A recommendation method, system and device based on weight calculation of multi-dimensional feature portrait

Abstract

This application provides a recommendation method, system, and apparatus for weight calculation based on multi-dimensional feature profiles. The method includes: using a demand analysis agent to extract task features, data features, user features, and environmental features from unstructured user natural language descriptions based on a preset prompt template; using a data parsing agent to match structured raw data with a preset data template to obtain a target data template, wherein the preset data template is set based on weight calculation methods in a weight method knowledge base; analyzing the raw data based on the target data template and adding the analysis results to the data features; and using a method recommendation agent to match the multi-dimensional feature profile with each weight calculation method in the weight method knowledge base to obtain at least one target weight calculation method for calculating weights.

Description

Technical Field

[0001] This application relates to the field of comprehensive evaluation technology, specifically to a recommendation method, system, and apparatus based on weight calculation of multi-dimensional feature profiles. Background Technology

[0002] Weight determination is a core step in multi-indicator comprehensive evaluation, decision analysis, and statistical modeling, and its results directly affect the scientific validity and reliability of the evaluation system. With the explosion of multi-source heterogeneous data in the era of big data, the scenarios of evaluation tasks are becoming increasingly complex.

[0003] While the theory of weight determination is relatively mature, existing technologies still have the following significant shortcomings in practical applications and system implementation: Lack of scientific basis for weighting method selection ("selection difficulty" problem): Existing software only provides a stack of calculation functions and lacks an intelligent method recommendation mechanism. Faced with more than 20 weighting methods, including subjective, objective, and combined methods, non-professional users find it difficult to make scientific choices based on task objectives (such as ranking, optimization), data characteristics (such as sample size, distribution pattern), and environmental constraints. They often rely on personal experience or blindly apply methods, leading to method abuse and distorted evaluation results. Summary of the Invention

[0004] The purpose of this application is to provide a recommendation method, system, and apparatus based on weight calculation of multi-dimensional feature profiles. The specific technical solution adopted is as follows: Firstly, a recommendation method based on weight calculation using multi-dimensional feature profiling is provided, the method comprising: The demand analysis agent extracts task features, data features, user features, and environmental features from the multi-dimensional feature profile based on a preset prompt template from unstructured user natural language descriptions. A data parsing agent is used to match structured raw data with a preset data template to obtain a target data template, wherein the preset data template is set based on the weight calculation method in the weight method knowledge base; the raw data is analyzed based on the target data template, and the analysis results are added to the data features; The method recommends that the agent match each weight calculation method in the weight method knowledge base with the multi-dimensional feature profile to obtain at least one target weight calculation method for calculating weights.

[0005] Secondly, a weighted recommendation system based on multi-dimensional feature profiling is provided. This system includes a demand analysis agent, a data parsing agent, and a method recommendation agent. The demand analysis agent is used to extract task features, data features, user features, and environmental features from the unstructured user natural language description based on a preset prompt template; The data parsing agent is used to match structured raw data with a preset data template to obtain a target data template, wherein the preset data template is set based on the weight calculation method in the weight method knowledge base; based on the target data template, the analysis results are added to the data features. The method recommends an intelligent agent, which is used to match each weight calculation method in the weight method knowledge base based on the multi-dimensional feature profile to obtain at least one target weight calculation method for calculating weights.

[0006] Thirdly, a weighted recommendation device based on multi-dimensional feature profiling is provided, the device comprising: The feature extraction module is used to extract task features, data features, user features, and environmental features from the multi-dimensional feature profile based on the unstructured user natural language description using a demand analysis agent and a preset prompt template. The data analysis module is used to match structured raw data with a preset data template using a data parsing intelligent agent to obtain a target data template. The preset data template is set based on the weight calculation method in the weight method knowledge base. Based on the analysis results of the target data template, the analysis results are added to the data features. The method matching module is used to match the method recommendation agent with each weight calculation method in the weight method knowledge base based on the multi-dimensional feature profile to obtain at least one target weight calculation method for calculating weights.

[0007] Fourthly, a computer program product is provided, comprising: computer program code, which, when run on a computer, causes the computer to perform the method in any of the possible implementations described above.

[0008] Fifthly, a computer-readable storage medium is provided that stores computer program code, which, when run on a computer, causes the computer to perform the methods in any of the possible implementations described above.

[0009] This application has the following beneficial effects: Achieving intelligent method recommendation based on "four-dimensional profile": By constructing a feature system covering task, data, user, and environment dimensions, and using an algorithm that combines rule matching and semantic reasoning, the system automatically recommends and matches the optimal weight calculation method for users, significantly improving the scientific nature and adaptability of weight method selection and addressing the problem of lack of scientific basis in method selection.

[0010] Achieving semantic-driven natural language interaction: By using a demand analysis agent, users' natural language demands are transformed into system instructions, enabling users to complete complex weighting tasks without needing to master complex programming or deep statistical principles, greatly reducing the barrier to entry for professional statistical analysis. Attached Figure Description

[0011] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A schematic diagram of the framework of a weighted recommendation system based on multi-dimensional feature profiling provided in an embodiment of this application; Figure 2 A schematic diagram of a multi-agent system based on a multi-agent cooperative mode is provided for an embodiment of this application; Figure 3 A schematic diagram illustrating the implementation process of a recommendation method based on multi-dimensional feature profiling for weight calculation, provided in an embodiment of this application; Figure 4 A schematic diagram illustrating the feature composition of a four-dimensional problem profile provided in this application embodiment; Figure 5 A schematic diagram illustrating a method for converting natural language using prompt words in a large language model, as provided in this application; Figure 6 A schematic diagram of a weight determination device based on a large language model and multi-agent collaboration provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0013] To further illustrate the technical means and effects adopted by this application to achieve the intended inventive objective, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a recommended method for weight calculation based on multi-dimensional feature profiling proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined from any suitable form.

