Method for reducing algorithm disgust degree based on user modification model

Abstract

The invention discloses a method for reducing algorithm disgust degree based on model modification by a user. The method comprises the following steps: providing the capability of deeply participating in algorithm design and customization for the user; the transparency and the interpretability of the algorithm are improved, so that a user can better understand the decision logic and the working principle of the algorithm; the user makes a decision by using an algorithm and modifies an algorithm model; collecting, sorting and analyzing feedback suggestions of users on algorithms and algorithm decision results, and establishing an effective user feedback mechanism; model parameter adjustment, feature selection optimization and training data updating are performed based on user data, modification traces and feedback, so that the performance and accuracy of the algorithm are improved.

Description

Technical field:

[0001] The present invention relates to a method for reducing algorithmic aversion based on user modification of a model. Background technology:

[0002] Artificial intelligence has huge advantages in data processing and analysis, and can provide decision makers with more comprehensive and accurate information support. It has become an important tool to assist human decision-making and will have a significant impact on the way humans make decisions.

[0003] Under the premise that algorithmic decisions are more accurate, users still tend to rely on their own decisions. Once they find that the algorithmic decision is wrong, they will quickly develop negative emotions or prejudices towards the algorithm, which will affect people's overall evaluation of and willingness to use the algorithmic decision. This phenomenon is called "algorithm aversion."

[0004] The phenomenon of algorithm aversion shows that in the era of artificial intelligence, there are still obstacles to effectively combining algorithmic decision-making with human decision-making to form good human-machine collaboration, which is not conducive to the promotion and application of artificial intelligence technology. Generally speaking, factors such as the transparency, degree of anthropomorphism, degree of autonomy, compatibility and complexity of the algorithm will affect people's tendency to averse to algorithms. Compared with mature and perfect algorithms, users are more inclined to choose modifiable algorithm models to make decisions.

[0005] Therefore, improving the design and implementation of algorithms can effectively reduce algorithm aversion and promote decision makers to use the comprehensive and accurate support provided by artificial intelligence technology. Summary of the invention:

[0006] The embodiment of the present invention provides a method based on a user modification model to reduce the degree of algorithm aversion. The method is reasonably designed, based on the mutual cooperation of multiple functional modules and user modification modeling, and operates and starts from the perspective of user participation. According to the user's control over the algorithm, the user modification right is added, which improves the transparency, anthropomorphism and user participation of the algorithm, thereby effectively reducing the user's aversion to the algorithm, making it more in line with human expectations and needs, promoting users to use the comprehensiveness and accuracy support provided by artificial intelligence technology, and solving the problems existing in the prior art.

[0007] The technical solution adopted by the present invention to solve the above technical problems is:

[0008] A method for reducing algorithmic aversion based on user modification of a model, the method comprising the following steps:

[0009] S1, providing users with the ability to deeply participate in algorithm design and customization;

[0010] S2, improve the transparency and explainability of the algorithm so that users can better understand the decision-making logic and working principles of the algorithm;

[0011] S3, users use the algorithm to make decisions and modify the algorithm model;

[0012] S4, collect, organize and analyze user feedback on algorithms and algorithm decision results, and establish an effective user feedback mechanism;

[0013] S5 adjusts model parameters, optimizes feature selection, and updates training data based on user data, modification traces, and feedback, thereby improving the performance and accuracy of the algorithm.

[0014] Providing users with the ability to deeply participate in algorithm design and customization includes the following steps:

[0015] S1.1, users select an algorithm mode that suits their own cultural background and values ​​from a variety of cultural modes;

[0016] S1.2, upload the user's own personal experience data collection to optimize and adjust the selected algorithm model;

[0017] S1.3, the degree of anthropomorphism of the algorithm selected by the user makes the expectations of the algorithm selected by the user meet the requirements and reduces the user's aversion to the algorithm.

