A Smart Library Management Method

Abstract

This invention discloses a smart library management method, relating to the field of data management technology. The method includes: aggregating local training parameters and behavior monitoring logs to an aggregation center; performing secure aggregation calculations using differential privacy methods and drift detection weights to generate a global parameter package and a unified drift signal; distributing the global parameter package and unified drift signal to distributed book data terminals; and generating personalized recommendation results and explanations by combining real-time collected search requests and borrowing behavior; monitoring the performance of personalized recommendation results and explanations in real time, and identifying changes in user interests through concept drift detection to generate drift alerts and affected feature sets; prioritizing the drift alerts and affected feature sets to generate updated local training parameters; and performing privacy compliance checks with behavior monitoring logs to generate a privacy compliance report and book recommendation optimization suggestions. This invention improves recommendation accuracy and enhances adaptability to concept drift.

Description

Technical Field

[0001] This invention relates to the field of data management technology, and in particular to a smart library management method. Background Technology

[0002] In the process of modern library informatization, library management technology is gradually developing towards digitalization and intelligence. Traditional library management mainly relies on manual registration, borrowing records, and simple database queries for business processing. However, with the increase in the number of collections and the diversification of user needs, data-driven smart library management technology is gradually emerging. Among related technologies, information collection, user behavior analysis, and the application of recommendation algorithms have become research hotspots. By collecting users' borrowing history, search logs, and interest tags, personalized recommendations and collection optimization have become feasible. At the same time, the development of distributed data processing and cloud computing technologies enables libraries to extract features, model, and analyze massive amounts of business data while ensuring data security and privacy, thereby improving library service efficiency and user experience.

[0003] However, existing technologies still have certain limitations in personalized recommendations and resource optimization. Dynamic changes in user interests may lead to recommendations that do not perfectly match actual needs. Existing methods typically lack real-time drift detection and rapid response mechanisms, making it difficult to adjust recommendation strategies promptly when user interests change. This limits the accuracy and real-time performance of personalized recommendations to some extent, and also impacts the efficiency of library resource scheduling and optimization. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a smart library management method to solve the problems of recommendation accuracy and resource optimization.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a smart library management method, which includes:

[0008] The original library business data is centrally collected and registered, and encrypted and anonymized in accordance with privacy protection rules to generate a manageable data list;

[0009] Using a manageable data list, feature extraction, format standardization, and sensitive information removal are performed on the original library business data. The local training model is initialized, and local training parameters and behavior monitoring logs are generated.

[0010] Local training parameters and behavior monitoring logs are aggregated to the aggregation center, and a global parameter package and a unified drift signal are generated by using differential privacy methods and drift detection weights for secure aggregation calculation.

[0011] The global parameter package and unified drift signal are distributed to the distributed book data terminal, and personalized recommendation results and explanations are generated by combining real-time collected search requests and borrowing behavior.

[0012] Real-time monitoring of personalized recommendation results and explanations; and identification of changes in user interests through concept drift detection, generating drift alerts and affected feature sets.

[0013] Prioritize the drift alerts and affected feature sets for training, generate updated local training parameters, and perform privacy compliance checks with behavior monitoring logs to generate a privacy compliance report and book recommendation optimization suggestions.

[0014] As a preferred embodiment of the smart library management method of the present invention, the privacy protection rules are obtained based on information security specifications, industry standards and library internal data management strategies.

[0015] In a preferred embodiment of the smart library management method of the present invention, the specific steps for generating a manageable data list are as follows:

[0016] The library collects information on its collections, borrowing records, search logs, and user interest tags in real time, and performs preliminary labeling, timestamp recording, and sensitivity classification according to a unified format to generate raw library business data.

[0017] In accordance with privacy protection rules, the original library business data is encrypted, anonymized, and desensitized. At the same time, duplicate records are removed, missing fields are filled, and integrity is verified to generate a qualified dataset.

[0018] The validated datasets are indexed and organized to generate a manageable data list.

[0019] As a preferred embodiment of the smart library management method of the present invention, the specific steps for generating local training parameters and behavior monitoring logs are as follows:

[0020] Based on the manageable data list, features are extracted from the original library business data and transformed into structured feature vectors to form a preliminary feature dataset;

[0021] The preliminary feature dataset is subjected to field standardization, numerical normalization, and sensitive information desensitization, and integrity verification is performed to generate a standardized feature dataset.

[0022] The local training model is initialized based on the standardized feature dataset, and local training and security verification are performed. Local training parameters are generated, and the behavior monitoring information generated during the training process is summarized to obtain behavior monitoring logs.

[0023] As a preferred embodiment of the intelligent library management method of the present invention, the specific steps for generating the global parameter package and the unified drift signal are as follows:

[0024] Send local training parameters and behavior monitoring logs to the aggregation center, and perform preliminary sorting and classification according to the source of the local training model to generate a node parameter summary table;

[0025] Based on the node parameter summary table, a secure aggregation operation is performed on the local training parameters. The weighted average combined with the differential privacy noise method is used to perform secure aggregation calculation on the local training parameters to generate a global parameter package.

[0026] By utilizing the drift weight information from local training parameters and behavior monitoring logs, the global parameter package is adjusted for drift sensitivity to generate a unified drift signal.

