A Label-Adaptive Contrastive Loss-Based Deep Hashing Method for Timely and Accurate Traffic Metadata Retrieval

Abstract

With the rapid development of the Internet of Things (IoT), the volume of network traffic data has increased exponentially. The high-dimensional, heterogeneous nature of such data makes efficient retrieval increasingly challenging, thereby degrading the performance of traffic processing systems. To address the inefficiency of high-dimensional traffic data retrieval, we propose a label-adaptive contrastive loss-based deep hashing (LCL-DH) model to compress high-dimensional traffic metadata and employ an approximate nearest neighbor (ANN) search method to enable efficient retrieval. The LCL-DH model leverages a convolutional neural network (CNN) together with a label-adaptive contrastive loss to generate highly discriminative hash codes while minimizing hash collisions. Furthermore, a hierarchical query strategy based on hierarchical navigable small world (HNSW) graphs is adopted to further improve hash-code query efficiency. Experimental results demonstrate that the proposed method achieves approximately 11% higher accuracy than deep polarization network (DPN) methods. It attains an average query time of 0.03 ms, which is only one-hundredth that of the sequential query method.