[0018] Improving the transparency and explainability of algorithms so that users can better understand the decision logic and working principles of the algorithms includes the following steps:

[0019] S2.1, provide algorithm model explanation, including input dimension explanation, output dimension explanation, input-output mapping relationship explanation and algorithm process explanation; the input dimension explanation includes the data type, data format and preprocessing rules received by the algorithm model; the output dimension explanation includes the result type, confidence assessment method and error range description generated by the algorithm model; the input-output mapping relationship explanation includes logic and function description; the algorithm process explanation includes the mathematical expression of the data processing flow, the mechanism of action of key parameters and the visual presentation of the decision path;

[0020] S2.2, provide training data explanation, including data structure explanation, data source explanation, data cleaning explanation and feature engineering explanation; the data structure explanation includes explaining the data structure used by the algorithm model to organize and operate data; the data source explanation includes databases, log files, user surveys or third-party data providers; the data cleaning explanation is used to correct errors in the data, remove duplicate and invalid data, and supplement missing values; the feature engineering explanation is used to convert raw data into features that can be used by the model and perform feature selection, feature extraction and feature conversion;

[0021] S2.3, provide explanations for the decision-making process, including reliability explanations, efficiency explanations and analytical explanations; the reliability explanation is used to illustrate the probability of the correct operation of the algorithm model, and to predict the robustness of the algorithm model in the face of noisy data, outliers or changes in data distribution; the efficiency explanation is used to illustrate the time parameters and space complexity required for the algorithm model to calculate the task under given conditions, that is, the amount of time required for the algorithm model to process the input data and the amount of additional storage space required during the operation; the analytical explanation is used to illustrate the performance evaluation results of the algorithm model on the training set and test set, and the generalization ability of the algorithm model.

[0022] The following steps are involved in users using algorithms to make decisions and modifying the algorithm model:

[0023] S3.1, dynamically allocates weights, rebalances feature weights through an interactive weight adjustment matrix, updates decision functions in real time using linear interpolation, provides a visual slider to adjust weight distribution, and adjusts weights to a precision of ±0.01;

[0024] S3.2, modify filtering conditions, build a dynamic filtering rule library, and support user-defined operations;

[0025] S3.3, modify the core parameters of the algorithm and automatically optimize the model according to the new parameter values ​​set by the user; the core parameters of the algorithm include the number of recommended neighbors for collaborative filtering, regularization strength and learning rate.

[0026] The user feedback mechanism includes correction operations on user recommendation results and a dynamic training set constructed based on user modification logs to achieve incremental data fusion, optimize algorithm design, and improve algorithm performance.

[0027] Adjusting model parameters, optimizing feature selection, and updating training data based on user data, modification traces, and feedback includes the following steps:

[0028] S5.1, establish the exponential decay mechanism of regularization parameters, design the periodic learning rate adjustment strategy and regulate the momentum coefficient in stages to achieve dynamic optimization of parameters;

[0029] S5.2, integrating statistical significance and business importance, and combining t-SNE visual dimensionality reduction to screen intelligent features;

[0030] S5.3, inject new data into the algorithm model, build a one-time verification mechanism for the new and old models, set an exponentially decaying freshness indicator, coordinately update the triple technologies of parameter adjustment, feature optimization and data dynamics, build a user-oriented algorithm evolution system, and effectively alleviate the phenomenon of algorithm aversion.

[0031] The system corresponding to the method includes:

[0032] Algorithm development module, which is used to provide users with the ability to deeply participate in algorithm design and customization;

[0033] An algorithm explanation module, which is used to improve the transparency and explainability of the algorithm so that users can better understand the decision logic and working principle of the algorithm;

[0034] An algorithm decision module, which is used to enable users to make decisions using algorithms and modify algorithm models;

[0035] A user feedback module, which is used to collect, organize and analyze user feedback on the algorithm and algorithm decision results, and establish an effective user feedback mechanism;

[0036] The algorithm optimization module is used to adjust model parameters, optimize feature selection and update training data based on user data, modification traces and feedback, so as to improve the performance and accuracy of the algorithm.

[0037] The present invention adopts the above method and structure, and provides users with the ability to deeply participate in algorithm design and customization through the algorithm development module, allowing users to select an algorithm model that conforms to their cultural background and values ​​from a variety of cultural models, and allowing users to select the degree of anthropomorphism of the algorithm, so that users' expectations of the algorithm meet the requirements and reduce the user's aversion to the algorithm; the algorithm explanation module improves the transparency and interpretability of the algorithm so that users can better understand the decision logic and working principle of the algorithm; the algorithm decision module enables users to use the algorithm to make decisions and modify the algorithm model; the user feedback module collects, organizes and analyzes user feedback on the algorithm and algorithm decision results, and establishes an effective user feedback mechanism; the algorithm optimization module adjusts model parameters, optimizes feature selection and updates training data based on user data, modification traces and feedback, thereby improving the performance and accuracy of the algorithm, which has the advantages of precision, practicality, safety and reliability. Description of the drawings:

[0038] Figure 1 It is a working principle diagram of the present invention.

