Broadband service dynamic compliance user portrait construction and recommendation system

Abstract

The invention discloses a broadband service dynamic compliance user portrait construction and recommendation system, and relates to the field of data analysis and management, and the system comprises a collection module which is used for collecting multi-source heterogeneous modal data of a user in a broadband service scene, and carrying out the desensitization of the collected data; the preprocessing module is used for performing missing value filling and abnormal value detection on the desensitized multi-source heterogeneous modal data, completing standardized regularization and cross-modal dimension adaptation based on semantic attributes of the modal data, and outputting semantic preprocessing data of adaptive feature fusion; according to the method, compliance desensitization processing is carried out on the multi-source heterogeneous data, data safety is guaranteed, associated semantics are reserved, data quality is effectively improved through missing value filling, abnormal value accurate detection and correction and standardized regularity, and cross-modal feature deep fusion and accurate weight distribution are combined, so that the data quality is improved. And comprehensively capturing user service preferences, behavior time sequence rules and compliance attributes, and constructing a dynamically adaptive user portrait.

Description

Technical Field

[0001] This invention relates to the field of data analysis and management technology, specifically to a system for constructing and recommending dynamic compliant user profiles for broadband services. Background Technology

[0002] With the booming development of the digital economy, broadband services, as a core component of digital infrastructure, have seen continuous expansion in user scale and enrichment of business scenarios, while users' demand for personalized services has become increasingly prominent.

[0003] The invention patent application with application number 201910951783.4 discloses a user profile construction, content recommendation method, device and equipment. The application aims to solve the problem that "since each child is an independent individual, they have different interests and books, which leads to their own reading preferences when reading picture books. Some children are interested in habit formation books, while others are interested in family affection books. It is necessary to understand each child's reading interests and provide them with targeted reading recommendations."

[0004] However, existing technologies often suffer from insufficient effectiveness in multimodal feature fusion. The heterogeneity of different modal data makes it easy for information redundancy or missing information to occur during feature fusion, affecting the accuracy and completeness of user profiles. At the same time, broadband service user behavior is highly dynamic, and user needs change rapidly with time and scenarios, making it difficult for existing profile update mechanisms to achieve real-time dynamic adjustment.

[0005] To address this, we propose a dynamic compliant user profile construction and recommendation system for broadband services. Summary of the Invention

[0006] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a dynamic compliant user profile construction and recommendation system for broadband services, which can effectively solve the problems of the existing technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions;

[0008] This invention discloses a system for constructing and recommending dynamic compliant broadband service user profiles, including:

[0009] The system comprises four modules: a data acquisition module for collecting multi-source heterogeneous modal data from users in broadband service scenarios and performing data anonymization; a preprocessing module for performing missing value imputation and outlier detection on the anonymized multi-source heterogeneous modal data, standardizing and adapting the data across modal dimensions based on the semantic attributes of each modality, and outputting semantically preprocessed data with adapted feature fusion; a fusion module for performing cross-modal feature association and deep fusion of the semantically preprocessed data based on Transformer encoding and attention weight allocation, generating a unified high-dimensional fusion feature set; a construction module for segmenting users based on the high-dimensional fusion feature set, extracting user business preferences, behavioral temporal patterns, and compliance attribute information, and constructing a dynamic compliant user profile; an early warning module for capturing dynamic changes in user business behavior data through LSTM temporal modeling, performing real-time early warning analysis of user churn risk, and synchronously generating profile update trigger signals; and a recommendation module for generating targeted business recommendations based on the latest dynamic compliant user profile.

[0010] The data acquisition module is interconnected with a preprocessing module via a wireless network. The preprocessing module is interconnected with a fusion module via a wireless network. The fusion module is interconnected with a construction module via a wireless network. The construction module is interconnected with an early warning module via a wireless network. The early warning module is interconnected with a recommendation module via a wireless network.

[0011] Among them, multi-source heterogeneous modal data includes user structured attribute data, network behavior log data, service usage trajectory data, and terminal-related interaction data.

[0012] Furthermore, the desensitization process in the acquisition module includes:

[0013] For core sensitive fields in user structured attribute data, a desensitization method combining feature masking and semantic replacement is used;

[0014] For network behavior log data, business usage trajectory data, and terminal-related interaction data, a dynamic perturbation desensitization method is used, and the perturbation amplitude is adaptively adjusted according to the data modality type.

[0015] Furthermore, the missing value imputation in the preprocessing module follows the following rules:

[0016] ;

[0017] In the formula: Fill in the missing values ​​for the j-th modality data of the i-th user; This represents the total number of multi-source heterogeneous modes; Let be the cross-modal semantic similarity between the j-th modality and the k-th modality; The confidence weight for the k-th mode; This represents the valid data value for the i-th user in the k-th modality.

