TWI859667B - Intelligent system for pluralistic interpretation general model, method and computer readable medium thereof - Google Patents

Abstract

The present invention is an intelligent system for pluralistic interpretation general model and a method thereof. The system comprises a model and data type analysis module, a model interpretation method application module, and an interpretation effect integration module. The received model file and analysis data are analyzed by the model and data type analysis module to obtain a model type and a data type. At least one interpretation method is queried according to the model type and the data type by the model interpretation method application module, and the model file and the analysis data is analyzed to obtain at least one interpretation result. The at least one interpretation result are integrated into a model interpretation result by the interpretation effect integration module and the model interpretation result is visualized for the user's reference. The user's difficulty in interpreting the model can be solved by the present invention. The present invention also provides a computer-readable medium for executing the method of the present invention.

Description

System, method and computer-readable medium for intelligent multi-interpretation general model

The present invention relates to the technology of interpretation model, and in particular to a system, method and computer-readable medium of an intelligent multi-interpretation universal model capable of integrating multiple interpretation methods.

In machine learning, it is usually necessary to build a model for processing data. Before the model processes data, data for training the model should be collected first. The model is trained by inputting the collected data, and the trained model is used to process data.

In recent years, thanks to the convenience of the Internet and the improvement of computer computing power, collecting data and training machine learning models are no longer just the work of data scientists and artificial intelligence (AI) experts, and machine learning-related applications are becoming more and more common. Most models focus on accuracy indicators when solving predictive problems, that is, the models applied usually focus on improving accuracy. Models with high accuracy often also represent high complexity, so it will be more difficult for people to understand their decision-making patterns. However, if you want to use the model for influential applications, such as medical diagnosis or automatic insurance claims, you must understand how the model works to ensure the trustworthiness of the model. Although there are many explanatory methods that can simulate and explain model behavior, due to the differences in model types and the different topics to which they are applied, it is still necessary to have relevant knowledge to choose an explanatory method that is suitable for the model and topic.

In view of the above problems, how to provide a method to intelligently explain the model, especially to reduce the analysis difficulties caused by the lack of background knowledge of users, and to be based on different types of models and avoid the limitation of non-single application scenarios, so that users can provide help to explain the model behavior when they have no idea about the model decision, will become the goal that people in this technical field are eager to pursue.

To solve the above problems of the prior art, the present invention discloses a system of intelligent multi-dimensional interpretation universal model, which includes: a model and data type analysis module, which is used to analyze the model file to be interpreted and the analysis data used to train the model, so as to obtain the model type and data type corresponding to the model file and the analysis data respectively; a model interpretation method application module, which is connected to the model and data type analysis module, and is used to query according to the model type and the data type to find at least one corresponding interpretation method, and then parse the model file and the analysis data through the at least one interpretation method to obtain at least one interpretation result; and an interpretation effect integration module, which is used to integrate the at least one interpretation result to obtain the model interpretation result.

In one embodiment, the model and data type analysis module comprises a model file analysis unit for determining the model type of the model file and a data type analysis unit for determining the data type of the analysis data.

In another embodiment, the model interpretation method application module includes an interpretable method database for storing multiple interpretation methods, a query method unit for querying the interpretable method database to obtain the at least one interpretation method according to the model type and the data type, a resource controller for confirming the computational resources for executing model analysis, and at least one execution method unit for analyzing the model file and the analysis data according to each interpretation method to obtain each interpretation result.

In another embodiment, each of the interpretation results includes multiple features and multiple feature weights.

In another embodiment, the interpretation effect integration module further includes a weight summing unit for averaging the feature weights of the same feature to obtain an average value, a register for storing the average value, and a weight sorting unit for arranging the multiple features according to the average value.

In another embodiment, the intelligent multivariate interpretation general model system further includes a user interface module that provides a user interface for the user to input the model file and the analysis data.

In another embodiment, the intelligent multi-dimensional interpretation system of the universal model further includes a visualization module for visually displaying the interpretation results of the model.

In another embodiment, the explanation effect integration module uses averaging and sorting, weighted averaging and sorting, or voting to integrate the at least one explanation result.

The present invention further discloses a method for intelligent multivariate interpretation of a universal model, which is executed by a computer device and includes the following steps: a model and data type analysis module is used to analyze a model file to be interpreted and analysis data used for training the model to obtain a model type and a data type corresponding to the model file and the analysis data, respectively; a model interpretation method application module is used to query according to the model type and the data type to find at least one corresponding interpretation method; the model interpretation method application module is used to parse the model file and the analysis data through the at least one interpretation method to obtain at least one interpretation result; and an interpretation effect integration module is used to integrate the at least one interpretation result to obtain a model interpretation result.