[0039] Figure 2 This is a diagram of the implementation steps of the algorithm development module of the present invention.

[0040] Figure 3 It is a diagram of the implementation steps of the algorithm explanation module of the present invention.

[0041] Figure 4 This is a diagram of the implementation steps of the algorithm decision module of the present invention.

[0042] Figure 5This is a diagram of the implementation steps of the user feedback module of the present invention.

[0043] Figure 6 This is a diagram of the implementation steps of the algorithm optimization module of the present invention. Specific implementation method:

[0044] In order to clearly illustrate the technical features of the present solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0045] like Figure 1-6 As shown in , a method for reducing algorithm aversion based on user modification model, the method comprises the following steps:

[0046] S1, providing users with the ability to deeply participate in algorithm design and customization;

[0047] S2, improve the transparency and explainability of the algorithm so that users can better understand the decision-making logic and working principles of the algorithm;

[0048] S3, users use the algorithm to make decisions and modify the algorithm model;

[0049] S4, collect, organize and analyze user feedback on algorithms and algorithm decision results, and establish an effective user feedback mechanism;

[0050] S5 adjusts model parameters, optimizes feature selection, and updates training data based on user data, modification traces, and feedback, thereby improving the performance and accuracy of the algorithm.

[0051] Providing users with the ability to deeply participate in algorithm design and customization includes the following steps:

[0052] S1.1, users select an algorithm mode that suits their own cultural background and values ​​from a variety of cultural modes;

[0053] S1.2, upload the user's own personal experience data collection to optimize and adjust the selected algorithm model;

[0054] S1.3, the degree of anthropomorphism of the algorithm selected by the user makes the expectations of the algorithm selected by the user meet the requirements and reduces the user's aversion to the algorithm.

[0055] Improving the transparency and explainability of algorithms so that users can better understand the decision logic and working principles of the algorithms includes the following steps:

[0056] S2.1, provide algorithm model explanation, including input dimension explanation, output dimension explanation, input-output mapping relationship explanation and algorithm process explanation; the input dimension explanation includes the data type, data format and preprocessing rules received by the algorithm model; the output dimension explanation includes the result type, confidence assessment method and error range description generated by the algorithm model; the input-output mapping relationship explanation includes logic and function description; the algorithm process explanation includes the mathematical expression of the data processing flow, the mechanism of action of key parameters and the visual presentation of the decision path;

[0057] S2.2, provide training data explanation, including data structure explanation, data source explanation, data cleaning explanation and feature engineering explanation; the data structure explanation includes explaining the data structure used by the algorithm model to organize and operate data; the data source explanation includes databases, log files, user surveys or third-party data providers; the data cleaning explanation is used to correct errors in the data, remove duplicate and invalid data, and supplement missing values; the feature engineering explanation is used to convert raw data into features that can be used by the model and perform feature selection, feature extraction and feature conversion;

[0058] S2.3, provide explanations for the decision-making process, including reliability explanations, efficiency explanations and analytical explanations; the reliability explanation is used to illustrate the probability of the correct operation of the algorithm model, and to predict the robustness of the algorithm model in the face of noisy data, outliers or changes in data distribution; the efficiency explanation is used to illustrate the time parameters and space complexity required for the algorithm model to calculate the task under given conditions, that is, the amount of time required for the algorithm model to process the input data and the amount of additional storage space required during the operation; the analytical explanation is used to illustrate the performance evaluation results of the algorithm model on the training set and test set, and the generalization ability of the algorithm model.

[0059] The following steps are involved in users using algorithms to make decisions and modifying the algorithm model:

[0060] S3.1, dynamically allocates weights, rebalances feature weights through an interactive weight adjustment matrix, updates decision functions in real time using linear interpolation, provides a visual slider to adjust weight distribution, and adjusts weights to a precision of ±0.01;

[0061] S3.2, modify filtering conditions, build a dynamic filtering rule library, and support user-defined operations;

[0062] S3.3, modify the core parameters of the algorithm and automatically optimize the model according to the new parameter values ​​set by the user; the core parameters of the algorithm include the number of recommended neighbors for collaborative filtering, regularization strength and learning rate.

