Automobile cloud exhibition hall personalized content pushing method based on entity store user portrait

Abstract

The application provides a car cloud showroom personalized content pushing method based on entity store user portrait, relates to the technical field of cloud digital display, and comprises the following steps: collecting multi-source heterogeneous behavior data of in-store users in real time; obtaining a store-side dynamic user portrait according to the multi-source heterogeneous behavior data; performing desensitization processing on the store-side dynamic user portrait to obtain a desensitized user portrait; extracting quantized values of multiple feature dimensions of the multi-source heterogeneous behavior data contained in the desensitized user portrait from the desensitized user portrait, and taking the quantized values as feature base points; and constructing a dynamic interest trend geometric configuration according to the feature base points. The application can realize accurate adaptation of cloud showroom content to real-time interests of users.

Description

Technical Field

[0001] This invention relates to the field of cloud-based digital display technology, and in particular to a method for personalized content delivery to a cloud-based car showroom based on user profiles of physical stores. Background Technology

[0002] With the deep integration of offline automotive services and cloud-based digital displays, building user profiles based on in-store customer behavior and delivering personalized content to cloud-based automotive showrooms have become important technological directions for improving the efficiency of scenario services and user interaction experience.

[0003] Most current automotive cloud showroom content push solutions in the industry adopt a generalized content distribution model or simply adapt content based on single, fragmented user data. They do not perform refined technical processing for user behavior characteristics in physical store scenarios, resulting in the following technical defects: it is difficult to achieve accurate cross-domain alignment between the dynamic user profiles on the offline store side and the content characteristics of the cloud automotive cloud showroom, and it is also difficult to deeply match the real interest trends of store users with cloud content.

[0004] This deficiency makes it difficult to transform heterogeneous behavioral data collected from offline stores, such as user dwell time in exhibition areas, attention settings, and consultation content, into quantifiable features that can be accurately matched with cloud content. This results in a significant discrepancy between the content to be pushed in the cloud exhibition hall and the actual interests of users. Even if store profiles are uploaded and cloud content is retrieved, only broad content adaptation can be achieved, which is difficult to match with the user's real-time interaction intent. Moreover, there is a lack of standardized adaptation logic between user profile data and cloud content features, making it difficult to dynamically adjust the matching weight based on the confidence level of the user's decision intent. In particular, the value of user behavior data is difficult to fully explore, resulting in inefficient use of cloud content resources and failing to meet the needs of personalized and precise content generation and matching. Summary of the Invention

[0005] This invention provides a method for personalized content push to a cloud-based car showroom based on user profiles of physical stores. This method can achieve accurate matching between cloud showroom content and users' real-time interests, effectively solving the problems of large content push deviations and low data value utilization in existing solutions.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for personalized content delivery to a cloud-based car showroom based on user profiles from physical stores, the method comprising: Collect multi-source heterogeneous behavioral data of in-store users in real time; obtain dynamic user profiles on the store side based on multi-source heterogeneous behavioral data; The dynamic user profiles on the store side are anonymized to obtain an anonymized user profile. The de-identified user profile is used to extract quantized values ​​of multiple feature dimensions derived from multi-source heterogeneous behavioral data, and these quantized values ​​are used as feature base points. A dynamic interest state geometric configuration is constructed based on these feature base points. The similarity of the dynamic interest pattern geometry with the baseline decision space pattern formed based on historical transaction user profiles is compared to obtain the comparison results; the confidence level of user decision intention is determined based on the comparison results. Based on the confidence level of the user's decision intent and the preset feature mapping model, the features in the current dynamic user profile on the store side are cross-domain aligned with the features of the cloud content to obtain a feature matching degree sequence; based on the feature matching degree sequence, personalized display content that matches the user's current decision intent is generated in real time. The personalized display content is pushed to users, and the user's deep interest behavior data is collected in real time during the interaction between the user and the cloud exhibition hall; The deep interest behavior data is fed back to the store side to update the dynamic user profile on the store side and synchronize it to the sales consultant terminal, so as to realize the closed-loop push of offline experience and online content linkage.

[0007] Secondly, the personalized content push system for car cloud showrooms based on user profiles of physical stores includes: The data collection module is used to collect multi-source heterogeneous behavioral data of in-store users in real time; and to obtain dynamic user profiles on the store side based on the multi-source heterogeneous behavioral data. The desensitization module is used to desensitize dynamic user profiles on the store side to obtain desensitized user profiles. The construction module is used to extract quantized values ​​of multiple feature dimensions from the de-identified user profile, which are derived from multi-source heterogeneous behavioral data, and use the quantized values ​​as feature base points; and to construct a dynamic interest situation geometric configuration based on the feature base points. The decision-making module compares the similarity of the dynamic interest pattern geometry with the baseline decision space based on historical transaction user profiles to obtain the comparison results; determines the confidence level of the user's decision intent based on the comparison results; aligns the features in the current store-side dynamic user profile with the cloud content features across domains based on the user's decision intent confidence level and a preset feature mapping model to obtain a feature matching degree sequence; and generates personalized display content that matches the user's current decision intent in real time based on the feature matching degree sequence.

[0008] Thirdly, a computing device includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0009] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0010] The above-described solution of the present invention has at least the following beneficial effects: By employing technologies such as multi-source behavioral data collection and fusion, dynamic user profile construction and desensitization, high-dimensional feature base point mapping, and dynamic interest status geometric configuration fitting, the technical problems of inaccurate cross-domain alignment between offline store profiles and cloud exhibition hall content, low content matching fit, and insufficient value mining of user behavior data have been overcome. This has resulted in the technical effects of accurate adaptation between cloud exhibition hall content and real-time user interests, more fitting content matching to user interaction intentions, and effective improvement in data resource utilization, thereby effectively enhancing the level of intelligent content matching and interaction in the exhibition hall. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the personalized content push method for a car cloud showroom based on user profiles of physical stores, provided by an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, embodiments of the present invention propose a method for personalized content push to a car cloud showroom based on user profiles of physical stores. The method includes the following steps: Step 1: Collect multi-source heterogeneous behavioral data of in-store users in real time; obtain dynamic user profiles on the store side based on the multi-source heterogeneous behavioral data. Step 2: De-identify the dynamic user profiles on the store side to obtain de-identified user profiles; Step 3: Extract the quantized values ​​of multiple feature dimensions from the de-identified user profile, which are derived from multi-source heterogeneous behavioral data, and use the quantized values ​​as feature base points; construct a dynamic interest situation geometric configuration based on the feature base points; Step 4: Compare the similarity between the dynamic interest pattern geometric configuration and the baseline decision space pattern formed based on historical transaction user profiles to obtain the comparison results; determine the confidence level of user decision intention based on the comparison results. Step 5: Based on the user's decision intent confidence level and the preset feature mapping model, perform cross-domain alignment between the features in the current dynamic user profile on the store side and the cloud content features to obtain a feature matching degree sequence; based on the feature matching degree sequence, generate personalized display content that matches the user's current decision intent in real time.

[0014] In this embodiment of the invention, a dynamic user profile is constructed by collecting multi-source heterogeneous behavioral data of users visiting physical stores. This user profile data is then processed using data anonymization standards. Feature dimension quantification values ​​are extracted from the anonymized profile, and a dynamic interest state geometric configuration is constructed. This configuration is further compared with the baseline decision space form of historical transaction user profiles to determine the confidence level of the user's decision intent. Finally, based on the confidence level, accurate cross-domain alignment of store-side profile features and cloud-based content features is achieved, generating personalized display content. This effectively overcomes the difficulty in accurately aligning offline store-side dynamic user profiles with cloud-based content features in existing automotive cloud showroom content adaptation solutions. This approach addresses the technical limitations of in-depth matching between alignment, real-world user interests, and cloud-based content, achieving precise adaptation between cloud-based showroom content and real-time user interests. Simultaneously, it fully leverages the value of multi-source, heterogeneous behavioral data from offline stores, transforming it into quantifiable features that can be precisely integrated with cloud content. This ensures that the generated personalized display content closely aligns with users' real-time interactive intentions. Furthermore, it establishes standardized feature adaptation logic based on the confidence level of user decision-making intentions, improving the utilization efficiency of cloud content resources. This effectively meets the personalized and precise content generation and matching needs of the automotive cloud showroom, enhancing its intelligent content matching level and user interaction experience.

