Longitudinal federated learning explainability control method, device and processing system
By combining neural networks and local linear models, the interpretability problem caused by data isolation in vertical federated learning is solved, enabling interpretability and visualization analysis across data sources, locating the causes of prediction failures, and evaluating data contribution.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 太保科技有限公司
- Filing Date
- 2022-08-12
- Publication Date
- 2026-06-23
AI Technical Summary
In longitudinal federated learning, due to data isolation, traditional ex-post interpretation methods such as LIME and SHAP cannot be directly applied, resulting in the inability to interpret prediction results across data sources.
Prediction is made using a neural network model, and a local linear model is trained to evaluate the feature contribution of each data source. The effect weights of the local models are then used to weight the results, generating a globally interpretable model.
It enables interpretable and visualized prediction results while protecting data privacy, quantitatively analyzes the reasons for prediction failures, and assesses the consistency of data contribution.
Smart Images
Figure CN115271100B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of big data, and specifically to the application of federated learning for external data collaboration within enterprises. It relates to a general-purpose vertical federated learning interpretability control method and corresponding control devices and processing systems. Background Technology
[0002] Machine learning applications in industries such as finance, healthcare, and law place high demands on the interpretability of prediction results. This is especially true regarding the analytical methods used by machine learning models to determine their predictions. Most complex machine learning models, such as neural networks and deep learning models, are black-box models, making it almost impossible to fully understand their internal workings. Currently, the interpretation of prediction results generally employs ex-post interpretation and model-independent methods. For example, the literature Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin, "Why should I trust you?" explaining the predictions of any classifier, in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, 1135–1144, 2016 (hereinafter referred to as: Prior Art Literature 3), describes LIME (Local Interpretable Model-agnostic Explanations). The literature Lundberg, Scott M., and Su-In Lee, "A unified approach to interpreting model predictions." Advances in Neural Information Processing Systems, 2017 (hereinafter referred to as: Prior Art Literature 4), describes methods such as SHAP (Shapely Additive Explanation). LIME utilizes interpretable models (such as linear models and decision trees) to locally approximate the prediction results of complex black-box models. SHAP is a method that uses approximate calculations of Shapley values from game theory to describe the marginal contribution of features and thus determine custom weights.However, in the field of federated learning, different technical solutions are described in documents such as the book "Federated Learning" (published by Electronic Industry Press, May 2020, hereinafter referred to as: Prior Art Document 1) and the document Peter Kairouz, H. Brendan McMahan, et al. "Advances and Open Problems in Federated Learning." arXiv:1912.04977 [cs.LG] 10 Dec 2019 (hereinafter referred to as: Prior Art Document 2). Under the premise of data privacy isolation between the two parties, methods such as LIME / SHAP cannot be used directly. This point can also be referenced in the aforementioned Prior Art Documents 3 and 4. Currently, there is no practical ex-post interpretation and model-independent method in the field of federated learning.
[0003] In recent years, vertical federated learning has played a significant role in jointly modeling important internal and external data within enterprises, and has been increasingly applied by companies in various business scenarios such as fraud prevention and marketing. However, in the process of enterprises using external data for federated learning, the data of each party cannot be exchanged with the other party's server. Figure 1 A schematic diagram of a federated learning technology solution in an existing technological scenario is shown. For example... Figure 1 As shown, in federated modeling between enterprise A11 and enterprise B12, data between enterprises A11 and B12 is not shared. However, gradient information for error learning between local models A13 and B14 is shared encrypted. Therefore, the training data itself cannot be shared with each other. Figure 1 This illustrates a dilemma of the aforementioned technical solution.
[0004] See again Figures 2 to 4 The situation illustrated further demonstrates the predicament encountered by existing technologies. For example... Figure 2 This is a diagram of the LIME algorithm in a non-federated learning scenario. When applying a longitudinal federated learning model, if you want to assess the model's interpretability, you cannot directly use a local post-interpretation algorithm. Local post-interpretation models, such as the LIME algorithm, require random sampling from the surrounding data, making predictions, and then feeding all the prediction results into a local interpretable model, such as a linear model, for training. Then, the trained samples are used to predict the data to be interpreted, and the contribution of each data point to the result is observed.
