ABSTRACT

High-order interaction events of multiple entities are ubiquitous, ranging from online advertising to commodity recommendation, from neural-signal transduction to gene regulation to disease spreading to international affairs. For example, online shopping behaviors are interaction events between customers, products and selling platforms. This project develops flexible, interpretable, and scalable Bayesian embeddings for massive high-order interaction events, in order to understand a variety of complex relationships between the events and discover the underlying rich patterns. The developed tools can fundamentally promote many important knowledge mining and prediction tasks. Examples include predicting the occurrence of hazardous online transactions to enhance financial security, predicting the outbreak and spreading of pandemic diseases to take effective preventive actions, early warnings of catastrophes, studying when and how rumors propagate through online social media, etc.


Current approaches for event data analysis are mostly restricted to binary interactions, and suffer from rough, over-simplified or opaque, uninterpretable modeling with limited computational efficiency. The goal of the project is to develop Bayesian embeddings that can efficiently process tremendous batch and fast streaming event data, capture both the static relationships of the entities and a variety of short-term, long-term, triggering, inhibition, and time varying influences among the events, and encode all of these into embedding representations to uncover rich temporal patterns. The research will be accomplished through four primary tasks: (1) using marked point processes to design highly expressive yet transparent Bayesian embedding models, (2) using variational transforms and composite Monte-Carlo approximations to fulfill stochastic mini-batch gradient and asynchronous stochastic learning on extremely large-scale batch data, (3) efficient posterior incremental learning for rapid event streams, and (4) comprehensive evaluations on synthetic and real-world applications. Moreover, using Bayesian frameworks, the developed tools are resilient to noises, provide posterior distributions to quantify uncertainties, and integrate all possible outcomes into robust predictions. The contribution is expected to dramatically promote the use of embedding as a means of temporal knowledge mining and predictive analytics.

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.

Please report errors in award information by writing to: awardsearch@nsf.gov.