[0018] Furthermore, during the preprocessing module's execution phase, a cross-modal semantic mapping matrix is ​​constructed based on the semantic attributes of each modality's data. Matrix elements This represents the semantic association strength between the q-th feature of the p-th modality and the target unified dimensional feature. Then, the dimensional importance score of each modality feature is calculated using information gain. Finally, core features are selected based on the comparison between dimensional importance scores and preset screening thresholds, and dimensional adaptation transformation is performed on the core features simultaneously.

[0019] ;

[0020] In the formula: The unified dimension feature matrix after adaptation; The feature matrix after standardization of each modality data; This is the dimension importance weight matrix.

[0021] Furthermore, the outlier detection of the preprocessing module follows the following rules:

[0022] ;

[0023] In the formula: This represents the anomaly determination result for the j-th modal data of the i-th user, where 1 indicates that it is an anomaly and 0 indicates that it is normal data. The statistical deviation of the j-th modal data of the i-th user relative to the feature distribution of that modality; Let be the semantic deviation coefficient of the j-th modality; Preset anomaly detection threshold;

[0024] like ∈( , ], To determine a condition as moderately abnormal based on a preset threshold, cross-modal semantic similarity is used. Modal confidence weights The top N modalities with the highest semantic relevance are selected to construct a correction baseline, and the correction value is calculated by weighted average; if > If the data is deemed severely abnormal, a modal confidence-weighted replacement method is used, where the replacement value is the average of the statistical mean of all normal data in that modality and the result of the filling logic.

[0025] Furthermore, the cross-modal feature fusion of the fusion module is based on a dual-driven attention mechanism of modality contribution and feature correlation, and performs deep fusion in conjunction with Transformer encoding, with attention weights for each modality. ;

[0026] In the formula: Let be the attention weights for the j-th modality; The contribution coefficient for the j-th mode; The average feature mutual information of the j-th mode with all other modes; This represents the total number of multi-source heterogeneous modes.

[0027] Furthermore, the user segmentation labeling in the construction module is based on compliance attribute information from a high-dimensional fusion feature set to determine compliance priority weights. Then, user business preference features and behavioral temporal pattern features are applied, and an adaptive density clustering algorithm is used to divide users into several groups with similar characteristics. For each group, core features are extracted to generate a multi-level label system, including compliance level labels, business preference labels, and behavioral pattern labels. The weights of each label are as follows:

[0028] ;

[0029] In the formula: The final weight of the label; Priority coefficient for compliance attributes; Compliance priority weight; The association weight between tags and compliance attributes; The association weight between the label and the business behavior characteristics.

[0030] Furthermore, the process for capturing the dynamic changes in user business behavior data in the early warning module is as follows:

[0031] S1 performs time-series normalization on the preprocessed user business behavior data and aligns the data according to a uniform time granularity;

[0032] S2 adaptively divides time-series windows based on user behavior activity, so that each window contains complete behavioral cycle features.

[0033] S3 extracts the core features of business behavior within each time window;

[0034] S4 uses linear regression to fit the core behavioral feature sequence within each time window and calculates the slope of the user behavior trend.

[0035] S5 calculates the Euclidean distance between the behavior feature vectors of the current time series window and the previous time series window, and performs normalization processing simultaneously, which is denoted as the behavior mutation coefficient.

[0036] S6 integrates the slope of behavioral trends with the coefficient of behavioral mutation to form a dynamic feature set of user business behavior data.

[0037] The trigger condition for the user churn risk warning is as follows:

[0038] ;

[0039] In the formula: This is the result of a risk warning for customer attrition. The slope of the user behavior trend within the current time series window; The preset minimum threshold for the trend slope; The coefficient of behavioral mutation; This is the preset maximum threshold for the behavioral mutation coefficient;

[0040] When the warning trigger conditions are met, a profile update trigger signal is generated, and churn risk warning information is output simultaneously. The warning information is based on a preset warning level threshold and... The comparison determines the warning level.

[0041] Furthermore, after receiving the latest dynamic compliance user profile, the recommendation module generates targeted business recommendations based on the user's business preferences, behavioral patterns, and compliance attribute information corresponding to the latest dynamic compliance user profile:

[0042] Each user's business preferences, behavioral patterns, and compliance attributes are manually pre-bound to a unique business list by the user on the system side. Based on the latest dynamic compliance user profile, the system queries the frequency of occurrence of each matching business list according to the user's business preferences, behavioral patterns, and compliance attributes, and pushes the business list with the most occurrences to the user as the targeted business recommendation target.

[0043] Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects:

[0044] This invention performs compliant anonymization processing on multi-source heterogeneous data, ensuring data security while preserving semantic relevance. Through filling missing values, accurately detecting and correcting outliers, and standardization, it effectively improves data quality. Furthermore, by combining deep fusion of cross-modal features and precise weight allocation, it comprehensively captures user business preferences, behavioral patterns, and compliance attributes, constructing dynamically adapted user profiles. It captures dynamic changes in behavior in real time to warn of churn risks and ensures information timeliness by synchronously updating the profiles. Based on these profiles, it generates targeted business recommendations, effectively improving user satisfaction and business operation efficiency. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0046] Figure 1 A schematic diagram of the structure of a dynamic compliant user profile construction and recommendation system for broadband services. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0048] The present invention will be further described below with reference to embodiments.