[0005] like Figure 2As shown, if we want to know how a trained model classifies or predicts data around a sample S, we would sample n new samples around S and submit them to a complex model for prediction, obtaining n prediction results. Then, we would submit the finite n samples X and n prediction results Y to a linear model for training, calculating the contribution of each feature. Based on this feature contribution, we would determine the contribution of each feature in the prediction of sample S. Therefore, this local training requires plaintext data aggregation for analysis. Thus, local post-hoc interpretation algorithms (e.g., the LIME algorithm) cannot be directly applied. Secondly, the real samples from one side cannot be directly transmitted to the other, making it difficult to explain the contribution of the data from both sides.
[0006] Figure 3 Furthermore, it demonstrates that in the prior art, in a vertical federated learning scenario, such as Figure 3 As shown, feature K consists of K1-dimensional features from company A and K2-dimensional features from company B. Due to data isolation, data from company B cannot be directly transmitted to company A, resulting in the inability to apply the LIME algorithm. Summary of the Invention
[0007] Those skilled in the art will understand that the technical solution provided by this invention at least addresses the problem that traditional ex post interpretation methods are not applicable in vertical federated learning and data isolation environments.
[0008] Therefore, to address the shortcomings of existing technologies, this invention provides a longitudinal federated learning interpretability control method for evaluating modeling effectiveness and analyzing data contribution in cross-data source privacy data collaboration, comprising the following steps:
[0009] a. Determine the sampled data X from data source A. A ={X1 A ,...,X n A} and validation data At, to determine the sampled data X of data source B. B ={X1 B ,...,X n B} and verification data Bt, where X i A =[x Ai 1 ,...,x Ai K1 ],X i B =[x Bi 1 ,...,x Bi K2 The data source A is represented as [x] A 1,...,x A K1 Data source B is represented as [x] B 1 ,...,x B K2 ];
[0010] b. Apply a neural network model to predict the above sampled data, and use a local model approach to predict the effect, and calculate the effect weight W of data source A. A =ACC A / (ACC A +ACC B The effect weight W of data source B B =ACC B / (ACC A +ACC B ), of which ACC A For the local model effect of data source A, ACC B This is the effect of a partial model of data source A;
[0011] c. Using a neural network model, predict the above sampled data to obtain the predicted result Y = {y1,...,y}. n}, in conjunction with X A ={X1 A ,...,X n A} and Y = {y1,...,y n Training a Local Linear Model (LineModel) A and through LineModel A Obtain the local contribution FW of K1 features of data source A A =[fw A 1 ,...,fw A K1 ]; United X B ={X1 B ,...,X n B} and Y = {y1,...,y n Training a Local Linear Model (LineModel) B and through LineModel B Analyze the local contribution of K2 features of data source B. B =[fw B 1 ,...,fw B K2 ];
[0012] d. Based on the effect weight W of data source AA and the effect weight W of data source B B We obtain the WFW by weighting the local contributions of K1 features from data source A and K2 features from data source B respectively. A =W A *FW A =[W A *fw A 1 ,...,W A *fw A K1 WFW B =W B *FW B =[W B *fw B 1 ,...,W B *fw B K2 ].
[0013] Preferably, the above method further includes the step of:
[0014] e. The effect weight W on data source A A , and the effect weight W of data source B B And the weighted feature contribution WFW of K1 features of data source A A The weighted feature contribution WFW of K2 features from data source B B Perform visualization processing to obtain visualization results.
[0015] Preferably, the following steps are included before step a:
[0016] i. Using vertical federated learning, the K1-dimensional features of data source A are jointly trained with the K2-dimensional features of data source B.
