Jewelry style prediction method and system based on user preference analysis

Abstract

The invention discloses a jewelry style prediction method and system for user preference analysis, and relates to the technical field of data analysis, and the method comprises the steps: collecting user jewelry interaction data, mining explicit features and implicit relationships, and constructing a preference feature matrix; decomposing the matrix and introducing a regularization item to train a prediction model; carrying out transfer learning through a dynamic self-attention layer in combination with the interaction scene to generate a style prediction model; according to the user recommendation task, adjusting the feature weight and generating a jewelry recommendation portrait; and through a platform retrieval engine matching resource, generating a recommendation result and visually displaying the recommendation result. The technical problem of insufficient user preference analysis and recommendation accuracy in the prior art is solved, and the technical effect of improving the user preference analysis accuracy and recommendation accuracy is achieved.

Description

Technical Field

[0001] The present application relates to the technical field of data analysis, and in particular to a jewelry style prediction method and system based on user preference analysis. Background Art

[0002] With the diversification and personalization of user needs, traditional recommendation systems face many challenges in capturing user behavior patterns and preferences. Especially in the field of jewelry style recommendations, user choices are often affected by many factors, including daily usage scenarios, personal aesthetic tendencies, and consumption habits. These factors include both explicit behaviors (such as browsing and purchase records) and implicit potential preferences that are difficult to directly observe. In the existing technology, due to the scattered data sources and single model analysis, it is difficult for the recommendation results to fully cover user needs, and at the same time lack flexibility and adaptability in dealing with dynamic changes in scenarios and diversified needs. To address this problem, the industry urgently needs a more comprehensive and accurate data analysis and modeling method to fully explore user behavior characteristics and provide an analysis framework that better meets the needs.

[0003] At the current stage, related technologies have technical problems such as insufficient accuracy in user preference analysis and recommendation. Summary of the invention

[0004] The present application solves the technical problem of insufficient accuracy of user preference analysis and recommendation in the prior art by providing a jewelry style prediction method and system based on user preference analysis.

[0005] The present application provides a jewelry style prediction method based on user preference analysis, comprising: Collect and call multi-source interaction records based on user jewelry interaction behaviors, build a preference feature matrix by mining explicit features and implicit relationships; decompose the preference feature matrix, introduce regularization terms, and supervise the training of the large prediction model; introduce a dynamic self-attention layer based on the interaction scenario, perform transfer learning on the large prediction model, and generate a style prediction model; receive user recommendation tasks, combine the style prediction model, perform feature attention reset and push style decisions based on the task scenario, and generate a jewelry recommendation portrait; according to the jewelry recommendation portrait, perform platform resource retrieval based on the platform search engine, determine the jewelry recommendation list and perform interface visualization.

[0006] The present application provides a jewelry style prediction system based on user preference analysis, comprising: A preference feature matrix construction module, which is used to collect and call multi-source interaction records based on user jewelry interaction behaviors, and build a preference feature matrix by mining explicit features and implicit relationships; a large model supervision training module, which is used to decompose the preference feature matrix, introduce regularization terms, and supervise the training of the prediction large model; a style prediction model generation module, which is used to introduce a dynamic self-attention layer based on an interaction scenario, perform transfer learning on the prediction large model, and generate a style prediction model; a jewelry recommendation portrait generation module, which is used to receive user recommendation tasks, combine the style prediction model, perform feature attention reset and push style decisions based on task scenarios, and generate a jewelry recommendation portrait; an interface visualization module, which is used to perform platform resource retrieval based on the platform retrieval engine according to the jewelry recommendation portrait, determine the jewelry recommendation list, and perform interface visualization.

[0007] The jewelry style prediction method and system for user preference analysis proposed in this application first collects user jewelry interaction data, mines explicit features and implicit relationships to construct a preference feature matrix; decomposes the matrix and introduces regularization terms to train the prediction model; combines the interaction scenario with transfer learning through a dynamic self-attention layer to generate a style prediction model; adjusts feature weights according to user recommendation tasks and generates a jewelry recommendation portrait; matches resources through a platform retrieval engine, generates recommendation results and displays them visually, thereby achieving the technical effect of improving the accuracy of user preference analysis and recommendation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solution of the embodiment of the present disclosure, the accompanying drawings of the embodiment of the present disclosure will be briefly introduced below. A flow chart is used in the present application to illustrate the operations performed by the system according to the embodiment of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps can be processed in reverse order or simultaneously as needed. At the same time, other operations can also be added to these processes, or one or more operations can be removed from these processes.

[0009] Figure 1 A flowchart of a jewelry style prediction method based on user preference analysis provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of a jewelry style prediction system based on user preference analysis provided in an embodiment of the present application.

[0010] Explanation of the accompanying drawings: preference feature matrix construction module 10, large model supervision training module 20, style prediction model generation module 30, jewelry recommendation portrait generation module 40, interface visualization module 50. DETAILED DESCRIPTION

[0011] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0012] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0013] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict, and the terms "first\second" involved are merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "including" and "having" and any variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by technicians in the technical field of this application. The terms used herein are for the purpose of describing the embodiments of the present application only.

[0014] The present application embodiment provides a jewelry style prediction method based on user preference analysis, such as Figure 1 As shown, the method includes: Step S100, collect and call multi-source interaction records based on user jewelry interaction behaviors, and build a preference feature matrix by mining explicit features and implicit relationships. Specifically, the user's jewelry interaction behavior records are obtained through the multi-source data acquisition module, including browsing, collection, and purchase data on online platforms and trial wearing and purchase behaviors in offline physical stores. At the same time, the interaction data of different platforms are integrated to build a panoramic portrait of the user. Through explicit feature mining, the system analyzes the user's explicit behavior (such as preferred style, material, price range) and quantifies it into explicit vectors; through implicit relationship mining, the system combines retrieval behavior, behavior patterns, time scenes and other information to extract hidden interests and needs and generate implicit vectors. Finally, the explicit vectors and implicit vectors are integrated to build a preference feature matrix, which fully reflects the user's personalized preferences and provides core data support for subsequent modeling and prediction.

[0015] In a possible implementation, multi-source interaction records based on user jewelry interaction behaviors are collected and called, and a preference feature matrix is ​​constructed by mining explicit features and implicit relationships. Step S100 further includes step S110, traversing the interaction records to divide explicit interaction information and implicit interaction information, wherein the division is based on preference intuitiveness. Specifically, by traversing the multi-source interaction records of the user, the interaction information is divided into explicit interaction information and implicit interaction information according to the correlation and intuitiveness between the interaction information and the user's style preference. Explicit interaction information includes behaviors in which the user directly expresses style preferences, such as browsing, clicking, collecting or purchasing jewelry of a specific style. This information has a high degree of intuitiveness and can directly reflect the user's preferences. Implicit interaction information contains data that is weakly associated with the user's style preferences or has more obscure characteristics, such as search keywords, page dwell time, etc. Most of the implicit interaction information is redundant, but may contain potential preference characteristics. In the process of segmentation, preference intuitiveness is taken as the main criterion, and classification is carried out through data labeling method. Explicit information is directly stored, while implicit information is redundantly filtered and potential features are extracted, providing a concise and efficient data basis for subsequent user preference feature analysis.

[0016] Step S120, extract and quantify style features for the explicit interaction information to determine an explicit vector. Specifically, for the explicit interaction information, by analyzing the user's browsing, collection, purchase and other behaviors that directly reflect style preferences, extract key style features including color, material, design style, brand, purpose and other dimensions. Subsequently, the extracted features are quantified according to factors such as the frequency, intensity and duration of the interaction behavior, and a weight is assigned to each feature. The explicit interaction information is converted into multi-dimensional numerical data to generate an explicit vector. The explicit vector can intuitively present the user's explicit preferences and provide accurate basic data for the subsequent construction of the preference matrix and the training of the prediction model.

[0017] Step S130, for the implicit interaction information, the implicit relationship is mined and quantified to determine the implicit vector. Specifically, for the implicit interaction information, the implicit vector is extracted by mining the implicit relationship in the user behavior and quantifying it. First, the implicit interaction information is defined, such as retrieval behavior, browsing frequency, and dwell time, and the potential correlation relationship of user behavior is extracted by using pattern mining algorithms and time series analysis methods to construct an implicit relationship map. Subsequently, through weight allocation and matrix decomposition technology, the complex implicit relationship is mapped into a low-dimensional feature vector, and it is normalized to ensure the consistency of data distribution. Finally, the feature vector is integrated into an implicit vector as a key input for user style preference analysis.

[0018] Step S140, integrating the explicit vector and the implicit vector to construct the preference feature matrix. Specifically, by integrating the explicit vector and the implicit vector to construct the preference feature matrix, the user's intuitive preferences and potential interests are fully integrated. The explicit vector extracts the preference features clearly expressed by the user, such as the style and material of the jewelry clicked and purchased; the implicit vector mines the user's hidden behavior patterns, such as browsing time and retrieval habits. By dimensional alignment, weight assignment and weighted fusion of the vectors, the user's preference feature vector is generated and arranged into a preference feature matrix. Each column of the matrix represents a feature dimension, and each row corresponds to a user, which ultimately lays a data foundation for user style prediction and personalized recommendations.