[0015] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Collect raw user behavior data. This raw user behavior data includes the user's experience area, function preference range, consultation question content, number of companions, frequency of visits, and function focus, resulting in multi-source heterogeneous raw data. Specifically, this includes: conducting full-domain collection of raw user behavior data in key locations across all scenarios of the physical store, such as the experience display area, user rest area, and service negotiation area. The collection process covers the complete conversation cycle of the user from entering the store and participating in in-store activities to leaving the store, strictly adhering to real-time, uninterrupted, and comprehensive collection rules to capture various types of user behavior data; for behavior related to attention in experience areas, accurately recording the user's dwell time in each display area, the specific experience area type, function zoning level, and dwell time period, and simultaneously marking whether the user touched or viewed the display content in that experience area; for behavior related to function preference ranges, clearly recording the user's verbal and written feedback on function preference ranges; for interactive behavior related to consultation question content, completely recording the original text of the questions the user asked the staff on-site, the core points of communication, the direction of the needs and questions, and the duration of the interaction through a structured form.

[0016] For behaviors related to the number of people in a group, the system accurately counts the number of people in the user's group, their composition, and the focus of their attention. For behaviors related to the frequency of visits to the store, the system simultaneously retrieves data from the store visitor record system to accumulate statistics on the user's historical number of visits to the store, the time period of the current visit, and the total duration of stay. For behaviors related to function interest, the system meticulously captures the types of functions that users pay special attention to, the details of the functions they actively ask about, and the frequency of their questions. All collected data retains the original data collection terminal identifier, data collection timestamp, data collection scenario code, and other attribute information without any processing or modification. Finally, the data integration module aggregates user behavior data of various types, formats, and dimensions, including visual (stay and observe), interactive (consultation and communication), and statistical (number of people or frequency), to form multi-source heterogeneous raw collected data.

[0017] Step 1.2 involves multi-source data fusion of the heterogeneous raw data collected from multiple sources. This includes time synchronization and correlation matching of data from different sources belonging to the same user session to generate a correlated user behavior dataset. Specifically, this includes: performing full-process multi-source data fusion processing on the collected heterogeneous raw data; assigning a unique and non-repeating session identification identifier (composed of date code + store code + random sequence) to each in-store user based on the store user identity information recognition mechanism; using this identifier as the core basis for full-process data correlation; and connecting all raw collected data to the store edge computing node (the store edge computing node is deployed on a physical...). This lightweight computing power platform, located locally in physical stores, features low-latency data processing, multi-terminal data access, local data caching, and secure transmission protection capabilities. It enables local data processing without relying on remote cloud services. The unified data cleaning module standardizes data according to preset cleaning rules: for duplicate data collection, the latest valid data is retained based on the collection timestamp; for data with format errors, such as garbled characters or inconsistent numerical formats, format correction is performed according to data type adaptation and repair rules; for data with missing information or untraceable data (such as data without user session identifiers), it is directly marked as invalid and redundant data and removed one by one.

[0018] After cleaning, the timestamp errors of data from different sources belonging to the same user session are calibrated. Using the unified clock in the store as a benchmark, the time accuracy of data collected from each terminal is calibrated to the second level to accurately synchronize the time sequence deviations generated in different collection stages, achieving complete alignment of the time dimensions of various types of data. Using user session identification identifiers as indexes, the dwell experience area, function preference range, consultation question content, number of companions, frequency of visits, and function attention data under the same user session are bound one-to-one, establishing a single user behavior data association matrix to ensure the correlation between data and the continuity of behavioral time sequence. A secondary verification mechanism is then initiated, using a combination of manual sampling and methodological verification to verify the accuracy and completeness of data association. After confirming that there are no data omissions, no association errors, and no time sequence disorder, a time sequence-coherent, clearly attributed, dimensionally complete, and directly usable for feature extraction associated user behavior dataset is generated.

[0019] Step 1.3: Extract the quantified values ​​corresponding to each data item from the associated user behavior dataset to form a user feature vector. Specifically, this includes: performing multi-dimensional feature decomposition and quantification analysis on the associated user behavior dataset according to pre-set user behavior feature extraction standards. The pre-set user behavior feature extraction standards are: customized based on offline store service scenarios, solidified after multiple rounds of scenario verification, and precisely define the extraction range, calculation method, outlier removal threshold (e.g., the outlier threshold for store visit frequency is >10 times per session) and text feature processing specifications for six feature dimensions: dwell experience area, function preference range, consultation question content, number of companions, store visit frequency, and function attention. The dataset is then structurally decomposed, and mixed format data is classified according to the six feature dimensions: for unstructured text data such as consultation question content, word segmentation is first performed using Chinese word segmentation methods, and then keywords are extracted based on the store service domain thesaurus, such as space, intelligent interaction, and control performance; at the same time, semantic quantification is performed to convert it into text feature values ​​that can participate in the calculation.

[0020] For numerical data such as the number of people in a group and the frequency of visits, a min-max normalization method is used to scale them to a unified numerical range of [0,1] to eliminate the differences in units between different data. For categorical data such as the area of ​​stay, the function focus, and the function preference range, a one-hot encoding method is used for digital conversion, such as encoding a large space experience area as [1,0,0] and a compact experience area as [0,1,0] to achieve a quantitative expression of categorical features. After completing the data format conversion, the standardized quantitative value corresponding to each data item is calculated one by one according to the preset calculation caliber of the extraction standard. At the same time, the rationality of the values ​​is checked by the outlier detection algorithm, and abnormal feature data exceeding the normal threshold is removed, such as the number of people in a group > 10 people being judged as abnormal. Finally, the standardized quantitative values ​​of all dimensions are arranged in an orderly manner according to a fixed order of stay experience area, function preference range, consultation question content, number of people in a group, frequency of visits, and function focus, forming a user feature vector with unified dimensions (6 dimensions), accurate values ​​(retaining 4 decimal places), standard format (array format), and can be directly connected to subsequent model calculations.

[0021] Step 1.4: Construct a dynamic user profile for the store side based on user feature vectors. The dynamic user profile for the store side includes the user's current store visit behavior interest tags and intent tendencies. Specifically, this includes: inputting standardized user feature vectors into the trained dynamic user profile generation model for the store side. The trained dynamic user profile generation model for the store side is an improvement on the Lightweight Gradient Boosting Tree (LightGBM) model architecture. Based on the original architecture, it simplifies the feature calculation branches and optimizes the gradient iteration strategy to adapt to the low computing power and low latency operation requirements of the store's edge computing environment. This improved architecture has the core advantages of fast inference response speed, high convergence efficiency of small sample training, and low computing power consumption per round of inference. It can accurately meet the real-time user behavior data processing needs of the store, and can complete effective training based solely on the daily accumulated data of the store without crowding out the resources of other data processing links in the store. At the same time, it can accurately capture the correlation between user behavior features and interest intents, and reasonably adapt to the exclusive scenario of store user behavior analysis.

[0022] The model construction phase is based on a compliant and anonymized historical user behavior dataset accumulated from stores over the past 6 months. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. For each sample data point, multi-dimensional profile tags are manually labeled, including scenario preferences, functional focus, demand inclinations, and experience intention levels. A four-layer network structure is built, comprising a feature input layer, a weighted fusion layer, an intent recognition layer, and a tag output layer. The feature input layer receives standardized user feature vectors and performs dimensional verification and outlier filtering. The weighted fusion layer aggregates the input features based on the influence weights of different feature dimensions on user experience intent. The intent recognition layer incorporates store-specific multi-scenario intent determination logic, outputting corresponding intent probability values ​​for typical in-store scenarios such as multi-person group experiences, single-person visits to learn about product categories, and key inquiries about handling performance-oriented test drives. The tag output layer maps these probability values ​​to standardized interest tags and intent inclination descriptions.