[0017] According to another aspect of the present invention, a longitudinal federated learning interpretability control device is provided for evaluating the modeling effectiveness and analyzing data contribution of cross-data source privacy data collaboration, comprising the following apparatus:
[0018] A first determining device is used to determine sampled data X from data source A. A ={X1 A ,...,X n A} and validation data At, to determine the sampled data X of data source B. B ={X1 B ,...,X n B} and verification data Bt, where X iA =[x Ai 1 ,...,x Ai K1 ],X i B =[x Bi 1 ,...,x Bi K2 The data source A is represented as [x] A 1 ,...,x A K1 Data source B is represented as [x] B 1 ,...,x B K2 ];
[0019] The first prediction device is used to apply a neural network model to predict the above-mentioned sampled data, and to predict the effect using a local model approach, and to calculate the effect weight W of data source A. A =ACC A / (ACC A +ACC B The effect weight W of data source B B =ACC B / (ACC A +ACC B ), of which ACC A For the local model effect of data source A, ACC B This is the effect of a partial model of data source A;
[0020] The second prediction device is used to predict the above-mentioned sampled data using a neural network model, and obtain the prediction result Y = {y1,...,y}. n}, in conjunction with X A ={X1 A ,...,X n A} and Y = {y1,...,y n Training a Local Linear Model (LineModel) A and through LineModel A Obtain the local contribution FW of K1 features of data source A A =[fw A 1 ,...,fw A K1 ]; United X B ={X1 B ,...,X n B} and Y = {y1,...,y n Training a Local Linear Model (LineModel) B and through LineModel B Analyze the local contribution of K2 features of data source B. B =[fw B 1 ,...,fw B K2 ];
[0021] The first calculation device is used to calculate the effect weight W according to the data source A. A and the effect weight W of data source B B We obtain the WFW by weighting the local contributions of K1 features from data source A and K2 features from data source B respectively. A =W A *FW A =[W A *fw A 1 ,...,W A *fw A K1 WFW B =W B *FW B =[W B *fw B 1 ,...,W B *fw B K2 ].
[0022] Preferably, the control device further includes:
[0023] The second calculation device is used to assign effect weights W to data source A. A , and the effect weight W of data source B B And the weighted feature contribution WFW of K1 features of data source A A The weighted feature contribution WFW of K2 features from data source B B Perform visualization processing to obtain visualization results.
[0024] Preferably, the control device further includes:
[0025] The first processing unit is used for joint training by combining the K1-dimensional features of data source A with the K2-dimensional features of data source B in a longitudinal federated learning manner.
[0026] Preferably, the joint training is implemented through a neural network model.
[0027] According to another aspect of this invention, a general-purpose longitudinal federated learning interpretability processing system is also provided, comprising:
[0028] One or more servers store multiple data sources.
[0029] At least one main processor includes the control device described above, and the control device processes multiple data sources according to the control method described above.
[0030] Preferably, the processing results of the multiple data sources by the above method, control device, and processing system can be displayed visually, and the displayed content includes any one or more of the following:
[0031] A weighted comparison of the evaluation results based on their effectiveness is presented.
[0032] A comparison of positive and negative sample contributions; and
[0033] A comparison of data contributions from different companies is presented.
[0034] The technical solution provided by this invention can achieve at least the following advantages:
[0035] 1. A general interpretability method for longitudinal federated learning, applicable to various longitudinal federated learning algorithms. It provides post-hoc visualization of interpretability, offering a basis for case-level data analysis. In federated learning, because real data cannot be shared between the two systems, it's impossible to determine the reasons for prediction failures at the case-level. This method allows for quantitative analysis of erroneous cases at the case level, pinpointing the causes of prediction failures.
[0036] 2. Achieve interpretability while protecting the privacy of real data. This addresses the issue of privacy concerns preventing interactive analysis of prediction results from real individual data. Because the characteristics of each party are randomly generated and not real data, and the prediction results are also not real data, it ensures that real data does not leave the local machine. Through a linear model learned from the generated data, interpretable interpretation of the real data is possible.