[0063] The user feedback mechanism includes correction operations on user recommendation results and a dynamic training set constructed based on user modification logs to achieve incremental data fusion, optimize algorithm design, and improve algorithm performance.

[0064] Adjusting model parameters, optimizing feature selection, and updating training data based on user data, modification traces, and feedback includes the following steps:

[0065] S5.1, establish the exponential decay mechanism of regularization parameters, design the periodic learning rate adjustment strategy and regulate the momentum coefficient in stages to achieve dynamic optimization of parameters;

[0066] S5.2, integrating statistical significance and business importance, and combining t-SNE visual dimensionality reduction to screen intelligent features;

[0067] S5.3, inject new data into the algorithm model, build a one-time verification mechanism for the new and old models, set an exponentially decaying freshness indicator, coordinately update the triple technologies of parameter adjustment, feature optimization and data dynamics, build a user-oriented algorithm evolution system, and effectively alleviate the phenomenon of algorithm aversion.

[0068] The system corresponding to the method includes:

[0069] Algorithm development module, which is used to provide users with the ability to deeply participate in algorithm design and customization;

[0070] An algorithm explanation module, which is used to improve the transparency and explainability of the algorithm so that users can better understand the decision logic and working principle of the algorithm;

[0071] An algorithm decision module, which is used to enable users to make decisions using algorithms and modify algorithm models;

[0072] A user feedback module, which is used to collect, organize and analyze user feedback on the algorithm and algorithm decision results, and establish an effective user feedback mechanism;

[0073] The algorithm optimization module is used to adjust model parameters, optimize feature selection and update training data based on user data, modification traces and feedback, so as to improve the performance and accuracy of the algorithm.

[0074] The working principle of a method based on user modification model to reduce algorithm aversion in an embodiment of the present invention is: based on the mutual cooperation of multiple functional modules and user modification modeling, operation and starting from the perspective of user participation, adding user modification rights according to the user's control over the algorithm, improving the transparency, anthropomorphism and user participation of the algorithm, thereby effectively reducing the user's aversion to the algorithm, making it more in line with human expectations and needs, and promoting users to use the comprehensiveness and accuracy support provided by artificial intelligence technology.

[0075] In the overall solution, the corresponding method includes the following steps: providing users with the ability to deeply participate in algorithm design and customization; improving the transparency and explainability of the algorithm so that users can better understand the decision-making logic and working principles of the algorithm; users use the algorithm to make decisions and modify the algorithm model; collecting, organizing and analyzing user feedback on the algorithm and algorithm decision results, and establishing an effective user feedback mechanism; adjusting model parameters, optimizing feature selection and updating training data based on user data, modification traces and feedback, thereby improving the performance and accuracy of the algorithm.

[0076] The corresponding system includes an algorithm development module, which is used to provide users with the ability to deeply participate in algorithm design and customization; an algorithm interpretation module, which is used to improve the transparency and explainability of the algorithm so that users can better understand the decision logic and working principles of the algorithm; an algorithm decision module, which is used to enable users to use the algorithm to make decisions and modify the algorithm model; a user feedback module, which is used to collect, organize and analyze user feedback on the algorithm and algorithm decision results, and establish an effective user feedback mechanism; an algorithm optimization module, which is used to adjust model parameters, optimize feature selection and update training data based on user data, modification traces and feedback, thereby improving the performance and accuracy of the algorithm.

[0077] Furthermore, the algorithm development module allows users to select an algorithm model that suits their cultural background and values ​​from a variety of cultural models. Users can upload their own personal experience data sets, which will be used to optimize and adjust the algorithm, allowing users to select the degree of anthropomorphism of the algorithm, such as character image, language interaction style, etc., so that users' expectations of the algorithm meet the requirements and reduce users' aversion to the algorithm; the algorithm explanation module enables users to better understand the decision-making logic and working principles of the algorithm by providing the functions of algorithm model explanation, training data explanation, and decision-making process explanation, thereby reducing algorithm aversion.