[0049] Example:

[0050] This embodiment presents a system for constructing and recommending dynamic compliant broadband service user profiles, such as... Figure 1 As shown, it includes:

[0051] The data acquisition module is used to collect multi-source heterogeneous modal data of users in broadband service scenarios and to perform de-identification processing on the collected data.

[0052] The desensitization process in the data acquisition module includes:

[0053] For core sensitive fields in user structured attribute data, a desensitization method combining feature masking and semantic replacement is used;

[0054] For network behavior log data, business usage trajectory data, and terminal-related interaction data, a dynamic perturbation desensitization method is used. The perturbation amplitude is adaptively adjusted according to the modality type of the data, so that the desensitized data not only meets the data security compliance requirements, but also retains the semantic features of cross-modal data. Modality types include numerical, text, and time series.

[0055] For numerical data, the perturbation amplitude is adaptively adjusted based on the data's own distribution characteristics to ensure that the data still conforms to the original statistical distribution pattern after perturbation. For textual data, the perturbation amplitude is constrained by the semantic preservation threshold, and perturbation is achieved through synonym replacement and sentence structure fine-tuning without changing the core business semantics. For time-series data, the perturbation amplitude is dynamically adapted based on the time series stationarity index, and a small offset adjustment is made in the time dimension to avoid destroying the temporal correlation characteristics of behavior.

[0056] The preprocessing module is used to perform missing value imputation and outlier detection on the desensitized multi-source heterogeneous modal data, complete the standardization and cross-modal dimension adaptation based on the semantic attributes of each modal data, and output semantic preprocessed data with adapted feature fusion.

[0057] Missing value imputation in the preprocessing module follows the following rules:

[0058] ;

[0059] In the formula: Fill in the missing values ​​for the j-th modality data of the i-th user; This represents the total number of multi-source heterogeneous modes; Let be the cross-modal semantic similarity between the j-th modality and the k-th modality; The confidence weight for the k-th mode; For the i-th user and k-th modality, the valid data value is...

[0060] The above formula introduces the semantic similarity between each modality and the target modality, and combines the data integrity ratio, data accuracy verification results and data noise content of the modality itself to determine the confidence weight of each modality in a weighted manner. Finally, based on these weights, the effective data is integrated and calculated to complete the accurate filling of missing values. This not only takes into account the correlation between data of different modalities, but also highlights the impact of data quality on the filling results.

[0061] in, , , , This represents the weighting coefficient of each influencing factor. , , Both are greater than zero, and , , The sum is 1. This represents the data integrity ratio of the k-th modality, calculated as the ratio of the amount of valid non-missing data in that modality to the total amount of data in that modality. This represents the data accuracy verification result for the k-th modality, which is calculated by combining the cross-validation accuracy rate between the data of this modality and the business benchmark data, and the pass rate of the data logic consistency verification. This represents the data noise content of the k-th mode;

[0062] During the preprocessing module's execution phase, a cross-modal semantic mapping matrix is ​​constructed based on the semantic attributes of each modality's data. Matrix elements This represents the semantic association strength between the q-th feature of the p-th modality and the target unified dimensional feature. Then, the dimensional importance score of each modality feature is calculated using information gain. The greater the information gain, the greater the contribution of the feature to business analysis and compliance judgment. Finally, core features are selected based on the comparison between the dimension importance score and the preset screening threshold, and the core features are simultaneously adapted and transformed for dimension matching.

[0063] ;

[0064] In the formula: The unified dimension feature matrix after adaptation; The feature matrix after standardization of each modality data; This is a dimension importance weight matrix, where the diagonal elements are the importance scores of the corresponding features. The normalized result has 0 for off-diagonal elements.

[0065] The above formula constructs a cross-modal semantic mapping matrix based on the semantic attributes of each modal data. It quantifies the contribution of each modal feature to business analysis and compliance judgment through information gain. The normalized feature importance score is used as the core element of the weight matrix and is operated with the standardized feature matrix. This not only achieves the dimensionality unification of different modal features, but also highlights the role of core features in subsequent analysis, ensuring the pertinence and effectiveness of feature adaptation.

[0066] In the preprocessing module, the standardization and normalization process uses Z-score standardization to eliminate dimensional differences for numerical data, L2 normalization of word embedding vectors to achieve scale uniformity for text data, and standardization based on the mean and variance of the sliding time window for time series data.

[0067] The outlier detection in the preprocessing module follows the following rules:

[0068] ;

[0069] In the formula: This represents the anomaly determination result for the j-th modal data of the i-th user, where 1 indicates that it is an anomaly and 0 indicates that it is normal data. The statistical deviation of the j-th modal data of the i-th user relative to the feature distribution of that modality; Let be the semantic deviation coefficient of the j-th modality; Preset anomaly detection threshold;

[0070] The above formula calculates the statistical deviation of data relative to the feature distribution of its modality, and introduces a semantic deviation coefficient based on the semantic consistency between this modality and other user modal data. The product of the two is compared with a preset anomaly judgment threshold to achieve accurate identification of outliers. For anomalous data of different degrees, a modified benchmark weighted average based on cross-modal semantic association and a modal confidence weighted replacement are adopted respectively to ensure the rationality of outlier judgment and to retain the effective information of the data to the greatest extent.

[0071] like ∈( , ], To determine a condition as moderately abnormal based on a preset threshold, cross-modal semantic similarity is used. Modal confidence weights The top N modalities with the highest semantic relevance are selected to construct a correction baseline, and the correction value is calculated by weighted average; if > If it is determined to be a serious anomaly, a modal confidence weighted replacement method is adopted, and the replacement value is the average of the statistical mean of all normal data under this modality and the result of the filling logic.

[0072] in, >0 indicates that the coefficient is calculated based on the semantic consistency between this modality data and other user modality data. The lower the semantic consistency, the larger the coefficient, reflecting the degree of deviation between the data and the user's overall behavioral characteristics.

[0073] The fusion module is used to perform cross-modal feature association and deep fusion of semantic preprocessed data based on Transformer encoding and attention weight allocation, and generate a high-dimensional fusion feature set with a unified dimension.

[0074] The cross-modal feature fusion module is based on a dual-driven attention mechanism of modality contribution and feature correlation, and performs deep fusion in conjunction with Transformer encoding, with attention weights for each modality. ;

[0075] In the formula: Let be the attention weights for the j-th modality; The contribution coefficient for the j-th mode; The average feature mutual information of the j-th mode with all other modes; This represents the total number of multi-source heterogeneous modes;

[0076] The above formula uses the product of the information entropy of each modality and the modality importance score as the modality contribution coefficient. At the same time, it calculates the average feature mutual information of each modality and all other modalities. The product of the two is used to quantify the comprehensive influence of each modality in feature fusion. Then, it is normalized based on the sum of the comprehensive influence of all modalities to obtain the attention weight of each modality. This realizes the dual driving force of modality contribution and feature relevance, making the weight allocation of each modality in the feature fusion process more in line with the needs of business analysis and compliance judgment.

[0077] in, It is obtained by multiplying the information entropy of the j-th modality with the modality importance score. The modality importance score is determined based on the degree of correlation between the modality data and the business analysis objectives and compliance attributes. The modality importance score is constrained to be within the range of (0,1).

[0078] The module is used to segment and label users based on a high-dimensional fusion feature set, extract user business preferences, behavioral time sequence patterns and compliance attribute information, and build dynamic compliant user profiles.

[0079] In the construction module, user segmentation labels are based on compliance attribute information from a high-dimensional fusion feature set. Compliance priority weights are determined based on the stringency of compliance requirements and the level of violation risk. Then, user business preference features and behavioral temporal patterns are applied, and an adaptive density clustering algorithm divides users into several groups with similar characteristics. For each group, core features are extracted to generate a multi-level label system, including compliance level labels, business preference labels, and behavioral pattern labels. The weights of each label are as follows:

[0080] ;

[0081] In the formula: The final weight of the label; Priority coefficient for compliance attributes; Compliance priority weight; The association weight between tags and compliance attributes; The association weight between tags and business behavior characteristics;

[0082] The above formula obtains the compliance priority weight by normalizing the weighted sum of the quantitative values ​​of the strictness of compliance requirements and the quantitative values ​​of the level of violation according to preset weights. Then, it combines the cosine similarity between the tag and the compliance attribute, and the normalized result of the mutual information between the tag and the business behavior characteristics, with the compliance attribute priority coefficient as the adjustment, and obtains the final tag weight by weighted summation. This ensures that the compliance attribute has the core position in the tag system, while also taking into account the impact of business behavior characteristics on the tag weight. This makes the tag weight allocation more in line with the core needs of user profile construction. The calculation of the compliance priority weight comprehensively considers the legal effect and standard effect of the compliance clauses, as well as the probability of violation and the severity of the consequences, ensuring the comprehensiveness and objectivity of the compliance priority determination.

[0083] Among them, user service preference features are a set of features extracted from multi-source heterogeneous modal data (service usage trajectory data, terminal-related interaction data, network behavior log data, etc.) that reflect users' core demand tendencies and selection preferences for broadband services; behavioral temporal pattern features are a set of features extracted from the temporal attributes of network behavior log data and service usage trajectory data that reflect the regularity of users' use of broadband services in the time dimension.

[0084] Among them, the core features of the compliance dimension include the user's compliance qualification status, the frequency and severity of historical violations, and the degree of matching between business usage behavior and compliance rules; the core features of the business preference dimension include the types of broadband services used frequently, the proportion of service usage time, service activation and renewal preferences, and the tendency to choose value-added services; the core features of the behavioral time sequence dimension include the distribution of peak service usage periods, the periodic patterns of behavior occurrence, the time sequence change trend of business interaction frequency, and the time sequence coordination characteristics of terminal-related usage.