[0027] In a preferred embodiment of the smart library management method of the present invention, the real-time collected search requests and borrowing behaviors are obtained through user operation logs automatically recorded by the library's back-end service.

[0028] In a preferred embodiment of the intelligent library management method of the present invention, the specific steps for generating personalized recommendation results and explanations are as follows:

[0029] The global parameter package and unified drift signal are distributed to the distributed book data terminal, and combined with the standardized feature dataset for mapping processing to generate a node-level parameter mapping table.

[0030] Based on the node-level parameter mapping table and real-time collected search requests and borrowing behavior, the system performs matching analysis on users' historical behavior and preferences to generate a preliminary personalized recommendation candidate set. Then, it uses a unified drift signal to adjust the drift sensitivity and generate personalized recommendation results and explanations.

[0031] In a preferred embodiment of the intelligent library management method of the present invention, the specific steps for generating drift alarms and affected feature sets are as follows:

[0032] Real-time collection of user interaction data with personalized recommendation results, and correlation with recommendation explanations to generate a recommendation performance dataset;

[0033] Perform indicator analysis and anomaly detection on the recommendation performance dataset, calculate the actual performance deviation, and generate a recommendation performance deviation report;

[0034] By combining recommendation performance deviation reports with historical behavioral characteristics, and using a concept drift detection algorithm to identify trends in user interest changes, interest drift signals are generated.

[0035] Based on the interest drift signal, identify the affected feature set and the affected recommendation category, and generate drift alerts and the affected feature set.

[0036] In a preferred embodiment of the intelligent library management method of the present invention, the specific steps for generating updated local training parameters are as follows:

[0037] Based on the drift alerts, the affected feature sets are prioritized according to feature impact weight and drift severity, and a feature priority list is generated.

[0038] Based on the feature priority list, adjust the high-priority features to perform incremental training on the local training model, and retain the original training parameters as a basis to generate preliminary updated model parameters.

[0039] The initial updated model parameters are subjected to security and integrity verification, abnormal parameters are removed, and qualified local training parameters are generated.

[0040] The verified local training parameters are correlated and analyzed with the behavior monitoring logs to generate updated local training parameters.

[0041] As a preferred embodiment of the smart library management method described in this invention, the specific steps for generating the privacy compliance report and book recommendation optimization suggestions are as follows:

[0042] The sensitive features and behavior monitoring logs in the updated local training parameters are matched to generate a preliminary privacy review dataset.

[0043] By using privacy protection rules, the preliminary privacy review dataset is minimized, anonymized, differentially private, and access rights verified to identify potential compliance risks and generate a privacy risk flag list.

[0044] Based on the privacy risk label list, the updated local training parameters and behavior monitoring logs are corrected and encrypted to generate compliant corrected parameter and log data.

[0045] Perform data integrity and compliance checks on the parameters and log data after compliance correction, and generate a privacy compliance report;

[0046] Based on privacy compliance reports and behavior monitoring logs, the response efficiency, recommendation accuracy, and cache usage of the locally trained model are analyzed to generate book recommendation optimization suggestions.

[0047] The beneficial effects of this invention are as follows: By securely aggregating the locally trained parameters independently generated by each distributed data terminal, and by controlling the aggregation process in conjunction with differential privacy noise and drift detection weights, privacy protection and drift sensitivity optimization of global model parameters are achieved; not only is knowledge sharing between different data sources guaranteed, but the impact of single-node data deviation on the global model is also avoided; the generated global parameter package and unified drift signal provide a unified model update reference for each distributed data terminal, enhancing the responsiveness of personalized recommendations to changes in user interests, improving recommendation accuracy, and optimizing the utilization efficiency of book resources. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of a smart library management method.

[0050] Figure 2 A flowchart for generating a manageable data list.

[0051] Figure 3 A flowchart generated for local training parameters and behavior monitoring logs.

[0052] Figure 4 A flowchart for generating the global parameter package and uniform drift signal. Detailed Implementation

[0053] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0054] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0055] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0056] Reference Figures 1-4As one embodiment of the present invention, this embodiment provides a smart library management method, including the following steps:

[0057] S1. Collect and register the original library business data in a centralized manner, and encrypt and anonymize it in accordance with privacy protection rules to generate a manageable data list.

[0058] S1.1 The privacy protection rules are derived from information security specifications, industry standards, and the library's internal data management strategies.

[0059] Specifically, the formulation of privacy protection rules is based on a comprehensive consideration of information security regulations, industry standards, and the library's internal data management strategies. The core purpose is to ensure that user privacy is protected and sensitive information is prevented from being leaked during the collection, storage, processing, and sharing of library business data.

[0060] Information security specifications provide general requirements such as data encryption, access control, and audit trails, while industry standards provide specific operational guidelines for data processing, borrowing record management, and user behavior analysis in the library field. Meanwhile, internal data management strategies for libraries, combined with actual business processes and management objectives, provide detailed constraints on data classification, sensitivity level division, anonymization, and access control.

[0061] S1.2. Collect library collection information, borrowing records, search logs and user interest tags in real time, and perform preliminary identification, timestamp recording and sensitivity level classification according to a unified format to generate raw library business data.

