A method and system for adaptive portrait generation based on hybrid factor full-dimensional fusion

Abstract

This invention relates to the field of portrait generation technology, specifically to a method and system for adaptive portrait generation based on multi-dimensional fusion of hybrid factors. To address the problems of large fitting deviations and low scene adaptability in traditional portrait generation, this invention first decomposes the target user information into multi-dimensional α-sensory factors and... β The rational factor is used to extract overlapping feature points through dual computation and assign initial weights. Then, the initial weight factors are classified and core factors are selected. Through bidirectional convergence verification of dual computation fusion and natural index fitting, the optimal weight factors are obtained. Finally, the optimal weight factors are weighted with dimensional information to obtain multi-dimensional features. The core factors are matched to generate personality profiles. Then, the profiles are graded and output through multi-dimensional aggregation. The generated standardized graded profiles not only fit the real characteristics of the target users, but also accurately match the needs of different application scenarios, which greatly improves the practical application value and adaptability of the personality profile generation method.

Description

Technical Field

[0001] This invention relates to the field of portrait generation technology, specifically to a method and system for adaptive portrait generation based on the full-dimensional fusion of hybrid factors. Background Technology

[0002] Accurate user profiles enable quantitative analysis of multi-dimensional characteristics of target users, providing a scientific basis for scenario-based decision-making. Their application value is increasingly prominent in areas such as interpersonal interaction, business competition, and organizational collaboration, and the requirements for the accuracy, scenario adaptability, and full-dimensional feature coverage of profiles are also increasing.

[0003] Current methods for generating character profiles mostly use a single calculation mode to achieve feature fitting, which is prone to fitting bias. Furthermore, the weight factors in the calculation process lack a precise convergence verification mechanism and rely solely on fixed thresholds, which lack dynamism and accuracy. This makes it difficult to output high-quality weight factors that fit the essence of the dimensions, resulting in low adaptability of the matching results to actual application scenarios. Consequently, it is impossible to further fit the corresponding standard application scenarios, thus limiting its practicality and applicability. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for generating adaptive portraits based on the full-dimensional fusion of hybrid factors.

[0005] The technical solution of this invention is as follows: A method for generating adaptive person profiles based on the full-dimensional fusion of hybrid factors includes the following operations: S1. Obtain target user information and structure it into multiple dimensions, including environment, person, and time. From each dimension, initially extract α-sensory factors and β-rational factors to form a basic factor set. Perform parallel operations of the first and second calculations on the basic factor set to extract the cross-overlapping feature points of the double calculation results. Based on the preset assignment rules, assign values ​​to the overlapping feature points according to the degree of feature overlap as the initial weight factors for the corresponding dimension information. S2. For the initial weight factors of each dimension of information, classify them according to environment, person, and time tags. Perform a second calculation on the initial weight factors under the same tag. In the calculation results, retain the weight factors under each tag whose weight is greater than the weight threshold as the core weight factors. Perform a global calculation on the core weight factors of all dimensions based on the first calculation. Using the natural exponential function as the fitting benchmark, and based on the approximation deviation term, use the results of the first and second calculations as two independent convergence curves for bidirectional convergence verification. When the deviation term value is less than the convergence threshold, take the average of the two convergence curves as the final fitted value of the core weight factor, perform weight correction, and obtain the optimal weight factors for each dimension of information. S3. Weight the optimal weight factor with the corresponding dimensional information to obtain the multi-dimensional features of the target user; based on the multi-dimensional features of the target user, perform core factor matching in the preset personality database to obtain the personality profile of the target user; aggregate the multi-dimensional features of the target user in multiple dimensions to obtain the final value of the multi-dimensional aggregation, match the output level of the personality profile with the corresponding standard application scenario, and obtain the standardized graded profile of the target user.

[0006] In S1, a first calculation is performed on the α-sensory factor, which is a non-linear calculation using Transform; a second calculation is performed on the β-rational factor, which is a linear calculation using an RNN network.

[0007] The adaptive portrait generation method based on hybrid factor full-dimensional fusion according to claim 2 is characterized in that, for each dimension of information, it is determined whether the α emotional factor is more than the β rational factor; if the α emotional factor is more than the β rational factor, the first calculation ratio is increased to the first update ratio; if the α emotional factor is less than the β rational factor, the first calculation ratio is decreased to the second update ratio; if the α emotional factor is equal to the β rational factor, the initial ratio remains unchanged.

[0008] In S1, the method for extracting cross-overlapping feature points is as follows: obtain the intersection points in the parallel operation of the first calculation and the second calculation, calculate the feature overlap of the intersection points based on feature similarity, semantic relevance, and dimension matching degree, and take the intersection points with feature overlap greater than the feature overlap threshold as cross-overlapping feature points.