[0023] During the training phase, supervised training is conducted using real user behavior data collected from offline stores. First, user feature vectors and corresponding profile labels from the training set are batch-input into the model. The backpropagation algorithm is used to calculate the deviation between the predicted and actual labels. Based on this deviation, the weight parameters of each layer of the model are iteratively optimized, and the weighting rules for interest labels and the intent judgment threshold are adjusted simultaneously. For example, the weight of the dwell experience area feature is set to 0.25, the weight of the function preference interval feature is set to 0.2, and the confidence threshold for judging the experience intention of users with high store visit frequency is appropriately lowered. Simultaneously, supplementary sample data is added based on typical store visit scenarios, such as multiple people visiting together focusing on large-space experience areas, single visitors focusing on compact experience areas, and key inquiries about handling performance biased towards test drive experiences. This allows for scenario-based fine-tuning of the model, making it more closely aligned with actual offline service scenarios. After each round of training, the model's recognition accuracy is verified using a validation set until the model converges and the recognition accuracy on the test set is ≥85%. The model training is then complete, and all parameters are solidified, forming a trained store-side dynamic profile generation model that can be directly used in stores.

[0024] Standardized user feature vectors are input into the trained store-side dynamic profile generation model. The feature input layer first verifies the completeness of the vector dimensions and the rationality of the values. The weighted fusion layer performs weighted fusion calculations on each feature dimension according to preset weights. Specifically, a preset weight coefficient is assigned to each feature dimension: dwell experience area W1=0.25, function preference interval W2=0.2, consultation question content W3=0.15, number of companions W4=0.15, store visit frequency W5=0.1, and function attention W6=0.15. Then, the standardized feature values ​​of each dimension are recorded as follows: dwell experience area feature value X1, function preference interval feature value X2, consultation question content feature value X3, number of companions feature value X4, store visit frequency feature value X5, and function attention feature value X6. These are then multiplied by the corresponding weight coefficients, i.e., X1×W1+X2×W2+X3×W3+X4× The process involves summing W4, X5, W5, and X6, and finally, summing all the product results to obtain the feature fusion value, thus completing the weighted aggregation calculation of multi-dimensional features. The intent recognition layer outputs the probability distribution of each experience intent dimension based on this fusion value, and then performs hierarchical parsing and annotation on the feature vectors to generate fine-grained interest tags covering dimensions such as scenario preferences, functional focus, demand priorities, and attention levels. Dynamic weighted analysis is then performed on various interest tags, and the intent probability value is adjusted in conjunction with features such as user store visit frequency and consultation depth to accurately determine the user's current demand tendency, focus, and experience intention level (high, medium, low). This ultimately generates a dynamic user profile on the store side that can dynamically reflect the user's behavioral preferences, interest characteristics, and experience intent tendencies throughout their entire store visit, and supports real-time updates based on newly added behavioral data. This dynamic user profile includes the user's current store visit behavior's interest tags and intent tendencies.

[0025] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1 involves parsing the dynamic user profiles on the store side to identify sensitive data fields containing direct personal identification information, thus obtaining the profile data to be anonymized. Specifically, this includes: retrieving the completed dynamic user profiles on the store side and performing profile parsing based on pre-defined sensitive data identification rules. These pre-defined sensitive data identification rules are formulated in conjunction with general data security management requirements and the offline service scenarios of stores, and include three core components: a keyword matching library, field format verification standards, and sensitivity level classification criteria. The profile is then decomposed field by field according to its data structure hierarchy, sequentially breaking it down into three levels: field groups, independent fields, and field content. Then, each... Independent fields are parsed character by character. First, the text content of the fields is searched and compared using a keyword matching library. Then, the character length, segmentation, and combination rules of the fields are checked for compliance using field format validation standards. This process accurately screens and identifies sensitive data fields in the profile that contain direct personal identity information such as user name, contact number, valid document number, residential address, and social media account. The text position, data type, and risk level of each sensitive field are recorded and marked in real time. Sensitive-related fields to be processed are retained, while ordinary behavioral feature fields without sensitive information are removed. Finally, the data is filtered and integrated to obtain complete, well-structured profile data with clearly marked sensitive fields.

[0026] Step 2.2 involves desensitizing sensitive data fields in the desensitized profile data. Direct personal identification information is removed through replacement or masking to obtain the desensitized user profile. Specifically, this includes: retrieving the marked desensitized profile data and calling a dedicated desensitization module deployed in the store's edge computing node to perform the desensitization operation. This dedicated desensitization module is adapted to local low-computing-power environments and integrates four core functional units: sensitive field reading, tiered desensitization execution, privacy compliance verification, and data structure protection. First, the sensitive field reading unit retrieves the risk level and corresponding data type of the marked sensitive fields. The categorized sensitive field information is then synchronously transmitted to the tiered desensitization execution unit. Simultaneously, the data structure protection unit connects to the tiered desensitization execution unit in real time, locking the overall data structure of the desensitized profile throughout the process to prevent interference from the desensitization operation on non-sensitive behavioral characteristic data.

[0027] For core sensitive fields such as contact phone numbers and valid document numbers, a multi-character masking process is initiated, retaining a small number of non-identifiable characters at the beginning and end, while the core characters in the middle are fully masked using dedicated placeholders. For general sensitive fields such as user names, residential addresses, and social media accounts, a full-field replacement process is initiated, completely replacing the field content with standardized universal placeholders. Throughout the desensitization process, the original structure of the profile data to be desensitized and the content and format of non-sensitive experience behavior characteristics, interest tags, intent tendencies, and other core data are preserved without any modifications. After desensitization is completed, the hierarchical desensitization execution unit pushes the processed profile data to the privacy compliance verification unit. Combined with the data structure protection log, a full-field review is conducted to check for any residual personal identity information and whether the data structure is complete and error-free. After confirming that there are no compliance issues, a desensitized user profile that meets data security compliance requirements and retains only user experience behavior and interest characteristics is finally generated.

[0028] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Read quantitative values ​​of multiple feature dimensions from the anonymized user profile. These multiple feature dimensions include dwell experience area, function preference range, consultation question content, number of companions, store visit frequency, and function focus. The function preference range refers to the user's inclination towards a specific function category, and function focus refers to the detailed function points that the user specifically focuses on. The function preference range and function focus are quantified independently. A set of user feature quantitative values ​​is generated, specifically including: retrieving the anonymized user profile that has passed privacy anonymization and verification, and reading feature quantitative values ​​according to pre-set behavioral feature extraction specifications. The pre-set behavioral feature extraction specifications are formulated based on offline store service scenarios, data security standards, and user behavior analysis needs, and include four core components: feature dimension definition standards, numerical standardization rules, outlier judgment thresholds, and data sorting format requirements. The feature dimension definition standards precisely identify six core dimensions: dwell experience area, function preference range, consultation question content, number of companions, store visit frequency, and function focus. The specifications define the numerical collection range and meaning of each dimension, accurately locate the data storage location of the six core behavioral feature dimensions within the profile, and extract standardized quantitative values ​​for the corresponding dimensions.

[0029] All extracted quantitative values ​​undergo a three-tiered, comprehensive, and refined screening process. The first tier involves line-by-line detection and filtering to remove invalid null values ​​caused by front-end data acquisition interruptions or transmission delays. The second tier compares the values ​​against the outlier thresholds specified in the guidelines, eliminating discrete outliers that exceed the normal range or deviate from the overall data distribution pattern. The third tier corrects data with inaccurate precision or inconsistent formats. Strictly adhering to the fixed dimensional order of the guidelines—staying experience area, function preference range, consultation question content, number of companions, store visit frequency, and function focus—the screened and corrected values ​​are systematically arranged according to these dimensions, while simultaneously unifying the data format and encoding method. This results in a complete set of user characteristic quantitative values ​​that is dimensionally comprehensive, numerically accurate, formatted correctly, free of redundancy and outliers, and directly accessible for high-dimensional spatial mapping calculations.