[0062] Specifically, when collecting library information, borrowing records, search logs, and user interest tags in real time, the fields are first mapped and preliminarily identified according to a unified data format. For example, the library information is identified by the book number, library category, and location in the library; the borrowing records are identified by the borrowing user ID, borrowing time, and return time; the search logs are identified by the search keywords, search time, and number of searches; and the user interest tags are identified by the tag type and associated book number.

[0063] Record the generated timestamp information, and record the operation time in the format YYYY-MM-DD HH:MM:SS;

[0064] Fields are categorized according to preset sensitivity level rules, such as marking user personal information as highly sensitive, borrowing behavior as moderately sensitive, and collection information as low sensitive;

[0065] After completing the identification, timestamp recording, and sensitivity level classification, the original library business data is generated by summarizing in a unified format.

[0066] It should also be noted that the preset sensitivity level rules are formulated based on information security specifications, industry standards, and the library's internal data management strategies. These rules are based on security protection requirements for data categories such as user personal identity information, borrowing behavior, search records, and interest tags. They also take into account the library's internal management strategies for access permissions, usage scenarios, and sharing scope. By assigning different sensitivity levels (high, medium, and low) to different types of data, the system achieves classified protection and risk control during the processing.

[0067] S1.3. In accordance with privacy protection rules, the original library business data is encrypted, anonymized, and desensitized. At the same time, duplicate records are removed, missing fields are filled, and integrity is verified to generate a qualified dataset.

[0068] Specifically, in accordance with privacy protection rules, symmetric encryption algorithms are used to encrypt user identity information and sensitive fields item by item, such as user name, account and contact information fields;

[0069] A general anonymization method is used to replace the identifiers in borrowing records and retrieval logs, converting user-identifiable identifiers into untraceable anonymous identifiers;

[0070] Fields involving personal privacy are anonymized. For example, only the first six and last four digits of an ID card number are retained, with the middle part covered by "*". Only the first few digits and the last digit of contact information are retained, with the rest covered by special symbols.

[0071] The original library business data is used to remove duplicate records by comparing user identifiers and book entry identifiers to delete completely duplicate records;

[0072] Fill in missing fields, for example, fill in missing interest tags with the average value or the default tag;

[0073] Perform integrity checks on the processed original library business data, checking each field type to ensure it matches the preset field type. For example, numeric fields must be integers or floating-point numbers, and string fields must be fixed-format text.

[0074] Check whether the data value range falls within the predetermined range. For example, the borrowing count field should be an integer greater than zero, and the interest weight field should be between 0 and 1. The predetermined range is determined based on the business semantics and historical distribution characteristics of the data fields. Reasonable boundaries are set by analyzing the actual values ​​of each field in the original library business data and combining the field definitions.

[0075] Check if the timestamp format conforms to the unified standard, and generate a qualified dataset after all verifications pass.

[0076] S1.4 Index and organize the qualified datasets to generate a manageable data list.

[0077] Specifically, a multidimensional index is built for the verified dataset based on collection information, borrowing records, retrieval logs, and user interest tags, for example, using book number, user ID, and timestamp as index keys;

[0078] The qualified dataset after indexing is classified and organized, and the entries of the same category are sorted in chronological order, and the sensitivity level and processing status of each entry are recorded.

[0079] Based on the organized index structure, a manageable data list is generated. The data entries and attribute information corresponding to the index keys are formed into a unified table to ensure that each record can be quickly located and tracked. At the same time, encryption, anonymization and desensitization processing marks are retained to generate a manageable data list.

[0080] S2. Using a manageable data list, perform feature extraction, format standardization, and sensitive information removal on the original library business data, initialize the local training model, and generate local training parameters and behavior monitoring logs.

[0081] S2.1 Based on the manageable data list, extract features from the original library business data and transform them into structured feature vectors to form a preliminary feature dataset.

[0082] Specifically, based on the manageable data list, feature extraction is performed on the collection information, borrowing records, search logs and user interest tags in the original library business data by field. For example, the borrowing frequency, the number of times the search keywords appear and the user preference tags are converted into numerical representations.

[0083] The extracted features of collection information, borrowing records, search logs and user interest tags are encoded and normalized in a unified format to generate structured feature vectors. Each structured feature vector contains collection number, user identifier, timestamp and corresponding feature value.

[0084] All structured feature vectors are aggregated to form a preliminary feature dataset for subsequent standardization processing and local model initialization.

[0085] S2.2 Perform field standardization, numerical normalization, and sensitive information desensitization on the preliminary feature dataset, and perform integrity verification to generate a standardized feature dataset.

[0086] Specifically, for the features of collection information, borrowing records, retrieval logs and user interest tags contained in the preliminary feature dataset, each field is standardized. For example, the categorical features are one-hot encoded, and the continuous numerical features are mapped to the range [0,1] to complete the numerical normalization.

[0087] For fields containing sensitive information, such as user identifiers, contact information, and borrowed content information, desensitization processing shall be carried out in accordance with privacy protection rules, including replacement, masking, or hashing.

[0088] The integrity of the standardized, normalized, and desensitized feature vectors is verified, including missing value checking, field quantity verification, and timestamp consistency verification, to generate a standardized feature dataset that passes the verification.

