Expense Anomaly Detection Using Segmented Isolation Forests
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting anomalies in expense reports are time-consuming, inaccurate, and require significant manual effort, especially for large datasets, and existing automated methods lack effectiveness in handling expense type, location, and seasonal trends.
Innovation Solution
An unsupervised machine learning model, specifically an isolation forest, is used to detect anomalies in expense reports by segregating data based on expense category, location, and season, followed by a feedback loop for supervised model training to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual checking and auditing methods are used for expense reports, then detection accuracy can be maintained through human judgment, but the process becomes time-consuming and inefficient for large datasets
Solution Approach 1:
The patent replaces manual mechanical review processes with automated machine learning systems. Specifically, it employs unsupervised learning models (isolation forests, autoencoders) and supervised learning models to automatically detect anomalies in expense reports, substituting human manual checking with computational algorithms that process data rapidly without sacrificing detection accuracy.
Solution Approach 2:
The patent introduces machine learning models as intermediary systems between raw expense data and final anomaly detection results. These models serve as mediators that transform unstructured expense data into structured anomaly predictions, enabling efficient processing while maintaining detection capabilities through learned patterns from historical data.
2Productivity
If automated anomaly detection methods are implemented, then processing speed and efficiency improve, but existing methods lack effectiveness in handling expense type, location, and seasonal trends
Solution Approach 1:
The patent segments the expense detection problem into multiple specialized models, each trained on specific expense categories (travel, meals, entertainment, etc.). This segmentation allows each model to develop expertise in detecting anomalies within its specific domain, improving overall detection accuracy while maintaining high processing efficiency through parallel model execution.
Solution Approach 2:
The patent implements a dynamic hybrid system that adapts between unsupervised and supervised learning approaches based on data characteristics. The system uses unsupervised models for initial anomaly detection and then applies supervised models for refined classification, creating a dynamic processing pipeline that optimizes both efficiency and accuracy for different expense patterns.
3Device complexity
If a single unified machine learning model is used for all expenses, then device complexity is reduced, but the model cannot effectively capture location-specific and category-specific expense patterns
Solution Approach 1:
The patent divides the expense detection system into multiple specialized machine learning models, each trained on specific expense categories and location patterns. This segmentation enables each model to capture nuanced local patterns while the overall system remains manageable through modular architecture, where each component's complexity is contained and well-defined.
Solution Approach 2:
The patent creates a universal framework that integrates multiple specialized models into a cohesive system. The hybrid architecture combines unsupervised and supervised learning capabilities across different expense categories and locations, providing multi-functional anomaly detection that adapts to various expense types while maintaining a unified processing interface.
Data Source
AI summary
Embodiments perform the anomaly detection of expense reports in response to receiving an expense report as input data, the expense report including a plurality of expenses. Embodiments create a plurality of groups of expenses, each group corresponding to a different combination of a category of the expense, a location of the expense and a season of the expense. Embodiments generate and train an unsupervised machine learning model corresponding to each group, and assign each of the expenses of the expense report into a corresponding group and input the expenses into the unsupervised machine learning model corresponding to the group. Embodiments then generate an anomaly prediction at each unsupervised machine learning model for each expense of the expense report.