[0009] In S2, the approximation deviation term is obtained through the following calculation formula: , For the first Dimension The approximation deviation term in the iteration. For the first The basic weights of the dimensions For the first Dimension 1 The result of the global feature fitting calculated in the first iteration. For the first Dimension 1 The derivation result of the second calculation in the next iteration. This is the absolute difference between the results of the first and second calculations.

[0010] The core factor matching operation in S3 is as follows: Extract core factors from the multidimensional features of the target user, and obtain the proportion of each core factor in all typical personality models in the preset personality database; based on the proportion of each factor in different personality models, calculate the information entropy and factor discrimination weight of each core factor; using the core factor discrimination weight as a weighting coefficient, perform weighted K-means clustering on the core factor matrix of the preset personality database to obtain the cluster centers of each personality model; obtain the core factor set of the target user and obtain the weighted Euclidean distance of each cluster center; based on the weighted Euclidean distance, obtain the dynamic anchor threshold specific to the target user; based on the core factor set of the target user, the cluster centers, and the factor discrimination weight, obtain the initial factor matching degree; combined with the preset scene adaptation matrix, perform scene adaptation correction on the initial factor matching degree to obtain the scene-corrected matching degree; traverse all scene-corrected matching degrees corresponding to all cluster centers, retain the corresponding personality models whose scene-corrected matching degrees are greater than the dynamic anchor threshold, and sort them in descending order according to the scene-corrected matching degrees, selecting the model ranked first as the optimal matching personality model for the target user, and the information corresponding to the model is used as the target user's personality profile.

[0011] In S3, multi-dimensional aggregation is achieved through weighted summation.

[0012] A hybrid factor-based full-dimensional fusion adaptive portrait generation system, used to implement the above-mentioned hybrid factor-based full-dimensional fusion adaptive portrait generation method, includes: The initial weight factor generation module is used to acquire target user information and structure it into multiple dimensions, including environmental, personal, and time dimensions. It initially extracts α-sensory factors and β-rational factors from each dimension to form a basic factor set. The basic factor set is then subjected to parallel computation of a first calculation and a second calculation to extract the overlapping feature points of the two calculation results. Based on preset assignment rules, the overlapping feature points are assigned values ​​according to their degree of overlap, serving as the initial weight factors for the corresponding dimension information. The optimal weight factor generation module is used to classify the initial weight factors of each dimension of information according to environment, person, and time tags. It performs a second calculation on the initial weight factors under the same tag and retains the weight factors under each tag with weights greater than the weight threshold as core weight factors. It then performs a global calculation on the core weight factors of all dimensions based on the first calculation. Using the natural exponential function as the fitting benchmark and based on the approximation deviation term, it uses the results of the first and second calculations as two independent convergence curves for bidirectional convergence verification. When the deviation term value is less than the convergence threshold, it takes the average of the two convergence curves as the final fitted value of the core weight factor and performs weight correction to obtain the optimal weight factors for each dimension of information. The user profile and standardized graded profile generation module is used to weight the best weight factors with the corresponding dimensional information to obtain the multi-dimensional features of the target user; based on the multi-dimensional features of the target user, core factors are matched in the preset user profile library to obtain the user profile; the multi-dimensional features of the target user are aggregated in multiple dimensions to obtain the final value of the multi-dimensional aggregation, which is matched with the user profile output level and corresponding standard application scenarios to obtain the standardized graded profile of the target user.

[0013] A device for generating adaptive portraits based on the full-dimensional fusion of hybrid factors includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the aforementioned method for generating adaptive portraits based on the full-dimensional fusion of hybrid factors.

[0014] A computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described adaptive portrait generation method based on hybrid factor full-dimensional fusion.

[0015] The beneficial effects of this invention are as follows: This invention provides a method for generating adaptive user profiles based on multi-dimensional fusion of hybrid factors. First, the target user information is decomposed into multi-dimensional and α / β dual-factor structure. Real overlapping feature points are extracted through parallel dual-calculation and assigned initial weights, avoiding information mixing and pseudo-feature interference from the source, ensuring that the initial weight factors accurately match the essential features of each dimension. Then, the initial weight factors are categorized by tags to select core factors. Combining the refined derivation of the second calculation with the global fusion of the first calculation, the weights are corrected through bidirectional convergence verification using natural index fitting, thus eliminating low-weight and ineffective features. The effectiveness factor ensures the validity of the calculation, while also taking into account the rational verification of the weight factors and the correlation of all-dimensional features. The final optimal weight factor is accurate, comprehensive and traceable. Finally, the optimal weight factor is weighted with dimensional information to obtain multi-dimensional features. The core factor is matched to generate a personality profile, and then the profile is output in a hierarchical manner through multi-dimensional aggregation. This completes the closed loop from factor weight to practical personality profile, so that the generated standardized hierarchical profile not only fits the real characteristics of the target user, but also accurately matches the needs of different application scenarios, which greatly improves the practical application value and adaptability of the personality profile generation method. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the exemplary embodiments of this application clearer, the technical solutions in the exemplary embodiments of this application are described clearly and completely below. Obviously, the described exemplary embodiments are only some embodiments of this application, and not all embodiments.