In one embodiment, the step of having the model and data type analysis module analyze the model file to be analyzed and the analysis data used to train the model includes: determining the model type of the model file through the model file analysis unit; and determining the data type of the analysis data through the data type analysis unit.

In another embodiment, the step of causing the model interpretation method application module to find at least one corresponding interpretation method and obtain at least one interpretation result includes: the query method unit searches from an interpretable method database storing multiple interpretation methods according to the model type and the data type to obtain the at least one interpretation method; causing the resource controller to confirm the computational resources for executing the model analysis; and causing multiple execution method units to analyze the model file and the analysis data according to each interpretation method to obtain each interpretation result.

In another embodiment, each of the interpretation results includes multiple features and multiple feature weights.

In another embodiment, the step of allowing the interpretation effect integration module to integrate the at least one interpretation result includes: allowing the weight summing unit to average the feature weights of the same feature to obtain an average value; storing the average value in a register; and allowing the weight sorting unit to arrange the multiple features according to the average value.

In another embodiment, the model file and the analysis data are input by the user through a user interface provided by the user interface module.

In another embodiment, the model interpretation result is visualized through a visualization module.

In another embodiment, the explanation effect integration module uses averaging and sorting, weighted averaging and sorting, or voting to integrate the at least one explanation result.

The present invention further discloses a computer-readable medium, which is applied to a computing device or a computer and stores instructions for executing the aforementioned intelligent multi-interpretation general model method.

In summary, the intelligent multi-dimensional interpretation general model system and method of the present invention analyzes the model file and the analysis data through the model and data type analysis module to obtain the model type and data type corresponding to the model file and the analysis data, so that the model method application module uses the multiple interpretation methods found to analyze the model file and the analysis data to obtain at least one corresponding interpretation result, and then the interpretation effect integration module integrates at least one interpretation result to generate a model interpretation result and display it to the user for reference, so as to not only help the user to interpret the model behavior, but also improve convenience and save time; furthermore, the present invention can automatically apply the interpretation method that is most suitable for the model to reduce the difficulty of interpretation; in addition, the present invention provides a visualization effect, which can achieve the purpose of making it easier for users to understand the model decision method.

1: Model and data type analysis module

11: Model file analysis unit

12: Data type analysis unit

2: Model interpretation method application module

21: Database of interpretable methods

22: Query method unit

23: Resource Controller

24: Execution method unit

3: Explain the effect integration module

31: Total weight unit

32: Register

33: Weight sorting unit

4: User interface module

41: User Interface

5: Visualization module

S510~S540: Steps

S600~S650: Steps

710-750: Process

Figure 1 is a schematic diagram of the system architecture of the intelligent multi-interpretation general model of the present invention.

Figure 2 is a diagram showing the internal structure of the model and data type analysis module of the present invention.

Figure 3 is an internal structure diagram of the model interpretation method application module of the present invention.

Figure 4 is a diagram showing the internal structure of the explanation effect integration module of the present invention.

Figure 5 is a step diagram of the method of the intelligent multi-interpretation general model of the present invention.

Figure 6 is a step diagram of another embodiment of the method of the intelligent multi-dimensional interpretation general model of the present invention.

Figure 7 is a flowchart of the specific implementation of the method of intelligent multi-dimensional interpretation of the general model of the present invention.

Figure 8 is a bar graph showing the model explanation results of the present invention in a visual manner.

Figure 9 is a visual display of the text emotion of the present invention.

The following describes the technical content of the present invention through a specific concrete implementation form. People familiar with this technology can easily understand the advantages and effects of the present invention from the content disclosed in this manual. However, the present invention can also be implemented or applied through other different specific implementation forms.

FIG1 is a schematic diagram of the system of the intelligent multivariate interpretation general model of the present invention. As shown in the figure, the system of the intelligent multivariate interpretation general model of the present invention includes a model and data type analysis module 1, a model interpretation method application module 2, and an interpretation effect integration module 3. Accordingly, the model and data type analysis module 1 is used to determine the model type and data type of the input model file and the analysis data, so that the model interpretation method application module 2 can query at least one interpretation method for interpreting the model file and the analysis data. After that, the model file and the analysis data are analyzed by the at least one interpretation method to obtain at least one interpretation result. After being integrated by the interpretation effect integration module 3, the model interpretation result can be finally obtained to provide the user with reference.

The specific description of the intelligent multi-interpretation general model system of the present invention is described as follows.

When receiving the model file to be interpreted and the analysis data used for training the model, the model and data type analysis module 1 analyzes the received model file and analysis data to obtain the model type and data type corresponding to the model file and the analysis data, respectively. That is, the model and data type analysis module 1 determines the model type and data type according to the model file format and internal parameters and the data extension and content. In specific implementation, the analysis data can be model training data (such as table format data, string data or graphic data, etc.) used to train the model, which can be pre-stored in the system with the model file or input by the user from the outside.

