Scholarlink

<p>In the fast paced world, digital payment systems have generated such massive retail transaction volumes that real-time anomaly detection pipelines must detect fraud while protecting consumers and minimizing operational losses. The effectiveness of anomaly detection systems which use statistical thresholds and batch-based machine learning decreases when operating under high-speed data streams and complex behavioral contexts. The paper introduces a Streaming Analytics LLM framework which includes three new modules: the Prismatic Alert Module (PAM) generates contextualized streaming embeddings through large language models (LLM) for alerting anomalous events, the Agile Response Module (ARM) uses reinforcement learning to detect high-risk transaction streams and the Trajectory Fusion Module (TFM) applies biologically inspired methods to combine multiple features from transaction streams. It involves domain-specific performance metrics that include Detection Lead Time (DLT), Streaming Throughput Stability (STS), Alert Relevance Index (ARI), and Fusion Coherence Score (FCS). The indicated pipeline demonstrates superior performance in detecting anomalies through experimental testing on large retail data collections because it reduces lead time by 4.7 times and improves alert relevance by 3.2 times and decreases false financial burden by 2.9 times when compared to LSTM, Transformer and Spectral FNO architectures.</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

Streaming analytics pipelines for LLM-based financial anomaly detection in Real-Time retail transaction flows

  • 期刊论文

  • 发表日期：

    2026年1月10日

查看全文
  • Venkat Sai Nageen Kanikanti,
  • Krishna Mula,
  • Kailasam Muthukumarasamy,
  • Chandra Sekhar Kubam,
  • Bidisha Goswami,
  • Harshini Gadam

摘要

In the fast paced world, digital payment systems have generated such massive retail transaction volumes that real-time anomaly detection pipelines must detect fraud while protecting consumers and minimizing operational losses. The effectiveness of anomaly detection systems which use statistical thresholds and batch-based machine learning decreases when operating under high-speed data streams and complex behavioral contexts. The paper introduces a Streaming Analytics LLM framework which includes three new modules: the Prismatic Alert Module (PAM) generates contextualized streaming embeddings through large language models (LLM) for alerting anomalous events, the Agile Response Module (ARM) uses reinforcement learning to detect high-risk transaction streams and the Trajectory Fusion Module (TFM) applies biologically inspired methods to combine multiple features from transaction streams. It involves domain-specific performance metrics that include Detection Lead Time (DLT), Streaming Throughput Stability (STS), Alert Relevance Index (ARI), and Fusion Coherence Score (FCS). The indicated pipeline demonstrates superior performance in detecting anomalies through experimental testing on large retail data collections because it reduces lead time by 4.7 times and improves alert relevance by 3.2 times and decreases false financial burden by 2.9 times when compared to LSTM, Transformer and Spectral FNO architectures.