[0085] The result is obtained by weighting and summing the quantitative values ​​of the stringency of compliance requirements and the quantitative values ​​of the violation risk level according to preset weights, and then normalizing the sum within the range of [0.6, 0.9].

[0086] Calculate the cosine similarity between the feature vector corresponding to the label and the core feature vector of the compliance attribute; The mutual information value of the combined feature vector of tag features and user business behavior is calculated and normalized to the [0,1] interval.

[0087] ;

[0088] In the formula: a and b are preset weighting ratios, and their sum is 1; S is the quantitative value of the strictness of compliance requirements, which is weighted according to the legal effect level of the compliance clauses (law > administrative regulations > industry standards > corporate norms) and the standard effect level (mandatory national standards > recommended national standards > group standards > enterprise standards); R is the quantitative value of the violation risk level, which is the product of the probability of violation P and the severity of the consequences of violation Q, where Q is assigned according to the degree of impact of the violation on business compliance and user rights;

[0089] The early warning module is used to capture the dynamic changes in user business behavior data through LSTM time series modeling, perform real-time early warning analysis of user churn risk, and synchronously generate profile update trigger signals.

[0090] The process for capturing dynamic changes in user business behavior data in the early warning module is as follows:

[0091] S1 performs time-series normalization on the preprocessed user business behavior data and aligns the data according to a uniform time granularity;

[0092] S2 adaptively divides time-series windows based on user behavior activity, so that each window contains complete behavioral cycle features.

[0093] S3 extracts the core features of business behavior within each time window, including behavior frequency, behavior duration, business type proportion, and compliance rate.

[0094] S4 uses linear regression to fit the core behavioral feature sequence within each time window and calculates the slope of the user behavior trend.

[0095] S5 calculates the Euclidean distance between the behavior feature vectors of the current time series window and the previous time series window, and performs normalization processing simultaneously, which is denoted as the behavior mutation coefficient.

[0096] S6 integrates the slope of behavioral trends with the coefficient of behavioral mutation to form a dynamic feature set of user business behavior data.

[0097] The conditions for triggering a user churn risk warning are:

[0098] ;

[0099] In the formula: This is the result of a risk warning for customer attrition. The slope of the user behavior trend within the current time series window; The preset minimum threshold for the trend slope; The coefficient of behavioral mutation; This is the preset maximum threshold for the behavioral mutation coefficient;

[0100] The above formula fits the core feature sequence of business behavior within the time window through linear regression to obtain the behavior trend slope that reflects the changing trend of user behavior activity. At the same time, it calculates the Euclidean distance between the behavior feature vectors of the current time window and the previous time window and normalizes it to obtain the behavior mutation coefficient. The two are compared with the preset minimum trend slope threshold and maximum behavior mutation coefficient threshold, respectively. When the behavior trend is declining and the mutation degree exceeds the limit, a churn risk warning is triggered. This formula can accurately capture the dynamic change characteristics of user business behavior and generate a timely profile update trigger signal, providing real-time optimization support for dynamic and compliant user profiles.

[0101] When the warning trigger conditions are met, a profile update trigger signal is generated, and churn risk warning information is output simultaneously. The warning information is based on a preset warning level threshold and... The comparison determines the warning level;

[0102] in, A negative value indicates a declining trend in behavioral activity. The content of the profile update trigger signal is to remove low-value historical data collected by the collection module up to the last stage of building the dynamic compliance user profile in the construction module. Low-value historical data is determined by the weighted result of the data timeliness score and the compliance feature correlation score. Data below the preset screening threshold is judged as low-value data. After the trigger signal is generated, it jumps to the preprocessing module. Based on the remaining historical data after removing low-value data and the new data accumulated in real time by the collection module, the construction of the dynamic compliance user profile is executed again. The cumulative time window of the new data is adaptively adjusted according to the user behavior update frequency.

[0103] The recommendation module is used to generate targeted business recommendations based on the latest dynamic and compliant user profiles;

[0104] After receiving the latest dynamic compliance user profiles, the recommendation module generates targeted business recommendations based on the user's business preferences, behavioral patterns, and compliance attributes corresponding to these profiles.

[0105] Each user's business preferences, behavioral patterns, and compliance attributes are manually pre-bound to a unique business list by the user on the system side. Based on the latest dynamic compliance user profile, the system queries the frequency of occurrence of each matching business list according to the user's business preferences, behavioral patterns, and compliance attributes. The business list with the most occurrences is used as the target business recommendation to be pushed to the user.

[0106] Among them, multi-source heterogeneous modal data includes user structured attribute data, network behavior log data, service usage trajectory data, and terminal-related interaction data;

[0107] The data acquisition module interacts with the preprocessing module via a wireless network. The preprocessing module interacts with the fusion module via a wireless network. The fusion module interacts with the construction module via a wireless network. The construction module interacts with the early warning module via a wireless network. The early warning module interacts with the recommendation module via a wireless network.