[0089] S2.3 Initialize the local training model based on the standardized feature dataset, perform local training and security verification, generate local training parameters, summarize the behavior monitoring information generated during the training process, and obtain behavior monitoring logs.

[0090] Specifically, based on the feature vectors of collection information, borrowing records, retrieval logs, and user interest tags contained in the standardized feature dataset, the local training model is initialized, initial weight parameters are set, and training hyperparameters are determined, such as learning rate, batch size, and number of iterations.

[0091] The local training model is trained using a standardized feature dataset, including forward propagation, error calculation, and backpropagation weight update. After each round of training, the parameters of the local training model are verified for safety, including parameter integrity check and outlier removal, and the parameters of the local training model are generated.

[0092] During training, user actions, feature updates, and changes in local training model parameters are recorded and summarized to generate behavior monitoring logs.

[0093] S3. The local training parameters and behavior monitoring logs are aggregated to the aggregation center, and the differential privacy method and drift detection weights are used to perform secure aggregation calculations to generate a global parameter package and a unified drift signal.

[0094] S3.1 Send the local training parameters and behavior monitoring logs to the aggregation center, and perform preliminary sorting and classification according to the source of the local training model to generate a node parameter summary table.

[0095] Specifically, the local training parameters and behavior monitoring logs generated by each distributed book data terminal are sent to the aggregation center according to a secure transmission protocol. After receiving the data, the local training parameters are classified according to the source information of the local training model, including sorting by data collection time, data terminal identifier and training version. The parameters are then preliminarily archived and organized according to the classification results to generate a node parameter summary table. Each record contains the local training parameter identifier, source data terminal identifier, training round and behavior monitoring log index.

[0096] S3.2. Based on the node parameter summary table, perform a secure aggregation operation on the local training parameters. Use a weighted average combined with differential privacy noise method to perform secure aggregation calculation on the local training parameters and generate a global parameter package.

[0097] Specifically, based on the node parameter summary table, the local training parameters are weighted and averaged according to the data terminal identifier and training round corresponding to each local training parameter. The contribution weight is allocated based on the data terminal contribution and the training round. For example, the proportion of data volume at the data terminal can be set as the contribution weight factor.

[0098] During the weighted averaging process, differential privacy noise is added to each local training parameter. The noise amplitude is adjusted according to the privacy budget to ensure that individual data parameters in the aggregation result cannot be inferred in reverse.

[0099] The weighted average and noise injection operations are performed on all local training parameters in the summary table in sequence to obtain the aggregated parameter vector. The parameters are then indexed and organized according to the training epoch and feature category corresponding to the parameter vector to generate a global parameter package.

[0100] S3.3 Utilize the drift weight information from local training parameters and behavior monitoring logs to adjust the drift sensitivity of the global parameter package and generate a unified drift signal.

[0101] Specifically, based on the global parameter package, the drift weight information corresponding to each local training parameter is associated with the behavior monitoring log, and the drift sensitivity index of each standardized feature vector and parameter vector under different training rounds is extracted.

[0102] The parameter vectors in the global parameter package are adjusted according to the drift weight, including increasing the response amplitude of parameters with high drift weight and reducing the influence of parameters with low drift weight, in order to reflect the sensitivity to changes in user interest.

[0103] After adjusting the drift weights of all parameter vectors in the global parameter package in sequence, the drift sensitivity information of each parameter vector is summarized and integrated into a time series to generate a unified drift signal.

[0104] Ideally, the local model parameters obtained by independent training at the distributed book data terminal are securely aggregated, while differential privacy noise and drift detection weights are introduced to optimize the concept drift sensitivity of the unified model while maintaining privacy protection.

[0105] Compared to existing technologies, traditional personalized recommendation or distributed training often simply summarizes node model parameters, lacking comprehensive processing of privacy protection and drift sensitivity. By combining differential privacy and drift weights, we achieve secure and reliable updates of global parameters.

[0106] It enables knowledge sharing between different data sources, while avoiding the impact of data leakage or deviation from a single node on the unified model. It provides updated references for each distributed data terminal, guides personalized recommendation algorithms to respond promptly to changes in user interests, thereby improving recommendation accuracy, enhancing adaptability to concept drift, and optimizing the utilization efficiency of book resources.

[0107] S4. Distribute the global parameter package and unified drift signal to the distributed book data terminal, and combine them with the real-time collected search requests and borrowing behavior to generate personalized recommendation results and recommendation explanations.

[0108] S4.1 The real-time collection of search requests and borrowing behaviors is obtained through user operation logs automatically recorded by the library's back-end service.

[0109] Specifically, through the library's back-end service, the search requests, book categories selected, click records, and borrowing operations entered by users on the book search interface are captured in real time. Each operation record is marked with a timestamp and associated with the user's unique identifier.

[0110] The captured operation logs are standardized in terms of fields and format to generate structured user operation logs. These structured user operation logs are then categorized and organized according to search request type and borrowing behavior category to form real-time collected search requests and borrowing behaviors that can be used to analyze user behavior. For example, information such as the keywords, search time, library collection number, and borrowing time of each user's search request is recorded.