[0017] This embodiment provides a method for generating adaptive person profiles based on the full-dimensional fusion of hybrid factors, including the following operations: S1. Obtain target user information and structure it into multiple dimensions, including environment, person, and time. From each dimension, initially extract α-sensory factors and β-rational factors to form a basic factor set. Perform parallel operations of the first and second calculations on the basic factor set to extract the cross-overlapping feature points of the double calculation results. Based on the preset assignment rules, assign values ​​to the overlapping feature points according to the degree of feature overlap as the initial weight factors for the corresponding dimension information. S2. For the initial weight factors of each dimension of information, classify them according to environment, person, and time tags. Perform a second calculation on the initial weight factors under the same tag. In the calculation results, retain the weight factors under each tag whose weight is greater than the weight threshold as the core weight factors. Perform a global calculation on the core weight factors of all dimensions based on the first calculation. Using the natural exponential function as the fitting benchmark, and based on the approximation deviation term, use the results of the first and second calculations as two independent convergence curves for bidirectional convergence verification. When the deviation term value is less than the convergence threshold, take the average of the two convergence curves as the final fitted value of the core weight factor, perform weight correction, and obtain the optimal weight factors for each dimension of information. S3. Weight the optimal weight factor with the corresponding dimensional information to obtain the multi-dimensional features of the target user; based on the multi-dimensional features of the target user, perform core factor matching in the preset personality database to obtain the personality profile of the target user; aggregate the multi-dimensional features of the target user in multiple dimensions to obtain the final value of the multi-dimensional aggregation, match the output level of the personality profile with the corresponding standard application scenario, and obtain the standardized graded profile of the target user.

[0018] S1. Obtain target user information and structure it into multiple dimensions, including environmental, personal, and time dimensions. Initially extract α-sensory factors and β-rational factors from each dimension to form a basic factor set. Perform parallel calculations (first and second calculations) on the basic factor set to extract the overlapping feature points of the double calculation results. Based on preset assignment rules, assign values ​​to the overlapping feature points according to the degree of feature overlap, as the initial weight factors for the corresponding dimension information.

[0019] First, multi-source heterogeneous information about the target user is collected, including visual modal information, such as micro-expressions, body movements, and scene images; voice modal information, such as tone of voice, speech rate, and communication skills; text modal information, such as social expressions, document content, and behavioral records; and numerical modal information, such as objective resumes, behavioral data, and time nodes, covering user-related data in various scenarios to form target user information.

[0020] Then, the target user information is structurally broken down into multiple dimensions, including environmental, personnel, and time dimensions, ensuring that each type of information belongs to a single core dimension and avoiding cross-dimensional information mixing. Specifically, the environmental dimension includes information such as the target user's context, associated resources, and external relationships; the personnel dimension includes information such as the target user's resume, behavioral characteristics, emotional expression, and language logic; and the time dimension includes information such as the target user's behavioral timeline, characteristic change cycles, and key milestones.

[0021] Next, α-sensory factors and β-rational factors are extracted from the information of each dimension to form a basic factor set. α-sensory factors are subjective information, such as unstructured features like micro-expressions, emotional details, and subtext in language; β-rational factors are objective information, such as structured features like objective resumes, organizational positioning, behavioral results, and time nodes.

[0022] Finally, the basic factor set is subjected to parallel computation of the first and second calculations, and the overlapping feature points of the two calculation results are extracted. The overlapping feature points are assigned the initial weight factors of the corresponding dimension information according to the feature overlap degree.

[0023] The first calculation is performed on the α-sensitivity factor. This first calculation is a non-linear calculation using Transform, which enables global feature extraction, flexible fitting, and fast convergence. The second calculation is performed on the β-rationality factor. This second calculation is a linear calculation using an RNN network, which provides stable fitting and resistance to bias.

[0024] Furthermore, for each dimension of information, determine whether the α-sensory factor is greater than the β-rational factor.

[0025] If the α-sensory factor is greater than the β-rational factor, the first calculation proportion is increased to the first update proportion to enhance the global fitting ability. The calculation formula is as follows: , , The percentage of the first update. The initial proportion is calculated first. , The first Dimension α Number of emotional factors β Number of rational factors To adjust the coefficient upwards, This is the updated second calculation percentage.

