Anomaly Detection in E-Commerce Multi-Table ETL Processes through Mamba-LSTM Collaborative Modeling

Abstract

To address the challenges of complex anomaly types, tight inter-table relationships, significant temporal dependencies, and the difficulty of traditional methods simultaneously considering global context and local dynamic changes in e-commerce multi-table ETL processes, this paper proposes a data anomaly detection method based on Mamba and LSTM collaborative modeling. First, this method constructs a unified sequence representation for multi-source business data such as orders, payments, logistics, products, and users, mapping heterogeneous field information, relational information, and process state information to a shared feature space to enhance the model's ability to characterize complex business semantics. Building upon this, a Mamba branch is introduced to model long-range dependencies and cross-stage contexts in the ETL chain, improving the ability to identify hidden anomaly propagation patterns. Simultaneously, an LSTM branch is used to capture local state transitions and short-term dynamic change features, enhancing the perception of fine-grained anomaly perturbations. Subsequently, an adaptive fusion mechanism dynamically coordinates the global and local representations to form a unified feature representation for anomaly detection, which is then combined with a classification head to complete anomaly sample identification. This method effectively adapts to the characteristics of multi-source heterogeneity, coupled relationships, and continuous process evolution in e-commerce ETL data, enabling efficient characterization and stable discrimination of abnormal patterns in complex business processes. The results show that the proposed method demonstrates good overall performance in anomaly identification, feature representation sufficiency, and model stability, providing effective support for data quality governance and anomaly monitoring in e-commerce scenarios.