[0019] Step S200, decompose the preference feature matrix, introduce regularization terms, and supervise the training of the prediction model. Specifically, decompose the preference feature matrix into a base matrix and a coefficient matrix, which represent the basic features and the user preference distribution, respectively. In the decomposition process, regularization terms are introduced to prevent the model from overfitting by constraining the matrix element values, and supervised learning is used to construct a prediction loss function using user historical behavior and style preference samples to optimize the matrix parameters. The model is continuously adjusted through iterative algorithms (such as gradient descent or alternating least squares) to make the prediction results more consistent with the actual data, and finally generate a style prediction model. The model can quickly predict the user's jewelry style preference, provide an accurate basis for personalized recommendations, and improve the generalization ability and stability of the model through validation set evaluation and parameter adjustment.

[0020] In a possible implementation, the preference feature matrix is ​​decomposed, and a regularization term is introduced to supervise the training of the large prediction model. Step S200 further includes step S210, traversing the preference feature matrix, decomposing and determining the base matrix and the coefficient matrix, wherein the product of the base matrix and the coefficient matrix is ​​the feature matrix. Specifically, the preference feature matrix is ​​decomposed into a base matrix and a coefficient matrix by traversal, wherein the base matrix represents a basic feature set, and the coefficient matrix reflects the strength of association between user preferences and basic features. During the decomposition process, B and C are optimized and adjusted based on a matrix decomposition algorithm (such as NMF or SVD) so that the product of the two can restore the original feature matrix as much as possible, ensuring the sparsity and interpretability of the results. The weights of the coefficient matrix are also used for dynamic attention adjustment, such as giving higher weights to important features in specific scenarios, thereby achieving accurate modeling and scene adaptation of user preferences, and providing data support for subsequent prediction models.

[0021] Step S220, based on the multi-source interaction records, determine the style preference samples. Specifically, by screening and cleaning the multi-source interaction records, extract the key data of the user's explicit and implicit behaviors such as browsing, purchasing, and collecting, and organize these data in chronological order and classify them into explicit features and implicit features. Explicit features directly mark the user's explicit preferences for materials, styles, etc., while implicit features infer possible interest tendencies through behavioral modeling. Combined with these feature information, a multi-dimensional preference vector is generated for the user, and a style preference sample set is constructed to ensure the integrity and accuracy of the samples, providing a high-quality data foundation for the subsequent training of the style prediction model.

[0022] Step S230, introduce the regularization term and construct the prediction loss function. Specifically, when constructing the prediction loss function, first calculate the difference between the model prediction output and the actual sample output through the mean square error (MSE) as the basic loss function to evaluate the model performance. At the same time, introduce the L2 regularization term to constrain the sum of squares of the model weight parameters to suppress excessive weight values ​​to improve the smoothness and generalization ability of the model. Finally, combine the basic loss function and the regularization term to form a complete prediction loss function: ,in is the regularization strength, is the overall loss function, Represents the first parameter (weight) of the model parameters, The basic loss function is optimized through cross-validation. During the model training process, the gradient descent algorithm is used to minimize the loss function to ensure the balance between the model prediction accuracy and complexity, and finally obtain a robust and efficient prediction model.

[0023] Step S240, according to the style preference samples and the prediction loss function, based on the base matrix and the coefficient matrix, perform prediction training based on time logic to generate the prediction large model. Specifically, by collecting user style preference samples, the data is grouped in chronological order to construct a time series data set. Based on the preference feature matrix, the base matrix and the coefficient matrix are obtained through matrix decomposition to capture the preference feature relationship. Introduce time logic, gradually update the base matrix and the coefficient matrix, so that the model dynamically adapts to the changing trend of user preferences. During the training process, the output results are evaluated in combination with the prediction loss function, where the basic loss measures the difference between the prediction and the actual, and the regularization loss suppresses the overfitting of the model. After multiple rounds of iterations, a large prediction model that can perceive time dynamics is generated to improve the prediction accuracy and adaptability of user style preferences.

[0024] Step S300, introduce a dynamic self-attention layer based on interactive scenes, perform transfer learning on the prediction model, and generate a style prediction model. Specifically, by introducing a dynamic self-attention layer based on interactive scenes, transfer learning is performed on the prediction model to generate a style prediction model. First, the usage scenarios (such as marriage, daily wear, formal occasions, etc.) are extracted according to the user interaction records, and the scene features are associated with the user preference features to form a scene feature vector. In the dynamic self-attention layer, by calculating the feature weights, the correlation between each feature and the scene is dynamically evaluated and the attention distribution is adjusted, such as focusing on the elegant style in the wedding scene. In the transfer learning process, the dynamically adjusted weight optimization model is used to adapt it to the needs of different scenes, generate a style prediction model that can perceive scene changes, and finally achieve accurate recommendations for scenes.