[0026] like α Less emotional factors β The rational factor, by reducing the proportion of the first calculation to the proportion of the second update, is equivalent to increasing the proportion of the second calculation, thus enhancing the accuracy of step verification. The calculation formula is as follows: , .

[0027] like α Sensitivity factor equals β The rational factor remains unchanged from its initial proportion: . This is the initial percentage for the second calculation.

[0028] In the parallel computation of the first computation (nonlinear computation) and the second computation (linear computation) mentioned above, the overlapping feature points of the two computation results are determined by feature similarity, semantic relevance, and dimensional matching. The overlapping feature points are the core source of the weighting factors.

[0029] Specifically, the process involves obtaining the intersection points in the parallel operations of the first and second calculations, calculating the feature overlap of the intersection points based on feature similarity, semantic relevance, and dimensional matching, and taking intersection points with feature overlap greater than the feature overlap threshold as overlapping feature points. This approach overcomes the limitation of relying solely on numerical similarity and effectively avoids pseudo-intersection points that are numerically close but semantically or dimensionally unrelated, ensuring that the intersection points represent the true common features of the two calculation results. Then, based on preset assignment rules, the overlapping feature points are assigned values ​​according to the feature overlap, serving as the initial weighting factors for the corresponding dimensional information.

[0030] Feature overlap is calculated as a weighted sum of feature similarity, semantic relevance, and dimensionality matching, using the following formula: , , Intersection k Feature overlap, K The total number of intersections. Intersection k The degree of overlap of comprehensive features , , Intersection points k Semantic relevance, feature similarity, and dimensional matching degree; , , These are the semantic relevance weight, feature similarity weight, and dimension matching weight, respectively.

[0031] The formula for calculating semantic relevance is as follows: , For the first i The α feature and the first j The semantic relevance of each β feature For the first iThe α feature and the first j When the original semantic similarity of each β feature in the SimBERT model output is less than 0.3... Assign 0 directly; the SimBERT model is an existing technology, and will not be described again here to save space.

[0032] The formula for calculating feature similarity is as follows: , For the first i The α feature and the first j The similarity of the β features, For the first d Dimension 1 t In the nth iteration i The normalized values ​​of the α features, For the first d Dimension 1 t In the nth iteration j indivual β Normalized value of the feature.

[0033] The dimension matching degree is obtained as follows: when the α and β feature dimension labels are the same, the dimension matching degree is 1; when the α and β feature dimension labels are different, the dimension matching degree is 0.

[0034] In this embodiment, S1 decomposes the target user information into multiple dimensions in a structured manner, and splits the α and β dual-factor adaptive feature attributes. The dual-calculation parallel operation fits the characteristics of the factors to achieve differentiated feature extraction. At the same time, the calculation ratio is dynamically adjusted based on the number of factors, and the true cross-overlapping feature points are extracted by combining multi-dimensional indicators and assigned initial weights. This avoids information mixing and pseudo-feature interference, and makes the initial weight factors fit the essential features of the dimensions, laying a precise and high-quality foundation for subsequent weight optimization.

[0035] S2. For the initial weight factors of each dimension of information, classify them according to environment, person, and time tags. Perform a second calculation on the initial weight factors under the same tag. In the calculation results, retain the weight factors under each tag whose weights are greater than the weight threshold as core weight factors. Perform a global calculation on the core weight factors of all dimensions based on the first calculation. Using the natural exponential function as the fitting benchmark, and based on the approximation deviation term, use the results of the first and second calculations as two independent convergence curves for bidirectional convergence verification. When the deviation term value is less than the convergence threshold, take the average of the two convergence curves as the final fitted value of the core weight factor, perform weight correction, and obtain the optimal weight factors for each dimension of information.

[0036] First, the initial weight factors for each dimension of information are categorized by environment tag, person tag, and time tag to ensure that the initial weight factors under the same tag belong to the same dimension of information. For example, the environment tag only contains the initial weight factors of the environment dimension, thus avoiding the mixing of factors across tags.

[0037] Then, for the initial weight factors under the same label, a second calculation (linear calculation) is performed separately. In the calculation results, weight factors with weights greater than a weight threshold under each label are retained as core weight factors. Low-weight, unrelated factors with weights not greater than the weight threshold under the same label are removed. This ensures the rational verification of the core weight factors, avoids invalid fitting, and provides high-quality input for subsequent global restoration. Simultaneously, the derivation results of the core weight factors from the second calculation are retained. .

[0038] Next, the core weight factors of all dimensions are subjected to global calculation based on the first calculation (non-linear calculation), and the global feature fitting result is output. By integrating the core factors of environment, people, and time, it makes up for the one-sidedness of refined deduction based on a single label, realizes the organic connection of the three core factors, constructs a full-dimensional feature association network of the target user, and restores the complete picture of the target user's characteristics.

