Contrastive Representation Learning on Temporal Point Process Data

The general goal of this project is to improve the downstream tasks by learning better
representations of the data, especially multimodal data pairs, like image/text pairs or user/item
interaction pairs. The user/item interaction data plays an important role in e-commerce, and
analyzing these data can help improve the banking system, e.g., recommendation, risk control, and
etc. The existing methods in this area mostly focusing on using deep neural networks, especially
graph neural network, to learn the connections and dependencies. Here, we want to leverage the
contrastive learning methods into this field, as it shows great power in learning image/text data
pairs representation and expect the learned representations to have better zero-shot performance
at prediction tasks, such as predict the future events.

Faculty Supervisor:

Arvind Gupta

Student:

Partner:

Layer 6 AI

Discipline:

Computer science

Sector:

Information and Communications Technology

University:

University of Toronto

Program: