Training Dataset Selection Using Graph Embeddings for Document Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional document classification systems face challenges in accurately clustering documents with varying templates due to reliance on image-based and word features, especially with scanned documents prone to OCR errors, leading to increased complexity and computational inefficiencies in training machine learning models.
Innovation Solution
A graph-based clustering approach using graph embedding neural networks to convert documents into graph objects, label entities as nodes, and generate embedding vectors for clustering, followed by optimizing the training dataset to select a finite set of samples for machine learning input.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If image-based processing algorithms are used to cluster documents, then documents can be grouped based on visual features, but OCR errors in scanned documents cause erratic clustering results and reduced reliability
Solution Approach 1:
The patent replaces image-based processing algorithms with graph-based clustering methods. Instead of relying on visual features extracted from scanned documents that are susceptible to OCR errors, the system converts documents into graph objects where nodes represent text elements and edges represent spatial relationships. This substitution eliminates dependence on optical character recognition and uses structural information that is invariant to OCR inaccuracies, thereby maintaining clustering reliability while preserving productivity.
Solution Approach 2:
The patent changes the feature parameters used for clustering from image-based visual features to graph-based structural features. By representing documents as graphs with nodes and edges capturing text element positions and relationships, the system transforms the clustering problem into one that operates on topological and structural parameters rather than pixel-based or OCR-dependent parameters. This parameter transformation makes clustering robust to OCR errors while maintaining efficiency.
2Measurement precision
If all sample documents from clustered templates are used for training, then the machine learning model can learn all variations, but computational time and complexity increase significantly
Solution Approach 1:
The patent extracts and selects only the most representative samples from each document cluster for training, rather than using all available documents. The graph-based clustering identifies template variations, and the system selects a finite subset of samples that best represent each template's characteristics. This extraction approach maintains model training accuracy by preserving essential variation patterns while dramatically reducing the training dataset size, thereby decreasing computational time and complexity.
Solution Approach 2:
The patent applies partial action by using a selected subset of training samples rather than the complete dataset. By identifying and utilizing only the necessary portion of documents that capture the essential template variations, the system achieves sufficient model accuracy without the excessive computational burden of processing all available samples. This selective approach optimizes the trade-off between training completeness and computational efficiency.
3Ease of manufacture
If conventional clustering methods are used, then the process is simple to implement, but the system cannot effectively handle documents with varying templates and structures
Solution Approach 1:
The patent implements a universal graph-based clustering framework that can handle multiple document types and template variations through a single unified approach. The graph object representation is generic and can represent any document structure, making the clustering system adaptable to diverse templates including application forms, financial statements, and receipts. This universal method maintains relative implementation simplicity while dramatically improving versatility compared to conventional template-specific approaches.
Data Source
AI summary
System for optimizing training dataset comprising sample documents. The system comprises one or more processors configured to create graph embedding vector for each of the sample documents of the training dataset and cluster the graph embedding vectors of the sample documents of the training dataset into clusters based on the similarity between the graph embedding vectors. Further, the processor is configured to select a first set of training data, using an optimization model, wherein the first set of training data comprises a finite number of graph embedding vectors of the sample documents from the clustered training dataset. Finally, the first set of training data is fed as for a machine learning model.