CAREER: Autonomous Tensor Analysis: From Raw Multi-Aspect Data to Actionable Insights

Real-world entities are often described by multiple aspects in data. For instance, a news article online can be expressed by its textual content, the images it may contain, its author, its publication date, and the number of times it has been shared in social media. Integration of those aspects has been shown to be beneficial in a number of data science tasks, such as extracting the topic of an article or ascertaining its trustworthiness.

Tensor decomposition, a class of multi-aspect data analytic methods, has been empirically shown to be particularly effective in such integration in a number of diverse applications, including: chemometrics, signal processing, social network analysis, and brain data analysis. Despite their effectiveness, current tensor methods have a crucial limitation that hinders their broad applicability: conducting tensor analysis is largely a manual endeavor that entails laborious trial-and-error tuning and requires both high familiarity with tensor methods and domain expertise in the target application.

As a result, the number of practitioners who can successfully apply tensor analysis is limited and, thus, tensor analysis has not enjoyed as broad an adoption despite being such a powerful tool. The goal of this project is to democratize the entire, currently laborious and highly inaccessible, process of unsupervised exploratory tensor analysis, towards producing actionable insights from raw multi-aspect data.

Towards broad convergence of research, practice, and education, the project will integrate the research outcomes to the investigator's educational and outreach activities. Those activities include the introduction of research outcomes to the development of undergraduate and graduate data science curricula, organization of workshops and tutorials at major research venues for disseminating the research outcomes to the scientific community, mentoring undergraduate students from underrepresented groups, and organizing summer workshops for teachers and capstone projects for students in collaboration with the local school district towards broadening the presence of data science in high school education.

The research activities are organized in two major tasks: (i) Algorithms for unsupervised autonomous tensor analysis and (ii) Real-world applications, in collaboration with domain experts. Towards democratizing tensor decomposition, in the first task, the project will be the first to formulate in a principled manner a number of very challenging problems which have traditionally been tackled manually and which are essential in unsupervised tensor analysis.

The project will develop methods of self-supervised learning in order to learn tensor datasets with exploitable and meaningful structure from raw data, solving a fundamental problem in data preparation for tensor analysis. The project has parallels with the emerging fields of meta-learning and auto-ML, which have recently gained traction in the data science and machine learning communities.

However, the focus of those fields has overwhelmingly been to automate the fine-tuning of supervised models. The lack of supervision in this endeavor, however, poses major challenges and presents a huge opportunity: success of the project has the potential to create a new field of unsupervised exploratory meta-learning where the goal is to extract actionable and interpretable insights from raw data in an unsupervised manner.

In order to maximize the potential benefit to the academic community, industry, and society at-large, and to stress-test the proposed methods on a wide array of diverse data (e.g., text, images, graphs, time-series, and scientific data), the second task of the project will focus on two major applications: (i) Misinformation detection on the Web: The project will provide the public with tools to better judge the trustworthiness of news online, towards more informed and active citizenry. (ii) Gravitational wave detection: the study of gravitational waves can help unravel current mysteries of the universe and analyze cosmic objects that do not emit light. The project aims at improving the detection and analysis of Gravitational Waves, with the potential to empower scientific discovery towards further understanding the universe.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.