High Dimensional Data Modeling Using Graphical Models

A defining feature of the current era is availability of massive data. This interest in data collection is thanks to methods that draw valuable insights and availability of automated responses to analytical results that bring down cost. Statistical methods and machine learning algorithms have added tremendous value to data that might have not been otherwise tapped into due to lack of resources. Models that distill and serve insights from data are being made commercially available and transforming our lives. Science has also been undergoing a transformation that welcomes data-driven approaches. While many disciplines are adopting a data-centric discovery approach, biomedical sciences have adapted and contributed early to data-driven research. Data produced by biomedical fields is huge and spans many levels of organism structure and function. The hopes and expectations from computational and data-centric methods are vast: computer-assisted diagnosis, drug discovery, personalized medicine/nutrition, lifestyle monitoring. However, challenges are equally grand. Biomedical data is high-dimensional, noisy, and incomplete. We propose data-driven models for biomedical problems. We approached these problems using probabilistic graphical models. First, we consider the problem of microRNA-gene interaction and interaction pattern discovery. A computational approach that is highly accurate is valuable given the huge number of possible pairings between microRNAs and genes and the high cost associated with verifying candidate interactions. We developed a non-parametric directed graphical model that was highly accurate in discovering these interactions. Then, we present our contributions to facilitate dietary monitoring. We use undirected graphical models to solve two problems in our efforts to build a smart mobile dietary application to track home made meals. These models were built to model food images and recipe texts and frequently seen ingredients. We compared deep learning architectures as the main image classifier and obtained highly accurate results. We incorporated co-occurrence frequencies using a conditional random field on the image predictions. We trained a highly accurate linear chain conditional random field to parse recipes to parse these recipes