A Brief Description of my PhD Research

Datum–wise Learning and Inference in Machine Learning: In many real–world applications (e.g., medical diagnosis, insurance recommendation) a data example is described by a set of related labels (e.g., physical activity and emotion, driving quality and accident severity). However, these variables are not directly observable in practice but can be inferred via noisy but costly features. At the same time, access to all features is prohibitive due to cost, invasiveness, or limited resources. On the other hand, example–wise adaptive decision–making is essential and required in many situations (e.g., person–centric diagnosis). To address these challenges, I have developed a novel feature acquisition and classification algorithm that selects the most informative features for classification in a structured environment. The proposed method outperforms existing classification and feature selection methods in terms of accuracy and the average number of features used. 

Moreover, it is essential to identify causal relationships between variables (e.g., physical activity and emotion) to reason on meaningful example–wise inferences. Toward this goal, I have developed an algorithm for learning underlying relationships between variables as a Bayesian network. Specifically, the proposed method speeds–up the variable relationship identification without compromising accuracy by sequential evaluation of each relationship. Still, an accurate estimate of variables is essential in datum–wise decision–making. To this end, note that traditional supervised classification relies on a single classifier that employs all of the available features to determine the classification outcomes of all examples in a dataset. However, the importance of features may differ across the entire dataset, while different classifiers may result in distinctively different classification outcomes. Therefore, going beyond standard classification, to further increase the accuracy, I have devised a dynamic feature and classifier selection strategy. The proposed approach improves overall accuracy by forwarding challenging examples to one of the powerful classifiers while balancing the average number of feature usage. In addition, the datum–wise nature of the proposed algorithms provides valuable insights into the underlying patterns of the data, enabling model interpretability, which is essential in real–world applications.

Moreover, in many applications, multiple experts (e.g., radiologists, primary care doctors) collaborate in addition to relying on feature observations (e.g., medical history, imaging scans). This is because, in critical situations, when arriving at a final diagnosis, collective decision ensures comprehensive and accurate medical decision–making. Similarly, in tandem with human experts, we observe a rapid advancement of Artificial Intelligence (AI) experts today. Thus, I believe collective yet accurate decision–making is important while balancing the associated usage costs for features and also experts. In such a setting, I have developed an algorithm to improve accuracy in tandem with real–world applications where multiple experts are involved in the example–wise decision–making.