[0025] Step S400, receiving the user's recommendation task, combining the style prediction model, performing feature attention reset and pushing style decisions based on the task scenario, and generating a jewelry recommendation portrait. Specifically, after receiving the user's recommendation task, the system analyzes the task scenario (such as wedding ceremony, daily wear, business occasions, etc.) in combination with the style prediction model, and reallocates feature weights through a dynamic attention mechanism to highlight key features related to the scenario, such as increasing the weight of the "luxury" feature in the wedding scene and weakening the influence of the "simple" feature. Based on the adjusted feature distribution, a recommended feature set is generated, and key style vectors are extracted. Finally, a jewelry recommendation portrait containing user preferences, scenario requirements and priorities is mapped and generated to provide precise guidance for subsequent recommendations.

[0026] In a possible implementation, a user recommendation task is received, and the style prediction model is combined to perform feature attention reset and push style decision based on the task scenario to generate a jewelry recommendation portrait. Step S400 further includes step S410, receiving the user recommendation task, and updating the coefficient matrix by resetting the self-attention. Specifically, after receiving the user recommendation task, the system parses the task scenario (such as wedding, daily wear or business occasion) and extracts explicit and implicit features related to the scenario. Based on the dynamic self-attention mechanism, the system first assigns a first weight according to the scene relevance, such as the wedding scene gives priority to luxury and exquisite features, while daily wear focuses more on lightness and comfort; secondly, according to the distribution of user preference features in the current task, a second weight of feature proportion is generated. Subsequently, the first weight and the second weight are averaged to generate a self-attention distribution for dynamically adjusting the eigenvalues ​​of the original coefficient matrix. Finally, by applying the self-attention weight to the coefficient matrix, the effective combination of user preferences and task scenarios is completed, providing an accurate feature basis for the generation of recommendation portraits.

[0027] Step S420, combining the updated coefficient matrix with the base matrix, performing an inverse matrix decomposition operation, and determining an updated feature matrix. Specifically, by combining the updated coefficient matrix with the base matrix, performing an inverse matrix decomposition operation, a user preference feature matrix in the task scenario is dynamically generated. First, the updated coefficient matrix adjusts the weights of specific preference features in the task scenario through the self-attention mechanism, and is combined with the original base matrix to input the inverse operation process. Subsequently, the system restores the complete feature matrix in the current task scenario through matrix operations, ensures the dynamic balance of explicit features and implicit features, and optimizes the embodiment of important features in specific scenarios. The updated feature matrix finally generated fully reflects the user's dynamic preferences in specific tasks, providing accurate data support for subsequent personalized recommendations.

[0028] Step S430, based on the updated feature matrix, make a jewelry style preference decision and generate the jewelry recommendation portrait. Specifically, based on the updated feature matrix, a jewelry style preference decision model is constructed, and the user's personalized preference is determined by comprehensively analyzing explicit features and implicit features. The model automatically assigns feature weights and emphasizes the priority of key features according to the task scenario. For example, wedding scenes focus on formal styles, while daily scenes focus on practicality. Combined with the multi-dimensional data of the feature matrix, the user's style preferences are classified to generate a recommendation portrait that includes preference ratios, representative features, and the matching degree of recommended jewelry. Finally, the recommendation portrait is presented in a visual way to intuitively show the user the basis for personalized recommendations, thereby improving the accuracy and acceptance of recommendations.

[0029] In a possible implementation, the user recommendation task is received, and the coefficient matrix is ​​updated by resetting the self-attention. Step S410 further includes step S411, based on the task scenario, determining the first weight according to the scene relevance, wherein the first weight corresponds to the preference feature one by one. Specifically, according to the task scenario, the specific requirements of the recommended task are first analyzed, such as the characteristics of scenes such as weddings, daily wear or business meetings, and the features related to the scene (such as material, color, style, etc.) are extracted. Subsequently, the above features are quantitatively analyzed to calculate their importance in the current scene. For example, in a wedding scene, luxury is more important than comfort, so a higher weight is assigned. Finally, a corresponding relationship between the feature and the first weight is established, so that the key features are given priority in the scene requirements, providing an accurate basis for subsequent self-attention distribution and personalized recommendations.

