Evolving graph convolutional networks for dynamic graphs

US11537852B2 · kind B2 · utility

Assignees

• International Business Machines Corporation · US
• Massachusetts Institute of Technology · US

Inventors

• Jie Chen · New Taipei, TW
• Aldo Pareja · Cambridge, US
• Giacomo Domeniconi · Miami, US
• Tengfei Ma · White Plains, US
• Toyotaro Suzumura · Tokyo, JP
• Timothy Kaler · Cambridge, US
• Tao B. Schardl · Cambridge, US
• Charles E. Leiserson · Pittsburgh, US

Key dates

• Filing date: Feb 13, 2020
• Grant date: Dec 27, 2022
• Expiry date: Mar 15, 2041

Classification

• Technology area (CPC G): Physics
• CPC primary: G06Q30/0201
• WIPO field: Computer technology
• WIPO sector: Electrical engineering

Abstract

A system includes a plurality of graph convolutional networks corresponding to a plurality of time steps, each network modelling a graph including nodes and edges, and in turn including a plurality of graph convolution units; an evolving mechanism; and an output layer. Each of the units, for a given one of the time steps, takes as input a graph adjacency matrix, a node feature matrix, and a parameter matrix for a current layer, and outputs a new node feature matrix for a next highest layer. The mechanism takes as input a parameter matrix for a prior time step updates the input parameter matrix, and outputs the parameter matrix for the given time step. The output layer obtains, as input, output of each of the units for a final time step, and based on the output of each of the units for the final time step, outputs a graph solution.