In a specific embodiment, as shown in FIG2 , it is an internal structure diagram of the model and data type analysis module of the present invention, wherein the model and data type analysis module 1 includes a model file analysis unit 11 and a data type analysis unit 12. In detail, the model and data type analysis module 1 determines the model type of the model file and the data type of the analysis data through the model file analysis unit 11 and the data type analysis unit 12, respectively. That is, after the user inputs the model file and the analysis data, the model analysis unit 11 automatically determines the model type (or model format) and internal parameters of the model file, and the data analysis unit 12 automatically determines the data type (or data format), and then outputs the analyzed model type and data type to the model interpretation method application module 2, so that the model interpretation method application module 2 can find a suitable interpretation method. Therefore, the model and data type analysis module 1 of the present invention can automatically determine the data type and model type, thereby achieving the effect of automatically reading various types of models and data.

In another embodiment, as shown in FIG. 2 , the intelligent multi-explanation universal model system of the present invention further includes a user interface module 4 having at least one user interface 41, so that the user can input the model file and analysis data to be analyzed through the user interface 41. When the user interface module 4 receives the model file and analysis data input by the user, the model file and analysis data are provided to the model and data type analysis module 1 for judging the model type and data type.

Please refer to Figure 1 again. The model interpretation method application module 2 is connected to the model and data type analysis module 1 by signal, and is used to receive the model type of the model file to be analyzed and the data type of the analysis data, and then query according to the model type and the data type. After the query, if the result only finds a single interpretation method, the model type and the data type are analyzed using the single interpretation method found; in addition, if the query result finds multiple interpretation methods corresponding to the model type and data type, the model type and the data type are analyzed using multiple interpretation methods to obtain multiple interpretation results.

In a specific embodiment, as shown in FIG. 3 , it is an internal structure diagram of the model interpretation method application module of the present invention. The model interpretation method application module 2 of the present invention includes an interpretable method database 21, a query method unit 22, a resource controller 23 and at least one execution method unit 24, wherein the interpretable method database 21 of the model interpretation method application module 2 stores a plurality of interpretation methods that can be used to interpret and analyze model types and data types. When the query method unit 22 searches the interpretable method database 21, it searches for the corresponding interpretation method according to the model type and data type; in other embodiments, the intelligent multi-interpretation universal model system of the present invention can set up a database for storing model interpretation methods outside the model interpretation method application module 2 according to needs, that is, the database is independent of the system, so that the model interpretation method application module 2 can query the required interpretation method from the database.

Furthermore, a model usually has an explanation method corresponding to the model type and data type of the model. However, as the application scenarios of the model are different, a single model may have multiple explanation methods that can be used to analyze and explain the model. Therefore, the query method unit 22 can query the explainable method database 21 according to the model type and data type to obtain at least one explanation method corresponding to the model type and the data type, and before performing model and data analysis based on the at least one explanation method obtained, the resource controller 23 is used to confirm how many computing resources are available in the system for analysis, so as to determine the number of method units 24 to be started. The resource controller 23 uses a Docker container (Docker container) to simultaneously activate multiple execution method units 24, and the activated execution method units 24 parse the model file and analysis data of one or more interpretation methods to obtain at least one interpretation result, that is, when there are multiple interpretation methods, multiple interpretation results will be obtained. In a specific embodiment, the model interpretation method includes SHAP (SHapley Additive exPlanations) and Local Interpretable Model-Agnostic Explanations (LIME) that can explain the influence between model features provided by the tree-based model.

Specifically, when there is only a single interpretation method, the resource controller 23 activates an execution method unit 24 to analyze the model file and the analysis data according to the single interpretation method, thereby obtaining a single interpretation result corresponding to the single interpretation method; if multiple interpretation methods are found, the resource controller 23 will first confirm the computing resource status of the execution method unit 24 that can be activated in the system to parse the model file and analyze the data. When the computing resources of the system are insufficient, only one execution method unit 24 can be activated to parse the model file and analyze the data in sequence through each interpretation method, thereby obtaining a single interpretation result corresponding to the single interpretation method. In order to obtain multiple interpretation results; in addition, when the system has sufficient computing resources, one or more execution method units 24 are activated according to the number of interpretation methods. That is, if three interpretation methods are found, only one or three execution method units 24 can be activated according to the needs. When multiple execution method units 24 are activated, each execution method unit 24 can be used to correspond to each interpretation method at the same time to analyze the model file and analysis data respectively, that is, one execution method unit analyzes the model file and analysis data in the manner of one interpretation method, thereby obtaining multiple interpretation results.