[0030] Step 3.2 maps each quantified value in the user feature quantification value set to a corresponding coordinate point in the high-dimensional feature space, obtaining a feature base point cluster. Specifically, this includes loading and activating dimension coordinate correspondence rules customized for the six major behavioral characteristics of store users and adapted to the needs of high-dimensional space analysis. These rules have been solidified after multiple rounds of verification in offline store behavior analysis scenarios. The core is establishing a three-level unique binding relationship between dimensions, coordinate axes, and coordinate values. This relationship is a one-to-one and unchangeable mapping relationship, clearly defining the stopping experience area corresponding to the first coordinate axis, the functional preference range corresponding to the second coordinate axis, and the consultation question content corresponding to the third coordinate axis. The system establishes a fixed correspondence between the coordinate axes: the fourth axis corresponds to the number of people traveling together, the fifth axis to the frequency of visits, and the sixth axis to the number of features followed. There are no overlapping dimensions or coordinate axes. Furthermore, each coordinate axis is subdivided into specific value ranges, determined by historical data distribution patterns for each feature dimension. The coordinate precision level is uniformly set to four decimal places, matching the precision standards of feature quantification values. A scaling factor is individually set for each coordinate axis, calculated based on the differences in the numerical magnitudes of the feature dimensions. A precise coordinate uniqueness verification logic is also established to ensure the compliance of the generated coordinate parameters.

[0031] To address the differences in numerical magnitude across different dimensions, a dedicated scaling ratio is assigned to the quantified values ​​of each feature dimension to achieve standardized adaptation of values ​​of different magnitudes. Simultaneously, the rules incorporate a coordinate duplication verification mechanism. This mechanism verifies the results by comparing the numerical combinations of coordinate parameters, effectively preventing coordinate conflicts, misalignments, and offsets at the source, thus ensuring the accuracy and uniqueness of the high-dimensional space mapping results. The system selectively retrieves a standardized set of user feature quantified values ​​that has undergone three levels of screening and format unification. It strictly follows a preset fixed dimension order based on the dwell experience area, function preference range, consultation question content, number of companions, store visit frequency, and function focus, and retrieves the standardized quantified values ​​of each dimension one by one through dimension indexing. Each quantified value is precisely matched with the corresponding coordinate axis bound in the dimension coordinate correspondence rule. The matching process is achieved through consistency verification between the dimension name and the coordinate axis identifier. Then, based on the preset scaling ratio and accuracy requirements of the coordinate axis, a linear normalization method is used to normalize and convert the quantified values, transforming feature values ​​of different magnitudes and units into a standard range of 0 to 1, effectively eliminating coordinate deviations caused by numerical differences between dimensions.

[0032] After the conversion, the result is sent to the uniqueness verification module for verification. The uniqueness verification module completes the verification by comparing it with the generated coordinate parameters. After the verification is successful, it generates independent coordinate parameters that are unique to this dimension, without repetition or conflict. Each set of coordinate parameters corresponds to a feature base point. This feature base point is the smallest spatial unit that carries single-dimensional behavioral feature information. It has fixed spatial coordinate attributes, unique feature dimension affiliation, and exclusive characteristics that cannot be repeated or replaced. Each base point corresponds to the quantification result of only one type of behavioral feature and is the basic point for constructing feature base point clusters and dynamic interest situation geometric configurations. The feature base points generated by the transformation of all six dimensions are as follows: Strictly adhering to the inherent logical connections of the original six dimensions and the fixed arrangement order of the stay experience area, function preference range, consultation question content, number of companions, frequency of visits, and function focus, the data is orderly clustered and arranged according to the orientation of their respective coordinate axes within the high-dimensional feature space. The arrangement process maintains a stable dimensional correlation spacing between the base points. This dimensional correlation spacing is set according to the correlation weight of the feature dimensions to ensure no point misalignment, no scattered offset, and no break in dimensional correlation. The final result is a feature base point cluster corresponding to the six feature dimensions, with spatial distribution patterns, accurate coordinate parameters, clear dimensional correlations, and stable data structure, which can be directly connected to the spatial situation analysis stage.

[0033] Step 3.3 involves analyzing the distribution of the feature base point cluster in the high-dimensional feature space to calculate its spatial distribution parameters. These parameters include the cluster center, dispersion, and the boundary of the clustered area. Specifically, this includes: performing a comprehensive, unrestricted traversal analysis of the complete feature base point cluster within the high-dimensional feature space; extracting the spatial coordinate data of all feature base points within the cluster according to the dimensional correspondence; classifying and organizing the scattered single-point coordinates according to six feature dimensions: dwell experience area, functional preference range, consultation question content, number of companions, frequency of visits, and functional focus; completing an orderly temporary aggregation; and verifying the number of points and the integrity of coordinates throughout the process to ensure no point omissions, no data corruption, and no coordinate deviations; and summarizing and statistically analyzing the coordinate values ​​of each coordinate axis of all feature base points. Let the total number of feature base points be N, and the coordinate of the i-th base point under a single coordinate axis be... Mean coordinates of this axis The arithmetic mean method is used for calculation, and the standard formula is: Repeat this operation to calculate the mean of all six coordinate axes. Combine the mean coordinates of each axis to form a complete mean point, which is the center coordinate of the feature base point cluster. .

[0034] Using the cluster center coordinates as a fixed measurement benchmark, for each independent feature point Its straight-line distance from the center coordinate C The spatial distance formula is used for calculation; the standard formula is: After calculating the distance to all base points using this logic, the average distance is... The standard formula is The overall coefficient of dispersion D is calculated using the standard deviation formula, which reflects the degree of dispersion of the cluster. The standard formula is shown below: The larger the coefficient of variation, the more dispersed the spatial distribution of the feature point clusters. Simultaneously, taking the cluster center coordinates as the origin, the high-dimensional feature space is radially divided into several standard spatial units of equal volume. Let the number of base points in a single unit be... The total number of elements is M, and the local distribution density of this element is... Overall cluster point distribution density The standard formula is .

[0035] A reasonable critical threshold for distribution density should be set in advance based on the overall distribution pattern of the cluster. Filter out those that meet the requirements All standard spatial units are defined as core regions with dense base points. The coordinates of all feature base points on the outer contour of the core region are extracted and connected sequentially in a clockwise continuous arrangement. The inflection points formed by the connecting lines are smoothed and corrected to eliminate sharp corner deviations. Finally, the clear boundary coordinates of the high-density clustered region are determined. Thus, the three types of core spatial distribution parameters—cluster center coordinates, dispersion, and clustered region boundary—are completely determined. All parameters are fixedly archived and stored, and a unique binding relationship is established between the parameters and the corresponding feature base point clusters to ensure that there is no data deviation, no information loss, and no parameter disorder in subsequent calling processes.

[0036] Step 3.4: Based on the spatial distribution parameters, fit the geometric contour surrounding the feature base point cluster as the dynamic interest situation geometric configuration. Specifically, this includes: retrieving three types of complete spatial distribution parameters that have been archived and verified: cluster center coordinates, dispersion, and cluster region boundaries. Using the cluster center coordinates as the core spatial positioning reference and the high-density cluster region boundary coordinates as the external constraint range for contour fitting, combined with the overall density characteristics of the feature base point cluster reflected by the dispersion, accurately follow the actual distribution situation of the feature base point cluster in the high-dimensional feature space, and formally start the entire process of smooth closed contour fitting: In the initial stage of fitting, accurately extract the coordinates of all feature base points on the boundary of the high-density cluster region from the spatial distribution parameters, and arrange them in a clockwise continuous order for all edges. The coordinates of the boundary base points are ordered and a preset coordinate deviation threshold is set. The preset coordinate deviation threshold is set based on 1.5 times the cluster dispersion coefficient. The larger the dispersion coefficient, the larger the threshold is. Each boundary base point is screened one by one by the coordinate deviation threshold judgment method: the vertical distance between this base point and the fitted line segment formed by the two adjacent base points is calculated. If the vertical distance exceeds the preset deviation threshold, it is judged as an isolated base point with abnormal coordinates and is directly eliminated. If the vertical distance is within the threshold range, it is retained as a candidate anchor point. After all boundary base points are screened, the candidate anchor points are verified a second time. The straight-line distance between the candidate anchor point and the cluster center is calculated. Extreme values ​​with distances deviating from the average value of candidate anchor points by ±20% are eliminated. Finally, the core anchor points with a moderate number, uniform distribution and close fit to the boundary trajectory are accurately locked.

