GraphDrift-net: a dynamic graph-based framework for concept drift detection in short unstructured text streams

Detecting concept drift in text streams is challenging due to the rapid evolution of language, shifting user behavior, and temporal dependencies. Issues like data sparsity, high dimensionality, lack of labeled data, and multimodal drift further complicate real-time detection and adaptation. This paper proposes GraphDrift-net, a novel dynamic graph-based framework for detecting and adapting to concept drift in evolving text streams. The model comprise of the following components: evolving Time BERT (EvoTimeBERT), which captures temporal language evolution via historical token memory and multi-scale temporal convolutions, hierarchical temporal graph network with dynamic topics and adaptive memory (HTGN-DTAM), a heterogeneous graph neural network that dynamically constructs topic-aware graphs to track changing semantics and Chronograph Detection, a time-series-based drift detection method leveraging graph statistics such as node centrality and clustering coefficient changes. In addition, graph neural reinforcement learning framework (GNRL), a reinforcement learning-based adaptive learning module, enables model adaptability by word embedding update, memory decay rate tuning, and few-shot adaptation. Experimental evaluations over various real-world datasets, including Twitter-1, Twitter-2, Enron, and News20, demonstrate that GraphDrift-net outperforms other methods in accuracy, F1-score, and drift detection sensitivity. The model achieves accuracy as high as 99.7%, is able to identify more drift points, and is more stable with computational efficiency, making it extremely appropriate for real-time text stream applications.