In other words, the model interpretation method application module 2 of the present invention uses the query method unit 22 to query the interpretation method that can be used to analyze the model from the interpretable method database 21 according to the model type and data type analyzed by the model and data type analysis module 1, and automatically adopts the corresponding interpretation method to perform analysis according to the model algorithm, and then confirms the number of computing resources of the current system that can activate the execution method unit 24 through the resource controller 23, so as to output the model interpretation result obtained after the execution of the execution method unit 24 to the interpretation effect integration module 3. Therefore, the model interpretation method application module 2 can automatically select the interpretation method used to analyze the model and perform model analysis based on it. Therefore, the user does not need to spend too much effort to understand various model interpretation methods, and can achieve the effect of automatically applying the most suitable model interpretation method.

In a specific embodiment, each of the interpretation results includes corresponding multiple features and multiple feature weights. For example, taking the application scenario of the model predicting the power generation of solar energy as an example, the features are related factors such as temperature, humidity or time used to judge the power generation of solar energy.

Please refer to Figure 1 again. The explanation effect integration module 3 is connected to the model explanation method application module 2 by signal, and is used to receive at least one explanation result analyzed by the model explanation method application module 2. After integrating the at least one explanation result, the model explanation result can be obtained. Specifically, the explanation result generated by the model explanation method application module 2 can be a model explanation vector, that is, the explanation effect integration module 3 receives one or more model explanation vectors to integrate them into a final vector as the model explanation result.

In one embodiment, the explanation effect integration module 3 adopts averaging and sorting, weighted averaging and sorting, or voting to integrate the at least one explanation result, wherein averaging and sorting refers to averaging the feature weights of the same features in each explanation result and arranging them in order according to the size of the average value; weighted averaging and sorting refers to adding a weighted value to each feature weight in the averaging and sorting, and then calculating the weighted average value for sorting; in addition, the voting method counts the number of times the feature weight value of each feature appears in the highest one in each explanation method. Taking the above model to predict solar power generation as an example, assuming that under the first explanation method, when the feature is time, the feature weight has the highest (or most important) value twice, when the feature is temperature, the feature weight has the highest value once, and humidity has 0 times, then the ranking results by voting are time, temperature and humidity.

In a specific embodiment, as shown in FIG4 , it is an internal structure diagram of the explanation effect integration module of the present invention. The explanation effect integration module 3 of the present invention includes a weight summing unit 31, a register 32, and a weight sorting unit 33. The explanation effect integration module 3 uses the weight summing unit 31 to average the feature weights of the same feature to obtain an average value, and stores the obtained average value in the register. The weight sorting unit 33 then arranges each feature according to the average value corresponding to each feature stored in the register 32 to generate a feature sorting result.

In another embodiment, as shown in FIG. 4 , the intelligent multi-explanatory universal model system of the present invention further includes a visualization module 5 for visualizing the model explanation results. Specifically, the visualization can present the model explanation results by, for example, a bar chart, a pie chart or other methods that can be understood by the user, so that the user can view the display effect presented by the visualization module 5 to understand the model explanation results.

Accordingly, the interpretation effect integration module 3 of the present invention receives one or more sets of feature weights through the weight summation unit 31, and calculates the average of the feature weights of the same feature and temporarily stores them in the register 32, so that the weight sorting unit 33 receives the weights of each feature and sorts them by size. Finally, the visualization module 5 presents the most suitable display screen according to the feature weight size and data type. In other words, the interpretation effect integration module 3 can automatically integrate multiple interpretation methods to obtain model interpretation results, and the visualization module 5 presents the best chart according to the data type, thereby achieving the effect of automatically presenting various types of model results.

In summary, the intelligent multivariate interpretation general model system of the present invention allows the user to input the model file and analysis data to the model and data type analysis module 1 through the user interface 41, so as to determine the model type and data type of the model file and the analysis data, and output the obtained results (i.e., model type and data type) to the model interpretation method application module 2, so that the model interpretation method application module 2 searches and finds the model type and data type from the method library according to the model type and data type. After the corresponding interpretation method is obtained, the corresponding interpretation method is executed to obtain one or more interpretation results. Then, the one or more interpretation results are output to the interpretation effect integration module 3, so that the interpretation effect integration module 3 can integrate the one or more interpretation results to analyze and obtain a unified model interpretation result. The received model interpretation result is automatically output as a visualization display result through the visualization module 5 and presented on the display screen for the user to view the model interpretation result.

Figure 5 is a step diagram of the method of the intelligent multi-dimensional interpretation general model of the present invention. As shown in the figure, the method of the intelligent multi-dimensional interpretation general model of the present invention includes the following steps.

