ML Payment Pattern Prediction via Frequency Bucketing

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional tree-based algorithms for predicting financial transaction patterns are inefficient and inaccurate due to lack of control over tree splitting during training, requiring significant time for large datasets and failing to provide high accuracy.

Innovation Solution

A Machine Learning-based system that includes modules for data reception, granularity level instance creation, payment frequency computation, payment pattern generation, and date prediction, using a regression-based ML model to dynamically generate date-shift-payment pattern features and select optimal patterns for predicting financial transaction payment dates.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If conventional tree-based algorithms are used for predicting payment patterns, then the system can process financial transaction data, but the training and prediction time becomes significantly long for large datasets

Engineering Contradiction:
Improveprediction speedVSAvoidtraining time
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The system performs preliminary actions by pre-processing invoice data and pre-computing payment frequency features before the actual prediction task. This includes extracting relevant features from historical invoice data and organizing them in a structured format that accelerates the subsequent prediction process, thereby reducing the time required during actual prediction operations.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent segments the payment prediction problem into distinct components: invoice data extraction, payment frequency computation, pattern feature generation, and final prediction. By dividing the large dataset into manageable segments and processing them through specialized modules, the system achieves faster training and prediction times compared to applying conventional tree-based algorithms to the entire dataset as a single unit.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If conventional tree-based algorithms are used for predicting payment patterns, then the system can provide payment date predictions, but the accuracy and control over prediction quality deteriorates

Engineering Contradiction:
Improveprediction accuracyVSAvoidalgorithm control
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system changes key parameters by computing payment frequency features and generating date-shift-payment-pattern features that capture temporal patterns in payment behavior. These transformed features provide richer information to the prediction model, improving accuracy while maintaining computational tractability. The ML model uses these engineered features to achieve higher prediction accuracy compared to conventional approaches.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces an intermediary layer between raw invoice data and final predictions, consisting of payment frequency computation and pattern feature generation modules. This intermediary processing extracts meaningful patterns from historical data and transforms them into features that the ML model can effectively utilize, thereby improving prediction accuracy and providing better control over prediction quality.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Ease of manufacture

If simple customer level average approach is used for predicting payment patterns, then the system is easy to implement, but the prediction accuracy remains insufficient

Engineering Contradiction:
Improveimplementation simplicityVSAvoidprediction accuracy
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The system achieves a balance between simplicity and accuracy by creating a multi-functional prediction framework that can handle multiple payment patterns (weekly, bi-weekly, monthly, quarterly) through a single unified ML model. The model universally processes different payment frequencies by using payment frequency features as inputs, eliminating the need for separate simple averaging calculations for each pattern type while maintaining implementation ease.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The system performs preliminary computation of payment frequency features from historical invoice data, organizing payment patterns into structured categories before feeding them to the ML model. This pre-processing step captures complex payment behaviors in advance, allowing the model to make accurate predictions without requiring complex real-time calculations, thus maintaining simplicity while improving accuracy.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4455973A1Machine learning (ML)-based system and method for predicting financial transaction patterns
Publication Date: 2024.10.30 HIGHRADIUS CORP
  • EP4455973A1 patent drawingFigure 1
  • EP4455973A1 patent drawingFigure 2
  • EP4455973A1 patent drawingFigure 3

AI summary

A system and method for predicting financial transaction patterns is disclosed. The method includes receiving invoice data of one or more customers. The method further includes receiving granularity levels, thereby generating a set granularity level instances based on various invoice data attributes. The method further computes payment frequency bucket features for all the generated set of granularity level instances and assigns invoices to clusters based on the set of payment frequency bucket features. The set of payment frequency bucket features assigns a customer to be one of: a weekly payer, alternative weekly payer, a monthly payer, a bi-monthly, a quarterly payer, a half yearly payer and an annual payer. The method further includes the generation of a set of payment pattern features based on the set of payment frequency bucket features of the cluster. Further, the method includes selection of an optimal pattern with highest probability of adherence.