Transient Concepts in Streaming Graphs

TL;DR

This paper introduces two novel frameworks, SGDD and SGDP, for detecting and predicting concept drift in streaming graphs, enabling timely responses without relying on data labels or payloads.

Contribution

It presents the first frameworks specifically designed for concept drift detection and prediction in streaming graphs, addressing limitations of previous methods.

Findings

SGDP predicts concept drift up to 0.19 milliseconds ahead.

SGDD detects concept drift with significant delays, highlighting the challenge in timely detection.

Frameworks do not require data payloads or labels, enhancing efficiency.

Abstract

Concept Drift (CD) occurs when a change in a hidden context can induce changes in a target concept. CD is a natural phenomenon in non-stationary settings such as data streams. Understanding, detection, and adaptation to CD in streaming data is (i) vital for effective and efficient analytics as reliable output depends on adaptation to fresh input, (ii) challenging as it requires efficient operations as well as effective performance evaluations, and (iii) impactful as it applies to a variety of use cases and is a crucial initial step for data management systems. Current works are mostly focused on passive CD detection as part of supervised adaptation, on independently generated data instances or graph snapshots, on target concepts as a function of data labels, on static data management, and on specific temporal order of data record. These methods do not always work. We revisit CD for the streaming graphs setting and introduce two first-of-its-kind frameworks SGDD and SGDP for streaming graph CD detection and prediction. Both frameworks discern the change of generative source. SGDD detects the CDs due to the changes of generative parameters with significant delays such that it is difficult to evaluate the performance, while SGDP predicts these CDs between 7374 to 0.19 milliseconds ahead of their occurrence, without accessing the payloads of data records.