[0037] 3. Analyzing the contribution of data from all parties helps to evaluate the validity of the data at the case-by-case level. This ensures consistency between the overall effect of the data and the analysis at the individual case level.
[0038] 4. Comprehensive visualization solution. It not only visualizes the contribution of features from both sides, but also visualizes the predicted contribution of the current data. Attached Figure Description
[0039] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:
[0040] Figure 1 A schematic diagram of a federated learning technology solution in an existing technology scenario is shown;
[0041] Figure 2 A schematic diagram of a federated learning technology solution in an existing technology scenario is shown;
[0042] Figure 3 A schematic diagram of a federated learning technology solution in an existing technology scenario is shown;
[0043] Figure 4 A schematic diagram comparing the general-purpose vertical federated learning technology solution provided by the present invention with existing vertical federated learning technology solutions in various scenarios is shown.
[0044] Figure 5 A flowchart of a longitudinal federated learning interpretability control method according to an embodiment of the present invention is shown;
[0045] Figure 6 It shows Figure 5 A schematic diagram of the longitudinal federated learning interpretability control method is shown.
[0046] Figure 7 A visualization of a typical application scenario of the present invention is shown; and
[0047] Figure 8 A block diagram of a longitudinal federated learning interpretability control device according to an embodiment of the present invention is shown. Detailed Implementation
[0048] The above Figure 4 The example shown illustrates a schematic diagram comparing the provided federated learning prediction result interpretable platform with a traditional federated learning platform in a preferred embodiment of the present invention. Specifically, compared with the traditional federated learning platform 41, the interpretable federated learning platform 42 provided by the present invention further includes a post-explanation model 421. Introducing a post-explanation model 421 into the federated learning platform is essential. After federated learning modeling, it is also necessary to analyze prediction error cases to determine the causes of the errors, accumulating experience for subsequent model improvement and data enhancement. However, due to the lack of data isolation in federated scenarios, local interpretability models are not applicable. The present invention provides an efficient post-explanation model (such as... Figure 6 As shown in the figure, it facilitates the interpretation and visualization of prediction results, enhancing users' understanding and trust in federated modeling.
[0049] Traditional ex-post interpretation methods such as LIME are not suitable for applications in data isolation scenarios. This invention is suitable for the interpretability of prediction results in a federated data isolation environment. It trains a linear model that locally interprets the prediction results for each data contributor, and then weights the models according to their prediction performance to obtain a globally interpretable model.
[0050] Furthermore, those skilled in the art will understand that the general federated learning provided by this invention can solve the problem of data isolation between multiple data sources. For example, in the case of multiple enterprises jointly conducting federated learning, the technical solution provided by this invention enables data comparison and joint computation among multiple enterprises, thereby providing a visual display of the sharing of an event by multiple enterprises. Figure 5 as well as Figure 6 The general federated learning method provided by this invention is described in detail.
