Scholarlink

<p>Detection of concept drift in time-varying short text streams has numerous challenges since the data are volatile. According to research, 30% to 40% of the traditional drift detection methods are not able to detect change of the concept in the text stream and, therefore, produce high false positives and slow response time. To address the above issues, the proposed Graph based Concept Drift Detection (GraCoD) method suggests a novel concept drift detection (CoD) framework. GraCoD uses ConvBERT with Hopfield layers and temporal convolution to capture linguistic context and temporal dependencies. The model constructs a graph representation of text data using a text GCN with Time Varying Spatio Temporal-Graph Attention Module (TVST-GAT) and uses the Graph Aware Drift Detection Algorithm (GADD) to classify the change in the graph metrics such as node centrality and edge density. The approach is more helpful and effective than the traditional approaches of detecting the occurrence of drift. To react to the detected drifts proactively, Deep Reinforcement Learning (DRL) is merged with Deep Q-Learning to automatically adapt parameters and behaviors based on the outcomes of detected drifts. The severity and classification modules detect the severity and classify the detected drifts for further investigation. The proposed model demonstrates exceptional performance in CoD across five diverse datasets: Twitter datasets 1 and 2, Enron, News 20, and Amazon Reviews. It achieves high accuracy (98.7%-99.5%) and F1-scores (96%-98%), with low false positive (0.02–0.04) and false negative (0.01–0.03) rates. The model effectively identifies 23–29 drifts, with drift indicators ranging from 81.3% to 86.6%, showcasing its robustness in handling dynamic data streams across various domains.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

GraCoD: a disruptive graph-aware drift detection algorithm with a GCN-based time-varying module for concept drift detection in short text streams

  • 期刊论文

  • 发表日期：

    2025年11月1日

查看全文
  • Megha Ashok Patil,
  • Sunil Kumar,
  • Sandeep Kumar

摘要

Detection of concept drift in time-varying short text streams has numerous challenges since the data are volatile. According to research, 30% to 40% of the traditional drift detection methods are not able to detect change of the concept in the text stream and, therefore, produce high false positives and slow response time. To address the above issues, the proposed Graph based Concept Drift Detection (GraCoD) method suggests a novel concept drift detection (CoD) framework. GraCoD uses ConvBERT with Hopfield layers and temporal convolution to capture linguistic context and temporal dependencies. The model constructs a graph representation of text data using a text GCN with Time Varying Spatio Temporal-Graph Attention Module (TVST-GAT) and uses the Graph Aware Drift Detection Algorithm (GADD) to classify the change in the graph metrics such as node centrality and edge density. The approach is more helpful and effective than the traditional approaches of detecting the occurrence of drift. To react to the detected drifts proactively, Deep Reinforcement Learning (DRL) is merged with Deep Q-Learning to automatically adapt parameters and behaviors based on the outcomes of detected drifts. The severity and classification modules detect the severity and classify the detected drifts for further investigation. The proposed model demonstrates exceptional performance in CoD across five diverse datasets: Twitter datasets 1 and 2, Enron, News 20, and Amazon Reviews. It achieves high accuracy (98.7%-99.5%) and F1-scores (96%-98%), with low false positive (0.02–0.04) and false negative (0.01–0.03) rates. The model effectively identifies 23–29 drifts, with drift indicators ranging from 81.3% to 86.6%, showcasing its robustness in handling dynamic data streams across various domains.