[0108] In this embodiment, the acquisition module collects multi-source heterogeneous modal data of users in broadband service scenarios and performs anonymization processing on the collected data. The preprocessing module runs afterward to perform missing value imputation and outlier detection on the anonymized multi-source heterogeneous modal data. Based on the semantic attributes of each modal data, it completes standardization and cross-modal dimension adaptation, and outputs semantic preprocessed data with adapted feature fusion. The fusion module simultaneously completes cross-modal feature association and deep fusion of the semantic preprocessed data based on Transformer encoding and attention weight allocation, generating a unified high-dimensional fusion feature set. The construction module further uses the high-dimensional fusion feature set to segment and label users, extract user business preferences, behavioral time series patterns and compliance attribute information, and construct a dynamic compliant user profile. Then, the early warning module captures the dynamic change features of user business behavior data through LSTM time series modeling, performs real-time early warning analysis of user churn risk, and synchronously generates profile update trigger signals. Finally, the recommendation module generates targeted business recommendations based on the latest dynamic compliant user profile.

[0109] The system described in the above embodiments can securely process multiple types of user data, accurately capture user compliance status, business preferences and usage patterns, provide timely warnings of churn risks, dynamically optimize user profiles, and push services that meet user needs. This not only ensures data compliance and security but also enhances user experience and helps improve business adaptability and conversion efficiency.

[0110] Application example:

[0111] To improve the accuracy of service recommendations, reduce customer churn, and ensure business compliance, broadband operators in XX province have adopted this system to optimize customer service. Specific applications are as follows:

[0112] I. Data Collection and Desensitization

[0113] The operator collects multi-source heterogeneous modal data from 5 million broadband users across the province through the system's data acquisition module: user structured attribute data (including name, ID number, package level, etc.), network behavior log data (such as daily internet access time, download speed, website categories visited, etc.), service usage trajectory data (package renewal records, value-added service activation status, etc.), and terminal-related interaction data (type of terminal connected to broadband, percentage of terminal usage time, etc.).

[0114] During the desensitization process, core sensitive fields such as ID card numbers and mobile phone numbers are handled using a combination of feature masking and semantic replacement (e.g., replacing the middle 8 digits of the ID card number with "********"). For network behavior logs, business usage trajectories, and terminal-related interaction data, dynamic perturbation desensitization is employed: numerical download speed data remains consistent with the original statistical distribution after perturbation; text-based website category data is perturbed by replacing "information" with synonyms such as "news" without altering the core semantics; and time-series internet access time data undergoes minor time offset adjustments to preserve the temporal correlation of behaviors.

[0115] II. Data Preprocessing

[0116] The preprocessing module first handles missing values. For example, if a user's value-added service usage data is missing, it combines the user's other modal data (package level, internet access time, terminal type) with the cross-modal semantic similarity of the value-added service modality, as well as the confidence weight of each modality (calculated by comprehensively considering data completeness, accuracy, and noise content, such as the confidence weight of the package level modality being 0.85), and finally calculates a reasonable filler value.

[0117] During outlier detection, a user was found to have a daily download speed of 1000Mbps, far exceeding the normal speed distribution corresponding to their data plan. After calculating the product of the statistical deviation and the semantic deviation coefficient, the user was determined to be moderately abnormal. A correction benchmark was constructed by selecting the three modal data with the highest semantic correlation: data plan level, terminal supported speed, and historical average speed. The corrected value was obtained by weighted averaging. Another user's compliance qualification status data had extremely low semantic consistency with other modal data, and the calculation results showed it to be severely abnormal. The statistical mean of all normal user data in this modality and the mean of the padding logic were used as replacement values.

[0118] In the standardization phase, numerical data such as internet usage duration and download speed are standardized using Z-score; textual data such as service type descriptions are normalized using L2 after word embedding; and time-series data such as daily internet usage time distribution are standardized based on the mean and variance of a sliding time window. Subsequently, a cross-modal semantic mapping matrix is ​​constructed, the information gain of each feature is calculated, and core features such as package type, peak internet usage times, and compliance status are selected. After dimensionality adaptation and transformation, a feature matrix with a unified dimension is output.

[0119] III. Feature Fusion and Profile Construction

[0120] The fusion module is based on Transformer encoding and attention mechanism to calculate the attention weight of each modality: the contribution coefficient of package data (the product of information entropy and modality importance score) is high, and the average feature mutual information with other modalities is strong, so it gets a high weight (e.g. 0.32). The weight of network behavior log data is 0.28, the weight of terminal association data is 0.21, and the weight of structured attribute data is 0.19. A high-dimensional fusion feature set is generated through weighted fusion.