[0014] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0015] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0016] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0017] Figure 1 This is a schematic diagram of the framework of a weighted recommendation system based on multi-dimensional feature profiling provided in an embodiment of this application, as shown below. Figure 1 As shown, the framework comprises two core stages: problem profiling (multi-dimensional feature profiling) 11 and method matching 12, wherein, Problem profiling stage 11: Transforming unstructured user needs and raw data characteristics into a standardized attribute vector encompassing four dimensions: task, data, user, and environment. P .

[0018] Method matching in 12 stages: based on knowledge graphs M By utilizing a dual reasoning mechanism of "rules + semantics", a problem profile is created. P The recommendations are generated by matching the results with weighted methods.

[0019] Figure 2 A schematic diagram of a multi-agent system based on a multi-agent collaborative mode is provided for embodiments of this application, such as... Figure 2 As shown, the multi-agent system 20 includes a central management agent 21 and three specialized function agents: Management Agent 21: Located at the center of the architecture, it is responsible for task scheduling, status monitoring, and information routing between agents.

[0020] Demand Analysis Agent 22 and Data Parsing Agent 23: As input terminals, they are responsible for parsing the user's natural language questionnaire and uploaded data files (Excel / CSV) respectively, and extracting features.

[0021] Method Recommendation Agent 24: Based on the parsed features, it outputs a recommended ranking using the built-in algorithm.

[0022] During implementation, the multi-agent system 20 first obtains requirement information, including user requirement questionnaires and data (Excel / CSV / PDF / Word), for requirement analysis; then it determines the target weight method that matches the requirements from the weight method library and custom weight methods; finally, it uses the target weight method to obtain the weight calculation results.

[0023] The weighted method library can store at least the following: Figure 2 The three main categories of weight calculation methods are shown below: Subjective weighting methods include: Delphi method, Analytic Hierarchy Process (AHP), and expert scoring method.

[0024] Objective weighting methods include: entropy weighting, coefficient of variation, and CRITIC method.

[0025] Combined weighting methods include: weighted method, product method, and minimum relative deviation method.

[0026] It can output as follows Figure 2 The weighting method recommendation ranking shown includes information such as the recommended method, four-dimensional matching score, reasons for recommendation, implementation suggestions, and method details.

[0027] In this application embodiment, four types of intelligent agents with specific functions and their data flow protocols are defined, realizing an automated closed loop for statistical analysis from qualitative to quantitative.

[0028] During implementation, the weight determination problem is abstracted into a standardized closed-loop process of "problem profiling - method matching". A multi-agent collaborative architecture is used to achieve fully automated decision support. The recommended method for calculating weights can be implemented through the following steps: Step S1: Multimodal input reception and preliminary analysis.

[0029] During implementation, methods such as Figure 2 The management agent 21 shown receives user input information, including a task description in natural language (such as evaluation purpose and background constraints) and structured / semi-structured data files (such as indicator data in CSV or Excel format).

[0030] Step S2: Construct a "problem profile" based on a four-dimensional framework.

[0031] It can be used as Figure 2 The demand analysis agent 22 and the data parsing agent 23 shown collaborate to extract features from the input information and construct a standardized "weight determination problem profile" (multi-dimensional feature profile).

[0032] During implementation, the system maps problem characteristics into a four-dimensional attribute vector. The four dimensions include: task dimension, data dimension, user dimension, and environment dimension. Task dimensions: include evaluation domain, evaluation purpose (ranking / optimization), problem complexity, etc. Data dimensions include the number of indicators, data types, data quality (missing / outliers), and distribution characteristics. User dimensions: These include user expertise, methodological preferences (subjective / objective), and accuracy requirements; Environmental dimension: including computing power constraints, time constraints, expert resources, etc.

[0033] For natural language descriptions, entity extraction and intent recognition are performed using a Large Language Model (LLM), which is then transformed into the aforementioned structured labels. For data files, descriptive statistics are automatically calculated to populate data dimensional features.

[0034] Step S3: Recommendation of methods based on a dual-driven approach of "rules + semantics".

[0035] During implementation, methods such as Figure 2 The method shown recommends agent 24, which matches the problem profile with a pre-built weighted method knowledge base. The matching process employs a hybrid strategy of "initial rule screening - semantic refinement": S3-1 Rule Matching (Initial Screening): Based on preset statistical applicability rules (such as "if the sample size is <30 and there are no experts, then the entropy weight method is excluded"), the rule matching score of each weighted method is calculated to select a set of candidate methods.

[0036] S3-2 Semantic Reasoning (Enhancement): Triggers the semantic reasoning mechanism of the large language model. Inputs the problem profile and method metadata into the LLM, analyzes their semantic similarity, and generates a semantic matching score.

[0037] S3-3 Hybrid Sorting: Calculate the comprehensive score according to the formula and output the optimal recommendation list.

[0038] To implement the above method, this application embodiment constructs a system architecture based on multi-agent cooperation, which mainly includes the following modules: 1. Central Management Module. As the system's central control center, it is responsible for parsing user requests, decomposing tasks into sub-tasks, and scheduling other specialized intelligent agents to work collaboratively. It monitors the status of each stage and manages the flow of context information. It can provide, for example... Figure 2 The management agent 21 is shown.

[0039] 2. Knowledge Base Module (Pre-built weighted method knowledge base). Stores a knowledge graph of weighted methods. Each method entity contains a four-tuple of information: category, mathematical principle, applicable rule set, and description of advantages and disadvantages. This module provides data support for rule matching.