[0051] Figure 5 A flowchart of a longitudinal federated learning interpretability control method 500 according to an embodiment of the present invention is shown. Figure 6 It shows Figure 5 A schematic diagram of the interpretability control method for vertical federated learning. The interpretability control method 500 for vertical federated learning (hereinafter referred to as interpretability control method 500) is used for evaluating the modeling effectiveness and analyzing data contribution in cross-data source privacy data collaboration. As shown in Figure 5, interpretability control method 500 includes the following steps:
[0052] a. Determine the sampled data X from data source A. A ={X1 A ,...,X n A} and validation data At, to determine the sampled data X of data source B. B ={X1 B ,...,X n B} and verification data Bt, where X i A =[x Ai 1 ,...,x Ai K1 ],X i B =[x Bi 1 ,...,x Bi K2 The data source A is represented as [x] A 1 ,...,x A K1 Data source B is represented as [x] B 1 ,...,xB K2 ];
[0053] b. Apply a neural network model to predict the above sampled data, and use a local model approach to predict the effect, and calculate the effect weight W of data source A. A =ACC A / (ACC A +ACC B The effect weight W of data source B B =ACC B / (ACC A +ACC B ), of which ACC A For the local model effect of data source A, ACC B This is the effect of a partial model of data source A;
[0054] c. Using a neural network model, predict the above sampled data to obtain the predicted result Y = {y1,...,y}. n}, in conjunction with X A ={X1 A ,...,X n A} and Y = {y1,...,y n Training a Local Linear Model (LineModel) A and through LineModel A Obtain the local contribution FW of K1 features of data source A A =[fw A 1 ,...,fw A K1 ]; United X B ={X1 B ,...,X n B} and Y = {y1,...,y n Training a Local Linear Model (LineModel) B and through LineModel B Analyze the local contribution of K2 features of data source B. B =[fw B 1 ,...,fw B K2 ];
[0055] d. Based on the effect weight W of data source A A and the effect weight W of data source B BWe obtain the WFW by weighting the local contributions of K1 features from data source A and K2 features from data source B respectively. A =W A *FW A =[W A *fw A 1 ,...,W A *fw A K1 WFW B =W B *FW B =[W B *fw B 1 ,...,W B *fw B K2 ].
[0056] In some embodiments, prior to step a, the longitudinal federated learning interpretability control method further includes the following step i: performing joint training by combining the K1-dimensional features of data source A with the K2-dimensional features of data source B according to longitudinal federated learning.
[0057] In some embodiments, the longitudinal federated learning interpretability control method further includes step e. Weighting the effect W of data source A. A , and the effect weight W of data source B B And the weighted feature contribution WFW of K1 features of data source A A The weighted feature contribution WFW of K2 features from data source B B Perform visualization processing to obtain visualization results.
[0058] refer to Figure 6 As shown, in a vertical federated learning scenario, this invention trains and interprets linear models for each isolated party separately, and then interprets the linear model by weighting the prediction results of each party. In a preferred embodiment, the general federated learning method provided by this invention can be implemented through the following steps:
[0059] Step 1: Training the model to be explained: Using vertical federated learning, the K1-dimensional features of company A are jointly trained with the K2-dimensional features of company B to obtain the federated model. The federated model can be a neural network model 61.
[0060] Step 2: Case-Related Data Preparation: Assume there is a case, and the data for Company A in this case is [x] A 1 ,...,x A K1 The data for Company B is [x] B1 ,...,x B K2 Two sets of data need to be prepared.
[0061] Data Set 1: Data used to train a locally interpretable linear model. Following a Gaussian distribution, n data points were randomly sampled from the surrounding dataset to generate n randomly generated samples. The randomly generated sample from Company A's data is X. A ={X1 A ,...,X n A}, where X i A =[x Ai 1 ,...,x Ai K1 The sample data randomly generated by company B is X. B ={X1 B ,...,X n B}, where X i B =[x Bi 1 ,...,x Bi K2 ].
[0062] Data set 2: Data to verify the effect. Based on the data from side A, find the distance from side A [x]. A 1 ,...,x A K1 The most recent t cases are used to obtain data to verify the effectiveness of the data, based on data provided by Party A for each of the t cases.
[0063] Step 3: Performance Evaluation: Then, the neural network model 61 is applied for prediction, and its performance is evaluated, for example, by using accuracy (ACC). Then, the performance is predicted using local models, and the performance of local model A611 for company A is obtained as ACC. A The local model B612 of company B has the effect of ACC. B Calculate the effect weight W for company A. A =ACC A / (ACC A +ACC B The effect weight W of enterprise B B =ACC B / (ACC A +ACC B ).
[0064] Step 4: Train a locally interpretable linear model: Let X be a randomly generated sample from the joint venture A and venture B. A and X B The neural network model 61 is used for prediction, and the prediction result Y = {y1,...,y} is obtained. n Then the predicted result Y = {y1,...,y} is calculated. n The data is then given to Company A and Company B respectively, and their respective local linear models are trained.