[0078] For the algorithm decision-making module, enabling users to have a certain degree of control over the algorithm can significantly alleviate aversion; the key measure is to provide users with options to adjust weight distribution, modify filtering conditions, and modify algorithm parameters.

[0079] For the user feedback module, it can effectively record user feedback on algorithm design, algorithm modification feedback and output result feedback, and use it to optimize algorithm design and improve algorithm performance after sorting and analysis. It can help reduce algorithm aversion and improve user experience.

[0080] For the algorithm optimization module, model parameter adjustment, feature selection optimization, and training data update can help the algorithm more accurately reflect the real needs of users, improve algorithm performance and user acceptance, and thus reduce algorithm aversion.

[0081] Preferably, providing users with the ability to deeply participate in algorithm design and customization includes the following steps: users select an algorithm model that suits their own cultural background and values ​​from a variety of cultural models; upload the user's own personal experience data collection to optimize and adjust the selected algorithm model; the user selects the degree of anthropomorphism of the algorithm so that the user's expectations of the algorithm selected meet the requirements and reduce the user's aversion to the algorithm.

[0082] Preferably, improving the transparency and explainability of the algorithm so that users can better understand the decision logic and working principle of the algorithm includes the following steps: providing an explanation of the algorithm model, including an explanation of the input dimension, an explanation of the output dimension, an explanation of the input-output mapping relationship, and an explanation of the algorithm process; the input dimension explanation includes the data type, data format, and preprocessing rules received by the algorithm model; the output dimension explanation includes the result type, confidence assessment method, and error range description generated by the algorithm model; the input-output mapping relationship explanation includes a description of logic and function; the algorithm process explanation includes a formalized description of the mathematical expression of the data processing flow, an explanation of the mechanism of action of key parameters, and a visual presentation of the decision path; providing an explanation of the training data, including an explanation of the data structure, data source, data cleaning, and feature engineering; the data structure explanation includes an explanation of the data structure used by the algorithm model to organize and operate data; The data source explanation includes databases, log files, user surveys or third-party data providers; the data cleaning explanation is used to correct errors in the data, remove duplicate and invalid data, and supplement missing values; the feature engineering explanation is used to convert the original data into features that can be used by the model and perform feature selection, feature extraction and feature conversion; provide decision-making process explanations, including reliability explanations, efficiency explanations and analytical explanations; the reliability explanation is used to illustrate the probability of the correct operation of the algorithm model and predict the robustness of the algorithm model in the face of noisy data, outliers or changes in data distribution; the efficiency explanation is used for the time parameters and space complexity required for the algorithm model to calculate the task under given conditions, that is, the amount of time required for the algorithm model to process the input data and the amount of additional storage space required during the operation; the analytical explanation is used to illustrate the performance evaluation results of the algorithm model on the training set and test set, and the generalization ability of the algorithm model.

[0083] Preferably, the user uses the algorithm to make decisions and modify the algorithm model, including the following steps: dynamically allocate weights, achieve feature weight rebalancing through an interactive weight adjustment matrix, use linear interpolation to update the decision function in real time, provide a visual slider to adjust the weight distribution, and adjust the weight adjustment range to a granularity of ±0.01; modify the filtering conditions, build a dynamic filtering rule library, and support user-defined operations; modify the core parameters of the algorithm, and automatically optimize the model according to the new parameter values ​​set by the user; the core parameters of the algorithm include the number of recommended neighbors for collaborative filtering, regularization strength, and learning rate.

[0084] Specifically, the interactive weight adjustment matrix is:

[0085] W=[w 1 , w 2 , ..., w n ] T

[0086] in,

[0087]

[0088] The decision function is:

[0089]

[0090] Preferably, model parameter adjustment, feature selection optimization and training data update based on user data, modification traces and feedback include the following steps: establishing an exponential decay mechanism for regularization parameters, designing a periodic learning rate adjustment strategy and regulating the momentum coefficient in stages to achieve dynamic optimization of parameters; integrating statistical significance and business importance, and combining t-SNE visual dimensionality reduction to screen intelligent features; injecting new data into the algorithm model, building a one-time verification mechanism for new and old models, setting an exponential decay freshness indicator, and collaboratively updating the triple technologies of parameter adjustment, feature optimization and data dynamics to build a user-oriented algorithm evolution system and effectively alleviate algorithm aversion.