In step S510, the model type and data type are determined. In the present invention, by setting up a model and data type analysis module, the received model file and analysis data are analyzed to obtain the model type and data type respectively.

In one embodiment, the model and data type analysis module includes a model file analysis unit for determining the model type of the model file and a data type analysis unit for determining the data type of the analysis data. Accordingly, the model and data type analysis module determines the model type of the model file through the model file analysis unit and determines the data type of the analysis data through the data type analysis unit.

In step S520, the corresponding explanation method is found. The present invention can query based on the model type and data type to find one or more corresponding explanation methods by setting up a model explanation method application module.

In step S530, one or more interpretation methods are used for analysis to obtain corresponding interpretation results. In this step, after searching for at least one interpretation method corresponding to the model type and data type, the present invention enables the model interpretation method application module to use these interpretation methods to parse the model file and analyze the data to obtain at least one interpretation result.

In one embodiment, each interpretation result obtained through analysis includes multiple features and multiple feature weights.

In one embodiment, the model interpretation method application module includes an interpretable method database, a query method unit, a resource controller, and at least one execution method unit, wherein the interpretable method database stores a plurality of interpretation methods for the query method unit to query according to the model type and the data type to obtain one or more corresponding interpretation methods. In other embodiments, the present invention can also set the interpretable method database outside the model interpretation method application module according to the needs, that is, independently set rather than set in the model interpretation method application module. In addition, the model interpretation method application module confirms the computing resources through the resource controller to determine the number of execution method units to start. That is, according to the number of interpretation methods, one or more execution method units can use each interpretation method to analyze the model file and the analysis data to obtain one or more corresponding interpretation results.

In step S540, one or more interpretation results are integrated to obtain a model interpretation result. The present invention sets an interpretation effect integration module to integrate one or more interpretation results to obtain a model interpretation result. In one embodiment, the interpretation effect integration module uses averaging and sorting, weighted averaging and sorting, or voting to integrate multiple interpretation results.

In one embodiment, the explanation effect integration module includes a weight summing unit, a register, and a weight sorting unit. Specifically, the weight summing unit averages the feature weights of the same feature to obtain an average value, and stores the average value in the register. Then, the weight sorting unit reads the average value in the register to arrange multiple features according to the average value.

FIG6 is a step diagram of another embodiment of the method of the intelligent multi-interpretation universal model of the present invention. As shown in the figure, steps S610 to S640 of this embodiment are the same as steps S510 to S540 of the first embodiment, and will not be repeated. The difference is that this embodiment further includes step S600 of providing a user interface and step S650 of providing a visual display, which are described in detail below.

In step S600, a user interface is provided. That is, the present invention further includes setting up a user interface module and providing the user interface to the user, so that the user can input the model file and analysis data of the model to be explained through the user interface.

In step S650, a visual display is provided. The present invention further includes setting a visual module to provide a visual display based on the model interpretation result. Accordingly, the model is analyzed by the method of the intelligent multi-interpretation general model of the present invention, and after the analysis is completed, an easy-to-understand visual display is provided to the user for viewing, so that the user can easily understand the model interpretation result.

FIG7 is a flowchart of the specific implementation of the method of the intelligent multivariate interpretation general model of the present invention. As shown in the figure, in process 710, the model file and analysis data are input. This process is to provide a user with a user interface such as a web page, allowing the user to input the model file to be interpreted and the analysis data used to train the model, and obtain the model type and data type accordingly.

In process 720, the matching model can be interpreted. In short, this process determines the interpretation method of the matching model file based on the model type and data type. There are different processing methods according to the number of interpretation methods found. If there is only one interpretation method, then enter process 730. If there are multiple interpretation methods, then enter process 740.

In process 730, an execution method unit is used to execute the interpretable method to output the interpretation result. If there is only one interpretable method, the interpretable method is directly executed to output the corresponding interpretation result, that is, the interpretable method is used to parse the model file and analyze the data to output the important feature weights of the model.

In process 740, multiple execution method units are used to execute the interpretable methods to output multiple interpretation results. If there are multiple interpretable methods, the interpretation methods can be analyzed by different execution method units respectively, thereby obtaining multiple interpretation results.

Before entering process 750, the interpretation results will be integrated first. Since process 730 is a single interpretation result, no integration is required. In addition, if it is process 740, multiple sets of model feature weights must be integrated through the interpretation effect integration module to obtain the final model feature weights as the model interpretation result.

In process 750, the visualization module generates results that meet the target type. This process is to present the feature weights and data types of the model on the user interface through the visualization module for the user to refer to.