[0065] Company A's Local Linear Model A 621: United X A ={X1 A ,...,X n A} and Y = {y1,...,y n Training a locally linear model (LineModel) A 621, then through LineModel A 621. Analysis of the local contribution of K1 features of company A yielded FW A =[fw A 1 ,...,fw A K1 ].
[0066] Company B's Local Linear Model B 622: United X B ={X1 B ,...,X n B} and Y = {y1,...,y n Training a locally linear model (LineModel) B 622, then through LineModel B 622 analysis of the local contribution of K2 features of enterprise B yielded FW B =[fw B 1 ,...,fw B K2 ].
[0067] Step 5: Weighted Interpretable Linear Model: Based on the effect weights W of Company A A And the weighting of the effect on enterprise B (W) B By weighting the local contributions of K1 features of company A and K2 features of company B respectively, we obtain the joint linear model 62:
[0068] WFW A =WA *FW A =[W A *fw A 1 ,...,W A *fw A K1 ];
[0069] WFW B =W B *FW B =[W B *fw B 1 ,...,W B *fw B K2 ];
[0070] Step 6: Weight the effect W of Company A A And the weighting of the effect on enterprise B (W) B And the weighted feature contribution WFW of K1 features of enterprise A. A The weighted feature contribution of K2 features of Company B (WFW) B Visualization. Specifically, the visualization results can be found in [reference needed]. Figure 7 The diagram shown is as follows. Figure 7 As shown, features A1 and B2 have the highest contribution to a prediction result of 0. Features A5 and A2 have the highest contribution to a prediction result of 1. Features of company A contribute 64% to the prediction result, while features of company B contribute 36%. This invention implements a weighted interpretation of the prediction result based on the contribution of each feature to the prediction effect, achieving not only interpretability but also allowing for comparative interpretation of the contributions of each feature based on the prediction effect. Therefore, compared to other interpretable strategies, this method is based on the contribution analysis of data from each party.
[0071] This invention uses a locally linearly weighted model as the base model for a post-hoc interpretable model. The linearly weighted model has excellent local interpretability and can display the prediction results in a weighted manner. Unlike interpretable modeling methods such as scorecard models, the post-hoc interpretable model implemented in this invention is a general-purpose interpretation method suitable for interpreting the prediction results of non-interpretable models such as neural networks and decision trees.
[0072] Figure 8A system block diagram of a longitudinal federated learning interpretability control device 800 according to an embodiment of the present invention is shown. The longitudinal federated learning interpretability control device 800 (hereinafter referred to as interpretability control device 800) is used for evaluating the modeling effect and analyzing data contribution of cross-data source privacy data collaboration. The interpretability control device 800 includes: a first determining device 81, a first predicting device 82, a second predicting device 83, and a first computing device 84.
[0073] The first determining device 81 is used to determine the sampled data X from data source A. A ={X1 A ,...,X n A} and validation data At, to determine the sampled data X of data source B. B ={X1 B ,...,X n B} and verification data Bt, where X i A =[x Ai 1 ,...,x Ai K1 ],X i B =[x Bi 1 ,...,x Bi K2 The data source A is represented as [x] A 1 ,...,x A K1 Data source B is represented as [x] B 1 ,...,x B K2 ];
[0074] The first prediction device 82 is used to apply a neural network model to predict the above-mentioned sampled data, and to predict the effect through a local model method, and to calculate the effect weight W of data source A. A =ACC A / (ACC A +ACC B The effect weight W of data source B B =ACC B / (ACC A +ACC B ), of which ACC A For the local model effect of data source A, ACC B This is the effect of a partial model of data source A;
[0075] The second prediction device 83 is used to predict the above-mentioned sampled data using a neural network model, and obtain the prediction result Y = {y1,...,y}. n}, in conjunction with X A ={X1 A ,...,X n A} and Y = {y1,...,y n Training a Local Linear Model (LineModel) A and through LineModel A Obtain the local contribution FW of K1 features of data source A A =[fw A 1 ,...,fw A K1 ]; United X B ={X1 B ,...,X n B} and Y = {y1,...,y n Training a Local Linear Model (LineModel) B and through LineModel B Analyze the local contribution of K2 features of data source B. B =[fw B 1 ,...,fw B K2 ];
[0076] The first computing device 84 is used to calculate the effect weight W of data source A. A and the effect weight W of data source B B We obtain the WFW by weighting the local contributions of K1 features from data source A and K2 features from data source B respectively. A =W A *FW A =[W A *fw A 1 ,...,W A *fw A K1 WFW B =W B *FW B =[W B *fw B 1 ,...,W B *fw B K2 ].