[0039] In this process, the natural exponential function is used as the fitting benchmark, based on the modified approximation deviation term. The results of the first and second calculations ( , As two independent convergence curves, bidirectional convergence verification is performed to ensure that the two curves smoothly approach the main function branch without oscillation or overshoot. When the deviation term value is less than the convergence threshold, the average value of the two convergence curves is taken as the final fitted value of the core weight factor, and the weight is corrected to obtain the optimal weight factor for each dimension of information.

[0040] The bidirectional convergence verification process is as follows: based on the results of the first and second calculations in the current iteration ( , ), basic weights for each dimension Calculate the deviation term value Set a convergence threshold ,like The results of the two calculations have been determined to have converged smoothly, and the difference between the two convergence curves has been controlled within an acceptable range; if Based on right , Adaptive correction is performed, and the correction formula is as follows: , , , The first Dimension 1 In the nth iteration, the th The global feature fitting result of the first calculation (nonlinear calculation) under +1 iterations. For the first Dimension 1 In the nth iteration, the th The derivation of the second calculation (linear calculation) under +1 iterations, For the first Dimension The approximation deviation term in the iteration. For the first The base weights of the dimensions, and the number of iterations after correction. Automatically increment, re-execute the global derivation of the first calculation and the fine-grained verification of the second calculation until the convergence condition is met.

[0041] Approximation deviation term The dynamic measure of the difference between the results of the first calculation (nonlinear calculation) and the second calculation (linear calculation) is calculated using the following formula: , For the first Dimension The approximation deviation term in the iteration has a range of values ​​of . As the number of iterations, dimensional weights, and differences between the two calculation results increase, the value gradually approaches 0. For the first The basic weight of a dimension, such as 0.5 for the character dimension, 0.3 for the environment dimension, and 0.2 for the time dimension, means that the higher the weight of the dimension, the faster the deviation correction speed. For the first Dimension 1 The global feature fitting result of the first calculation (nonlinear calculation) in the next iteration. For the first Dimension 1 The derivation of the second calculation (linear calculation) in the next iteration. The absolute difference between the first and second calculation results represents the degree of dispersion of the two convergence curves; the larger the difference, the more dispersed the convergence curves. The smaller the value, the stronger the correction.

[0042] In this embodiment, S2 filters core weight factors through label classification and restores the full picture of features through dual calculation (α nonlinear globalization and β linear refinement). Dynamic and stable convergence is achieved by using an approximation deviation term that fits the difference between the two calculations. It avoids single calculation deviation and invalid fitting through threshold screening and bidirectional verification, and improves efficiency through dimensional difference convergence. The final output of the best weight factor has accuracy, comprehensiveness and traceability.

[0043] S3. Weight the optimal weight factor with the corresponding dimensional information to obtain the multi-dimensional features of the target user; based on the multi-dimensional features of the target user, perform core factor matching in the preset personality database to obtain the personality profile of the target user; aggregate the multi-dimensional features of the target user in multiple dimensions to obtain the final value of the multi-dimensional aggregation, match the output level of the personality profile with the corresponding standard application scenario, and obtain the standardized graded profile of the target user.

[0044] First, the optimal weighting factor is weighted with the corresponding dimensional information to obtain the multidimensional features of the target user.

[0045] Then, based on the multidimensional characteristics of the target user, core factor matching is performed in the preset personality database to obtain the personality profile of the target user. The specific steps of core factor matching are as follows.

[0046] Step 1: Extract the core factors from the multidimensional features of the target user. The types of core factors can be preset according to the actual scenario requirements. Obtain the proportion of each core factor in all typical personality models in the preset personality database, and clarify the distribution of each factor in different personality models. Based on the proportion of each factor in different personality models, calculate the information entropy of each core factor. The lower the entropy value, the stronger the factor's ability to distinguish between different personality models. Based on the information entropy of each core factor, calculate its respective factor discrimination weight. Core factors with strong discrimination receive higher weights and participate in the matching calculation first.

[0047] The formula for calculating information entropy is as follows: , , For the first k Information entropy of each core factor For the first k The core factor in the first i The proportion of each personality model m To preset the total number of personality models in the personality database, For the first character in the preset character personality database i In the first personality model k Normalized values ​​of the core factors When, define To avoid calculation errors.

[0048] The formula for calculating the factor discrimination weight: , For the first k The discrimination weights of the core factors K The total number of core factors, The factor discrimination index is used; the higher the discrimination index, the greater the weight ratio.