[0037] Using the cluster center coordinates as the origin of radiation, two adjacent core anchor points are initially connected by a straight line. An initial contour fitting control point is selected at the midpoint of the line connecting the two points. The calculated cluster dispersion coefficient is retrieved and divided into three density levels: low (<0.2), medium (0.2 to 0.5), and high (>0.5). Differentiated offset adjustment coefficients are set for different levels (0.1 for low density, 0.3 for medium density, and 0.5 for high density). The base point distribution density of the area where this fitting control point is located is then calculated, and the corresponding offset adjustment coefficient is matched. The spatial offset distance of the control point is dynamically adjusted based on this coefficient: high density... The control points in the low-density areas are slightly offset towards the cluster center, while the control points in the medium-density areas are moderately offset away from the cluster center. The control points in the medium-density areas maintain a neutral offset. The initial straight line segments are corrected by natural arcs along the positions of the adjusted control points: that is, using two adjacent core anchor points as curve endpoints and the adjusted control points as curvature control points, smooth transition curve segments are generated to replace the rigid straight line segments. After completing the curve fitting segment by segment, the curvature is checked (curvature value controlled in the range of 0.05 to 0.8) to ensure that the arcs are not excessively curved or straight, effectively eliminating sharp corners on the contour and making the contour lines highly consistent with the actual distribution of the base points.

[0038] After completing curve fitting between all adjacent core anchor points segment by segment, the core anchor points at the beginning and end of the contour are precisely connected in a closed loop. The arc at the connection point is then smoothed and fine-tuned to eliminate gaps and line abruptness, ensuring the overall contour lines are continuous without breaks, abrupt turns, or trajectory deviations. This generates a closed geometric contour that highly matches the spatial distribution trajectory of the feature base point cluster, exhibiting a regular and smooth shape. This closed geometric contour incorporates proprietary dynamic adaptation rules, which possess complete operational logic for real-time monitoring, automatic triggering, and precise correction. Through a high-dimensional spatial data perception module, it captures in real-time the addition, deletion, and coordinate shifts of feature base points, as well as various changes in the overall distribution of the feature base point cluster. Once a change in the number, spatial location, or regional distribution density of base points that meets the trigger threshold is detected, it immediately... The dynamic contour adaptation and correction process involves: first, re-collecting all the updated feature base point cluster data, and simultaneously calling the spatial distribution parameter calculation logic in step 3.3 to recalculate and generate the latest spatial distribution parameters such as cluster center coordinates, dispersion, and cluster region boundaries; then, extracting the updated high-density cluster region boundary core anchor points from the latest parameters, and strictly repeating the entire process of smooth closed contour fitting. Simultaneously, the overall size of the contour, local line curvature, curve bending angle, and boundary coverage are adaptively adjusted according to the new base point distribution pattern to ensure that the corrected geometric contour always fits the latest feature base point cluster distribution pattern without lag deviation or morphological distortion, and always maintains a high degree of matching with the feature base point cluster. Finally, a dynamic interest pattern geometric configuration that can intuitively and accurately reflect the user's offline experience behavior preferences and experience intention tendencies is formed.

[0039] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: Extract geometric feature parameters from the dynamic interest situation geometric configuration to obtain the geometric feature set to be compared. Specifically, this includes: retrieving the full data of the generated dynamic interest situation geometric configuration; extracting feature parameters according to the preset geometric feature extraction specifications. These specifications are formulated based on the requirements of high-dimensional space geometric analysis and the store user behavior analysis scenario, and include three core aspects: feature type definition, parameter calculation caliber, and accuracy retention standards; first, analyze the overall morphological features of the geometric configuration, and extract the contour-enclosed volume V of the configuration. Its standard calculation formula is: ,in The maximum span of the configuration profile along the k-th coordinate axis. The symbol for cumulative multiplication in mathematics indicates a continuous multiplication operation on subsequent variables; k is the coordinate axis dimension number (ranging from 1 to 6), corresponding to six feature dimensions: dwell experience area, function preference range, consultation question content, number of companions, store visit frequency, and function focus; then, the contour surface area S of the configuration is extracted, and the polyhedral surface area approximation method is used, with the formula being... ,in Let n be the area of ​​the p-th polygonal facet on the configuration surface, and n be the total number of facests; simultaneously, extract the center offset of the configuration. That is, the spatial straight-line distance between the geometric configuration center and the center of the feature point cluster, which is expressed by the formula: in The coordinates of the geometric configuration center are, The coordinates of the feature base point cluster center are used; in addition, the mean contour curvature of the configuration is extracted. The formula is (in (This refers to the curvature value of the q-th sampling point of the contour, where q is the total number of sampling points), and the contour compactness. Core geometric feature parameters are extracted; the integrity of all extracted geometric feature parameters is checked, invalid and abnormal parameter values ​​are removed, the precision of the parameters is unified to four decimal places, and the parameters are arranged in a fixed order of contour bounding volume, contour surface area, center offset, mean contour curvature, and contour compactness, so as to form a geometric feature set to be compared with complete dimensions, accurate values, and uniform format.

[0040] Step 4.2: Invoke the pre-generated baseline decision space form. The baseline decision space form is formed by fitting the feature base point cluster of historical transaction user profiles. Specifically, it includes: retrieving the full data of the baseline decision space form that has been archived and stored and verified for data integrity through the form retrieval interface of the store data platform. The baseline decision space form is a standardized reference form constructed based on the complete profile data of the store's transaction users in the past 12 months. Its generation process is completely consistent with the fitting logic of the dynamic interest status geometric configuration in Step 3, and has undergone multiple rounds of effectiveness verification: First, collect all six types of core behavioral data of historical transaction users from the first visit to the store for consultation, experience interaction to the final transaction, namely, the dwell experience area, the functional preference range, the content of consultation questions, and so on. The original behavioral data collected, including the number of pedestrians, frequency of visits, and function preferences, undergoes the same anonymization process as in step 2, removing direct personal identification information such as names and phone numbers. Following the feature extraction specifications in step 3.1, the quantification values ​​of six core features are obtained, and a unique feature base point cluster for each transacting user is generated through high-dimensional space mapping. The spatial distribution parameter calculation logic from step 3.3 is then invoked to calculate core parameters such as the cluster center, dispersion, and aggregation region boundary of each feature base point cluster. Finally, the smooth closed contour fitting process in step 3.4 generates the decision space form for a single transacting user. All transacting users' decision space forms are then clustered and integrated according to their experience preferences, forming a benchmark decision space form library covering different transaction types and experience preferences.

[0041] The benchmark decision space morphology library adopts a three-level storage architecture of primary category, secondary label, and morphology instance: the primary category is divided according to the core experience preference dimension (such as preference for large space experience area, preference for intelligent function, preference for cost-effectiveness, etc.), and each primary category is set with secondary labels (such as subdividing intelligent function preference into voice interaction function, intelligent control function, in-vehicle entertainment function, etc.). Each secondary label is associated with several benchmark decision space morphology instances, and each morphology instance is labeled with corresponding auxiliary attributes such as conversion rate, feature matching weight, user profile label, and transaction cycle. In the benchmark morphology call stage, the feature dimension weight analysis method is used to extract the 2 to 3 dominant feature dimensions with the highest weight from the dynamic interest trend geometry of the user to be compared, so as to determine the core experience preference direction of the user to be compared. Based on this preference direction, the benchmark decision space morphology library is initially screened to select the top 10 benchmark morphology instances with the highest matching degree with the feature dimension of the user to be compared in the same primary category.

[0042] Simultaneously, the feature distance between the user's features to be compared and these 10 candidate instances is calculated, and the instance with the smallest feature distance is selected as the single benchmark decision space form with the highest matching degree. If the number of candidate instances after the initial screening is insufficient (less than 3), it is determined that there is no single matching form. At this time, the mean geometric feature of all benchmark form instances under the corresponding first-level category is retrieved to generate the mean benchmark form under that category. After the call is completed, the data integrity of the retrieved benchmark form is verified to confirm that its geometric feature parameters, auxiliary attributes and other information are not missing or abnormal, ensuring that the retrieved benchmark form is highly consistent with the experience feature dimension of the user to be compared, providing an accurate and effective reference benchmark for subsequent similarity calculation.

[0043] Step 4.3: Calculate the similarity between the geometric feature set to be compared and the corresponding geometric features of the baseline decision space morphology to obtain a geometric morphology similarity value. The geometric features include contour bounding volume, contour surface area, center offset, mean contour curvature, and contour compactness. Specifically, this includes: extracting geometric feature parameters from the retrieved baseline decision space morphology that are identical to those in Step 4.1 to obtain a baseline geometric feature set. This feature set contains core parameters such as contour bounding volume, contour surface area, center offset, mean contour curvature, and contour compactness that correspond one-to-one with the geometric feature set to be compared; and using a cosine similarity algorithm to calculate the overall similarity between the geometric feature set to be compared and the baseline geometric feature set. Let the geometric feature set to be compared be a vector. The baseline geometric feature set is a vector. cosine similarity The standard calculation formula is: in, Let r be the value of the r-th feature parameter in the set of geometric features to be compared; It is the value of the r-th feature parameter in the reference geometric feature set; numerator It is the dot product of two eigenvectors, representing the directional similarity of the vectors; the denominator is the dot product of two eigenvectors. It is the product of the magnitudes of two feature vectors, used to normalize the dot product result; before calculation, the two feature vectors are normalized, scaling all parameter values ​​to the [0,1] interval to eliminate calculation bias caused by different parameter magnitudes; during the calculation, differentiated weights are set for different geometric feature parameters: contour bounding volume weight 0.25, contour surface area weight 0.2, center offset weight 0.2, contour curvature mean weight 0.15, and contour compactness weight 0.2. The weighted similarity is... The calculation formula is: in is the weight value of the r-th feature parameter; after completing the weighted similarity calculation, the validity of the result is verified to ensure that the value is in the range of [0,1], and finally the precise quantified geometric similarity value is obtained. The closer the value is to 1, the higher the matching degree between the geometric configuration to be compared and the baseline decision space morphology.

[0044] Step 4.4: Based on the geometric similarity value, calculate the user's decision intent confidence level using a pre-set confidence mapping function. This includes: first, retrieving a pre-trained and fixed confidence mapping function based on historical store transaction data. This function is a non-linear mapping function, and its core function is to establish a precise correspondence between the geometric similarity value and the user's decision intent confidence level. Its standard expression is: in Confidence level of user decision intent The weighted geometric similarity value obtained in step 4.3 is used, where m is the sensitivity adjustment coefficient and t is the threshold parameter. Both m and t are determined based on regression analysis of historical transaction data, with default values ​​of m=5 and t=0.6. The calculated geometric similarity value is substituted into this mapping function to complete the non-linear transformation of the value: when the similarity value is below the threshold t, the confidence level increases slowly with increasing similarity; when the similarity value is above the threshold t, the confidence level increases rapidly with increasing similarity, ensuring that the confidence level accurately reflects the correlation between similarity and decision-making intent. After calculation, the obtained user decision... The confidence score of the user's decision intent is validated within a range. Values ​​outside the [0,1] range are corrected to the range boundary values. Values ​​below 0 are corrected to 0, and values ​​above 1 are corrected to 1. All values ​​are kept to three decimal places. The final confidence score of the user's decision intent can intuitively represent the degree of matching between the current user's behavioral characteristics and the decision characteristics of historical users who have made transactions. The closer the value is to 1, the higher the probability that the user has the intention to make a transaction. The closer the value is to 0, the more likely the user has not yet shown a clear intention to make a transaction. This confidence score can be directly used for the formulation of user service strategies and the prediction of decision tendencies in stores.

[0045] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Obtain the user's decision intent confidence level. Determine the corresponding cross-domain alignment weight adjustment coefficient based on the numerical range of the confidence level. Specifically, this includes: retrieving the user's decision intent confidence level value from the output and completing precision calibration, range normalization, and anomaly removal; performing a two-layer pre-validation check on the retrieved confidence level value; the first layer checks whether the value has null values, missing values, incorrect format, or broken data links; the second layer checks whether the value is within the standard value range of 0 to 1 and whether the decimal precision is uniformly up to four decimal places. If any type of abnormal value is found, a multi-level automatic backtracking mechanism is immediately triggered, and the confidence level calculation, calibration, and archiving nodes in step 4.4 are checked level by level. The calibration cache data is re-extracted and the two-layer check is repeated until a compliant and usable confidence level value is output.

[0046] The system begins by introducing a pre-defined three-tiered confidence interval threshold for interval determination. These thresholds are standardized, tiered intervals solidified based on statistical analysis of massive historical user behavior data and multiple rounds of business scenario verification. Specifically, they are divided as follows: High confidence interval: values ​​greater than or equal to 0.7 and less than or equal to 1.0, representing clear user decision-making intent and well-defined behavioral characteristics; Medium confidence interval: values ​​greater than or equal to 0.3 and less than 0.7, representing ambiguous user decision-making intent and some fluctuation in behavioral characteristics; Low confidence interval: values ​​greater than or equal to 0 and less than 0.3, representing weak user decision-making intent and no obvious behavioral characteristics. The compliant confidence values ​​are precisely compared and determined against the three-tiered thresholds level by level to accurately pinpoint the unique confidence interval to which the value belongs, avoiding misjudgments due to interval overlap.

[0047] Simultaneously, it invokes an encrypted and archived interval-coefficient one-to-one mapping rule base: This rule base is a standardized mapping dataset specifically customized for cross-domain alignment weight control, stored in an encrypted and archived manner. It contains a unique correspondence between three levels of confidence intervals and cross-domain alignment weight adjustment coefficients. The coefficient values ​​have undergone multiple rounds of testing and calibration, with numerical precision uniformly set to four decimal places. There are no duplicate mappings or coefficient conflicts, and it has read-only anti-tampering attributes. It only supports precise invocation by the system according to intervals and does not support real-time modification. In the rule base, high-confidence intervals correspond to high control coefficients, medium-confidence intervals correspond to neutral control coefficients, and low-confidence intervals correspond to low control coefficients. The corresponding exclusive coefficients are accurately retrieved according to the interval matching results. After retrieval, the compliance, dimensional adaptability, and precision consistency of coefficient values ​​are checked simultaneously to investigate coefficient overflow and adaptation misalignment issues, ensuring that the final adjustment coefficients used are consistent with the strength of the user's current decision intention and can accurately adapt to the fine-grained dynamic control needs of cross-domain feature alignment.

[0048] Step 5.2: Apply the cross-domain alignment weight adjustment coefficient to the initial weights between the store-side profile features and cloud content features in the preset feature mapping model to generate the adjusted cross-domain alignment weight matrix, thereby obtaining the feature mapping model carrying the adjusted weight matrix. The preset feature mapping model is a neural network model based on the cross-domain attention mechanism. Specifically, it includes: applying the determined cross-domain alignment weight adjustment coefficient to the initial weights between the store-side profile features and cloud content features in the preset feature mapping model. This preset feature mapping model is an improved and optimized version of the classic Transformer cross-domain attention mechanism architecture. It adds a cross-domain feature adaptation head and a scene bias term on the basis of the native multi-head attention module, abandoning the generalization adaptation defects of the general cross-domain model. It is specifically designed for the differentiated feature alignment needs of offline physical store service scenarios and cloud content distribution scenarios, and is fully customized throughout the process.

[0049] The specific construction process involves using the five geometric feature dimensions of the dynamic user profile generated and normalized in step 4.1 on the store side, and the refined content feature dimensions such as product parameters, service descriptions, experience guides, guarantee policies, and store activities built into the cloud content library as dual input nodes to build a three-layer progressive feature alignment network. The first layer is responsible for uniformly reducing and standardizing the heterogeneous features of the store and the cloud, eliminating the difference in the magnitude of cross-domain features. The second layer realizes fine-grained cross-domain association between store-side user features and cloud content features. The third layer completes the initial weight allocation and preliminary calibration. In the middle layer of the network, a standardized initial association weight layer for store-side profile features and cloud content features is set up. At the same time, offline scenario-specific feature adaptation modules such as user preference tags, key experience areas, function preference ranges, frequency of in-store consultations, experience records, and function focus points are embedded to open up a precise feature association channel between user decision intentions and cloud content.

