Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Online Drift Detection with Maximum Concept Discrepancy

Ke Wan, Yi Liang, Susik Yoon

TL;DR

This paper tackles unsupervised online concept drift detection in high-dimensional data streams. It introduces Maximum Concept Discrepancy (MCD), learned via contrastive objectives and time-aware sampling, to quantify distributional shifts between consecutive sub-windows. The proposed method, MCD-DD, combines a sample-set encoder with a dynamically thresholded drift detector and a robust optimization objective including positive/negative sampling and a Lipschitz gradient penalty. Empirical results on 11 datasets show state-of-the-art detection accuracy and interpretability, with theoretical bounds and complexity analysis supporting practical deployment and real-time operation.

Abstract

Continuous learning from an immense volume of data streams becomes exceptionally critical in the internet era. However, data streams often do not conform to the same distribution over time, leading to a phenomenon called concept drift. Since a fixed static model is unreliable for inferring concept-drifted data streams, establishing an adaptive mechanism for detecting concept drift is crucial. Current methods for concept drift detection primarily assume that the labels or error rates of downstream models are given and/or underlying statistical properties exist in data streams. These approaches, however, struggle to address high-dimensional data streams with intricate irregular distribution shifts, which are more prevalent in real-world scenarios. In this paper, we propose MCD-DD, a novel concept drift detection method based on maximum concept discrepancy, inspired by the maximum mean discrepancy. Our method can adaptively identify varying forms of concept drift by contrastive learning of concept embeddings without relying on labels or statistical properties. With thorough experiments under synthetic and real-world scenarios, we demonstrate that the proposed method outperforms existing baselines in identifying concept drifts and enables qualitative analysis with high explainability.

Online Drift Detection with Maximum Concept Discrepancy

TL;DR

This paper tackles unsupervised online concept drift detection in high-dimensional data streams. It introduces Maximum Concept Discrepancy (MCD), learned via contrastive objectives and time-aware sampling, to quantify distributional shifts between consecutive sub-windows. The proposed method, MCD-DD, combines a sample-set encoder with a dynamically thresholded drift detector and a robust optimization objective including positive/negative sampling and a Lipschitz gradient penalty. Empirical results on 11 datasets show state-of-the-art detection accuracy and interpretability, with theoretical bounds and complexity analysis supporting practical deployment and real-time operation.

Abstract

Continuous learning from an immense volume of data streams becomes exceptionally critical in the internet era. However, data streams often do not conform to the same distribution over time, leading to a phenomenon called concept drift. Since a fixed static model is unreliable for inferring concept-drifted data streams, establishing an adaptive mechanism for detecting concept drift is crucial. Current methods for concept drift detection primarily assume that the labels or error rates of downstream models are given and/or underlying statistical properties exist in data streams. These approaches, however, struggle to address high-dimensional data streams with intricate irregular distribution shifts, which are more prevalent in real-world scenarios. In this paper, we propose MCD-DD, a novel concept drift detection method based on maximum concept discrepancy, inspired by the maximum mean discrepancy. Our method can adaptively identify varying forms of concept drift by contrastive learning of concept embeddings without relying on labels or statistical properties. With thorough experiments under synthetic and real-world scenarios, we demonstrate that the proposed method outperforms existing baselines in identifying concept drifts and enables qualitative analysis with high explainability.
Paper Structure (48 sections, 1 theorem, 15 equations, 12 figures, 4 tables, 1 algorithm)

This paper contains 48 sections, 1 theorem, 15 equations, 12 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background and Motivation
  3. Main Idea and Challenges
  4. Summary
  5. Related Work
  6. Concept Drift Detection
  7. Contrastive Learning in Data Streams
  8. Maximun Mean Discrepancy
  9. Preliminaries
  10. Concept Drift
  11. Problem Setting
  12. Maximum Mean Discrepancy
  13. Proposed Method
  14. Overview
  15. Sample Set Encoder
  16. ...and 33 more sections

Key Result

Theorem 1

Assume that the sets $\{X_i\}_{i=1}^n$ and $\{Y_i\}_{i=1}^n$ are independently and identically distributed (i.i.d.), both drawn from the probability distribution $p(x)$ with a mean $\mu$ and variance $\sigma$. If $f$ is a Lipschitz continuous function with Lipschitz constant $L$, we have: where $G(\cdot)$ is the standard Gaussian distribution function. Therefore, for a given significance level $\

Figures (12)

  • Figure 1: Unsupervised online concept drift detection over sliding window $\mathbb{W}$ with sub-windows $\mathbb{S}$.
  • Figure 2: Overall procedure of MCD-DD. Sub-windows are encoded and compared to derive MCD for drift detection.
  • Figure 3: Heatmaps of MCD between sub-windows for primary synthetic data sets with drift indicators (red lines and arrows).
  • Figure 4: Heatmap for INSECTS_Sud with drift indicators.
  • Figure 5: Ablation study of contrastive learning strategies.
  • ...and 7 more figures

Theorems & Definitions (1)

  • Theorem 1