[0091] Specifically, taking the hospital's artificial intelligence consultation platform as an example, users have a high degree of aversion to algorithm recommendations, mainly because users do not trust the diagnosis and treatment plans provided by the algorithm and still want to obtain treatment plans provided by the doctor himself or obtain opinions from experts.

[0092] In order to improve the above problems, the platform adopts the method of this application to reduce the degree of algorithm aversion; the following is an explanation based on the characteristics of each functional module.

[0093] In the algorithm development module, users are allowed to adjust and optimize the consultation platform algorithm through cultural mode selection, personal experience upload and personification selection; in this embodiment, a "cultural mode selection" function can be added to the user interface of the artificial intelligence consultation platform, and users can choose a suitable consultation mode according to their cultural background and preferences; two types are designed in the treatment plan preference, Chinese medicine treatment and Western medicine treatment; for users who focus on traditional medical culture, the platform can provide consultation suggestions combined with Chinese medicine theory; for users who prefer Western medicine, suggestions based on modern medical knowledge are provided; two types are designed in the cultural background, individualism and collectivism; in collectivist countries that focus on interpersonal relationships, people tend to trust the personal experience and professional judgment of doctors, and believe that the emotional connection and trust relationship between doctors and patients are the key factors for successful treatment. More caring statements can be added to the algorithm decision results and actively guided to find doctors' advice; in contrast, in individualistic countries, people may pay more attention to the objectivity and efficiency of algorithms, and believe that these technologies can provide faster and more accurate medical advice. The algorithm can be displayed in the algorithm decision results through charts, data, etc. in disease diagnosis and treatment plans, so the corresponding suggestion content can be pushed according to the cultural background selected by the user in real time.

[0094] A "personal experience upload" function has been added to the user interface of the artificial intelligence consultation platform, allowing users to upload their own experience and knowledge to the system so that the algorithm can better learn and understand the user's decision-making habits, including the user's health status, medical history, treatment experience, treatment data, response to specific drugs or treatments, side effects experience, and feelings during the recovery process.

[0095] In the user interface of the artificial intelligence medical consultation platform, an "anthropomorphic selection" function is added; the portrait factor of the interactive interface also affects user trust. People tend to rely on trustworthy faces. Users can choose portraits based on gender, title, and familiarity. Gender is designed into two types, male and female; four types of titles are designed, divided into resident physicians (junior), attending physicians (intermediate), deputy chief physicians (deputy senior), chief physicians (senior), and the portrait of the algorithm decision result output page is selected based on the doctors the user is familiar with.

[0096] In the algorithm explanation module, the decision-making process of the algorithm model can be explained to users, including algorithm model explanation, training data explanation and decision-making process explanation; by providing detailed explanatory text and visualization tools, users can better understand how the algorithm makes decisions, thereby enhancing their trust in the algorithm.

[0097] Algorithm model explanation refers to clearly explaining the structure, principles and working methods of the algorithm model to users. This includes but is not limited to: explaining which model the algorithm belongs to, such as linear regression, decision tree, random forest, neural network, etc.; explaining the key parameters in the model and their functions, such as weights and bias terms in linear regression; explaining how the model generates output based on input data, such as how decision trees make decisions by splitting nodes, so that people can better understand how the algorithm makes decisions, thereby enhancing their trust in the algorithm. Training data explanation refers to showing users the training data set used by the algorithm and explaining how this data affects the training results of the model.

[0098] Training data explanation refers to showing the training data set used by the algorithm to the user and explaining how this data affects the training results of the model. In the user interface, it is explained whether the training data comes from a public data set or user-uploaded data, and the pre-processing operations performed on the data before training, such as data cleaning and feature selection, are explained.

[0099] Explanation of the decision process refers to showing the user in detail the steps and factors that the algorithm took into account when making a decision. This includes visualizing the input data received by the algorithm and its format, showing the flowchart of the algorithm's decision making, and explaining which features play a key role and their contribution.