In addition, the present invention also discloses a computer-readable medium, which is applied to a computing device or computer having a processor (e.g., CPU, GPU, etc.) and/or a memory, and stores instructions, and the computing device or computer can execute the computer-readable medium through the processor and/or memory to execute the above-mentioned method and each step when executing the computer-readable medium.

The modules, units, devices, etc. of the present invention include a microprocessor and a memory, and algorithms, data, programs, etc. are stored in the memory or chip. The microprocessor can load data or algorithms or programs from the memory to perform data analysis or calculation, etc., which will not be elaborated here. In other words, the intelligent multi-interpretation general model system and method of the present invention can be executed on electronic devices, such as general computers, tablets or servers, and perform data analysis and calculations after receiving data. Therefore, the program performed by the intelligent multi-interpretation general model system and method can be designed and constructed on electronic devices with components such as processors and memories through software to run on various types of electronic devices; in addition, each module or unit of the intelligent multi-interpretation general model system and method can be composed of independent components, such as calculators, memories, storages or firmware with processing units, which can all become components for realizing the present invention, and can also be presented in the form of software programs, hardware or firmware architectures.

The following uses the system, method and computer-readable medium of the intelligent multi-dimensional interpretation general model of the present invention to predict the power generation of solar panels, and takes the application theme of providing information on solar panel maintenance as an example to illustrate the practical application of the present invention, as detailed below.

The power generation efficiency of solar panels is related to factors such as the angle of sunlight and whether the solar panels are dirty or shaded. In the past, the power generation efficiency of solar panels was checked manually, so the labor cost was very heavy. In recent years, due to the rapid development of artificial intelligence (AI), the power generation efficiency can be checked by collecting a large amount of sensor data generated by sensing devices to train a model for predicting power generation. Based on this, it is only necessary to check the difference between actual power generation and expected power generation every day to help determine whether the solar panels need maintenance, thereby effectively reducing the burden on personnel. However, although the power generation prediction model can provide information about which solar panel has an abnormality, maintenance personnel still cannot immediately determine the cause of the abnormality, and there is a certain false alarm rate. If the model can provide an explanation of the model while providing decisions on the maintenance of solar panels, it will be able to save maintenance personnel's time and energy. Based on this, combined with the system and method of the present invention, the following specifically describes the process steps of solar power generation prediction.

Assume that the solar panels and related sensor data owned by Company A are shown in Table 1, where the data includes data corresponding to time, temperature, humidity, and power generation. Through each data, time, temperature, and humidity are used as features, and power generation is used as a label to train a power generation prediction model using a decision tree algorithm. Company A can currently accurately predict the power generation of the day through AI, and Company A hopes to further understand what factors the model uses as the basis for judging power generation.

The power generation prediction model in this embodiment is trained using the eXtreme Gradient Boosting (xgboost) algorithm, and its model file is stored in Mlflow format. Mlflow is an open source machine learning platform that can manage and build the required models to solve the problem of the lack of a standard model format. Decision tree-based algorithms such as xgboost can output the weights of each feature during model training. As shown in Table 2, the importance of the features of the model calculated by the algorithm is ranked as temperature>time >humidity. The above description is only one of the machine learning algorithms and model file formats. Those who are familiar with machine learning can use other model algorithms to represent.

Through the intelligent multi-dimensional interpretation general model system and method of the present invention, users are provided with the ability to input model files and analysis data through a user interface such as a web page, wherein the analysis data can be training data for training the model, and then the model and data type analysis module is used to analyze the input model file and analysis data to obtain the corresponding model type and data type. Specifically, the model file analysis unit determines that the model for power generation prediction is the xgboost model by analyzing the file extension of the input model file and the parameters in the file, and then the data type analysis unit determines that the data type is execution table data through the file extension and content of the training data.

Furthermore, in order to confirm which interpretability methods are suitable for the model and data type, the following uses numerical data as an example, but it is not limited to this. For example, the same approach can also be applied to other types of data such as pictures and text. Specifically, the results judged by xgboost, that is, xgboost and tabular data, are input into the model interpretation method application module. Through the various interpretation methods and corresponding data type tables in the pre-established interpretable method database, the model interpretation method application module can query the interpretability method of the model that can match the xgboost algorithm from the interpretable method database. Among them, the interpretability method can include SHAP (SHapley Additive exPlanations) that can explain the influence between model features and the feature importance of the algorithm itself. Based on this, it is determined that the number of methods required to be executed is 2. After that, the model interpretation method application module can automatically check the resource usage of the current computing environment, and use docker or other equivalent parallel computing methods to accelerate the execution efficiency, thereby obtaining the results in Tables 2 and 3.

