Files
Abstract
To achieve high performance in a data analysis pipeline, complexity has to be expanded from data collection to the models' architectures, their evaluation, and interpretation. The deep learning models created from the data collected and integrated from heterogeneous but semantically related data sources can achieve excellent performance. The results, especially without expert feedback, may not be inherently interpretable. A model's results should be visualized and interpreted using Machine Learning in order to build trust since they provide insight into the reasoning behind complex automated systems. Explainable AI and interpretable Machine Learning aim to provide explanations, such as the significance of the key features contributing to a model's results. However, without domain experts' feedback, interpretations do not indicate whether the model is aligned with ground truth and domain knowledge, especially when the results are based on unstructured data such as images and text documents. In response to such uncertainty, we need to visualize data compilations, explain the models' decisions, and evaluate the results systematically. Despite progress in the research, it remains challenging to quantify the quality of interpretations systematically and measure their relevance with domain knowledge. In this research, we first discuss visualizing interpretations of predictions based on models fitted to data collected from heterogeneous data sources in the bioinformatics domain. Then, we propose a novel method for data gathering and integration to build a model that predicts whether a social media user propagates fake news and explains that by providing a mapping for the results, a text analysis tool's findings. Furthermore, we propose a method for systematically assessing a "black-box" text classification model's interpretation. Finally, we discuss the pipeline of creating a model-agnostic interpretation from heterogeneous data sources and guidelines for evaluating interpretations.