviors and exploring more user actions. We also investigate different manual and trained approaches to set up user action probabilities. We introduce both task-dependent and task-independent measures to compare these simulations. We show that a simulated user which mimics realistic user behaviors is not always necessary for the dialog strategy learning task. For the dialog system testing task, a user simulation which simulates user behaviors in a statistical way can generate both objective and subjective measures of dialog system performance similar to human users. Our research examines the strengths and weaknesses of user simulations in spoken dialog system development. Although our results are constrained to our task domain and the resources available, we provide a general framework for comparing user simulations in a task-dependent context. In addition, we summarize and validate a set of evaluation measures that can be used in comparing different simulated users as well as simulated versus human users.

In this dissertation, I investigated applying a form of machine learning, reinforcement learning, to induce tutorial tactics from pre-existing data collected from real subjects. Tutorial tactics are policies as to how the tutor should select the next action when there are multiple ones available at each step. In order to investigate whether micro-level tutorial decisions would impact students' learning, we induced two sets of tutorial tactics: the ``Normalized Gain' tutorial tactics were derived with the goal of enhancing the tutorial decisions that contribute to the students' learning while the Inverse Normalized Gain ones were derived with the goal of enhancing those decisions that contribute less or even nothing to the students' learning. The two sets of tutorial tactics were compared on real human participants. Results showed that when the contents were controlled so as to be the same, different tutorial tactics would indeed make a difference in students' learning gains. The Normalized Gain students out-performed their ``Inverse Normalized Gain' peers. This dissertation sheds some light on how to apply reinforcement learning to induce tutorial tactics in natural language tutoring systems.

This thesis investigates whether reflection after tutoring with the Itspoke qualitative physics tutoring system can improve both near and far transfer learning and retention. This question is formalized in three major hypotheses. H1: that reading a post-tutoring reflective text will improve learning compared to reading a non-reflective text. H2: that a more cohesive reflective text will produce higher learning gains for most students. And H3: that students with high domain knowledge will learn more from a less cohesive text. In addition, this thesis addresses the question of which mechanisms affect learning from a reflective text. Secondary hypotheses H4 and H5 posit that textual cohesion and student motivation, respectively, each affect learning by influencing the amount of inference performed while reading. These hypotheses were tested by asking students to read a reflective/abstractive text after tutoring with the Itspoke tutor. This text compared dialog parts in which similar physics principles had been applied to different situations. Students were randomly assigned among two experimental conditions which got ``high' or ``low' cohesion versions of this text, or a control condition which read non-reflective physics material after tutoring. The secondary hypotheses were tested using two measures of cognitive load while reading: reading speeds and a self-report measure of reading difficulty. Near and far transfer learning was measured using sets of questions that were mostly isomorphic vs. non-isomorphic the tutored problems, and retention was measured by administering both an immediate and a delayed post-test. Motivation was measured using a questionnaire. Reading a reflective text improved learning, but only for students with a middle amount of motivation, confirming H1 for that group. These students also learned more from a more cohesive reflective text, supporting H2. Cohesion also affected high and low knowledge students significantly differently, supporting H3, except that high knowledge students learned best from high, not low cohesion text. Students with higher amounts of motivation did have higher cognitive load, confirming hypothesis H5 and suggesting that they engaged the text more actively. However, secondary hypothesis H4 failed to show a role for cognitive load in explaining the learning interaction between knowledge and cohesion demonstrated in H3.

The challenge of translation varies from one sentence to another, or even between phrases of a sentence. We investigate whether variations in difficulty can be located automatically for Statistical Machine Translation (SMT). Furthermore, we hypothesize that customization of a SMT system based on difficulty information, improves the translation quality. We assume a binary categorization for phrases: easy vs. difficult. Our focus is on the Difficult to Translate Phrases (DTPs). Our experiments show that for a sentence, improving the translation of the DTP improves the translation of the surrounding non-difficult phrases too. To locate the most difficult phrase of each sentence, we use machine learning and construct a difficulty classifier. To improve the translation of DTPs, we introduce customization methods for three components of the SMT system: I. language model; II. translation model; III. decoding weights. With each method, we construct a new component that is dedicated for the translation of difficult phrases. Our experiments on Arabic-to-English translation show that DTP-specific system customization is mostly successful. Overall, we demonstrate that translation difficulty is an important source of information for machine translation and can be used to enhance its performance.

Biomarkers are a critical tool for the detection, diagnosis, monitoring and prognosis of diseases, and for understanding disease mechanisms in order to create treatments. Unfortunately, finding reliable biomarkers is often hampered by a number of practical problems, including scarcity of samples, the high dimensionality of the data, and measurement error. An important opportunity to make the most of these scarce data is to combine information from multiple related data sets for more effective biomarker discovery. Because the costs of creating large data sets for every disease of interest are likely to remain prohibitive, methods for more effectively making use of related biomarker data sets continues to be important. This thesis develops TRL, a novel framework for integrative biomarker discovery from related but separate data sets, such as those generated for similar biomarker profiling studies. TRL alleviates the problem of data scarcity by providing a way to validate knowledge learned from one data set and simultaneously learn new knowledge on a related data set. Unlike other transfer learning approaches, TRL takes prior knowledge in the form of interpretable, modular classification rules, and uses them to seed learning on a new data set. We evaluated TRL on 13 pairs of real-world biomarker discovery data sets, and found TRL improves accuracy twice as often as degrading it. TRL consists of four alternative methods for transfer and three measures of the amount of information transferred. By experimenting with these methods, we investigate the kinds of information necessary to preserve for transfer learning from related data sets. We found it is important to keep track of the relationships between biomarker values and disease state, and to consider during learning how rules will interact in the final model. If the source and target data are drawn from the same distribution, we found the performance improvement and amount of transfer increase with increasing size of the source compared to the target data.