Risk Feature Extraction for Interpretable Transaction Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks used in risk prediction models lack interpretability, making it difficult to understand and effectively manage risk transactions.
Innovation Solution
A method and apparatus for extracting risk feature descriptions by inputting risk and random transaction data into a pre-trained neural network model to determine risk feature descriptions based on the importance of transaction representations, using linear models and partial derivatives to classify and interpret risk transactions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network-based risk prediction model is used, then risk prediction accuracy is improved, but model interpretability deteriorates
Solution Approach 1:
The patent extracts feature importance information from the neural network model by analyzing the weights and activations of neurons in hidden layers. Specifically, it computes the absolute values of weights connecting neurons to the output layer and aggregates these across multiple neurons to identify which input features contribute most to risk predictions. This extraction process reveals interpretable feature importance rankings while maintaining the model's predictive accuracy.
Solution Approach 2:
The patent introduces an intermediary analysis layer between the neural network model and the final risk assessment. This intermediary component calculates and reports feature importance metrics that serve as a bridge, translating the black-box neural network outputs into human-readable explanations. The intermediary computes weight-based importance scores and presents them in a structured format that stakeholders can understand and act upon.
2Loss of information
If feature description extraction is performed to interpret risk predictions, then model interpretability is improved, but computational complexity increases
Solution Approach 1:
The patent applies partial action by focusing computational resources only on the specific task of feature importance extraction rather than analyzing every possible aspect of the model. It selectively computes weight-based importance metrics for features that have non-zero weights in the neural network, avoiding unnecessary computations for irrelevant features. This partial approach provides sufficient interpretability without excessive computational overhead.
Solution Approach 2:
The patent changes the parameter representation from raw neural network activations to transformed feature importance scores. By converting the complex multi-dimensional activation patterns into simplified importance rankings based on weight magnitudes, the system reduces computational complexity while enhancing interpretability. The parameter transformation makes the model behavior more transparent without requiring exhaustive computational analysis.
Data Source
AI summary
Embodiments of this specification describe a risk feature description extraction method and apparatus. In the method in the embodiments, risk transaction data for risk transaction prediction are obtained, and random transaction data are randomly obtained from transaction record data. Then, the risk transaction data and the random transaction data are separately input into a risk transaction prediction model, and respective transaction representation is output by a neuron layer that is not an output layer of the risk transaction prediction model. Further, a risk feature description capable of being used to perform risk determining can be determined based on importance of each obtained transaction representation in determining whether a risk transaction has a risk.