[0049] Step 2: Using the core factor discrimination weight as the weighting coefficient, perform weighted K-means clustering on the core factor matrix of the preset personality database to obtain the cluster centers of each personality model; obtain the core factor set of the target user (formed by all core factors of the target user), and obtain the weighted Euclidean distance of each cluster center to measure the degree of fit between the user factors and each personality model category; based on the weighted Euclidean distance, obtain a dynamic anchor point threshold exclusive to the target user to avoid the problem of high-matching models being missed due to setting the fixed threshold too high, and low-matching models being misselected due to setting it too low. The smaller the weighted Euclidean distance, the lower the threshold, and the more lenient the matching threshold, so as to realize personalized judgment with different matching benchmarks for different user factor distributions.

[0050] The formula for calculating the weighted Euclidean distance is as follows: , Set the core factors for the target user to the first g The weighted Euclidean distance of cluster centers. For the first g The first cluster center k One core factor, For the first k The discrimination weights of the core factors K This represents the total number of core factors.

[0051] The formula for calculating the dynamic anchor point threshold is as follows: , The dynamic anchor threshold for the k-th core factor. The minimum distance from the target user to all cluster centers. G This represents the total number of clusters. The average distance from the core factor set of the target user to all cluster centers.

[0052] Step 3: Based on the target user's core factor set, cluster center, and factor discrimination weight, obtain the initial factor matching degree; combined with the preset scenario adaptation matrix, which is composed of factor correction coefficients for different application scenarios, the initial factor matching degree is modified for scenario adaptation. The scenario correction coefficients are used to strengthen the weight of key factors in the current scenario and weaken the influence of secondary factors to obtain the scenario-corrected matching degree.

[0053] The formula for calculating the initial factor matching degree is as follows: , For the core factor set of target users and the first g Initial matching degree of the personality model.

[0054] Scene adaptation correction is achieved through the following formula: , The matching degree is adjusted according to the specific context. The scene adaptability matrix represents the current scene. s Next k Correction coefficients for each core factor.

[0055] Step 4: Iterate through the context-corrected matching scores of all cluster centers, retaining personality models with a context-corrected matching score greater than the dynamic anchor threshold. Sort these models in descending order of context-corrected matching score, selecting the top-ranked model as the optimal matching personality model for the target user. The information corresponding to this model is used as the target user's personality profile. If the context-corrected matching score of all personality models is less than or equal to the dynamic anchor threshold, a fallback rule is triggered: the top 3 matching scores are merged to generate a hybrid personality model. This avoids invalid scenarios with no matching results and ensures the integrity of the technical solution.

[0056] At the same time, the multi-dimensional features of the target user can be aggregated in multiple dimensions, which can be achieved by weighted summation to obtain the final value of the multi-dimensional aggregation.

[0057] Finally, based on the final value aggregated from multiple dimensions, the user profile output level and corresponding standard application scenarios are matched to obtain a standardized hierarchical profile of the target user.

[0058] The image output level and corresponding standard application scenario classification standards are as follows: When At that time, a high-confidence, full-dimensional profile is generated, corresponding to standard application scenarios such as high-pressure games, business negotiations, and core interactions within the system; when At that time, a medium-confidence profile corresponds to standard application scenarios such as routine business interactions and daily interpersonal analysis; when At that time, the basic feature profile corresponds to the standard application scenario of preliminary character investigation and information collection; when If no valid matching image is found, the calculation is retried.

[0059] In this embodiment, S3 uses the core rule of weight × factor to obtain multi-dimensional features, ensuring the consistency and accuracy of feature quantification. At the same time, it adopts an innovative core factor matching method with entropy weighting, dynamic anchor threshold, and scenario adaptation correction, so that the personality profile not only matches the user's real characteristics but also adapts to actual application scenarios. Furthermore, it accurately matches the needs of different application scenarios through standardized hierarchical profiles, completing the closed loop from factor weights to practical personality profiles and enhancing the practical application value of the method.

[0060] This embodiment also provides a hybrid factor full-dimensional fusion adaptive portrait generation system to implement the above-mentioned hybrid factor full-dimensional fusion adaptive portrait generation method, including: The initial weight factor generation module is used to acquire target user information and structure it into multiple dimensions, including environmental, personal, and time dimensions. It initially extracts α-sensory factors and β-rational factors from each dimension to form a basic factor set. The basic factor set is then subjected to parallel computation of a first calculation and a second calculation to extract the overlapping feature points of the two calculation results. Based on preset assignment rules, the overlapping feature points are assigned values ​​according to their degree of overlap, serving as the initial weight factors for the corresponding dimension information. The optimal weight factor generation module is used to classify the initial weight factors of each dimension of information according to environment, person, and time tags. It performs a second calculation on the initial weight factors under the same tag and retains the weight factors under each tag with weights greater than the weight threshold as core weight factors. It then performs a global calculation on the core weight factors of all dimensions based on the first calculation. Using the natural exponential function as the fitting benchmark and based on the approximation deviation term, it uses the results of the first and second calculations as two independent convergence curves for bidirectional convergence verification. When the deviation term value is less than the convergence threshold, it takes the average of the two convergence curves as the final fitted value of the core weight factor and performs weight correction to obtain the optimal weight factors for each dimension of information. The user profile and standardized graded profile generation module is used to weight the best weight factors with the corresponding dimensional information to obtain the multi-dimensional features of the target user; based on the multi-dimensional features of the target user, core factors are matched in the preset user profile library to obtain the user profile; the multi-dimensional features of the target user are aggregated in multiple dimensions to obtain the final value of the multi-dimensional aggregation, which is matched with the user profile output level and corresponding standard application scenarios to obtain the standardized graded profile of the target user.