[0111] S4.2 Distribute the global parameter package and unified drift signal to the distributed book data terminal, and perform mapping processing in conjunction with the standardized feature dataset to generate a node-level parameter mapping table.

[0112] Specifically, the global parameter package and unified drift signal are sent to each distributed book data terminal. Based on the standardized feature dataset of each distributed book data terminal, the parameter items of the global parameter package and unified drift signal are matched and mapped with the standardized feature vector, the corresponding parameter positions and mapping relationships are recorded, and a node-level parameter mapping table is generated.

[0113] The index is organized according to the node identifier of the node-level parameter mapping table to ensure that each mapping relationship can uniquely correspond to the specific distributed book data terminal and standardized feature dataset. For example, the mapping index of each parameter on the node feature dimension and the sensitive feature dimension corresponding to the drift signal are recorded.

[0114] S4.3. Based on the node-level parameter mapping table and the real-time collected search requests and borrowing behavior, perform matching analysis on the user's historical behavior and preferences to generate a preliminary personalized recommendation candidate set, and use a unified drift signal to adjust the drift sensitivity to generate personalized recommendation results and recommendation explanations.

[0115] Specifically, based on the node-level parameter mapping table, the global parameter items of each distributed book data terminal are matched with the user's historical behavior and preference data corresponding to the standardized feature vector, and book items related to the user's search request and borrowing behavior are extracted to form a preliminary personalized recommendation candidate set;

[0116] Based on the unified drift signal, the weights of book items in the initial personalized recommendation candidate set are adjusted for drift sensitivity, the recommendation priority of each candidate book item is updated, and personalized recommendation results and explanations are generated; for example, the degree of matching between each recommended item and the user's preference features and the sensitivity feature adjustment value corresponding to the drift signal are recorded.

[0117] S5. Monitor the performance of personalized recommendation results and explanations in real time, and identify changes in user interests through concept drift detection, generating drift alerts and affected feature sets.

[0118] S5.1 Collect user interaction data on personalized recommendation results in real time, and associate it with the recommendation explanation to generate a recommendation performance dataset.

[0119] Specifically, based on personalized recommendation results and explanations, the system records users' actions on the library interface in real time, including clicking, borrowing, collecting, and rating recommended items. These actions are then linked according to timestamps, recommended item identifiers, and user identifiers. Each interaction is matched with its corresponding explanation, and features such as interaction type, frequency, and duration are extracted to generate a recommendation performance dataset containing user actions and corresponding explanations. For example, the dataset records the number of times a user clicks on a book recommendation and the corresponding recommendation reason identifier.

[0120] S5.2 Perform indicator analysis and anomaly detection on the recommendation performance dataset, calculate the actual performance deviation, and generate a recommendation performance deviation report.

[0121] Specifically, for metrics such as click-through rate, conversion rate, and response time in the recommendation performance dataset, the actual values ​​of each metric are cumulatively summed and counted according to the time series. The actual performance deviation is calculated using the following expression:

[0122] ;

[0123] in, Indicates at a point in time Up indicators Actual performance deviation. Indicates at a point in time Up indicators The actual value, Indicates at a point in time Up indicators The expected value, Indicates the type of indicator. Indicates a point in time;

[0124] For the time series of each indicator, first, sum and count the actual values ​​of each indicator at all time points to calculate the mean. Then, calculate the sum of squares of the differences between each time point value and the mean, and average them to obtain the standard deviation. Based on the mean and standard deviation, calculate the standardized value of the indicator at each time point. The expression is as follows:

[0125] ;

[0126] in, Indicates at a point in time Up indicators The standardized value, Indicators The mean over a period of time, Indicators Standard deviation over a period of time;

[0127] when When this happens, the indicators become abnormal, among which, This represents the anomaly detection threshold. It is set based on the statistical distribution characteristics of historical recommendation performance data. By calculating the mean and standard deviation of various indicators in the recommendation performance dataset over a long period of time, and combining the actual business requirements for anomaly tolerance, a preset deviation range is used as the anomaly detection threshold. For example, the mean plus or minus three times the standard deviation can be used as the anomaly detection threshold. At the same time, the fluctuation range of indicators can also be set according to empirical rules or commonly used industry standards, and the effectiveness of the threshold can be verified by comparing with historical anomaly cases, thereby determining the final anomaly detection threshold.

[0128] when At that time, the indicator fluctuations are within a reasonable range; for example, if set This means that when the standardized score of an indicator deviates from the mean by more than three times the standard deviation, it will be marked as an anomaly.

[0129] The actual performance deviation and the anomaly detection results are summarized to generate a recommended performance deviation report.

[0130] S5.3 Combine the recommendation performance deviation report with historical behavioral characteristics, and use the concept drift detection algorithm to identify the trend of user interest changes and generate interest drift signals.

[0131] Specifically, the actual performance deviation values ​​of each recommended item in the performance deviation report are aligned with the user click behavior, borrowing behavior, and interest tags in the historical behavioral characteristics through time series analysis. The aligned data is then input into a concept drift detection algorithm. Based on a sliding time window or cumulative deviation index, the algorithm uses finite difference or rolling difference methods to calculate the rate of change of user interest characteristics, identify significantly changed interest characteristics and their direction of change, and generate an interest drift signal that includes the trend and intensity of user interest changes. For example, cumulative deviation analysis is performed on the frequency changes of user clicks on book categories over the past 30 days, and drift characteristics are labeled.

[0132] S5.4 Based on the interest drift signal, identify the affected feature set and the affected recommendation category, and generate drift alerts and the affected feature set.

[0133] Specifically, based on the drift intensity of each interest feature in the interest drift signal, the drift intensity is compared with a preset drift threshold. If the drift intensity is greater than or equal to the drift threshold, the interest feature is determined to be affected, and the corresponding user history behavior field and book category are recorded. All interest features determined to be affected and their corresponding book categories are organized into an affected feature set. At the same time, a drift alarm and an affected feature set containing interest feature identifiers, book categories, and drift level information are generated. For example, when the drift intensity is equal to or greater than 0.3, the corresponding interest feature and book category are included in the affected feature set.

[0134] It should also be noted that when setting the drift threshold, the drift amplitude distribution of each interest feature in different time windows is statistically analyzed from historical recommendation performance data and user behavior characteristics. Based on the statistical results, a representative drift amplitude is selected as the benchmark value. For example, the average or median of the drift amplitude of the interest feature can be used as the initial threshold. Combining the sensitivity of book category and the requirements of recommendation strategy, the benchmark value is weighted and adjusted to obtain the drift threshold. The drift threshold is then mapped to the corresponding interest feature and book category for subsequent comparison of interest drift signals and identification of affected features.

[0135] Ideally, real-time performance monitoring of personalized recommendation results should be closely integrated with concept drift detection. By continuously tracking the correlation between user interaction data and recommendation explanations, the changing trends of user interests can be dynamically identified, thereby generating drift alerts and affected feature sets.

[0136] Compared with existing technologies that rely solely on static user profiles or periodic behavior analysis, this method achieves immediate perception and rapid response to interest shifts by collecting and processing user behavior data in real time.

[0137] It can adjust recommendation strategies in a timely manner to ensure that recommended content is highly matched with users' current interests, improve recommendation accuracy and user satisfaction, and provide accurate basis for incremental updates of subsequent local training parameters, thereby optimizing the allocation and utilization efficiency of book resources and enhancing the dynamic adaptability and intelligence level of smart library management.

[0138] S6. Prioritize the training of drift alarms and affected feature sets, generate updated local training parameters, and perform privacy compliance checks with behavior monitoring logs to generate a privacy compliance report and book recommendation optimization suggestions.

[0139] S6.1 Based on the drift alert, prioritize the affected feature sets according to the feature impact weight and drift severity, and generate a feature priority list.

[0140] Specifically, based on the drift alert, the corresponding feature influence weight and drift severity are read for each interest feature in the affected feature set, and the feature influence weight and drift severity are weighted to obtain the comprehensive influence value.

[0141] The affected feature sets are sorted from largest to smallest based on their overall impact value, generating a feature priority list arranged by priority. The priority of each interest feature is then associated with its corresponding book category, forming a feature priority list.

[0142] S6.2. Based on the feature priority list, adjust the high-priority features to perform incremental training on the local training model, and retain the original training parameters as a basis to generate preliminary updated model parameters.

[0143] Specifically, based on the feature priority list, high-priority interest features and corresponding book category data are extracted from the standardized feature dataset. The extracted book category data is then combined with the original training parameters of the local training model. The local training model is iteratively updated using an incremental training method. For example, in each iteration, the gradient is calculated using the high-priority interest features and the original training parameters of the local training model are updated, while the original training parameters are retained as a basis. This process continues until all high-priority interest features have completed incremental training, generating preliminary updated model parameters that include the original training parameters and the incremental training results.

[0144] S6.3 Perform security and integrity verification on the initially updated model parameters, remove abnormal parameters, and generate qualified local training parameters.

[0145] Specifically, for the initial updated model parameters, each initial updated model parameter value is compared with the allowable range according to the security verification rules. For example, it is determined whether the initial updated model parameters exceed the upper and lower limits of the example or whether there are illegal values. Parameters that do not meet the security verification conditions are marked as abnormal parameters.

[0146] Perform integrity verification on all preliminary updated model parameters, such as comparing the consistency of parameter dimensions, numerical types, and gradient calculation results, and remove preliminary updated model parameters that fail the integrity verification.

[0147] After removing outlier parameters, the remaining preliminary updated model parameters that have passed security and integrity checks are aggregated to form qualified local training parameters.

[0148] It should also be noted that the sources of security verification rules include the original local training parameter example range and the upper and lower limits of the standardized feature dataset values, as well as historical training parameter distribution statistics, such as minimum, maximum and mean variance, while also referring to similar parameter constraints in existing industry standards.

[0149] The allowable range refers to the numerical boundaries or legal value intervals set for the parameters of the initial updated model, used to determine whether the parameters are within an acceptable range. The allowable range can be obtained in two ways: First, by conducting statistical analysis based on historical training data or the maximum and minimum values ​​of the validated initial updated model, for example, calculating the mean ± 3 times the standard deviation of the initial updated model in past training rounds as example upper and lower limits; second, by setting a theoretically reasonable value interval for the parameters based on domain knowledge or business rules, and the resulting numerical interval is used as the allowable range for security verification.

