Graph-Based Event Classification for Fraud Detection

The patent transitions from analyzing events in isolation (one-dimensional) to analyzing events within their temporal sequences and relational contexts (multi-dimensional). Event sequences capture the chronological dimension, while graph structures add the relational dimension by connecting users, devices, and events. This dimensional expansion enables the model to consider holistic user interactions without proportionally increasing computational complexity through efficient sequence and graph processing techniques.

The patent segments the analysis into distinct components: event sequences for temporal patterns, graph structures for relational patterns, and feature embeddings for attribute patterns. Each component processes specific aspects of user interactions independently, then their results are integrated to form comprehensive predictions. This segmentation allows complex holistic analysis to be achieved through modular, manageable components that can be processed efficiently.

The patent extracts essential features from large volumes of raw interaction data and represents them as compact event sequences and graph structures. Instead of processing all raw data, the system extracts key events, their temporal relationships, and relational patterns into condensed representations. This extraction maintains the essential information needed for accurate predictions while dramatically reducing the data volume that requires intensive processing.

The patent transforms raw interaction data into different parameter representations through feature embedding, where categorical and numerical attributes are converted into dense vector representations. This parameter transformation compresses large volumes of discrete data into continuous, lower-dimensional feature spaces that retain semantic meaning and can be processed more efficiently by machine learning models while preserving predictive power.