[0050] The specific training process is as follows: The training sample set is composed of anonymized historical valid user behavior data from offline physical stores nationwide over the past three years, cloud content matching data, and user content interaction data. Invalid visitor data, duplicate data, abnormal interaction data, and test data are removed from the samples beforehand using a data cleaning module to ensure the purity and representativeness of the training samples. Training, validation, and test sets are divided and stratified according to proportions. Multiple rounds of iterative training are conducted using supervised learning. Positive user behaviors such as valid clicks, browsing, in-depth searches, proactive inquiries, and appointment bookings on cloud content are used as the optimization target for the loss function. The cross-domain feature association parameters, attention allocation parameters, and scene bias parameters within the model are fine-tuned round by round. Early stopping is used to prevent overfitting and continuously reduce model matching error and generalization bias. After each fixed number of training rounds, the model stability is verified using a validation set until the model's cross-domain feature matching accuracy and content alignment precision both reach the preset technical thresholds. Training is then stopped and the model parameters are fixed.

[0051] The core advantage of this improved cross-domain attention mechanism model architecture lies in its ability to effectively capture the implicit correlation between offline user behavior characteristics and cloud content characteristics, adapt to the cross-domain differences between offline behavior data and cloud content data, effectively avoid problems such as feature mismatch and generalized push, and improve the targeting and accuracy of cross-domain feature matching. Furthermore, through a refined calculation logic of weighted adaptation, the cross-domain alignment weight adjustment coefficient is fused and corrected item by item with the initial weight. The specific weighted calculation formula is as follows: In the formula To integrate the corrected single-set cross-domain alignment weights, The model is built with initial cross-domain weights. The cross-domain alignment weight adjustment coefficients for the output. A static smoothing factor (a constant ranging from 0.05 to 0.1, used to prevent abrupt weight changes) is built into the model. After each item is calculated using this formula, the weight proportion corresponding to low confidence is dynamically weakened, while the weight proportion corresponding to high confidence is strengthened. This effectively eliminates generalization bias and historical redundancy interference in the initial weights, generating a customized cross-domain aligned weight matrix that is more suitable for the strength of the current user's decision intent. This adjusted weight matrix is ​​then precisely embedded into the core computation chain of the original feature mapping model, replacing the original initial weight module. Full model parameter verification and compatibility testing are then conducted to ensure that the model operation is stable and error-free after the weight update. Finally, a feature mapping model carrying the adjusted weight matrix is ​​obtained, which is suitable for the current user's decision intent and can be directly used for cross-domain feature calculation.

[0052] Step 5.3: Based on the feature mapping model carrying the adjusted weight matrix, perform cross-domain similarity calculation between each feature dimension in the current store-side dynamic user profile and the feature dimensions of each content item in the cloud content library to obtain a feature matching degree sequence. Specifically, this includes: building a dedicated cross-domain alignment channel between store-side user behavior features and cloud content features based on the verified feature mapping model carrying the adjusted weight matrix; the channel is a standardized data transmission and computation link specifically designed for cross-domain adaptation of offline store-side behavior features and cloud content features, adopting an isolated encrypted transmission architecture, and consisting of four layers: feature access layer, format adaptation layer, encrypted transmission layer, and computation docking layer. The feature access layer is responsible for receiving the features decomposed from the store side. The feature data extracted from the cloud and the feature data are standardized in terms of data format, dimensionality, and precision at the format adaptation layer. The encrypted transmission layer uses static encryption and dynamic verification to ensure data transmission security. The computational interface layer achieves seamless integration between feature data and feature mapping models. After the system is built, a three-layer specialized test is conducted on the channel. The first layer is connectivity testing to verify whether the nodes at each layer of the channel are connected normally and without link interruption. The second layer is data compatibility testing to check whether the feature data on the store side and the cloud side can be transferred normally without format conflicts. The third layer is transmission stability testing to monitor data transmission rate, packet loss rate, and latency to ensure that cross-domain feature data transmission is lossless, latency-free, and distortion-free. The channel is officially launched after the test is passed.

[0053] The five core geometric feature dimensions (contour bounding volume, contour surface area, center offset, mean contour curvature, and contour compactness) of the current dynamic user profile at the store side, which are normalized and free of redundancy and constructed based on six behavioral characteristics—staying experience area, function preference range, consultation question content, number of companions, store visit frequency, and function attention—are decomposed one by one to ensure that each feature dimension is independently computable, without dimension cross-interference or parameter confusion. At the same time, the standardized corresponding feature dimensions of various content items in the cloud content library (this content library is a standardized content storage pool built and centrally operated by the platform, which gathers various compliant content items that have been reviewed and archived, covering categories such as product introductions, service descriptions, experience guides, and rights and benefits information. All content has completed standardized feature labeling and unique number binding, supports quick retrieval by dimension, real-time retrieval and dynamic updates, and ensures stable and standardized content supply) are extracted. After dimension cleaning, outlier removal and format regularization, a precise one-to-one correspondence is established with the feature dimensions decomposed at the store side, and dimension mismatch or omission is strictly prohibited.

[0054] Cross-domain similarity calculation is performed item-by-item based on the cosine similarity algorithm. The specific calculation process is as follows: Single-set feature data from the store side and corresponding feature data from cloud-based content items are converted into standardized feature vectors. Null values ​​and noise data are removed from the vectors. The store-side feature vector is denoted as... The feature vector of cloud content items is denoted as Substituting into the basic formula for cosine similarity: ,in It is the dot product of two sets of eigenvectors; The magnitudes of the two sets of vectors are given respectively. The cosine value of the angle between the two sets of feature vectors is calculated. The closer the angle is to 0 degrees and the closer the cosine value is to 1, the higher the similarity between the two sets of features. The closer the angle is to 90 degrees and the closer the cosine value is to 0, the lower the similarity between the two sets of features. After completing the vector transformation and cosine value calculation for all corresponding feature dimensions for each set, the initial similarity results of a single set of content items are obtained.

[0055] The entire calculation process utilizes the adjusted weight matrix to dynamically weight and calibrate each set of feature similarity results. First, customized weight values ​​corresponding to the current feature dimension are extracted from the weight matrix. These weight values ​​are then multiplied and fused with the initial similarity results for the corresponding dimension. For irrelevant and interfering features with low relevance to the user's decision-making intent, low-value weights from the weight matrix are used to dilute them, automatically weakening their impact on the overall similarity. For core features highly aligned with the user's decision-making intent (such as features related to the user's dwell time area or functional preference range), high-value weights from the weight matrix are used to strengthen them, automatically increasing their dominant role in the overall similarity. After fusion, the calculation results are verified a second time to remove abnormal values ​​caused by noise interference or calculation bias, correcting similarity results that deviate from the normal range. This ensures that each set of weighted similarity values ​​truly reflects the degree of feature matching. Finally, the features are categorized and sorted according to the unique ID and category of the cloud content item, forming an ordered, accurate, uniformly formatted, and non-redundant feature matching sequence.

[0056] Step 5.4 involves selecting one or more content items with the highest feature matching degree sequence from the cloud content library as personalized display content that aligns with the user's current decision-making intent. Specifically, this includes: based on the feature matching degree sequence that has passed both completeness and accuracy checks, initiating a hierarchical closed-loop content filtering process; retrieving a preset matching degree benchmark (the benchmark is a minimum matching degree threshold solidified after extensive historical data experiments and scenario verification; only content items reaching or exceeding this value can be considered valid matches that meet the user's needs); and reordering all valid content items in the sequence according to a descending order of matching degree values. The sorting process simultaneously removes low-matching content items that do not meet the acceptable baseline, marks high-quality content items with excellent matching scores and adds hierarchical labels. Then, combined with a fixed filtering threshold (this threshold is a fixed value preset by the platform based on user terminal display specifications and visual browsing experience, limiting the maximum number of content items output in a single filtering session, balancing display completeness and loading efficiency, avoiding excessive quantity causing interface congestion and insufficient quantity causing information shortage), and the user terminal display interface's capacity and layout limitations, it accurately filters out content items with high matching scores and appropriate quantities, strictly controlling the total filtering volume to avoid terminal display errors and loading lag due to excessive quantity.