[0040] 3. Requirements Analysis and Data Parsing Module.

[0041] Requirements Analysis Unit: This unit is configured with dedicated prompt templates for extracting task, user, and contextual features from natural language dialogue. It can be used for, for example... Figure 2 The requirement analysis agent 22 is shown.

[0042] The data parsing unit integrates statistical tools such as Pandas / NumPy to read uploaded files and automatically detect data dimensional features (such as correlation matrix and missing rate). It can be used for... Figure 2 The data parsing agent 23 is shown.

[0043] 4. Intelligent Recommendation Module. This module incorporates a hybrid recommendation algorithm engine. It connects to the knowledge base module to perform rule matching operations; simultaneously, it connects to a large language model interface to perform semantic similarity analysis, ultimately outputting a ranked recommendation list and the reasons for each recommendation. It can provide, for example... Figure 2 The method shown recommends proxy 24.

[0044] This application's embodiments utilize the semantic understanding and reasoning capabilities of a large language model, combined with a multi-agent collaborative architecture, to achieve intelligent and automated completion of the entire weighted method recommendation process. This achieves the following beneficial technical effects: Achieving intelligent method recommendation based on "four-dimensional profile": By constructing a feature system covering task dimension, data dimension, user dimension, and environment dimension, and using an algorithm that combines rule matching and semantic reasoning, the system automatically recommends and matches the optimal weight calculation method for users, thus solving the problem of lack of scientific basis for solution selection.

[0045] Achieving semantic-driven natural language interaction: Utilizing large language models to transform users' unstructured natural language needs into structured system instructions, enabling users to complete complex empowerment tasks without needing to master complex programming or deep statistical principles.

[0046] This application provides a recommendation method based on weight calculation of multi-dimensional feature profiles, such as... Figure 3As shown, this can be achieved through the following steps: Step S310: Using a demand analysis agent based on a preset prompt template, extract task features, data features, user features, and environmental features from the unstructured user natural language description in the multi-dimensional feature profile; Here, the demand analysis agent can be like this: Figure 2 The required analysis agent 22 shown is used to parse user natural language input and structured requirement descriptions, and identify features in the task dimension, user dimension and environment dimension.

[0047] The task dimensions include the evaluation domain, evaluation purpose, and problem complexity; The data dimensions include the number of indicators, data types, data quality, and distribution characteristics; The user dimensions include the user's professional level, methodological preferences, and accuracy requirements; The environmental dimension includes computing power constraints, time constraints, and expert resource availability.

[0048] Figure 4 This application provides a schematic diagram illustrating the feature composition of a four-dimensional problem profile, as shown in the embodiments. Figure 4 The "problem profile" shown includes attribute vectors for four key dimensions: task dimension 41, data dimension 42, user dimension 43, and environment dimension 44. These four dimensions are used to comprehensively characterize the task and determine its weights. Task Dimension 41: Describe the evaluation field (such as financial investment, education evaluation, medical and health care), the nature of the evaluation objectives (such as descriptive, predictive, comparative), the complexity of the problem (such as simple ranking, multi-attribute decision-making, group decision-making), the evaluation level (strategic level, tactical level, operational level), and the scope of application of the results (internal management, external reports, academic publications).

[0049] Data Dimension 42: Describes data scale (sample size, number of indicators), data quality (completeness, accuracy, consistency), data distribution (distribution pattern, outliers), and data correlation (degree of correlation between indicators).

[0050] User Dimension 43: Describes the user's cognitive preferences (decision-making style, method complexity), experience level (weight determination experience, decision analysis experience), and special needs.

[0051] Dimension 44: Describe resource constraints (number of experts, difficulty of data acquisition) and time constraints (real-time, fast, long-term).

[0052] Here, the proposed "weighted intelligent profiling" theory maps fuzzy statistical evaluation tasks into a four-dimensional feature vector composed of "task-data-user-environment". In this way, the system can map problem features into a four-dimensional attribute vector. .

[0053] In some embodiments, the preset prompt template includes: recognizing language descriptions related to the task features, the data features, the user features, and the environment features, and outputting the multidimensional feature vector of the structured representation.

[0054] Figure 5 This application provides a schematic diagram of a large language model conversion method that utilizes prompt words to convert natural language, such as... Figure 5 As shown, the input natural language for the large language model is: "We want to make a rough comprehensive ranking of some counties in our jurisdiction. The indicators are some economic and social development data, such as population, GDP, and public service status, which are all publicly available and have specific figures. We don't understand complex algorithms and it's inconvenient to ask experts to score them, so we want a relatively simple and transparent way to calculate it, without too much subjectivity." The prompt is: "You are a weighted method recommendation expert, skilled at converting natural language task descriptions into structured 'problem profile' attribute vectors. Based on the following user input, identify the four dimensions of its task-related attributes (task dimension, data dimension, user dimension, and environment dimension), and output the prompt as a standard attribute vector in JSON format." Output a JSON-formatted attribute vector, with the following content: { "Task Dimension": { "Evaluation Area": ​​"Resource Allocation" "Evaluation Target Type": "Comparability" "Problem Complexity": "Multidimensional Indicators" "Does it have a hierarchical structure?": "No" "Scope of application of results": "Management decision-making" }, "Data Dimensions": { Sample size: Medium Data precision: Medium "Variable consistency": "Yes" "Distribution characteristics": "Unknown" "Indicator Correlation": "Unknown" }, "User Dimension": { "Cognitive bias": "Objective data" "Experience Level": "Low" "Special Needs": "Avoiding Expert Scoring" }, "Environmental Dimension": { Number of experts: None "Difficulty of data acquisition": "Low" "Time Constraint": "Unknown" } } like Figure 5 As shown, the large language model can transform the natural language description input by the user into a structured attribute vector based on prompt words.

