Outlier Removal in Machine Learning Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer systems are unable to effectively identify and control access to restricted information within documents, relying on inefficient keyword searches that consume significant memory and processing resources, limiting system throughput.
Innovation Solution
A document analysis system employing artificial intelligence and machine learning to analyze text, tag specific sentence types, train machine learning models, determine sentence similarities, and identify outlier sentences, thereby improving the accuracy and efficiency of information access control.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If keyword searches are used for controlling data access, then data access control can be implemented, but memory resources and processing resources are significantly consumed
Solution Approach 1:
The patent extracts the essential meaning from documents by identifying and removing outlier sentences, keeping only the most representative sentences for training the machine learning model. This reduces the quantity of training data while maintaining the core information needed for accurate classification, thereby reducing memory resources required for storage and processing.
Solution Approach 2:
The patent changes the approach from keyword-based search to machine learning-based classification. By training a model on cleaned training data (with outliers removed), the system achieves more efficient data access control that requires less memory and processing power compared to maintaining and searching large keyword sets.
2Reliability
If keyword searches are used for controlling data access, then data access control can be implemented, but processing resources are significantly consumed
Solution Approach 1:
The patent performs preliminary cleaning of training data by identifying and removing outlier sentences before training the machine learning model. This preprocessing step ensures that the model is trained on high-quality, representative data, leading to more efficient inference during data access control operations and reduced processing resources during runtime.
Solution Approach 2:
The patent transitions from keyword-based search to machine learning-based classification. The machine learning model, trained on cleaned data, provides more efficient processing during data access control compared to searching through large keyword sets, thereby reducing processing resource consumption.
3Measurement precision
If outlier sentences are removed from training data, then training accuracy is improved, but training data quantity is reduced
Solution Approach 1:
The patent extracts and removes outlier sentences from the training data that do not represent the typical characteristics of the document corpus. By eliminating these atypical sentences, the training data becomes more consistent and representative, improving model training accuracy without requiring an increase in overall data quantity.
Solution Approach 2:
The patent changes the quality parameter of training data by removing outliers, transforming the dataset from containing mixed-quality sentences to containing only high-quality, representative sentences. This parameter change improves training accuracy while the reduced data quantity is compensated by the higher quality of remaining sentences.
Data Source
AI summary
A document analysis device that includes an artificial intelligence (AI) processing engine configured to receive training data, to select a sentence from the training data, and to compute a first set of similarity scores between the selected sentence and other sentences from the training data. The AI processing engine is further configured to determine a set count that is equal to a number of similarity scores in the first set of similarity scores that exceed a similarity score threshold value and to compare the set count to a set outlier threshold value. The AI processing engine is further configured to keep the selected sentence in the training data when the set count is greater than or equal to the set outlier threshold value and to remove the selected sentence from the training data when the set count is less than the set outlier threshold value.