[0077] In some embodiments, the longitudinal federated learning interpretability control device further includes a second computing device for assigning effect weights W to data source A. A , and the effect weight W of data source B B And the weighted feature contribution WFW of K1 features of data source A A The weighted feature contribution WFW of K2 features from data source B B Perform visualization processing to obtain visualization results.
[0078] In some embodiments, the longitudinal federated learning interpretability control apparatus further includes a first processing unit for jointly training K1-dimensional features of data source A and K2-dimensional features of data source B by longitudinal federated learning.
[0079] The present invention also provides a general-purpose longitudinal federated learning interpretability processing system, comprising: one or more servers storing multiple data sources, at least one main processor including the control device described above, and the control device processing the multiple data sources according to the control method described above.
[0080] In some embodiments, the processing results of the multiple data sources by the above-described method, control device, and processing system can be displayed in a visual manner. The display content includes any one or more of the following: a weighted comparison display of the evaluation results based on effectiveness; a comparison display of the contributions of positive and negative samples; and a comparison display of the data contributions of different enterprises.
[0081] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.
Claims
1. A longitudinal federated learning interpretability control method for evaluating modeling effectiveness and analyzing data contribution in cross-data source privacy data collaboration, characterized in that, Includes the following steps: a. Determine the sampled data X from data source A. A = {X1 A , ..., X n A } and validation data At, to determine the sampled data X of data source B. B = {X1 B , ..., X n B } and verification data Bt, where X i A =[x Ai 1 ,...,x Ai K1 ], X i B =[x Bi 1 ,...,x Bi K2 The data source A is private data with business characteristics from enterprise A, represented as [x]. A 1 ,...,x A K1 ], where feature dimension K1 corresponds to the business data dimension of enterprise A, and data source B is private data with business characteristics from enterprise B, represented as [x B 1 ,...,x B K2 The feature dimension K2 corresponds to the business data dimension of enterprise B, and the data source A and the data source B are in a data isolation state; b. Apply a neural network model to predict the above sampled data, and use a local model approach to predict the effect, and calculate the effect weight of data source A. W A =ACC A / (ACC A +ACC B ), the effect weight of data source B W B =ACC B / (ACC A +ACC B ), of which ACC A For the local model effect of data source A, ACC B This is the effect of a partial model of data source A; c. Using a neural network model, predict the above sampled data to obtain the predicted result Y = {y1, ..., y...} n }, in conjunction with X A = {X1 A , ..., X n A } and Y = {y1, ..., y n Training a Local Linear Model (LineModel) A and through LineModel A Obtain the local contribution FW of K1 features of data source A A =[fw A 1 ,...,fw A K1 ]; United X B = {X1 B ,..., X n B } and Y = {y1, ..., y n Training a Local Linear Model (LineModel) B and through LineModel B Analyze the local contribution of K2 features of data source B. B =[fw B 1 ,...,fw B K2 ]; d. Weighting based on data source A W A and the effect weight of data source B W B We obtain the WFW by weighting the local contributions of K1 features from data source A and K2 features from data source B respectively. A = W A FW A =[ W A fw A 1 ,..., W A fw A K1 WFW B = W B FW B =[ W B fw B 1 ,..., W B fw B K2 ]; e. Effect weighting of data source A W A , and the effect weight of data source B W B And the weighted feature contribution WFW of K1 features of data source A A The weighted feature contribution WFW of K2 features from data source B B Perform visualization processing to obtain visualization results.