[0055] This approach employs a "four-dimensional problem profiling" modeling technique to acquire features across four dimensions. It considers not only objective data-level characteristics (such as sample size and distribution) but also qualitative task-level objectives (such as evaluation purpose and application scenario) and user preferences. This multi-dimensional matching mechanism avoids the drawbacks of traditional methods that rely solely on experience or blind application. It provides a foundation of feature inputs for subsequent methods that determine target weights based on problem profiling.

[0056] Step S320: Use a data parsing agent to match the structured raw data with a preset data template to obtain a target data template, wherein the preset data template is set based on the weight calculation method in the weight method knowledge base; analyze the raw data based on the target data template and add the analysis results to the data features; Here, the preset data template can be set based on the weight calculation methods in the weight method knowledge base. For example, a corresponding data template can be set for multiple weight calculation methods with similar characteristics.

[0057] After users fill in the data based on the data template, the data parsing agent transforms the unstructured data file into a structured feature vector (data feature) that can be understood by a large language model through a two-stage processing flow (structural compliance verification and deep extraction of content features).

[0058] In this way, the system does not directly process files of arbitrary formats, but instead has a pre-built data template library corresponding to the weighting method knowledge base (including subjective scoring templates, objective indicator templates, etc.). For CSV / Excel data files uploaded by users based on data templates, the data parsing agent transforms the unstructured data files into structured feature vectors that can be understood by large models.

[0059] The system establishes a precise mapping relationship between its pre-built standardized template library (pre-defined data templates) and the weight knowledge method library. Each category of weight method can be assigned one or more standard data templates, as shown in Table 1 below: Table 1

[0060] As shown in Table 1 above, schema constraints explicitly define the structure that data should have. This includes which fields (or attributes) the data should contain, the data types of these fields (such as strings, numbers, booleans, etc.), and possibly the hierarchical relationships between fields.

[0061] Different weighting methods can be configured with corresponding standard templates and schema constraints. For example, the objective weighting method can correspond to the original indicator data model, with the schema constraint being "an n×m pure numerical matrix, with the first column being the object identifier".

[0062] The system's pre-built standardized template library establishes a precise mapping relationship with the weighting method library. For objective weighting methods such as entropy weighting and CRITIC methods, it provides raw indicator data templates. These templates require the data to be presented as an n x m pure numerical matrix, where the first column is the evaluation object identifier, and the remaining columns are the numerical data for each indicator. For subjective weighting methods such as expert evaluation and Delphi methods, it provides expert scoring templates, requiring the data to be a scoring matrix composed of experts and indicators, with values ​​typically ranging from one to ten. For the Analytic Hierarchy Process (AHP) and Analytic Network Analysis (ANP), it provides pairwise comparison judgment templates, requiring the data to be an n-order symmetric matrix, with diagonal elements all being 1, and satisfying reciprocity constraints. For the Data Envelopment Analysis (DEA), it provides multi-input multi-output templates, requiring a structured table with clearly labeled input and output columns.

[0063] The technical advantages of this design are: first, it eliminates the uncertainty caused by arbitrary format recognition; second, it significantly reduces the error rate of data parsing; and third, it provides users with clear operating guidelines and specifications.

[0064] During implementation, a visual interface can be provided to users, displaying the data required for different data models. Users can fill in the data based on the data template to obtain structured raw data.

[0065] Utilize Figure 2 The data parsing agent 23 shown combines structured parsing rules and semantic understanding capabilities to determine the target data template based on the original data, or based on the template identifier of the original data filled in by the user.

[0066] After determining the target data template, the original data is analyzed using the target data template. This allows for the analysis of the data characteristics of the original data, including data size, data quality, and data distribution. The analysis results are then added to the data characteristics obtained in step S310 to update the multidimensional feature profile.

[0067] Here, you can set up weighting methods in the pre-built weighting method knowledge base, including the following three major systems: 1. Subjective Weighting Methods: Represented by the Analytic Hierarchy Process (AHP) and the Delphi method, these methods primarily rely on expert experience, determining weights through the construction of judgment matrices or multiple rounds of scoring. These methods are widely used in scenarios with a limited number of indicators and available expert resources. 2. Objective Weighting Methods: Represented by entropy weighting, the CRITIC method, and the coefficient of variation method, these methods primarily mine the statistical characteristics of the data itself (such as dispersion, information content, and correlation) to automatically calculate weights, avoiding human interference. 3. Combined Weighting Methods: These methods integrate subjective and objective weights through linear weighting or nonlinear models (such as game theory combination or multiplicative synthesis), attempting to balance expert experience with data information.

[0068] Step S330: The recommended agent uses the method to match each weight calculation method in the weight method knowledge base with the multi-dimensional feature profile to obtain at least one target weight calculation method for calculating weights.

[0069] Here, the recommended agent can be as follows: Figure 2 The method recommended by agent 24, based on a pre-set weight method knowledge base and method matching algorithm, determines at least one target weight calculation method for calculating weights based on multi-dimensional feature profiles, thereby achieving accurate matching between features and methods.

[0070] The multidimensional feature profile can be the updated multidimensional feature profile obtained in step S320, which adds data features.

[0071] This application proposes a method for receiving user input, constructing a multi-dimensional feature profile, performing rule-based and semantic hybrid matching, and outputting a recommendation. This method offers the following beneficial technical effects: Achieving intelligent method recommendation based on "four-dimensional profile": By constructing a feature system covering task, data, user, and environment dimensions, and using an algorithm that combines rule matching and semantic reasoning, the system automatically recommends and matches the optimal weight calculation method for users, significantly improving the scientific nature and adaptability of weight method selection and addressing the problem of lack of scientific basis in method selection.