，可計算出時間之特徵於統整後之權重為(0.35+0.4)/2=0.375，溫度之特徵於統整後之權重為(0.6+0.1)/2=0.35，以及濕度之特徵於統整後之權重為(0.05-0.05)/2=0。將前述之計算結果自大至小排序後得到統整後的模型解釋為時間>溫度>濕度。 From Table 2 and Table 3, it can be found that the feature importance (i.e., feature weight) obtained by the two interpretation methods may be different. In Table 2, it is temperature>time>humidity, and in Table 3, it is time>temperature>humidity. In order to avoid confusion for users and improve the reliability of interpretation, the present invention designs a method for integrating the results of model interpretation, wherein the method for integrating the results of model interpretation is not limited to average or other similar weighting methods. This embodiment takes average as an example. The current number of methods is 2, according to the formula: Important _i =

, we can calculate that the weight of the time characteristic after integration is (0.35+0.4)/2=0.375, the weight of the temperature characteristic after integration is (0.6+0.1)/2=0.35, and the weight of the humidity characteristic after integration is (0.05-0.05)/2=0. After sorting the above calculation results from large to small, the integrated model is interpreted as time>temperature>humidity.

FIG8 is a bar graph of the model interpretation result of the present invention in a visual presentation. As shown in the figure, a bar graph is presented in which the feature weight of time is 0.375, the feature weight of temperature is 0.35, and the feature weight of humidity is 0 in the aforementioned embodiment. Since the data type of this embodiment is numerical, for the numerical data type, the visualization module chooses to present the integrated result in the form of a bar graph, so that the user can see at a glance which feature is the most important. In this embodiment, the visualization module can use a database or other regular methods to define the visualization method corresponding to various data types, and is not limited to the bar graph of this embodiment.

In addition, FIG. 9 is a visualization diagram of the text emotion of the present invention. As shown in the figure, the present invention uses a model for analyzing the text emotion in a string through visualization as an example. Specifically, the model analyzes the text emotion of each word in the string "What is on the table? When I got home I saw my cat on the table, and my frog on the floor." Among them, if the text position falls on the left side of the figure, the text emotion is getting more angry, and conversely, if the text position falls on the right side of the figure, it means that the text emotion is getting less angry.

In summary, the system, method and computer-readable medium of the intelligent multi-dimensional interpretation universal model of the present invention allow users to input model files and analysis data through a user interface, so that the model and data type analysis module determines the corresponding model type and data type based on the model file and the analysis data. When the model interpretation method application module receives the model type and data type, it first searches for one or more corresponding interpretation methods, and then analyzes the one or more interpretation methods found to obtain one or more interpretation results. After being integrated by the interpretation effect integration module, the final model interpretation result is generated. Finally, the visualization module provides the corresponding display result to the user, and the user can easily understand the model interpretation result. As can be seen from the above, the present invention can analyze data without being limited by the input format, and can integrate multiple results to find the final model explanation result. Therefore, the present invention is not affected by the application scenario, and can universally analyze and explain the model so that users can understand the model explanation result of the model used. In contrast, in traditional methods, if image recognition is desired, a model interpretability presentation method suitable for the image must be adopted, and for numerical classification models, another method must be presented, causing difficulty for users in understanding. Secondly, the present invention can intelligently adopt corresponding explanation methods to analyze the model according to different model algorithms, and the present invention can integrate the results of multiple analysis methods to recommend the most suitable analysis result.

Therefore, the present invention can reduce the analysis difficulties caused by insufficient background knowledge of users. It is different from the previous technology that requires writing programs when analyzing different types of models and can only be used for a single application scenario. That is, the present invention provides users with a user interface that only needs to click and enter model information, and then according to the model type and application theme, a suitable interpretation method can be selected in an intelligent way to achieve the effect of improving convenience and saving time. Therefore, when users have no idea about model decisions, the present invention can help users explain model behavior. In other words, the present invention proposes a brand-new method to automatically analyze and explain models and use appropriate visualization to present results, so as to reduce the difficulty of analyzing model decisions and reduce the time users spend on exploration and trial.

In addition, the present invention can automatically apply the most suitable model interpretation method to reduce the difficulty of interpreting the results. Unlike the previous technology that only uses one interpretable method, the present invention can automatically apply the most suitable model interpretation method and integrate multiple methods to obtain interpreted results, and recommend more reliable interpretability for users to refer to.

In addition, the present invention also presents the model interpretation results using corresponding visualization effects according to different data types. Compared with the learning method, the present invention can also achieve the purpose of making it easier for users to understand the model's decision-making method, and automatically present various types of model interpretation results.

The above embodiments are only illustrative and not intended to limit the present invention. Anyone familiar with this technology may modify and change the above embodiments without violating the spirit and scope of the present invention. Therefore, the scope of protection of the present invention is defined by the scope of the patent application attached to the present invention. As long as it does not affect the effect and implementation purpose of the present invention, it should be covered by this public technical content.