2. The method according to claim 1, characterized in that, The following steps are included before step a: i. Using vertical federated learning, the K1-dimensional features of data source A are jointly trained with the K2-dimensional features of data source B.
3. The method according to claim 2, characterized in that, The joint training is achieved through a neural network model.
4. A longitudinal federated learning interpretability control device for evaluating modeling effectiveness and analyzing data contribution in cross-data source privacy data collaboration, characterized in that, Includes the following devices: A first determining device is used to determine sampled data X from data source A. A = {X1 A , ..., X n A } and validation data At, to determine the sampled data X of data source B. B = {X1 B , ..., X n B } and verification data Bt, where X i A =[x Ai 1 ,...,x Ai K1 ],X i B =[x Bi 1 ,...,x Bi K2 The data source A is private data with business characteristics from enterprise A, represented as [x]. A 1 ,...,x A K1 ], where feature dimension K1 corresponds to the business data dimension of enterprise A, and data source B is private data with business characteristics from enterprise B, represented as [x B 1 ,...,x B K2 ], its feature dimension K2 corresponds to the business data dimension of enterprise B, and the data source A and the data source B are in a data isolation state; The first prediction device is used to apply a neural network model to predict the above-mentioned sampled data, and to predict the effect using a local model approach, and to calculate the effect weight of data source A. W A =ACC A / (ACC A +ACC B ), the effect weight of data source B W B =ACC B / (ACC A +ACC B ), of which ACC A For the local model effect of data source A, ACC B This is the effect of a partial model of data source A; The second prediction device is used to predict the above-mentioned sampled data using a neural network model, and obtain the prediction result Y = {y1, ..., y}. n }, in conjunction with X A = {X1 A , ..., X n A } and Y = {y1, ..., y n Training a Local Linear Model (LineModel) A and through LineModel A Obtain the local contribution FW of K1 features of data source A A =[fw A 1 ,...,fw A K1 ]; United X B = {X1 B , ..., X n B } and Y = {y1, ..., y n Training a Local Linear Model (LineModel) B and through LineModel B Analyze the local contribution of K2 features of data source B. B =[fw B 1 ,...,fw B K2 ]; The first calculation device is used to calculate the effect weights of data source A. W A and the effect weight of data source B W B We obtain the WFW by weighting the local contributions of K1 features from data source A and K2 features from data source B respectively. A = W A FW A =[ W A fw A 1 ,..., W A fw A K1 WFW B = W B FW B =[ W B fw B 1 ,..., W B fw B K2 ]; The second computing device is used to perform visualization processing on the effect weight WA of data source A, the effect weight WB of data source B, the weighted feature contribution WFWA of K1 features of data source A, and the weighted feature contribution WFWB of K2 features of data source B, to obtain visualization results.
5. The control device according to claim 4, characterized in that, The control device shown also includes: The first processing unit is used for joint training by combining the K1-dimensional features of data source A with the K2-dimensional features of data source B in a longitudinal federated learning manner.
6. The control device according to claim 5, characterized in that, The joint training is achieved through a neural network model.
7. A general-purpose longitudinal federated learning interpretability processing system, comprising: One or more servers store multiple data sources. At least one main processor includes a control device according to any one of claims 4 to 6, and the control device processes a plurality of data sources according to the control method according to any one of claims 1 to 3; The processing results of the multiple data sources are displayed visually, and the displayed content includes any one or more of the following: A weighted comparison of the evaluation results based on their effectiveness is presented. A comparison of positive and negative sample contributions; and A comparison of data contributions from different companies is presented.