[0072] Achieving semantic-driven natural language interaction: By using a demand analysis agent, users' natural language demands are transformed into system instructions, enabling users to complete complex weighting tasks without needing to master complex programming or deep statistical principles, greatly reducing the barrier to entry for professional statistical analysis.

[0073] In some embodiments, the step S320 above, "using a data parsing agent to match structured raw data with a preset data template to obtain a target data template," can be achieved through the following steps: Step 321: The data parsing agent reads the header information of the original data; During implementation, after the user uploads the completed template file, the data parsing agent first reads the header fingerprint (header information) of the file.

[0074] The data parsing agent first parses the header area of ​​the file, typically reading the first three lines to extract metadata such as template identification information, version number, and schema declaration. Then, it performs a hash calculation on the data table header to generate a unique identifier, which takes into account the column name sequence, the number of columns, and the data type declaration.

[0075] Step 322: Compare the header information with the header information pre-stored in the standard template library to determine the target data template corresponding to the original data.

[0076] During implementation, the header information entered by the user can be compared with the pre-stored header information to determine the target data template to use for entering the original data. For example, if the header entered by the user matches the header of the original indicator data template, it can be determined that the user is using the original indicator data template to enter the original data.

[0077] By comparing the generated header fingerprint with the fingerprints pre-stored in the standard template library, the system can quickly determine the type of template used by the user in constant time complexity.

[0078] In this embodiment of the application, the header information is compared with the header information pre-stored in the standard template library, which can effectively and accurately determine the target data template used to fill in the original data.

[0079] In some embodiments, the step S320 above, "based on the analysis results of the original data according to the target data template, add the analysis results to the data features", can be achieved through the following steps: Step 323: The data parsing agent reads the metadata of the original data; During implementation, after the user uploads the completed template file, the data parsing agent can also read the metadata filled in the file's metadata area.

[0080] Step 324: Determine the matching degree between the metadata and the target data template; During implementation, filling in the matching degree can be used to verify whether the uploaded data conforms to the system's predefined schema constraints (e.g., AHP templates must contain a judgment matrix with a diagonal of 1; entropy weight method templates must contain pure numerical columns).

[0081] Here, we provide a multi-dimensional validation method for schema constraints as follows: Verification Dimension 1: Structural Constraints.

[0082] The system verifies whether the actual number of rows and columns conforms to the template definition specifications, checks whether key columns such as object identifier columns and indicator columns are complete and exist, and verifies whether the data entry area exceeds the boundary or intrudes into the metadata reservation area.

[0083] Verification Dimension Two: Data Type Constraints.

[0084] For columns declared as numeric, the system will scan every cell to check for the presence of text characters or illegal symbols. For text columns, especially those with a fixed vocabulary such as "Yes / No" or "High / Medium / Low", the system will verify whether the values ​​are within the predefined allowed range.

[0085] Verification Dimension 3: Method-Specific Constraints.

[0086] This is one of the core innovations of this invention. For different weighting methods, the system embeds domain-specific verification rules.

[0087] Taking the AHP pairwise comparison matrix as an example, the system performs triple verification: First, diagonal element verification, ensuring that all elements on the main diagonal of the matrix are 1; Second, reciprocity verification, which is the key innovation point, the system checks whether the product of any pair of elements in row i and column j and row j and column i in the matrix is ​​equal to 1, allowing a numerical error of one-thousandth; Third, value range verification, ensuring that all elements are within the scale range of one-ninth to nine.

[0088] Taking the DEA data template as an example, the system verifies the clear labeling of input and output columns, checks for negative or zero values, and verifies whether the ratio of the number of decision-making units to the number of input and output indicators meets the statistical requirements.

[0089] During implementation, standard templates can only regulate data structure and cannot guarantee content quality. Data entered by different users using the same standard template may vary significantly in quality. Therefore, even with a known structure, in-depth content analysis is still necessary. Matching accuracy can be analyzed across the following four feature dimensions: Feature Dimension 1: Statistical Distribution Feature Extraction.

[0090] A full statistical analysis is performed on the data matrix that has passed schema validation. For each column of indicator data, central tendency indicators, including the arithmetic mean and median, are calculated, and dispersion indicators, including standard deviation, variance, and coefficient of variation, are calculated. The coefficient of variation is defined as the ratio of the standard deviation to the mean, and is used to measure the relative volatility of the data.

[0091] Further calculations are performed on the distribution characteristics, including skewness and kurtosis. Skewness measures the asymmetry of the data distribution, while kurtosis measures the steepness of the data distribution. Based on these two indicators, the system automatically determines the distribution type: when the absolute value of skewness is less than 0.5 and the absolute value of the difference between kurtosis and 3 is less than 1, it is determined to be an approximately normal distribution; when skewness is greater than 1, it is determined to be a right-skewed distribution, indicating the existence of a maximum value; when skewness is less than -1, it is determined to be a left-skewed distribution, indicating the existence of a minimum value; other cases are determined to be non-normal distributions.

[0092] Identifying the distribution pattern provides a crucial basis for subsequent method recommendations. For example, when the data exhibits a severely skewed distribution, some statistical methods based on the assumption of normality, such as principal component analysis, may not be applicable, and the system will adjust its recommendation strategy accordingly.

[0093] Feature Dimension Two: Quantification of Data Quality Features.

[0094] Completeness analysis: Calculate the percentage of missing cells in the system's statistical data matrix.