[0150] S6.4. Perform correlation analysis between the verified local training parameters and the behavior monitoring logs to generate updated local training parameters.

[0151] Specifically, the system performs correlation analysis between the verified local training parameters and the behavior monitoring logs. This includes matching the behavior monitoring logs with the local training parameter records by timestamp, mapping the operation type, frequency, and deviation value in the behavior monitoring logs to the corresponding local training parameters, adjusting the weights of the local training parameters affected by the behavior, and generating updated local training parameters based on the adjusted weights and mapping relationships.

[0152] S6.5 Match the sensitive features and behavior monitoring logs in the updated local training parameters to generate a preliminary privacy review dataset.

[0153] Specifically, a matching analysis is performed on the sensitive features in the updated local training parameters and the behavior monitoring logs. This includes extracting the corresponding parameter items from the updated local training parameters according to the sensitive feature identifiers, mapping the relevant operation records in the behavior monitoring logs to the sensitive feature parameter items by timestamp, mapping the operation type, operation frequency and deviation information to the sensitive feature parameters, and generating a preliminary privacy review dataset containing sensitive feature parameters and corresponding behavior information.

[0154] S6.6. Using privacy protection rules, perform data minimization, anonymization, differential privacy, and access permission verification on the preliminary privacy review dataset to identify potential compliance risks and generate a privacy risk flag list.

[0155] Specifically, the preliminary privacy review dataset is processed, including extracting sensitive feature fields from the dataset according to privacy protection rules and performing data minimization operations, encoding or masking identifiable user information using anonymization methods, adding noise to sensitive feature parameters using differential privacy methods, verifying access permissions for each record using access permission verification methods, statistically analyzing sensitive features that do not comply with privacy protection rules at any processing step, and generating a privacy risk label list containing sensitive feature identifiers, processing methods, and potential compliance risk levels.

[0156] S6.7 Based on the privacy risk label list, correct and encrypt the updated local training parameters and behavior monitoring logs to generate compliant corrected parameter and log data.

[0157] Specifically, based on the privacy risk label list, the sensitive features labeled in the updated local training parameters are corrected. This includes removing redundant features using data minimization methods, encoding or masking identifiable information using anonymization methods, adding noise to sensitive feature parameters using differential privacy methods, and encrypting the corresponding sensitive event fields in the behavior monitoring logs. This generates compliant corrected parameters and log data containing the corrected local training parameters and encrypted behavior monitoring logs. For example, sensitive features marked as high-risk are replaced with example encoded values ​​after anonymization, and ciphertext is generated for the relevant log fields using symmetric encryption methods.

[0158] S6.8 Perform data integrity and compliance verification on the parameters and log data after compliance correction, and generate a privacy compliance report.

[0159] Specifically, integrity checks are performed on the parameters and log data after compliance correction. This includes calculating the hash values ​​of each field of the updated local training parameters and behavior monitoring logs based on the hash verification method and comparing them with the preset example hash values. At the same time, the data is checked to see if it meets the requirements of minimization, anonymization, differential privacy, and access control according to privacy protection rules. Anomaly markers are generated for fields that do not meet the requirements. All verification results are summarized to form a privacy compliance report.

[0160] It should also be noted that the preset example hash value is a fixed hash result calculated by using a hash algorithm on the fields in the initial legitimate local training parameters and behavior monitoring logs respectively. For example, the example sensitive feature field is generated using the SHA-256 algorithm to generate a hash value as the example hash value.

[0161] The comparison process involves recalculating the hash value of the corresponding fields in the compliant and corrected parameters and log data using the same hash algorithm, and then comparing the calculated result with the example hash value field by field. If they are equal, the field integrity check is considered to be qualified; otherwise, it is marked as abnormal.

[0162] S6.9. Based on the privacy compliance report and behavior monitoring logs, analyze the response efficiency, recommendation accuracy, and cache usage of the locally trained model, and generate book recommendation optimization suggestions.

[0163] Specifically, based on privacy compliance reports and behavior monitoring logs, the system extracts the response time of the locally trained model for each recommendation request, the matching between the recommendation results and the user's actual clicks or borrowing behavior, as well as the cache hit count and cache usage. It calculates the average response latency and response fluctuation metrics for the response time data, and scores the accuracy of the recommendation results against user behavior, such as calculating the hit rate or Top-N coverage. It also calculates the cache hit rate for cache usage and analyzes frequently accessed data hotspots. Finally, it summarizes and sorts these metrics in conjunction with abnormal fluctuations or low performance conditions to generate book recommendation optimization suggestions.

[0164] This embodiment also provides a computer device applicable to the smart library management method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the smart library management method proposed in the above embodiment.

[0165] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0166] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the smart library management method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0167] In summary, this invention achieves privacy protection and drift sensitivity optimization of global model parameters by: securely aggregating locally generated training parameters from each distributed data endpoint and controlling the aggregation process using differential privacy noise and drift detection weights; ensuring knowledge sharing between different data sources while avoiding the impact of single-node data bias on the global model; and providing a unified model update reference for each distributed data endpoint with the generated global parameter package and unified drift signal, thereby enhancing the responsiveness of personalized recommendations to changes in user interests, improving recommendation accuracy, and optimizing the utilization efficiency of book resources.