[0061] This embodiment also provides a hybrid factor full-dimensional fusion adaptive portrait generation device, including a processor and a memory, wherein the processor executes the computer program stored in the memory to implement the above-mentioned hybrid factor full-dimensional fusion adaptive portrait generation method.

[0062] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described adaptive portrait generation method based on hybrid factor full-dimensional fusion.

[0063] This embodiment provides a method for generating adaptive user profiles based on multi-dimensional fusion of hybrid factors. First, the target user information is decomposed into multi-dimensional and α / β dual-factor structure. Real overlapping feature points are extracted through parallel dual-calculation and assigned initial weights, avoiding information mixing and pseudo-feature interference from the source, ensuring that the initial weight factors accurately match the essential features of each dimension. Then, the initial weight factors are categorized by tags to select core factors. Combining the refined derivation of the second calculation with the global fusion of the first calculation, the weights are corrected through bidirectional convergence verification using natural exponent fitting, thus eliminating low-weight factors. The invalid factor ensures the validity of the calculation while also taking into account the rational verification of the weight factors and the correlation of all-dimensional features. The final optimal weight factor is accurate, comprehensive and traceable. Finally, the optimal weight factor is weighted with dimensional information to obtain multi-dimensional features. The core factor is matched to generate a personality profile, and then the profile is output in a hierarchical manner through multi-dimensional aggregation. This completes the closed loop from factor weight to practical personality profile. The generated standardized hierarchical profile not only fits the real characteristics of the target user, but also accurately matches the needs of different application scenarios, which greatly improves the practical application value and adaptability of the personality profile generation method.

[0064] While exemplary embodiments of the invention have been described herein, many other variations or modifications conforming to the principles of the invention can be directly determined or derived from the disclosure of this invention without departing from its spirit and scope. Therefore, the scope of the invention should be understood and recognized to cover all such other variations or modifications.

Claims

1. A method for generating adaptive portraits based on the full-dimensional fusion of hybrid factors, characterized in that, This includes the following operations: S1. Obtain target user information and structure it into multiple dimensions, including environmental, personal, and time dimensions; initially extract α-sensory factors and β-rational factors from each dimension to form a basic factor set; The basic factor set is subjected to parallel computation of the first and second calculations, and the overlapping feature points of the two calculation results are extracted. Based on the preset assignment rules, overlapping feature points are assigned values ​​according to the degree of feature overlap, which serve as the initial weighting factors for the corresponding dimension information. S2. For the initial weight factors of each dimension of information, classify them according to environment, people, and time tags. Perform a second calculation on the initial weight factors under the same tag. In the calculation results, retain the weight factors under each tag whose weight is greater than the weight threshold as the core weight factors. Perform a global calculation based on the first calculation for the core weight factors of all dimensions; Using the natural exponential function as the fitting benchmark, and based on the approximation deviation term, the results of the first and second calculations are used as two independent convergence curves for bidirectional convergence verification. When the deviation term value is less than the convergence threshold, the average value of the two convergence curves is taken as the final fitting value of the core weight factor, and the weight is corrected to obtain the optimal weight factor for each dimension of information. S3. Weight the optimal weight factor with the corresponding dimensional information to obtain the multidimensional features of the target user; Based on the multidimensional characteristics of the target user, core factors are matched in a pre-set personality database to obtain the personality profile of the target user. By aggregating the multidimensional features of the target user in multiple dimensions, the final value of the multidimensional aggregation is obtained. This value is then matched with the user profile output level and corresponding standard application scenarios to obtain a standardized hierarchical profile of the target user.

2. The adaptive portrait generation method based on full-dimensional fusion of hybrid factors as described in claim 1, characterized in that, In S1, a first calculation is performed on the α-sensory factor, which is a non-linear calculation using Transform; a second calculation is performed on the β-rational factor, which is a linear calculation using an RNN network.

3. The adaptive portrait generation method based on full-dimensional fusion of hybrid factors as described in claim 2, characterized in that, For each dimension of information, determine whether the α-sensory factor is greater than the β-rational factor; if the α-sensory factor is greater than the β-rational factor, increase the first calculated proportion to the first updated proportion; if the α-sensory factor is less than the β-rational factor, decrease the first calculated proportion to the second updated proportion; if the α-sensory factor is equal to the β-rational factor, keep the initial proportion unchanged.