[0030] Step S412, identifying the coefficient matrix, and determining the second weight, wherein the second weight is determined by the relative proportion of the coefficients of the matrix items. Specifically, by identifying the coefficient matrix, the coefficient values ​​in the matrix are parsed item by item to reflect the user's preference for a specific feature, and the importance of the feature is determined based on the relative proportion of each coefficient in the overall matrix. A larger coefficient value represents a higher user preference for the feature, thereby giving it a greater weight to form a second weight. The weight dynamically reflects the characteristics of the user in the interaction, and is adjusted in real time with new interaction data to accurately capture changes in user preferences and provide an accurate basis for subsequent analysis and recommendation.

[0031] Step S413, calculate the mean of the first weight and the second weight to determine the self-attention distribution. Specifically, in order to generate the self-attention distribution, the first weight is first determined from the task scenario analysis to reflect the relevance of the user's demand for each feature in a specific scenario; at the same time, the second weight is extracted from the decomposed coefficient matrix and calculated based on the importance ratio of the user's historical preference features. After standardizing the first weight and the second weight, their mean is calculated feature by feature to comprehensively reflect the degree of combination of the current task requirements and historical preferences to generate the self-attention distribution. This distribution is used to dynamically adjust the feature weights to ensure that the model can flexibly adapt to changes in user needs and achieve accurate recommendations.

[0032] Step S414, calculate the product of the self-attention distribution and the coefficient matrix to determine the coefficient matrix. Specifically, the updated coefficient matrix is ​​generated by calculating the product of the self-attention distribution and the original coefficient matrix. First, the self-attention distribution generated based on the task scenario is multiplied one by one with the feature columns of the original coefficient matrix to achieve dynamic weight adjustment. Specifically, the self-attention distribution reflects the correlation weights of each preference feature in the current scenario. It is matched column by column with the original coefficient matrix and then weighted calculated to generate a new coefficient matrix. The updated coefficient matrix integrates the user's historical preferences and real-time scenario requirements, strengthens the features related to the task, weakens the irrelevant features, provides optimized input for subsequent recommendation decisions, and improves the scene adaptability and accuracy of the recommendation.

[0033] Step S500, based on the jewelry recommendation portrait, platform resources are searched based on the platform search engine, and the jewelry recommendation list is determined and the interface is visualized. Specifically, based on the jewelry recommendation portrait, the platform search engine is used to filter jewelry that meets the user's preferences from the comprehensive resource pool. The resource pool is composed of platform resources and interactive platform data, including jewelry categories, styles, evaluations and other information. Combined with user portrait features, such as style, material, budget, etc., a comparison is performed item by item to filter out jewelry that meets the conditions, and a recommendation list is generated based on the preference matching degree, evaluation, inventory status and other factors. Finally, the recommendation results are displayed visually through the interface. Each item contains jewelry pictures, prices, feature descriptions and purchase links, achieving accurate recommendations and convenient operations.

[0034] In a possible implementation, according to the jewelry recommendation portrait, platform resources are searched based on the platform search engine, the jewelry recommendation list is determined and the interface is visualized, and step S500 further includes step S510, based on the platform resources and the interactive platform resources, a search resource pool is determined, wherein the interactive platform is a platform that has information interaction with the search platform. Specifically, a unified search resource pool is constructed by integrating local platform resources and interactive platform resources. First, jewelry information (such as style, material, price, etc.) of the local platform is collected and classified and labeled; secondly, resource data (such as user behavior data, popular trends, etc.) of the interactive platform is obtained in real time through the information interaction interface, and deduplication, denoising and format unification are performed. Subsequently, the dynamic update mechanism is used to synchronize resources in real time, eliminate redundant information, and ensure the accuracy and timeliness of the resource pool data. Finally, a comprehensive resource pool is formed, integrating high-quality resources from multiple platforms, providing diversified options and efficient data support for jewelry recommendations.

[0035] Step S520, based on the jewelry recommendation portrait, search the search resource pool to filter a preset number of recommended jewelry sets. Specifically, according to the user preference information (such as style, material, color, price range, etc.) in the jewelry recommendation portrait, set search conditions and match data in the search resource pool, and filter jewelry items that meet the conditions through database query or index search. According to the priority rules of the matching results (such as matching score), a preset number of recommended jewelry sets are filtered out, and a small amount of jewelry with similar but not completely matched styles can be supplemented to enrich the recommended content. Finally, the filtered recommendation set is saved as structured data as the basis for subsequent sorting and interface display.

