Interleaved Sequence RNNs for Fraud Detection (2002.05988v2)

Abstract: Payment card fraud causes multibillion dollar losses for banks and merchants worldwide, often fueling complex criminal activities. To address this, many real-time fraud detection systems use tree-based models, demanding complex feature engineering systems to efficiently enrich transactions with historical data while complying with millisecond-level latencies. In this work, we do not require those expensive features by using recurrent neural networks and treating payments as an interleaved sequence, where the history of each card is an unbounded, irregular sub-sequence. We present a complete RNN framework to detect fraud in real-time, proposing an efficient ML pipeline from preprocessing to deployment. We show that these feature-free, multi-sequence RNNs outperform state-of-the-art models saving millions of dollars in fraud detection and using fewer computational resources.

• The paper presents a novel approach using Interleaved Sequence RNNs (specifically GRUs) for real-time fraud detection, eliminating the need for extensive feature engineering by modeling transaction histories as interleaved sequences.
• This method employs a three-block architecture (Feature, Recurrent, Classification) allowing GRUs to encode individual card transaction histories into fixed-size states for efficient real-time prediction.
• Experimental results on financial datasets show superior recall and cost savings compared to traditional tree-based models, demonstrating the method's effectiveness and practicality within strict production latency requirements.

Interleaved Sequence RNNs for Fraud Detection

The paper, "Interleaved Sequence RNNs for Fraud Detection," presents a novel approach to real-time fraud detection in financial transactions using recurrent neural networks (RNNs). This approach circumvents traditional reliance on computationally expensive feature engineering techniques by leveraging RNNs to model transactions as interleaved sequences, where each card's transaction history forms an unbounded, irregular sub-sequence.

Overview

Payment card fraud represents a significant financial burden to businesses globally, with billions lost to unauthorized transactions annually. The conventional solution involves tree-based machine learning models, such as Random Forests or XGBoost, which depend heavily on feature engineering to incorporate historical transaction data. Such methods necessitate complex data augmentation processes, contributing to increased resource consumption and latency.

This paper introduces a framework wherein RNNs, particularly Gated Recurrent Units (GRUs), manage transaction sequences without the need for extensive feature engineering. By treating each card's transaction history as interleaved sequences, the model can apply GRUs to encode these histories into fixed-size recurrent states, capable of making real-time predictions. The proposed system consists of offline components for model training and preprocessing, and online components to deploy the model in real-time applications.

Model Architecture and Methodology

The model comprises three primary blocks:

• Feature Engineering and Transformation: Transforms raw transaction data into a suitable format for processing by neural networks. This includes embedding categorical features and scaling numerical features.
• Recurrent Block: Utilizes GRUs to analyze sequences, emphasizing irregular intervals between transactions as indicative of fraud. The recurrent state encodes the transaction history.
• Classification Block: Generates predictions using the encoded state and current transaction features.

The approach allows for efficient compute during training by preprocessing data as sequences sorted by transaction time, and in production by saving and restoring GRU states, utilizing caching strategies to ensure computational efficiency and adherence to latency requirements.

Experimental Validation

Experiments conducted across two datasets from European financial institutions demonstrated the effectiveness of the interleaved sequence RNN approach. The model achieved superior recall and fraud detection metrics compared to feature-heavy tree-based models, with substantial improvements in recall and cost savings attributable to fraud detection.

Key performance findings included:

• Dataset A showed an increase of 9.7% in recall compared to LightGBM models using engineered features, correlating to significant financial savings.
• Dataset B highlighted an increase of 3.2% in recall and additional millions saved, showcasing the robustness of the model across varied use-cases.

Additionally, engineering metrics confirmed the practicality of the system within production environments governed by strict latency SLAs. The use of HaloDB for state management alongside an embedded approach substantially reduced the overhead associated with external storage solutions, optimizing speed and efficiency.

Conclusion and Future Work

The paper profoundly impacts fraud detection methodologies by demonstrating that RNNs can outperform traditional models without the need for engineered features. It emphasizes the practicality of GRUs to encapsulate transaction histories within real-time systems efficiently.

The authors suggest exploring the integration of engineered profiles within RNN architectures to further enhance fraud detection capabilities. Moreover, an adjustment in GRU architectures to inherently accommodate transaction interval irregularities could yield improvements in sequence prediction accuracy.

The research presents a critical step in evolving fraud detection strategies towards more streamlined, resource-effective, and high-performance solutions, adaptable to dynamic financial environments.