[0121] In the construction module, the compliance priority weight is first set: combining the legal requirements for real-name authentication (a high level of strictness quantification) and the user's historical overdue payment and violation records (the risk level quantification is obtained by multiplying the probability of violation and the severity of the consequences), the compliance priority weight is calculated to be 0.83. An adaptive density clustering algorithm is used to divide users into six groups: "high compliance, high frequency audio-visual users," "compliant, stable office users," and "low compliance, occasional internet users." For the "high compliance, high frequency audio-visual users" group, core features are extracted to generate multi-level labels: compliance level label (Level A compliance, weight 0.91), business preference label (4K movie preference, weight 0.82), and behavioral pattern label (peak internet usage between 8-10 PM, weight 0.87). The label weights are calculated using the compliance attribute priority coefficient (0.75, obtained by normalizing the quantification of compliance strictness and risk level), the compliance priority weight, and the correlation between the label and compliance / business behavior.

[0122] IV. Churn Warning and Business Recommendation

[0123] The early warning module organizes user business behavior data by day and time sequence, adaptively dividing it into 7-day time windows based on behavioral activity. It extracts core features such as internet frequency, duration, and compliance rate for each window, fits the behavioral trend slope, calculates the Euclidean distance between the behavioral feature vectors of the current window and the previous window, and normalizes it to obtain the behavioral mutation coefficient. For example, a user with "Level A compliance + 4K movie preference" has a 7-day internet duration trend slope of -0.3 (below the preset minimum threshold of -0.2) and a behavioral mutation coefficient of 0.85 (above the preset maximum threshold of 0.7). The system triggers a churn warning, determines it as a high-risk churn user, generates a profile update trigger signal, removes low-value historical login data from three months ago, and combines it with the latest data on non-renewed 4K membership data from the past month to reconstruct the user profile.

[0124] The recommendation module receives the updated user profile and queries the pre-bound service list: the service list corresponding to the combination of "Level A compliance + 4K movie preference + peak internet usage time of 8-10 pm" appears most frequently. The system then pushes this targeted service recommendation to the user and simultaneously displays the compliance rights protection statement to improve user acceptance.

[0125] In summary, the system in the above embodiments ensures data security while preserving semantic relationships by performing compliant de-identification processing on multi-source heterogeneous data. Through filling missing values, accurately detecting and correcting outliers, and standardizing and regularizing, it effectively improves data quality. Furthermore, by combining deep fusion of cross-modal features and precise weight allocation, it comprehensively captures user business preferences, behavioral temporal patterns, and compliance attributes, constructs dynamically adapted user profiles, captures dynamic changes in behavior in real time to warn of churn risks, and ensures information timeliness by synchronously updating the profiles. Based on the profiles, it generates targeted business recommendations, effectively improving user satisfaction and business operation efficiency.

[0126] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A system for constructing and recommending dynamic compliant user profiles for broadband services, characterized in that, include: The data acquisition module is used to collect multi-source heterogeneous modal data of users in broadband service scenarios and to perform de-identification processing on the collected data. The preprocessing module is used to perform missing value imputation and outlier detection on the desensitized multi-source heterogeneous modal data, complete the standardization and cross-modal dimension adaptation based on the semantic attributes of each modal data, and output semantic preprocessed data with adapted feature fusion. The fusion module is used to perform cross-modal feature association and deep fusion of semantic preprocessed data based on Transformer encoding and attention weight allocation, and generate a high-dimensional fusion feature set with a unified dimension. The module is used to segment and label users based on a high-dimensional fusion feature set, extract user business preferences, behavioral time sequence patterns and compliance attribute information, and build dynamic compliant user profiles. The early warning module is used to capture the dynamic changes in user business behavior data through LSTM time series modeling, perform real-time early warning analysis of user churn risk, and synchronously generate profile update trigger signals. The recommendation module is used to generate targeted business recommendations based on the latest dynamic and compliant user profiles; Among them, multi-source heterogeneous modal data includes user structured attribute data, network behavior log data, service usage trajectory data, and terminal-related interaction data.

2. The broadband service dynamic compliance user profile construction and recommendation system according to claim 1, characterized in that, The desensitization process in the acquisition module includes: For core sensitive fields in user structured attribute data, a desensitization method combining feature masking and semantic replacement is used; For network behavior log data, business usage trajectory data, and terminal-related interaction data, a dynamic perturbation desensitization method is used, and the perturbation amplitude is adaptively adjusted according to the data modality type.

3. The broadband service dynamic compliance user profile construction and recommendation system according to claim 1, characterized in that, The missing value imputation in the preprocessing module follows the following rules: ； In the formula: Fill in the missing values ​​for the j-th modality data of the i-th user; This represents the total number of multi-source heterogeneous modes; Let be the cross-modal semantic similarity between the j-th modality and the k-th modality; The confidence weight for the k-th mode; This represents the valid data value for the i-th user in the k-th modality.