4. The adaptive portrait generation method based on full-dimensional fusion of hybrid factors according to claim 1, characterized in that, In S1, the method for extracting cross-overlapping feature points is as follows: obtain the intersection points in the parallel operation of the first calculation and the second calculation, calculate the feature overlap of the intersection points based on feature similarity, semantic relevance, and dimension matching degree, and take the intersection points with feature overlap greater than the feature overlap threshold as cross-overlapping feature points.

5. The adaptive portrait generation method based on full-dimensional fusion of hybrid factors according to claim 1, characterized in that, In S2, the approximation deviation term is obtained through the following calculation formula: ， For the first Dimension The approximation deviation term in the iteration. For the first The basic weights of the dimensions For the first Dimension 1 The result of the global feature fitting calculated in the first iteration. For the first Dimension 1 The derivation result of the second calculation in the next iteration. This is the absolute difference between the results of the first and second calculations.

6. The adaptive portrait generation method based on full-dimensional fusion of hybrid factors according to claim 1, characterized in that, The core factor matching operation in S3 is as follows: Extract the core factors from the multidimensional features of the target user and obtain the proportion of each core factor in all typical personality models in the preset personality database; based on the proportion of each factor in different personality models, calculate the information entropy and factor discrimination weight of each core factor. Using the core factor discrimination weight as the weighting coefficient, weighted K-means clustering is performed on the core factor matrix of the preset personality database to obtain the cluster center of each personality model; Obtain the core factor set of the target user and get the weighted Euclidean distance of each cluster center; based on the weighted Euclidean distance, obtain the dynamic anchor point threshold specific to the target user; The initial factor matching degree is obtained based on the core factor set of the target user, cluster center, and factor discrimination weight; By combining the preset scene adaptation matrix, the initial matching degree of the factors is corrected for scene adaptation, and the matching degree after scene-adapted correction is obtained. Traverse all the context-corrected matching scores corresponding to the cluster centers, retain the corresponding personality models with context-corrected matching scores greater than the dynamic anchor threshold, sort them in descending order according to the context-corrected matching scores, select the model ranked first as the optimal matching personality model for the target user, and use the information corresponding to the model as the personality profile of the target user.

7. The adaptive portrait generation method based on full-dimensional fusion of hybrid factors according to claim 1, characterized in that, In S3, multi-dimensional aggregation is achieved through weighted summation.

8. A hybrid factor full-dimensional fusion adaptive portrait generation system, used to implement the hybrid factor full-dimensional fusion adaptive portrait generation method of claim 1, characterized in that, include: The initial weight factor generation module is used to obtain target user information and structure it into multiple dimensions, including environmental dimension, person dimension, and time dimension. The α-sensory factor and β-rational factor are initially extracted from information from various dimensions to form a basic factor set; The basic factor set is subjected to parallel computation of the first and second calculations, and the overlapping feature points of the two calculation results are extracted. Based on the preset assignment rules, overlapping feature points are assigned values ​​according to the degree of feature overlap, which serve as the initial weighting factors for the corresponding dimension information. The optimal weight factor generation module is used to classify the initial weight factors of each dimension of information according to environment, people, and time tags. It performs a second calculation on the initial weight factors under the same tag and retains the weight factors under each tag with a weight greater than the weight threshold in the calculation results as the core weight factors. Perform a global calculation based on the first calculation for the core weight factors of all dimensions; Using the natural exponential function as the fitting benchmark, and based on the approximation deviation term, the results of the first and second calculations are used as two independent convergence curves for bidirectional convergence verification. When the deviation term value is less than the convergence threshold, the average value of the two convergence curves is taken as the final fitting value of the core weight factor, and the weight is corrected to obtain the optimal weight factor for each dimension of information. The user profile and standardized hierarchical profile generation module is used to weight the best weight factors with the corresponding dimensional information to obtain the multi-dimensional characteristics of the target user. Based on the multidimensional characteristics of the target user, core factors are matched in a preset personality database to obtain the personality profile of the target user; the multidimensional characteristics of the target user are aggregated in multiple dimensions to obtain the final value of the multidimensional aggregation, and the output level of the personality profile and the corresponding standard application scenario are matched to obtain the standardized graded profile of the target user.

9. A device for generating adaptive portraits based on the full-dimensional fusion of hybrid factors, characterized in that, It includes a processor and a memory, wherein when the processor executes a computer program stored in the memory, it implements the adaptive portrait generation method based on full-dimensional fusion of hybrid factors as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the adaptive portrait generation method based on full-dimensional fusion of hybrid factors as described in any one of claims 1-7.

Images