Rule Agnostic Reject Inferencing for Transaction Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for transaction classification tasks are limited by incomplete training data, as declined transactions lack verified labels, leading to suboptimal and inaccurate predictions.
Innovation Solution
A multi-model framework using non-rule-based reject inferencing generates training data by inferring labels and weights for declined transactions based on scores from a first machine learning model, allowing for the inclusion of both approved and declined transactions in the training dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training data includes only approved transactions with verified labels, then label accuracy is ensured, but training data completeness deteriorates
Solution Approach 1:
The patent introduces a multi-model framework where an intermediary labeling system uses multiple machine learning models to infer labels for declined transactions. This intermediary approach allows the system to include previously excluded declined transactions in the training data while maintaining label quality through cross-model validation and confidence scoring mechanisms.
Solution Approach 2:
The patent changes the parameter of label certainty by introducing confidence scores and threshold-based filtering. Instead of binary label inclusion/exclusion, the system uses continuous confidence metrics to determine which inferred labels are reliable enough to include in training data, thus improving both completeness and maintaining accuracy.
2Reliability
If declined transactions are excluded from training data, then label reliability is maintained, but model accuracy deteriorates
Solution Approach 1:
The patent applies preliminary labeling actions to declined transactions before they are used in training. Multiple machine learning models pre-infer labels for declined transactions, and these preliminary labels are then validated through confidence scoring and threshold filtering. This preliminary action ensures that when declined transactions are added to training data, they come with reliable inferred labels rather than missing or unreliable labels.
3Productivity
If rule-based methods are used to infer labels for declined transactions, then processing speed is improved, but subjectivity and error rates increase
Solution Approach 1:
The patent substitutes mechanical rule-based labeling systems with machine learning model-based inference systems. Instead of using predetermined rigid rules that are prone to subjectivity, the system employs trained machine learning models that learn patterns from data, providing more objective and accurate label inference for declined transactions while maintaining efficient processing speeds.
Data Source
AI summary
There are provided systems and methods for rule agnostic reject inferencing. An example method may receive a request for processing a transaction, and determine, using a machine learning model, a classification for the transaction based on data associated with the transaction. The machine learning model may be trained using first training data having verified labels and second training data having inferred labels, and the inferred labels of the second training data may be generated based on a distribution of classifications associated with the first training data. The example method may further process the request based on the classification.