[0100] In the algorithm decision module, the process of users using algorithms to make decisions allows the algorithm model to be modified, including adjusting weight distribution, modifying filtering conditions, and modifying algorithm parameters; specifically, weight distribution adjustment is defined as: each user has different levels of emphasis on each factor affecting the decision, allowing users to flexibly adjust the weights of different factors in the algorithm. Allow users to adjust the weights of different diagnostic factors in the algorithm based on their personal health status, medical history or preferences. For patients who are concerned about drug side effects, the weight of drug safety assessment can be increased; for users who focus on lifestyle, the weight of lifestyle adjustment recommendations can be increased.

[0101] Modifying filter conditions is defined as: providing flexible filter condition settings, refining filter conditions, and allowing users to exclude factors that do not match their situation or are not of interest to them. This includes but is not limited to filtering conditions such as age, gender, specific disease history, health status, lifestyle habits, allergy history, etc., to ensure that the recommendations provided by algorithmic decision-making are more in line with the actual needs of users.

[0102] Modifying algorithm parameters is defined as: For users with a certain technical background, open advanced settings of some algorithm parameters, such as learning rate, regularization strength, etc., allowing them to fine-tune according to their own understanding, in order to obtain more personalized diagnosis results. At the same time, the system should provide a safety range prompt to avoid improper parameter settings leading to a decline in model performance.

[0103] In the user feedback module, a "user feedback" option is provided to encourage users to make suggestions and comments on algorithm modifications in order to continuously improve and optimize the functions and performance of the platform; it can collect user feedback on algorithm design through various channels. These feedbacks may include positive comments, negative comments, improvement suggestions, etc.; it can organize and classify the relevant information of users modifying the algorithm model, and the sorting process includes removing duplicate modifications, classifying the modified content by topic, and extracting key information; it can identify the commonalities and differences of users' modifications to the algorithm model, and discover problems and deficiencies in the system or algorithm; it can collect user satisfaction feedback on the output results of the algorithm, and verify the accuracy of the user's decision to modify the output of the algorithm model.

[0104] In the algorithm optimization module, model parameter adjustment is a basic task in algorithm optimization. Fine-tuning the model parameters can significantly improve the accuracy and efficiency of the algorithm. Feature selection optimization is another important part of algorithm optimization. By selecting the most representative features based on user behavior and feedback, the accuracy and efficiency of the model can be significantly improved. Training data update is an important part of algorithm optimization that cannot be ignored. With the continuous generation of new data and the obsolescence of old data, timely updating of training data is crucial to maintaining algorithm performance. At the same time, it is necessary to update all training data in a timely manner in combination with user-uploaded data sets.

[0105] To sum up, a method based on user modification model to reduce algorithm aversion in an embodiment of the present invention is based on the mutual cooperation of multiple functional modules and user modification modeling, operates and starts from the perspective of user participation, and adds user modification rights based on the user's control over the algorithm, thereby improving the transparency, anthropomorphism and user participation of the algorithm, thereby effectively reducing the user's aversion to the algorithm, making it more in line with human expectations and needs, and promoting users to use the comprehensiveness and accuracy support provided by artificial intelligence technology.

[0106] The above specific implementation manner cannot be used as a limitation on the protection scope of the present invention. For those skilled in the art, any substitution, improvement or change made to the implementation manner of the present invention falls within the protection scope of the present invention.

[0107] The matters not described in detail in the present invention are all known technologies to those skilled in the art.

Claims

1. A method based on user modification of a model to reduce algorithmic aversion, characterized in that: The method comprises the following steps: S1, providing users with the ability to deeply participate in algorithm design and customization; S2, improve the transparency and explainability of the algorithm so that users can better understand the decision-making logic and working principles of the algorithm; S3, users use the algorithm to make decisions and modify the algorithm model; S4, collect, organize and analyze user feedback on algorithms and algorithm decision results, and establish an effective user feedback mechanism; S5 adjusts model parameters, optimizes feature selection, and updates training data based on user data, modification traces, and feedback, thereby improving the performance and accuracy of the algorithm.

2. A method for reducing algorithm aversion based on user modification of a model according to claim 1, characterized in that: Providing users with the ability to deeply participate in algorithm design and customization includes the following steps: S1.1, users select an algorithm mode that suits their own cultural background and values ​​from a variety of cultural modes; S1.2, upload the user's own personal experience data collection to optimize and adjust the selected algorithm model; S1.3, the degree of anthropomorphism of the algorithm selected by the user makes the expectations of the algorithm selected by the user meet the requirements and reduces the user's aversion to the algorithm.

