Hard-Negative Sampling for Imbalanced Transaction Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large training datasets for machine learning models are costly, time-consuming, and resource-intensive, leading to challenges in rapid development and experimentation, especially in scenarios with imbalanced transactions.
Innovation Solution
A server computer implements a sampling framework that encodes transaction data into N-dimensional vectors, performs hard-negative and diversity sampling processes, and iteratively adjusts sample sizes to optimize the performance of machine learning models on imbalanced transaction data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If large training datasets are used to improve machine learning model performance, then model accuracy and generalization are improved, but storage costs, computational resources, and training time increase significantly
Solution Approach 1:
The patent extracts and removes redundant or low-value data points from the training dataset through automated filtering criteria. By identifying and eliminating unnecessary data, the system reduces the overall dataset size while preserving the essential information needed for model training, thus resolving the contradiction between model performance and data quantity
Solution Approach 2:
The system dynamically adjusts data sampling parameters and filtering thresholds based on the specific characteristics of the dataset and model requirements. By changing these parameters, the system optimizes the balance between retaining sufficient data for accurate model training and reducing excessive data that would consume unnecessary storage and computational resources
2Reliability
If large training datasets are collected and labeled, then model training quality is improved, but data collection and labeling costs and time increase
Solution Approach 1:
The patent implements preliminary automated filtering and validation of data before the time-consuming labeling process. By pre-processing the dataset to remove obviously incorrect or low-quality entries early in the pipeline, the system reduces the amount of data that requires manual labeling, thereby decreasing data preparation time while maintaining training quality
Solution Approach 2:
The system employs automated quality assessment algorithms that can independently evaluate and filter data without requiring extensive manual review. This self-service approach to data quality control reduces dependency on manual labeling efforts, decreasing both time and cost while preserving the essential training signal
3Productivity
If powerful hardware is used to train on large datasets, then training speed and model capability are improved, but hardware costs and resource requirements increase
Solution Approach 1:
The patent extracts and removes redundant computations and data processing steps from the training pipeline. By eliminating unnecessary computational operations and reducing the volume of data that needs to be processed, the system achieves efficient training with reduced hardware resources while maintaining acceptable training speed and model capability
Solution Approach 2:
The system applies partial processing strategies where not all data points require the same level of computational attention. By focusing computational resources on the most informative data subsets and using less intensive processing for other data, the system achieves effective training with reduced hardware requirements
4Productivity
If dataset size is reduced to lower storage and computational costs, then resource efficiency is improved, but model performance and generalization may deteriorate
Solution Approach 1:
The system dynamically adjusts data sampling parameters, filtering thresholds, and selection criteria to optimize the balance between dataset size and model performance. By carefully tuning these parameters, the system identifies the minimum sufficient data size required to maintain acceptable model performance while achieving significant reductions in storage and computational resource requirements
Data Source
AI summary
A server computer that implements an efficient sampling technique for imbalanced transactions. Transaction data comprising unlabeled transaction data and fraudulent transaction data is encoded to form (i) a first set of encoded data associated with fraudulent transactions and unlabeled transactions that are similar to fraudulent transactions, and (ii) a second set of encoded data associated with unlabeled transactions. A first sampling process is executed with respect to the first set of encoded data to obtain a first sampled set of encoded data, and a second sampling process is executed with respect to the second set of encoded data to obtain a second sampled set of encoded data. An optimal sampling size of the transaction data is determined based on whether a performance of a machine learning model that classifies the first sampled set of encoded data and the second sampled set of encoded data satisfies a condition.