4. The broadband service dynamic compliance user profile construction and recommendation system according to claim 3, characterized in that, During the preprocessing module's execution phase, a cross-modal semantic mapping matrix is ​​constructed based on the semantic attributes of each modality's data. Matrix elements This represents the semantic association strength between the q-th feature of the p-th modality and the target unified dimensional feature. Then, the dimensional importance score of each modality feature is calculated using information gain. Finally, core features are selected based on the comparison between dimensional importance scores and preset screening thresholds, and dimensional adaptation transformation is performed on the core features simultaneously. ； In the formula: The unified dimension feature matrix after adaptation; The feature matrix after standardization of each modality data; This is the dimension importance weight matrix.

5. The broadband service dynamic compliance user profile construction and recommendation system according to claim 3, characterized in that, The outlier detection in the preprocessing module follows the following rules: ； In the formula: This represents the anomaly determination result for the j-th modal data of the i-th user, where 1 indicates that it is an anomaly and 0 indicates that it is normal data. The statistical deviation of the j-th modal data of the i-th user relative to the feature distribution of that modality; Let be the semantic deviation coefficient of the j-th modality; Preset anomaly detection threshold; like ∈( , ], To determine a condition as moderately abnormal based on a preset threshold, cross-modal semantic similarity is used. Modal confidence weights The top N modalities with the highest semantic relevance are selected to construct a correction baseline, and the correction value is calculated by weighted average; if > If the data is deemed severely abnormal, a modal confidence-weighted replacement method is used, where the replacement value is the average of the statistical mean of all normal data in that modality and the result of the filling logic.

6. The broadband service dynamic compliance user profile construction and recommendation system according to claim 1, characterized in that, The cross-modal feature fusion of the fusion module is based on an attention mechanism driven by both modality contribution and feature correlation, and performs deep fusion in conjunction with Transformer encoding, with attention weights for each modality. ; In the formula: Let be the attention weights for the j-th modality; The contribution coefficient for the j-th mode; The average feature mutual information of the j-th mode with all other modes; This represents the total number of multi-source heterogeneous modes.

7. The broadband service dynamic compliance user profile construction and recommendation system according to claim 1, characterized in that, In the construction module, the user group label is based on the compliance attribute information in the high-dimensional fusion feature set to determine the compliance priority weight. Then, user business preference features and behavioral time sequence pattern features are applied, and users are divided into several groups with similar features through an adaptive density clustering algorithm. For each subgroup, core features are extracted to generate a multi-level tagging system, including compliance level tags, business preference tags, and behavioral pattern tags. The weights of each tag are as follows: ； In the formula: The final weight of the label; Priority coefficient for compliance attributes; Compliance priority weight; The association weight between tags and compliance attributes; The association weight between the label and the business behavior characteristics.

8. The broadband service dynamic compliance user profile construction and recommendation system according to claim 1, characterized in that, The process for capturing dynamic changes in user business behavior data in the early warning module is as follows: S1 performs time-series normalization on the preprocessed user business behavior data and aligns the data according to a uniform time granularity; S2 adaptively divides time-series windows based on user behavior activity, so that each window contains complete behavioral cycle features. S3 extracts the core features of business behavior within each time window; S4 uses linear regression to fit the core behavioral feature sequence within each time window and calculates the slope of the user behavior trend. S5 calculates the Euclidean distance between the behavior feature vectors of the current time series window and the previous time series window, and performs normalization processing simultaneously, which is denoted as the behavior mutation coefficient. S6 integrates the slope of behavioral trends with the coefficient of behavioral mutation to form a dynamic feature set of user business behavior data. The trigger condition for the user churn risk warning is as follows: ； In the formula: This is the result of a risk warning for customer attrition. The slope of the user behavior trend within the current time series window; The preset minimum threshold for the trend slope; The coefficient of behavioral mutation; This is the preset maximum threshold for the behavioral mutation coefficient; When the warning trigger conditions are met, a profile update trigger signal is generated, and churn risk warning information is output simultaneously. The warning information is based on a preset warning level threshold and... The comparison determines the warning level.

9. A broadband service dynamic compliance user profile construction and recommendation system according to claim 1, characterized in that, After receiving the latest dynamic compliance user profile, the recommendation module generates targeted business recommendations based on the user's business preferences, behavioral patterns, and compliance attribute information corresponding to the latest dynamic compliance user profile. Each user's business preferences, behavioral patterns, and compliance attributes are manually pre-bound to a unique business list by the user on the system side. Based on the latest dynamic compliance user profile, the system queries the frequency of occurrence of each matching business list according to the user's business preferences, behavioral patterns, and compliance attributes, and pushes the business list with the most occurrences to the user as the targeted business recommendation target.

10. A broadband service dynamic compliance user profile construction and recommendation system according to claim 1, characterized in that, The data acquisition module is interconnected with a preprocessing module via a wireless network. The preprocessing module is interconnected with a fusion module via a wireless network. The fusion module is interconnected with a construction module via a wireless network. The construction module is interconnected with an early warning module via a wireless network. The early warning module is interconnected with a recommendation module via a wireless network.

Images