[0095] Outlier detection employs a strategy that integrates three algorithms. The first is the Three Sigma principle, which marks data points exceeding three standard deviations above or below the mean as outliers. The second is the IQR interquartile range method, marking data points below the first quartile minus 1.5 times the IQR or above the third quartile plus 1.5 times the IQR as outliers. The third is the LOF (Local Outlier Factor) algorithm, a density-based anomaly detection method that identifies outliers within local regions. The system combines the detection results from these three algorithms to obtain a comprehensive list of outliers and calculates the proportion and specific locations of each outlier.

[0096] For expert rating data, the system calculates the degree of consistency among expert opinions. The Kendall coefficient of concordance or intragroup correlation coefficient (ICC) is used as a quantitative indicator to assess whether the ratings among different experts exhibit high consistency.

[0097] Through multi-dimensional quality quantification, the system can comprehensively assess the reliability of the data, providing a basis for subsequent recommendations.

[0098] Feature Dimension 3: In-depth analysis of correlation structure.

[0099] Pearson correlation coefficient matrix calculation: Calculate the Pearson correlation coefficients between all pairs of indicators to construct a complete correlation coefficient matrix.

[0100] Overall correlation strength assessment: The average absolute values ​​of the upper triangular elements of the correlation coefficient matrix are calculated to obtain the overall correlation strength. A correlation coefficient greater than 0.7 is considered strong, greater than 0.4 is considered moderate, and otherwise it is considered weak.

[0101] Multicollinearity diagnosis: Calculate the variance inflation factor (VIF) for each indicator. A VIF value greater than 10 indicates severe multicollinearity between that indicator and other indicators. This diagnosis is crucial for recommending subsequent methods.

[0102] Significantly correlated pairs identification: Traverse the correlation coefficient matrix, identify index pairs with an absolute correlation coefficient greater than 0.7, mark them as highly correlated, and add the following warning message: "Highly correlated, information redundancy may exist".

[0103] Correlation structure analysis enables the system to convey key information to large language models, such as "Although this is raw index data that meets the format requirements of the entropy weight method, there is serious multicollinearity. It is recommended to first use principal component analysis for dimensionality reduction, or to use a method that is not sensitive to multicollinearity, such as the CRITIC method."

[0104] Feature Dimension Four: Calculation of Method-Specific Validation Metrics.

[0105] AHP template consistency ratio (CR) (Key Quality Indicator): First, calculate the largest eigenvalue λmax of the judgment matrix. Then, calculate the consistency index CI, which is equal to λmax minus the matrix order n, and then divided by n minus 1. Next, look up the random consistency index RI, which is a predefined constant table based on the matrix order. Finally, calculate the consistency ratio CR, which is equal to CI divided by RI.

[0106] Quality is determined based on the CR value: a CR value less than 0.05 is considered excellent, less than 0.1 is considered good, less than 0.15 is considered acceptable, and a CR value greater than or equal to 0.15 is considered unacceptable and requires reassessment.

[0107] It not only verifies the data structure, but also automatically verifies whether the data content satisfies the mathematical assumptions of the AHP method, a capability that traditional data processing systems do not possess.

[0108] The DEA template pre-calculates efficiency values: Using the CCR model for linear programming solutions, it pre-calculates the efficiency value for each decision unit (DMU). If the efficiency value of all DMUs is 1.0, the system issues a warning: "All decision unit efficiency values ​​are 1, the data may have problems or the model may be overfitting." If the efficiency values ​​are generally below 0.5, the system suggests: "Efficiency values ​​are generally too low; it is recommended to check whether the selection of input-output indicators is reasonable."

[0109] By embedding domain knowledge from statistics and operations research into validation rules in a procedural manner, the automated application of expert experience is achieved.

[0110] Step 325: If the matching degree is greater than or equal to the matching degree threshold, analyze the metadata based on the target data template and add the analysis results to the data features.

[0111] During implementation, if the verification is successful, i.e. the matching degree is greater than or equal to the matching degree threshold, the distribution characteristics (such as skewness and kurtosis), quality characteristics (such as missing rate and outlier ratio) and correlation structure of the data are calculated based on the target data template, and the analysis results are added to the data features.

[0112] All the above calculation results (structure, quality, distribution) are encapsulated into a standardized JSON object, namely a structured feature vector. This vector contains metadata such as data source, template type, validation status, and timestamps, as well as substantive content such as structural features, quality features, distribution features, correlation features, and method applicability scores. For example... Figure 5 The data dimensions shown have the following characteristics.

[0113] Step 326: If the matching degree is less than the matching degree threshold, the corresponding metadata is identified as abnormal metadata; If the verification fails (e.g., the user changed the table header or entered text), the system can directly intercept and report the specific error, identifying the corresponding metadata as abnormal metadata, thus achieving pre-processing of data cleaning.

[0114] Step 327: Output a prompt message to indicate that the abnormal metadata should be modified to match the target data template.

[0115] When the verification fails, the system generates a detailed error report, pinpointing the specific row and column numbers, comparing the expected value with the actual value, and providing actionable repair suggestions. For example, when the reciprocity of the AHP matrix is ​​not satisfied, the system will indicate that "the matrix element in the second row and third column does not satisfy the reciprocity with the element in the third row and second column: 5 multiplied by 0.5 equals 2.5, which is not equal to 1. It is recommended to change the value in the third row and second column to 0.2."

[0116] In this embodiment, a pre-built standardized data template library (including subjective scoring templates, objective indicator templates, etc.) corresponding to the weighting method library is utilized. For CSV / Excel data files uploaded by users based on standardized templates, the data parsing agent can filter out abnormal metadata and transform the metadata into structured feature vectors that can be understood by large models.