3. A method for reducing algorithm aversion based on user modification of a model according to claim 1, characterized in that: Improving the transparency and explainability of algorithms so that users can better understand the decision logic and working principles of the algorithms includes the following steps: S2.1, provide algorithm model explanation, including input dimension explanation, output dimension explanation, input-output mapping relationship explanation and algorithm process explanation; the input dimension explanation includes the data type, data format and preprocessing rules received by the algorithm model; the output dimension explanation includes the result type, confidence assessment method and error range description generated by the algorithm model; the input-output mapping relationship explanation includes logic and function description; the algorithm process explanation includes the mathematical expression of the data processing flow, the mechanism of action of key parameters and the visual presentation of the decision path; S2.2, provide training data explanation, including data structure explanation, data source explanation, data cleaning explanation and feature engineering explanation; the data structure explanation includes explaining the data structure used by the algorithm model to organize and operate data; the data source explanation includes databases, log files, user surveys or third-party data providers; the data cleaning explanation is used to correct errors in the data, remove duplicate and invalid data, and supplement missing values; the feature engineering explanation is used to convert raw data into features that can be used by the model and perform feature selection, feature extraction and feature conversion; S2.3, provide explanations for the decision-making process, including reliability explanations, efficiency explanations and analytical explanations; the reliability explanation is used to illustrate the probability of the correct operation of the algorithm model, and to predict the robustness of the algorithm model in the face of noisy data, outliers or changes in data distribution; the efficiency explanation is used to illustrate the time parameters and space complexity required for the algorithm model to calculate the task under given conditions, that is, the amount of time required for the algorithm model to process the input data and the amount of additional storage space required during the operation; the analytical explanation is used to illustrate the performance evaluation results of the algorithm model on the training set and test set, and the generalization ability of the algorithm model.

4. A method for reducing algorithmic aversion based on user modification of a model according to claim 1, characterized in that: The following steps are involved in users using algorithms to make decisions and modifying the algorithm model: S3.1, dynamically allocates weights, rebalances feature weights through an interactive weight adjustment matrix, updates decision functions in real time using linear interpolation, provides a visual slider to adjust weight distribution, and adjusts weights to a precision of ±0.01; S3.2, modify filtering conditions, build a dynamic filtering rule library, and support user-defined operations; S3.3, modify the core parameters of the algorithm and automatically optimize the model according to the new parameter values ​​set by the user; the core parameters of the algorithm include the number of recommended neighbors for collaborative filtering, regularization strength and learning rate.

5. The method according to claim 1, characterized in that: The user feedback mechanism includes correction operations on user recommendation results and a dynamic training set constructed based on user modification logs to achieve incremental data fusion, optimize algorithm design, and improve algorithm performance.

6. A method for reducing algorithmic aversion based on user modification of a model according to claim 1, characterized in that: Adjusting model parameters, optimizing feature selection, and updating training data based on user data, modification traces, and feedback includes the following steps: S5.1, establish the exponential decay mechanism of regularization parameters, design the periodic learning rate adjustment strategy and regulate the momentum coefficient in stages to achieve dynamic optimization of parameters; S5.2, integrating statistical significance and business importance, and combining t-SNE visual dimensionality reduction to screen intelligent features; S5.3, inject new data into the algorithm model, build a one-time verification mechanism for the new and old models, set an exponentially decaying freshness indicator, coordinately update the triple technologies of parameter adjustment, feature optimization and data dynamics, build a user-oriented algorithm evolution system, and effectively alleviate the phenomenon of algorithm aversion.

7. A method for reducing algorithmic aversion based on user modification of a model according to claim 1, characterized in that: The system corresponding to the method includes: Algorithm development module, which is used to provide users with the ability to deeply participate in algorithm design and customization; An algorithm explanation module, which is used to improve the transparency and explainability of the algorithm so that users can better understand the decision logic and working principle of the algorithm; An algorithm decision module, which is used to enable users to make decisions using algorithms and modify algorithm models; A user feedback module, which is used to collect, organize and analyze user feedback on the algorithm and algorithm decision results, and establish an effective user feedback mechanism; The algorithm optimization module is used to adjust model parameters, optimize feature selection and update training data based on user data, modification traces and feedback, so as to improve the performance and accuracy of the algorithm.

Images