Multi-source knowledge fusion and learning

Real-world complex systems can be observed from many different angles or perspectives, and datasets collected from various perspectives often emphasize different types of features. This results in inconsistent beliefs about what is relevant to the system, how relevant features are related to one another, and what statistical properties these features possess. Many methods have been proposed to combine such diverse information sources. However, current algorithms only learn from each dataset separately and then combine individual outputs since this is easier to do with heterogeneous datasets with unknown feature correlations. This approach, although convenient and intuitive, cannot capture the logical linkages between various datasets. To understand variables’ interactions learned from datasets with noise and incompleteness, we are exploring algorithms that naturally fuse these datasets based on their shared variables and induce new variable relationships.

Faculty contact: Eugene Santos Jr.