[0036] Step S530, sort the recommended jewelry set, determine the jewelry recommendation list and visualize the interface, wherein the recommended jewelry on the interactive platform is displayed as a link. Specifically, according to the user preference characteristics (such as style, material, color, brand, price range, etc.) in the jewelry recommendation portrait, the recommended jewelry set is quantified by a weighted scoring algorithm, and the items with high matching degree are ranked first. At the same time, the sorting is optimized in combination with the user's historical interaction data to ensure that the recommendation results are more in line with the user's interests. According to the sorting results, the top several high-matching jewelry are selected to generate a recommendation list, and the number of recommendations is dynamically adjusted according to the task requirements, and innovative style exploration items are appropriately added. Finally, the recommendation list is intuitively presented in the user interface, showing the picture, title, price and purchase link of each recommended item, wherein the recommended items on the interactive platform are displayed in the form of jump links to ensure that users can quickly view or purchase, thereby improving the accuracy of recommendations and user experience.

[0037] In a possible implementation, according to the jewelry recommendation portrait, platform resources are searched based on the platform search engine, the jewelry recommendation list is determined and the interface is visualized, and step S500 further includes step S540, determining user interaction behavior and generating preference feedback information, wherein the user interaction behavior refers to user operation information based on the jewelry recommendation list. Specifically, by real-time monitoring of the user's interaction behavior on the jewelry recommendation list, including operations such as clicking, browsing, collecting, sharing and purchasing, the system collects detailed behavior data, such as operation time, frequency and duration, and classifies it into strong preference behavior (such as purchase, collection), moderate preference behavior (such as repeated browsing) and negative behavior (such as quick skipping) according to the intensity of the behavior. Subsequently, the style features (such as material, design style, price range) and behavior characteristics (such as operation frequency, repeated interaction) of the user's preference are extracted from the interaction behavior, and weights are assigned based on the intensity of the behavior to generate structured preference feedback information. For example, if a user frequently collects simple-style silver jewelry, the weight of the style will be significantly increased. The feedback information is dynamically updated to ensure that the user's latest interests are captured, and noise interference is avoided by filtering abnormal behaviors, providing accurate support for subsequent model optimization and personalized recommendations.

[0038] Step S550, adding the preference feedback information into a temporary database, and updating and learning the style prediction model based on the temporary database. Specifically, the user's preference feedback information (including the characteristics of the interaction behavior and the preference weight) is stored in a temporary database, and the data is denoised, standardized, and fused with historical preference information to form a complete training sample set. The style prediction model is updated through an incremental learning strategy, and the model parameters are optimized with the latest interaction data to ensure that the model can dynamically adapt to changes in user preferences. The updated model is deployed to the recommendation system after verification to achieve more accurate personalized recommendations.

[0039] The embodiment of the present application collects user jewelry interaction data, mines explicit features and implicit relationships to construct a preference feature matrix; decomposes the matrix and introduces regularization terms to train the prediction model; combines the interaction scenario with transfer learning through a dynamic self-attention layer to generate a style prediction model; adjusts feature weights and generates a jewelry recommendation portrait based on user recommendation tasks; matches resources through a platform search engine, generates recommendation results and displays them visually, thereby achieving the technical effect of improving the accuracy of user preference analysis and recommendation accuracy.

[0040] In the above, refer to Figure 1 A jewelry style prediction method based on user preference analysis according to an embodiment of the present invention is described in detail. Figure 2 A jewelry style prediction system based on user preference analysis according to an embodiment of the present invention is described.

[0041] A jewelry style prediction system based on user preference analysis according to an embodiment of the present invention is used to solve the technical problem of insufficient accuracy of user preference analysis and recommendation in the prior art, and achieves the technical effect of improving the accuracy of user preference analysis and recommendation. A jewelry style prediction system based on user preference analysis includes: a preference feature matrix construction module 10, a large model supervision training module 20, a style prediction model generation module 30, a jewelry recommendation portrait generation module 40, and an interface visualization module 50.

[0042] The preference feature matrix building module 10 is used to collect and call multi-source interaction records based on user jewelry interaction behaviors, and build a preference feature matrix by mining explicit features and implicit relationships.

[0043] The large model supervised training module 20 is used to decompose the preference feature matrix, introduce regularization terms, and supervise the training to predict the large model.

[0044] The style prediction model generation module 30 is used to introduce a dynamic self-attention layer based on an interactive scenario, perform transfer learning on the large prediction model, and generate a style prediction model.

[0045] The jewelry recommendation portrait generation module 40 is used to receive user recommendation tasks, combine the style prediction model, perform feature attention reset and push style decision based on the task scenario, and generate a jewelry recommendation portrait.

[0046] The interface visualization module 50 is used to perform a platform resource search based on the platform search engine according to the jewelry recommendation portrait, determine the jewelry recommendation list and perform interface visualization.

