Feedback-driven causal contribution node and explainable pathway analysis system

Abstract

This invention discloses a feedback-based causal contribution node and explanatory path analysis system applied to a horizontal federated learning architecture involving multiple data providers. Its aim is to quantify the contribution of each data node and its internal data to model performance and causality without sharing the original data, and to provide traceable, visualized, and strategically valuable contribution sharing and data optimization suggestions. The system includes modules for confirming the dataset, confirming contribution rules, and providing feedback and explanatory path analysis. It incorporates methods such as Shapley scores, propensity score, and influence functions to integrate model performance, data heterogeneity, and cost assessment, generating structured contribution quantification and contribution sharing suggestions. The system also features a feedback module that can provide optimization directions and high-contribution suggestions based on data types, promoting fair cooperation and sustainability among nodes, and providing a concrete solution for privacy-preserving artificial intelligence applications.

Claims

1. A feedback-based causal contribution node and interpretive path analysis system, applied to a horizontal federated learning architecture, comprising: A dataset verification module group is used to process and verify the dataset of federated learning. The dataset verification module group includes: a dataset allocation and selection module, which is used to divide the dataset according to the data purpose set by the model initiator, and can set the initial profit sharing weight according to the data characteristics of the participants. And a data governance module, which is used to verify and process whether the data provided by participants meets the format, structure and consistency requirements of federated learning; a group of contribution confirmation rules, which is used to establish and reveal the contribution evaluation rules of federated learning, including: a contribution influencing factor selection module, which is used to set cost items and weighting factors to adjust data contributions; and a causal feature selection module, which is used to determine the features required for bias score modeling through statistical analysis, domain knowledge or feature interaction selection. The system includes a contribution rule transparency module to compile and reveal the settings for participants to confirm; and a profit sharing feedback and interpretive path analysis module group to calculate and provide profit sharing suggestions for each participant. This module group includes: a model performance contribution module to analyze the marginal improvement of model performance from single data and assess the potential value of new data types for model training; and a causal feature contribution calculation module to treat data from specific participants as the treatment group and data from other participants as the non-treatment group, using the causal feature selection module to determine the feature items. The system employs a propensity score model to calculate a propensity score for each data point, assesses the causal contribution of individual data points based on these scores, and generates suggested data patterns with high potential value. A contribution quantification module integrates model performance, causal analysis, and weighting factors to calculate the contribution ratio of each participant. An interpretive path analysis module integrates the information output from each module, calculates the final contribution results, and generates reports for participants to review. Finally, a profit-sharing feedback module provides profit-sharing suggestions based on the final contribution results and provides feedback on the training value of the data patterns for the model, supporting subsequent data strategy optimization.

2. The system as described in claim 1, wherein, The dataset allocation and selection module is further used to classify data according to different data usage types.

3. The system as described in claim 1, wherein, The contribution factor selection module is further used to set at least one of the following cost factors: raw data assetization cost, additional human resource cost, computing resource cost, or other quantifiable cost.

4. The system as described in claim 1, wherein, The contribution influence factor selection module is further used to set at least one of the following weighting mechanisms: model initiator data weighting, missing value data weight reduction, or special group data weighting.

5. The system as described in claim 1, wherein, The model performance contribution module further includes: a baseline model generator, which generates a model without specific participant data; and an incremental contribution analyzer, which calculates the performance improvement of the model after adding specific participant data, and can output Shapley values ​​or influence functions as contribution evaluation metrics.

6. The system as described in claim 1, wherein, The causal feature contribution calculation module uses logistic regression to model the propensity score.

7. The system as described in claim 1, wherein, The contribution quantification module is used to sum up the contribution values ​​of individual data points and calculate the contribution ratio of each participant's overall data to the final model.

8. The system as described in claim 1, wherein, This explanatory path analysis module is further used to integrate profit results, cost items, individual contribution values, total contribution ratio of participants, and distributable profit information to generate corresponding reports and profit-sharing suggestions.

9. The system as described in claim 1, wherein, This revenue sharing feedback module is used to feed back information on the potential contribution of data types to model training and suggested data types to data providers, in order to optimize subsequent data collection strategies and support integration into blockchain smart contracts as part of a dynamic revenue sharing mechanism.

Images