[0117] In some embodiments, step S330 above, "using the method recommendation agent to match each weight calculation method in the weight method knowledge base based on the multi-dimensional feature profile to obtain at least one target weight calculation method for calculating weights," can be implemented through the following steps: Step 331: Using the statistical applicability rules integrated in the method recommendation agent, match the problem profile with each weight calculation method in the pre-set weight method knowledge base to obtain the rule matching score corresponding to each weight calculation method. Here, each method entity in the pre-built weighted method knowledge base contains a quadruple of information: category, mathematical principle, set of applicable condition rules, and description of advantages and disadvantages.

[0118] During implementation, statistical applicability rules integrated into the method recommendation agent can be used (e.g., "if the sample size is <30 and there are no experts, then exclude the entropy weight method"). Based on the above four-tuple information, the rule matching score of each weight method can be calculated. .

[0119] Step 332: Based on sorting the rule matching scores as the total matching scores, obtain the at least one target weight calculation method for calculating weights.

[0120] In this embodiment of the application, sorting based on rule matching scores can yield at least one target weight calculation method that matches the rule.

[0121] In some embodiments, step 332 above, "obtaining the at least one target weight calculation method for calculating weights by sorting the rule matching scores as the total matching scores," can be implemented through the following steps: Step 3321: Analyze the semantic similarity between each target weight calculation method and the multi-dimensional profile using a large language model to obtain the semantic matching score of each weight calculation method. During implementation, the semantic reasoning mechanism of the large language model can be triggered. The metadata of the multi-dimensional feature profile and the weight calculation method is input into the LLM to analyze their semantic similarity and generate a semantic matching score for each target weight calculation method. .

[0122] Step 3322: Perform a weighted summation of the rule matching score and the semantic matching score to obtain the total matching score for each weight calculation method; During implementation, the matching total score can be determined using the following formula (1). Score final : (1); in, Score for rule matching. For semantic matching scores, α These are the weighting coefficients.

[0123] Here, by adjusting the weighting coefficients αIt allows for flexible control over the contribution of rule matching score and semantic matching score to the final matching score. If the rule matching score... The weight α is set to 0.7, and the semantic matching score is... If the weight is set to 0.3, the total matching score will depend more on the strictness of the rule matching.

[0124] Step 3323: Sort the total matching score by row to obtain at least one target weight calculation method for calculating weight.

[0125] In this embodiment, a more comprehensive matching score is generated by combining the strictness of the rule matching with the flexibility of the semantic matching through weighted summation. Then, the optimal method is selected based on the score ranking, thus avoiding the one-sidedness of a single indicator.

[0126] This application provides a weight determination system based on a large language model and multi-agent collaboration. The weight determination system includes, for example: Figure 2 The diagram shows a requirement analysis agent 22, a data parsing agent 23, and a method recommendation agent 24, among which... The demand analysis intelligent agent 22 is used to extract task features, data features, user features and environmental features from the unstructured user natural language description based on a preset prompt template; The data parsing agent 23 is used to match structured raw data with a preset data template to obtain a target data template, wherein the preset data template is set according to the weight calculation method in the weight method knowledge base; based on the target data template, the analysis results are added to the data features. The method recommends agent 24, which is used to match the multi-dimensional feature profile with each weight calculation method in the weight method knowledge base to obtain at least one target weight calculation method for calculating weights.

[0127] This application provides a weight determination device based on a large language model and multi-agent collaboration. Please refer to [link to relevant documentation]. Figure 6 The system 600 includes: The feature extraction module 610 is used to extract task features, data features, user features and environmental features from the unstructured user natural language description based on a preset prompt template using a demand analysis intelligent agent. The data analysis module 620 includes a target data template, wherein the preset data template is set based on the weight calculation method in the weight method knowledge base; the analysis results of the original data are added to the data features based on the analysis results of the target data template; The method matching module 630 is used to match the method recommendation agent with each weight calculation method in the weight method knowledge base based on the multi-dimensional feature profile to obtain at least one target weight calculation method for calculating weights.

[0128] Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. For example, as shown... Figure 7 As shown, the computer device 700 includes: a memory 701, a processor 702, and a computer program 703 stored in the memory 701 and running on the processor 702, wherein when the processor 702 executes the computer program 703, the computer device can execute any of the aforementioned recommendation methods based on multi-dimensional feature profile weight calculation.

[0129] Furthermore, this application also protects a control device, which may include a memory and a processor. The memory stores executable program code, and the processor is used to call and execute the executable program code to perform a recommendation method for weight calculation based on multi-dimensional feature profiles provided in this application. This application can divide the control device into functional modules based on the above method examples. For example, each module may correspond to a specific function, or two or more functions may be integrated into a single processing module. The integrated module can be implemented in hardware. It should be noted that the module division in this application is illustrative and only represents a logical functional division; other division methods may exist in actual implementation. It should also be noted that all relevant content of each step involved in the above method embodiments can be referenced to the functional description of the corresponding functional module, and will not be repeated here. It should be understood that the control device provided in this application is used to execute the above-mentioned recommendation method for weight calculation based on multi-dimensional feature profiles, and therefore can achieve the same effect as the above-described implementation method. When using integrated units, the control device may include a processing module and a storage module. When the control device is applied to a block device, the processing module can be used to control and manage the actions of the block device. The storage module can be used to support block devices in executing mutual program code, etc. The processing module can be a processor or controller, which can implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. The processor can also be a combination of functions that implement computing capabilities, such as a combination of one or more microprocessors, a combination of digital signal processing (DSP) and microprocessors, etc., and the storage module can be a memory.