[0047] The specific configuration of the preference feature matrix construction module 10 will be described in detail below. As described above, the multi-source interaction records based on the user's jewelry interaction behavior are collected and called, and the preference feature matrix is ​​built by mining explicit features and implicit relationships. The preference feature matrix construction module 10 further includes: an interaction information division unit, which is used to traverse the interaction records and divide explicit interaction information and implicit interaction information, wherein the division is based on preference intuitiveness; an explicit vector determination unit, which is used to extract and quantify style features for the explicit interaction information and determine explicit vectors; an implicit vector determination unit, which is used to mine implicit relationships and quantify implicit vectors for the implicit interaction information and determine implicit vectors; a preference feature matrix construction unit, which is used to integrate the explicit vectors and the implicit vectors to construct the preference feature matrix.

[0048] The specific configuration of the large model supervised training module 20 will be described in detail below. As described above, the preference feature matrix is ​​decomposed, and the regularization term is introduced to supervise the training and predict the large model. The large model supervised training module 20 further includes: a preference feature matrix traversal unit, the preference feature matrix traversal unit is used to traverse the preference feature matrix, decompose and determine the base matrix and the coefficient matrix, wherein the product of the base matrix and the coefficient matrix is ​​the feature matrix; a style preference sample determination unit, the style preference sample determination unit is used to determine the style preference sample based on the multi-source interaction record; a prediction loss function construction unit, the prediction loss function construction unit is used to introduce the regularization term and construct the prediction loss function; a prediction large model generation unit, the prediction large model generation unit is used to perform prediction training based on time logic based on the style preference sample and the prediction loss function, based on the base matrix and the coefficient matrix, to generate the prediction large model.

[0049] The specific configuration of the jewelry recommendation portrait generation module 40 will be described in detail below. As described above, the user recommendation task is received, and the style prediction model is combined to perform feature attention reset and push style decision based on the task scenario to generate the jewelry recommendation portrait. The jewelry recommendation portrait generation module 40 further includes: a coefficient matrix update unit, the coefficient matrix update unit is used to receive the user recommendation task, and update the coefficient matrix by resetting the self-attention; a matrix decomposition inverse operation unit, the matrix decomposition inverse operation unit is used to combine the updated coefficient matrix with the base matrix, perform matrix decomposition inverse operation, and determine the updated feature matrix; a jewelry style preference decision unit, the jewelry style preference decision unit is used to make a jewelry style preference decision based on the updated feature matrix to generate the jewelry recommendation portrait.

[0050] Wherein, the user recommended task is received, and the coefficient matrix is ​​updated by resetting self-attention. The coefficient matrix updating unit further includes: a first weight determination subunit, the first weight determination subunit is used to determine a first weight based on the task scenario and according to the scenario relevance, wherein the first weight corresponds to the preference feature one-to-one; a second weight determination subunit, the second weight determination subunit is used to identify the coefficient matrix and determine the second weight, wherein the second weight is determined by the relative proportion of the coefficients of the matrix items; a mean calculation subunit, the mean calculation subunit is used to perform mean calculation on the first weight and the second weight to determine the self-attention distribution; a multiplication calculation subunit, the multiplication calculation subunit is used to calculate the product of the self-attention distribution and the coefficient matrix to determine the coefficient matrix.

[0051] The specific configuration of the interface visualization module 50 will be described in detail below. As described above, according to the jewelry recommendation portrait, the platform resource search is performed based on the platform search engine, the jewelry recommendation list is determined and the interface visualization is performed, and the interface visualization module 50 further includes: a search resource pool determination unit, the search resource pool determination unit is used to determine the search resource pool based on the platform resources and the interactive platform resources, wherein the interactive platform is a platform that has information interaction with the search platform; a recommended jewelry set screening unit, the recommended jewelry set screening unit is used to search the search resource pool according to the jewelry recommendation portrait, and screen a preset number of recommended jewelry sets; a jewelry set sorting unit, the jewelry set sorting unit is used to sort the recommended jewelry set, determine the jewelry recommendation list and perform interface visualization, wherein the recommended jewelry of the interactive platform is displayed as a link.

[0052] Among them, the interface visualization module 50 further includes: a preference feedback information generation unit, which is used to determine user interaction behavior and generate preference feedback information, wherein the user interaction behavior refers to user operation information based on the jewelry recommendation list; a style prediction model update unit, which is used to add the preference feedback information into a temporary database, and update and learn the style prediction model based on the temporary database.

[0053] The jewelry style prediction system based on user preference analysis provided in the embodiment of the present invention can execute the jewelry style prediction method based on user preference analysis provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0054] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules may be used and run on the user terminal and / or server, and the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.

[0055] The above specific implementations do not constitute a limitation on the protection scope of this application. It should be understood by those skilled in the art that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of this application should be included in the protection scope of this application.

