# Neural subgraph counting on stream graphs via localized updates and monotonic learning

## TL;DR

This paper introduces StreamSC, a new framework that efficiently estimates subgraph counts in dynamically changing stream graphs using machine learning.

## Contribution

StreamSC is the first learning-based framework for subgraph counting on stream graphs, addressing dynamic changes caused by edge insertions and deletions.

## Key findings

- StreamSC outperforms existing methods in both accuracy and efficiency on real-world stream graphs.
- The framework effectively handles dynamic updates in stream graphs through localized updates and monotonic learning.

## Abstract

Graphs are a representative type of fundamental data structures. They are capable of representing complex association relationships in diverse domains. For large-scale graph processing, the stream graphs have become efficient tools to process dynamically evolving graph data. When processing stream graphs, the subgraph counting problem is a key technique, which faces significant computational challenges due to its #P-complete nature. This work introduces StreamSC, a novel framework that efficiently estimate subgraph counting results on stream graphs through two key innovations: (i) It’s the first learning-based framework to address the subgraph counting problem focused on stream graphs; and (ii) this framework addresses the challenges from dynamic changes of the data graph caused by the insertion or deletion of edges. Experiments on 5 real-word graphs show the priority of StreamSC on accuracy and efficiency.