[0130] Furthermore, the control device provided in the embodiments of this application may specifically be a chip, component, or module. The chip may include a connected processor and a memory. The memory stores instructions, and when the processor calls and executes the instructions, the chip can execute the recommendation method based on multi-dimensional feature profile weight calculation provided in the above embodiments. The embodiments of this application also provide a computer-readable storage medium storing computer program code. When the computer program code is run on a computer, it causes the computer to execute the aforementioned method steps to implement the recommendation method based on multi-dimensional feature profile weight calculation provided in the above embodiments.

[0131] This application also provides a computer program product. When the computer program product is run on a computer, it causes the computer to perform the above-mentioned related steps to implement the recommendation method based on multi-dimensional feature profile weight calculation provided in the above embodiments. The control device, computer-readable storage medium, computer program product, or chip provided in this application embodiment are all used to execute the corresponding methods provided above. Therefore, the beneficial effects they achieve can be referred to in the beneficial effects of the corresponding methods provided above, and will not be repeated here. Through the description of the above embodiments, those skilled in the art can understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the control device can be divided into different functional modules to complete all or part of the functions described above. In the embodiments provided in this application, it should be understood that the disclosed control device and method can be implemented in other ways. For example, the control device embodiments described above are merely illustrative. For example, the division of modules or units is merely a logical functional division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another control device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, control device or unit, and can be electrical, mechanical or other forms.

[0132] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multiple task processing and parallel processing are possible or may be advantageous. The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. The above content is only a specific implementation of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application.

Claims

1. A recommendation method based on weight calculation of multi-dimensional feature profiles, characterized in that, The method includes: The demand analysis agent extracts task features, data features, user features, and environmental features from the multi-dimensional feature profile based on a preset prompt template from unstructured user natural language descriptions. A data parsing agent is used to match structured raw data with a preset data template to obtain a target data template, wherein the preset data template is set based on the weight calculation method in the weight method knowledge base; the raw data is analyzed based on the target data template, and the analysis results are added to the data features; The method recommends that the agent match each weight calculation method in the weight method knowledge base with the multi-dimensional feature profile to obtain at least one target weight calculation method for calculating weights.

2. The method as described in claim 1, characterized in that, The preset prompt template includes: recognizing language descriptions related to the task features, data features, user features, and environmental features, and outputting the multidimensional feature vector of structured representation.

3. The method as described in claim 1, characterized in that, The task characteristics include the evaluation domain, evaluation purpose, and problem complexity; The data characteristics include the number of indicators, data type, data quality, and distribution characteristics; The user characteristics include the user's professional level, methodological preferences, and accuracy requirements; The environmental characteristics include computing power constraints, time constraints, and expert resource availability.

4. The method as described in claim 1, characterized in that, The process of using a data parsing agent to match structured raw data with a preset data template to obtain a target data template includes: The data parsing agent reads the header information of the original data; The header information is compared with the header information pre-stored in the standard template library to determine the target data template corresponding to the original data.

5. The method as described in claim 1, characterized in that, The process of analyzing the original data based on the target data template and adding the analysis results to the data features includes: The data parsing agent reads the metadata of the original data; Determine the matching degree between the metadata and the target data template; If the matching degree is determined to be greater than or equal to the matching degree threshold, the metadata is analyzed based on the target data template, and the analysis results are added to the data features.

6. The method as described in claim 5, characterized in that, The method further includes: If the matching degree is determined to be less than the matching degree threshold, the corresponding metadata is identified as abnormal metadata. Output a prompt message to indicate that the abnormal metadata should be modified to match the target data template.

7. The method according to any one of claims 1 to 6, characterized in that, The method recommends that the agent match the multi-dimensional feature profile with each weight calculation method in the weight method knowledge base to obtain at least one target weight calculation method for calculating weights, including: The statistical applicability rules integrated in the method recommendation agent are matched with each weight calculation method in the pre-set weight method knowledge base based on the problem profile to obtain the rule matching score corresponding to each weight calculation method. Based on sorting the rule matching scores as the total matching scores, the at least one target weight calculation method for calculating weights is obtained.

8. The method as described in claim 7, characterized in that, The method for calculating at least one target weight for weight calculation, which sorts the results based on the rule matching scores as the total matching scores, includes: The semantic similarity between each weighting method and the multi-dimensional profile is analyzed using a large language model to obtain the semantic matching score of each weighting calculation method. The rule matching score and the semantic matching score are weighted and summed to obtain the total matching score for each weight calculation method; Based on the total matching score, the rows are sorted to obtain at least one target weight calculation method for calculating weights.

9. A statistical weighted recommendation system based on multi-dimensional feature profiling, characterized in that, The system includes a demand analysis agent, a data parsing agent, and a method recommendation agent, wherein... The demand analysis agent is used to extract task features, data features, user features, and environmental features from the unstructured user natural language description based on a preset prompt template; The data parsing agent is used to match structured raw data with a preset data template to obtain a target data template, wherein the preset data template is set based on the weight calculation method in the weight method knowledge base; based on the target data template, the analysis results are added to the data features. The method recommends an intelligent agent, which is used to match each weight calculation method in the weight method knowledge base based on the multi-dimensional feature profile to obtain at least one target weight calculation method for calculating weights.

10. A statistical weighted recommendation device based on multi-dimensional feature profiling, characterized in that, The device includes: The feature extraction module is used to extract task features, data features, user features, and environmental features from the multi-dimensional feature profile based on the unstructured user natural language description using a demand analysis agent and a preset prompt template. The data analysis module is used to match structured raw data with a preset data template using a data parsing intelligent agent to obtain a target data template. The preset data template is set based on the weight calculation method in the weight method knowledge base. Based on the analysis results of the target data template, the analysis results are added to the data features. The method matching module is used to match the method recommendation agent with each weight calculation method in the weight method knowledge base based on the multi-dimensional feature profile to obtain at least one target weight calculation method for calculating weights.

Images