STRAP: a deep learning-based predictive framework with low-latency evaluation for real-time streaming data
TL;DR
Results demonstrate that STRAP delivers both high predictive accuracy and ultra-low-latency responsiveness in highly dynamic, high-concurrency streaming environments, exhibiting strong practical deployability for latency-sensitive real-time forecasting scenarios such as smart grid anomaly detection, financial risk control, and edge-based monitoring systems.
Abstract
With the advancement of data science technologies and the growing demand for real-time applications, streaming data has assumed an increasingly prominent role in modern data processing systems, continuously challenging traditional batch-processing paradigms. To address key challenges in streaming learning—such as extreme class imbalance, concept drift, and the requirement for millisecond-level latency response—this paper proposes STRAP (Streaming Temporal deep learning with Robust Attentive Prediction), an end-to-end deep streaming prediction framework. STRAP constructs temporal inputs using a sliding-window mechanism combined with causal slicing, ensuring predictions rely solely on current and historical information. For representation learning, it integrates bidirectional gated recurrent units (BiGRUs) with a multi-head attention mechanism to capture complex contextual dependencies. Furthermore, Focal Loss is incorporated to dynamically adjust class weights, thereby enhancing the model’s ability to learn from minority classes. Recognizing the lack of standardized evaluation protocols in streaming learning, this paper introduce WASE (Window- Adaptive Streaming Evaluation), a comprehensive assessment framework that jointly evaluates deployment performance along both accuracy and efficiency dimensions. WASE combines online AUC, latency percentiles (p95, p99), and system throughput, featuring a window-level adaptive evaluation mechanism that aligns with evolving data distributions and real-time operational requirements. In experiments on a streaming-reconstructed version of the IEEE-CIS Fraud Detection dataset, STRAP achieves a 1.5% absolute improvement in test-set AUC over an MLP baseline while maintaining a p99 latency below 7 ms, and attains a throughput of 1,258.2 samples/s. Ablation studies confirm that the multi-head attention mechanism and the dynamic reweighting strategy jointly improve minority-class recognition. These results demonstrate that STRAP delivers both high predictive accuracy and ultra-low-latency responsiveness in highly dynamic, high-concurrency streaming environments, exhibiting strong practical deployability for latency-sensitive real-time forecasting scenarios such as smart grid anomaly detection, financial risk control, and edge-based monitoring systems.