Claims

1. A jewelry style prediction method based on user preference analysis, characterized in that: The method comprises: Collect and call multi-source interaction records based on user jewelry interaction behaviors, and build a preference feature matrix by mining explicit features and implicit relationships; Decomposing the preference feature matrix and introducing a regularization term to supervise the training of a large prediction model; A dynamic self-attention layer based on interactive scenarios is introduced to perform transfer learning on the large prediction model to generate a style prediction model; Receive user recommendation tasks, combine the style prediction model, perform feature attention reset and push style decision based on task scenarios, and generate jewelry recommendation portraits; According to the jewelry recommendation portrait, platform resource search is performed based on the platform search engine, the jewelry recommendation list is determined and the interface is visualized.

2. The jewelry style prediction method based on user preference analysis according to claim 1, characterized in that: The construction of the preference feature matrix includes: Traversing the interaction records, dividing explicit interaction information and implicit interaction information, wherein the division is based on preference intuitiveness; Extracting and quantifying style features based on the explicit interaction information to determine an explicit vector; For the implicit interaction information, mining and quantifying the implicit relationship to determine the implicit vector; The explicit vector and the implicit vector are integrated to construct the preference feature matrix.

3. The jewelry style prediction method based on user preference analysis as claimed in claim 1, characterized in that: The supervised training prediction model includes: Traversing the preference feature matrix, decomposing and determining a base matrix and a coefficient matrix, wherein the product of the base matrix and the coefficient matrix is ​​the feature matrix; Determining a style preference sample based on the multi-source interaction records; Introduce regularization terms and construct prediction loss function; According to the style preference samples and the prediction loss function, prediction training based on time logic is performed based on the base matrix and the coefficient matrix to generate the prediction large model.

4. The jewelry style prediction method based on user preference analysis as claimed in claim 3, characterized in that: Perform feature attention reset and push style decision based on task scenarios to generate jewelry recommendation portraits, including: Receiving the user recommended task, and updating the coefficient matrix by resetting self-attention; Combining the updated coefficient matrix with the base matrix, performing an inverse matrix decomposition operation to determine an updated feature matrix; Based on the updated feature matrix, a jewelry style preference decision is made to generate the jewelry recommendation portrait.

5. The jewelry style prediction method based on user preference analysis as claimed in claim 4, characterized in that: By resetting the self-attention, the coefficient matrix is ​​updated, including: Based on the task scenario, determining a first weight according to the scenario relevance, wherein the first weight corresponds to the preference feature one by one; Identify the coefficient matrix and determine the second weight, wherein the second weight is determined based on the relative proportion of the coefficients of the matrix items; Calculate the mean of the first weight and the second weight to determine the self-attention distribution; The product of the self-attention distribution and the coefficient matrix is ​​calculated to determine the coefficient matrix.

6. The jewelry style prediction method based on user preference analysis as claimed in claim 1, characterized in that: Search platform resources, determine jewelry recommendations and visualize the interface, including: Determine the search resource pool based on the platform resources and the interactive platform resources, wherein the interactive platform is a platform that interacts with the search platform; According to the jewelry recommendation portrait, searching in the search resource pool, and screening a preset number of recommended jewelry sets; The recommended jewelry set is sorted, the jewelry recommendation list is determined and interface visualization is performed, wherein the recommended jewelry on the interactive platform is displayed as a link.

7. The jewelry style prediction method based on user preference analysis according to claim 1, characterized in that: After the interface visualization, including: Determine user interaction behavior and generate preference feedback information, wherein the user interaction behavior refers to user operation information based on the jewelry recommendation list; The preference feedback information is added into a temporary database, and based on the temporary database, the style prediction model is updated and learned.

8. A jewelry style prediction system based on user preference analysis, characterized in that: The system is used to implement the jewelry style prediction method based on user preference analysis as described in any one of claims 1 to 7, and the system comprises: A preference feature matrix construction module, which is used to collect and call multi-source interaction records based on user jewelry interaction behaviors, and build a preference feature matrix by mining explicit features and implicit relationships; A large model supervised training module, which is used to decompose the preference feature matrix, introduce regularization terms, and supervise the training to predict the large model; A style prediction model generation module, wherein the style prediction model generation module is used to introduce a dynamic self-attention layer based on an interactive scene, perform transfer learning on the large prediction model, and generate a style prediction model; A jewelry recommendation portrait generation module, which is used to receive a user recommendation task, perform feature attention reset and push style decision based on the task scenario in combination with the style prediction model, and generate a jewelry recommendation portrait; The interface visualization module is used to search platform resources based on the platform search engine according to the jewelry recommendation portrait, determine the jewelry recommendation list and perform interface visualization.

Images