Multi-Level ML Labeling for ERP Benchmark Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Labeling training data for machine learning algorithms in ERP systems is time-consuming and challenging due to the need for domain knowledge, making it difficult to efficiently identify anomalies and perform benchmark comparisons.
Innovation Solution
A multi-level machine learning process is used to automate the labeling of training data, employing MLP, KNN regression, isolation forest, and random forest classifier models to predict and classify data points as inliers, outliers, or uncertain based on density and trends, with preprocessing and correlation analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling of training data is performed by domain experts, then labeling accuracy is improved, but time consumption and cost increase significantly
Solution Approach 1:
The system performs self-service by using the ERP system's own historical data and business logic to automatically generate labels for training data. The multi-level machine learning process leverages the system's inherent structures (organizational hierarchy, business processes, KPI relationships) to create accurate labels without external human intervention, thus achieving both high accuracy and efficiency
Solution Approach 2:
The patent introduces an intermediary multi-level machine learning process that acts as a bridge between raw ERP data and labeled training data. This intermediary system uses multiple ML models (regression, classification, clustering) to transform unlabelled data into accurately labeled training datasets, eliminating the need for manual expert labeling while maintaining high label quality
2Measurement precision
If manual labeling by domain experts is used, then domain knowledge is properly applied, but productivity and scalability deteriorate
Solution Approach 1:
The ERP system automatically generates labels by leveraging its own business logic, data structures, and historical patterns. The multi-level ML process extracts meaningful labels from unlabelled data using the system's inherent domain knowledge embedded in its processes, hierarchies, and relationships, achieving both high quality and high productivity
Solution Approach 2:
The system performs preliminary analysis and feature extraction before final label generation. The multi-level process prepares data through preprocessing, feature engineering, and intermediate modeling steps that capture domain knowledge in advance, enabling efficient and accurate label creation at scale
3Device complexity
If traditional single-level machine learning is used, then model complexity is reduced, but labeling accuracy and anomaly detection capability worsen
Solution Approach 1:
The labeling process is segmented into multiple independent but coordinated levels, each handling specific aspects of the data. Different ML models (regression, classification, clustering) are applied at different levels to capture various patterns and relationships, with results integrated to produce comprehensive labels that improve detection accuracy
Solution Approach 2:
The patent implements a nested multi-level architecture where simpler models are embedded within more complex ones. The hierarchical structure allows basic patterns to be detected at lower levels while more complex relationships are captured at higher levels, with each level building upon and refining the outputs of previous levels
Data Source
AI summary
In an example embodiment, a multi-level machine learning process is used to automate labelling data and training and fine-tuning a number of benchmarking classifier machine learning models. Automatic labeling is performed partially based on density and trends among data points in a training set. This approach may be used with different types of performance or telemetry data types.