Ultimate Reason Code Generation for Reliable Nonlinear Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating reason codes for non-linear complex models in big data applications, such as neural networks, rely on assumptions that the probability consistency holds when input variables are modified without retraining, which is not always valid, leading to inconsistent results.
Innovation Solution
A four-step approach involving training a base model and subsequent models with knocked-out variables, comparing score differences to identify the ultimate reason code for high scores, ensuring consistency with probability of the target.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single base model is used to compute reason codes by modifying input variables, then the process is simple and fast, but the probability consistency assumption does not hold and results are unreliable
Solution Approach 1:
The patent segments the single model approach into multiple sub-models, each trained with specific knocked-out variables. This segmentation allows each sub-model to be optimized for specific reason code scenarios while maintaining overall system reliability through comparative analysis of their outputs.
Solution Approach 2:
The patent applies preliminary action by pre-training multiple sub-models with different variable knockouts before actual inference. This preliminary training ensures that when reason codes are generated, the system already has optimized models ready, avoiding the need for on-the-fly retraining and ensuring probability consistency assumptions are met.
2Measurement precision
If multiple sub-models are trained with knocked-out variables, then probability consistency is ensured and reason codes are more reliable, but the training process becomes more complex and time-consuming
Solution Approach 1:
The patent applies partial action by training sub-models with selective variable knockouts rather than all possible combinations. This partial approach focuses computational resources on the most critical variables for each business scenario, achieving sufficient precision without the exponential time cost of exhaustive model training.
Solution Approach 2:
The patent changes parameters by systematically varying which input variables are knocked out in each sub-model training iteration. This parameter change strategy allows the system to explore different model configurations and their impact on reason codes, achieving precise measurements while managing training time through structured parameter variation.
3Reliability
If model parameters are optimized by maximum-likelihood training, then the knocked-out model maintains probability consistency, but retraining each sub-model increases computational cost
Solution Approach 1:
The patent applies preliminary action by performing maximum-likelihood training of multiple sub-models during the offline preparation phase. This preliminary optimization ensures probability consistency is built into the model structures before deployment, allowing real-time inference to reuse these pre-optimized models without incurring repeated training energy costs.
Data Source
AI summary
A system and method for generating ultimate reason codes for computer models is provided. The system for generating ultimate reason codes for computer models comprising a computer system for receiving a data set, and an ultimate reason code generation engine stored on the computer system which, when executed by the computer system, causes the computer system to train a base model with a plurality of reason codes, wherein each reason code includes one or more variables, each of which belongs to only one reason code, train a subsequent model using a subset of the plurality of reason codes, determine whether a high score exists in the base model, determine a scored difference if a high score exists in the base model, and designate a reason code having a largest drop of score as an ultimate reason code.