1: Model and data type analysis module

2: Model interpretation method application module

3: Explain the effect integration module

Claims (17)

A system of intelligent multi-dimensional interpretation of universal models includes: a model and data type analysis module, which is used to analyze the model file to be interpreted and the analysis data used to train the model, so as to obtain the model type and data type corresponding to the model file and the analysis data respectively; a model interpretation method application module, which is connected to the model and data type analysis module, and is used to query according to the model type and the data type to find at least one corresponding interpretation method, and then parse the model file and the analysis data through the at least one interpretation method to obtain at least one interpretation result; and an interpretation effect integration module, which is connected to the model interpretation method application module, and is used to integrate the at least one interpretation result to obtain the model interpretation result. A system of intelligent multivariate interpretation universal model as described in claim 1, wherein the model and data type analysis module comprises: a model file analysis unit for determining the model type of the model file; and a data type analysis unit for determining the data type of the analysis data. The system of intelligent multi-interpretation general model as described in claim 1, wherein the model interpretation method application module comprises: an interpretable method database for storing multiple interpretation methods; a query method unit for querying the interpretable method database according to the model type and the data type to obtain the at least one interpretation method; a resource controller for confirming the computational resources for executing model analysis; and multiple execution method units for analyzing the model file and the analysis data according to each interpretation method to obtain each interpretation result. A system of intelligent multi-explanation general model as described in claim 1, wherein each of the explanation results includes multiple features and multiple feature weights. A system of intelligent multivariate explanation general model as described in claim 4, wherein the explanation effect integration module further comprises: a weight summing unit, which is used to average the feature weights of features with the same features to obtain an average value; a register, which is used to store the average value; and a weight sorting unit, which is used to arrange the multiple features according to the average value. The system of intelligent multivariate interpretation general model as described in claim 1 further includes a user interface module that connects the model with the data type analysis module and provides a user interface for the user to input the model file and the analysis data. The system of the intelligent multi-explanatory general model as described in claim 1 further includes a visualization module connected to the explanation effect integration module and used to visualize the explanation results of the model. A system of intelligent multivariate explanation general model as described in claim 1, wherein the explanation effect integration module adopts averaging and sorting, weighted averaging and sorting, or voting to integrate the at least one explanation result. A method for intelligent multivariate interpretation of a general model is implemented by a computer device, and includes the following steps: a model and data type analysis module is used to analyze a model file to be interpreted and analysis data used for training the model to obtain a model type and a data type corresponding to the model file and the analysis data, respectively; a model interpretation method application module is used to query according to the model type and the data type to find at least one corresponding interpretation method; the model interpretation method application module is used to analyze the model file and the analysis data through the at least one interpretation method to obtain at least one interpretation result; and an interpretation effect integration module is used to integrate the at least one interpretation result to obtain a model interpretation result. The method of intelligent multivariate interpretation of universal model as described in claim 9, wherein the step of allowing the model and data type analysis module to analyze the model file to be analyzed and the analysis data used for training the model includes: determining the model type of the model file through the model file analysis unit; and determining the data type of the analysis data through the data type analysis unit. The method of intelligent multi-explaining universal model as described in claim 9, wherein the step of making the model explanation method application module find at least one corresponding explanation method and obtain the at least one explanation result includes: the query method unit searches from the explainable method database storing multiple explanation methods according to the model type and the data type to obtain the at least one explanation method; the resource controller confirms the computational resources for executing model analysis; and the multiple execution method units analyze the model file and the analysis data according to each of the explanation methods to obtain each of the explanation results. A method for intelligent multivariate interpretation of a general model as described in claim 9, wherein each of the interpretation results comprises multiple features and multiple feature weights. The method of the intelligent multivariate interpretation general model as described in claim 12, wherein the step of allowing the interpretation effect integration module to integrate the at least one interpretation result includes: allowing the weight summation unit to average the feature weights of the same feature to obtain an average value; storing the average value in a register; and allowing the weight sorting unit to arrange the multiple features according to the average value. The method of intelligent multivariate interpretation general model as described in claim 9, wherein the model file and the analysis data are input by the user through the user interface provided by the user interface module. A method for intelligent multivariate interpretation of a general model as described in claim 9, wherein the model interpretation result is visualized through a visualization module. The method of intelligent multivariate explanation general model as described in claim 9, wherein the explanation effect integration module adopts averaging and sorting, weighted averaging and sorting, or voting to integrate the at least one explanation result. A computer-readable medium, used in a computing device or a computer, stores instructions for executing a method of intelligent multi-interpretation general model as described in any one of claims 9 to 16.

Images