[0057] Simultaneously, a three-tiered in-depth verification process is conducted on the shortlisted content items. The first tier is compliance verification, which checks whether the text, materials, and format of each content item conform to the platform's general specifications, identifies and eliminates non-compliant items due to issues such as illegal expressions and abnormal materials. The second tier is duplication verification, which compares the core characteristics, thematic information, and display elements of each content item, eliminating redundant items that are repeatedly pushed, highly similar, or have homogenized information. The third tier is content validity verification, which checks the timeliness and integrity of the content items, determines whether there are any expired, invalid, abnormal links, or missing materials, and filters out abnormal items that do not meet the display standards. The final remaining content items first undergo automatic review to verify the matching degree, compliance, and validity again, and then are supplemented by manual random checks for final confirmation. After double verification without any abnormalities, the content is considered personalized display content that matches the user's current decision-making intention, adapts to the user's historical behavioral preferences (stay experience area, function preference range, consultation question content, number of companions, store visit frequency, function attention), is free of redundancy and deviation, and can be directly pushed to the terminal for stable display.

[0058] A personalized content push system for car cloud showrooms based on user profiles from physical stores includes: The data collection module is used to collect multi-source heterogeneous behavioral data of in-store users in real time; and to obtain dynamic user profiles on the store side based on the multi-source heterogeneous behavioral data. The desensitization module is used to desensitize dynamic user profiles on the store side to obtain desensitized user profiles. The construction module is used to extract quantized values ​​of multiple feature dimensions from the de-identified user profile, which are derived from multi-source heterogeneous behavioral data, and use the quantized values ​​as feature base points; and to construct a dynamic interest situation geometric configuration based on the feature base points. The decision-making module compares the similarity of the dynamic interest pattern geometry with the baseline decision space based on historical transaction user profiles to obtain the comparison results; determines the confidence level of the user's decision intent based on the comparison results; aligns the features in the current store-side dynamic user profile with the cloud content features across domains based on the user's decision intent confidence level and a preset feature mapping model to obtain a feature matching degree sequence; and generates personalized display content that matches the user's current decision intent in real time based on the feature matching degree sequence.

[0059] A computing device, comprising: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0060] A computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0061] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for personalized content push to a cloud-based car showroom based on user profiles of physical stores, applied to a cloud-based car showroom, characterized in that... The method includes: Step 1: Collect multi-source heterogeneous behavioral data of in-store users in real time; obtain dynamic user profiles on the store side based on the multi-source heterogeneous behavioral data. Step 2: De-identify the dynamic user profiles on the store side to obtain de-identified user profiles; Step 3: Extract quantized values ​​of multiple feature dimensions from the anonymized user profile, which originate from multi-source heterogeneous behavioral data, and use these quantized values ​​as feature base points; construct a dynamic interest state geometry configuration based on the feature base points, including: The quantitative values ​​of multiple feature dimensions are read from the de-identified user profile. These multiple feature dimensions include the dwell experience area, function preference range, consultation question content, number of companions, frequency of visits to the store, and function focus. Among them, the function preference range refers to the user's inclination towards a specific function category, and the function focus refers to the detailed function points that the user focuses on. The function preference range and function focus are quantified independently; a set of user feature quantitative values ​​is generated. Each quantized value in the set of quantized user feature values ​​is mapped to a corresponding coordinate point in a high-dimensional feature space to obtain a cluster of feature base points; The distribution pattern of the feature point cluster in the high-dimensional feature space is analyzed to calculate its spatial distribution parameters; the spatial distribution parameters include the cluster center, the degree of dispersion, and the boundary of the clustering region. Based on the spatial distribution parameters, a geometric contour surrounding the cluster of feature base points is formed, which serves as the geometric configuration of the dynamic interest situation. Step 4: Compare the dynamic interest pattern geometry with the baseline decision space pattern formed based on historical transaction user profiles to obtain the comparison results; determine the confidence level of user decision intent based on the comparison results, including: Geometric feature parameters are extracted from the dynamic interest situation geometric configuration to obtain the set of geometric features to be compared; The pre-generated baseline decision space form is invoked, which is formed by fitting a cluster of feature base points based on the historical transaction user profile. The similarity between the geometric feature set to be compared and the corresponding geometric features of the baseline decision space morphology is calculated to obtain the geometric morphology similarity value; wherein, the geometric features include the contour bounding volume, contour surface area, center offset, mean contour curvature, and contour compactness. Based on the geometric similarity value, the confidence level of the user's decision intention is calculated through a preset confidence mapping function; Step 5: Based on the user's decision intent confidence level and the preset feature mapping model, perform cross-domain alignment between the features in the current dynamic user profile on the store side and the cloud content features to obtain a feature matching degree sequence; based on the feature matching degree sequence, generate personalized display content that matches the user's current decision intent in real time.

2. The method for personalized content push to a car cloud showroom based on user profiles of physical stores according to claim 1, characterized in that, Real-time collection of multi-source heterogeneous behavioral data from in-store users; Based on multi-source heterogeneous behavioral data, a dynamic user profile is obtained on the store side, including: Collect raw user behavior data, which includes dwell time experience area, function preference range, consultation question content, number of companions, frequency of store visits, and function attention, to obtain multi-source heterogeneous raw data. Multi-source data fusion is performed on heterogeneous raw data collected from multiple sources. Data from different sources belonging to the same user session are synchronized in time and matched for correlation to generate a correlated user behavior dataset. Extract the quantified values ​​corresponding to each data item from the correlated user behavior dataset to form a user feature vector; A dynamic user profile for the store side is constructed based on user feature vectors. The dynamic user profile for the store side includes the user's interest tags and intention tendencies for the current store visit behavior.

3. The method for personalized content push to a car cloud showroom based on user profiles of physical stores as described in claim 2, characterized in that, The dynamic user profiles on the store side are anonymized to obtain an anonymized user profile, including: Analyze the dynamic user profiles on the store side to identify sensitive data fields containing direct personal identification information, and obtain profile data to be anonymized. Sensitive data fields in the de-identified profile data are de-identified by replacing or masking direct personal identification information to obtain a de-identified user profile.

4. The method for personalized content push to a car cloud showroom based on user profiles of physical stores according to claim 3, characterized in that, Based on the confidence level of user decision intent and the preset feature mapping model, the features in the current dynamic user profile on the store side are cross-domain aligned with the cloud content features to obtain a feature matching degree sequence. Based on the feature matching degree sequence, personalized display content that matches the user's current decision intention is generated in real time, including: Obtain the confidence level of the user's decision intent, and determine the corresponding cross-domain alignment weight adjustment coefficient based on the numerical range of the confidence level; The cross-domain alignment weight adjustment coefficient is applied to the initial weight between the store-side profile features and the cloud content features in the preset feature mapping model to generate the adjusted cross-domain alignment weight matrix, thereby obtaining the feature mapping model carrying the adjusted weight matrix. The preset feature mapping model is a neural network model based on the cross-domain attention mechanism. Based on the feature mapping model carrying the adjusted weight matrix, cross-domain similarity calculation is performed between each feature dimension in the dynamic user profile on the store side and the feature dimension of each content item in the cloud content library to obtain a feature matching degree sequence. One or more content items with the highest value in the feature matching sequence are selected from the cloud content library and used as personalized display content that matches the user's current decision intention.

5. A personalized content push system for a car cloud showroom based on user profiles of physical stores, wherein the system implements the method as described in any one of claims 1 to 4, characterized in that, include: The data acquisition module is used to collect multi-source heterogeneous behavioral data of in-store users in real time. Based on multi-source heterogeneous behavioral data, dynamic user profiles are obtained on the store side; The desensitization module is used to desensitize dynamic user profiles on the store side to obtain desensitized user profiles. The construction module is used to extract quantized values ​​of multiple feature dimensions from the de-identified user profile, which are derived from multi-source heterogeneous behavioral data, and use the quantized values ​​as feature base points; and to construct a dynamic interest situation geometric configuration based on the feature base points. The decision module is used to compare the similarity between the dynamic interest status geometric configuration and the baseline decision space form formed based on the historical transaction user profile to obtain the comparison result; determine the confidence level of the user's decision intention based on the comparison result; and perform cross-domain alignment between the features in the current store-side dynamic user profile and the cloud content features based on the user's decision intention confidence level and the preset feature mapping model to obtain the feature matching degree sequence. Based on the feature matching degree sequence, personalized display content that matches the user's current decision intention is generated in real time.

6. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 4.

Images