[0168] It should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A smart library management method, characterized in that: include, The original library business data is centrally collected and registered, and encrypted and anonymized in accordance with privacy protection rules to generate a manageable data list; Using a manageable data list, feature extraction, format standardization, and sensitive information removal are performed on the original library business data. The local training model is initialized, and local training parameters and behavior monitoring logs are generated. Local training parameters and behavior monitoring logs are aggregated to the aggregation center, and a global parameter package and a unified drift signal are generated by using differential privacy methods and drift detection weights for secure aggregation calculation. The global parameter package and unified drift signal are distributed to the distributed book data terminal, and personalized recommendation results and explanations are generated by combining real-time collected search requests and borrowing behavior. Real-time monitoring of personalized recommendation results and explanations; and identification of changes in user interests through concept drift detection, generating drift alerts and affected feature sets. Prioritize the drift alerts and affected feature sets for training, generate updated local training parameters, and perform privacy compliance checks with behavior monitoring logs to generate a privacy compliance report and book recommendation optimization suggestions.

2. The smart library management method as described in claim 1, characterized in that: The privacy protection rules are derived from information security regulations, industry standards, and the library's internal data management strategies.

3. The smart library management method as described in claim 1, characterized in that: The specific steps for generating a manageable data list are as follows: The library collects information on its collections, borrowing records, search logs, and user interest tags in real time, and performs preliminary labeling, timestamp recording, and sensitivity classification according to a unified format to generate raw library business data. In accordance with privacy protection rules, the original library business data is encrypted, anonymized, and desensitized. At the same time, duplicate records are removed, missing fields are filled, and integrity is verified to generate a qualified dataset. The validated datasets are indexed and organized to generate a manageable data list.

4. The smart library management method as described in claim 1, characterized in that: The specific steps for generating local training parameters and behavior monitoring logs are as follows: Based on the manageable data list, features are extracted from the original library business data and transformed into structured feature vectors to form a preliminary feature dataset; The preliminary feature dataset is subjected to field standardization, numerical normalization, and sensitive information desensitization, and integrity verification is performed to generate a standardized feature dataset. The local training model is initialized based on the standardized feature dataset, and local training and security verification are performed. Local training parameters are generated, and the behavior monitoring information generated during the training process is summarized to obtain behavior monitoring logs.

5. The smart library management method as described in claim 1, characterized in that: The specific steps for generating the global parameter package and the unified drift signal are as follows: Send local training parameters and behavior monitoring logs to the aggregation center, and perform preliminary sorting and classification according to the source of the local training model to generate a node parameter summary table; Based on the node parameter summary table, a secure aggregation operation is performed on the local training parameters. The weighted average combined with the differential privacy noise method is used to perform secure aggregation calculation on the local training parameters to generate a global parameter package. By utilizing the drift weight information from local training parameters and behavior monitoring logs, the global parameter package is adjusted for drift sensitivity to generate a unified drift signal.

6. The smart library management method as described in claim 1, characterized in that: The real-time collected search requests and borrowing behaviors are obtained through user operation logs automatically recorded by the library's back-end service.

7. The intelligent library management method as described in claim 1, characterized in that: The specific steps for generating personalized recommendation results and explanations are as follows. The global parameter package and unified drift signal are distributed to the distributed book data terminal, and combined with the standardized feature dataset for mapping processing to generate a node-level parameter mapping table. Based on the node-level parameter mapping table and real-time collected search requests and borrowing behavior, the system performs matching analysis on users' historical behavior and preferences to generate a preliminary personalized recommendation candidate set. Then, it uses a unified drift signal to adjust the drift sensitivity and generate personalized recommendation results and explanations.

8. The intelligent library management method as described in claim 1, characterized in that: The specific steps for generating drift alarms and affected feature sets are as follows. Real-time collection of user interaction data with personalized recommendation results, and correlation with recommendation explanations to generate a recommendation performance dataset; Perform indicator analysis and anomaly detection on the recommendation performance dataset, calculate the actual performance deviation, and generate a recommendation performance deviation report; By combining recommendation performance deviation reports with historical behavioral characteristics, and using a concept drift detection algorithm, user interest change trends can be identified, generating interest drift signals. Based on the interest drift signal, identify the affected feature set and the affected recommendation category, and generate drift alerts and the affected feature set.

9. The intelligent library management method as described in claim 1, characterized in that: The specific steps for generating the updated local training parameters are as follows: Based on the drift alert, the affected feature sets are prioritized according to the feature impact weight and drift severity, and a feature priority list is generated; Based on the feature priority list, high-priority features are adjusted to incrementally train the local training model, while retaining the original training parameters as a basis to generate preliminary updated model parameters. The initial updated model parameters are subjected to security and integrity verification, abnormal parameters are removed, and qualified local training parameters are generated. The verified local training parameters are correlated and analyzed with the behavior monitoring logs to generate updated local training parameters.

10. The smart library management method as described in claim 1, characterized in that: The specific steps for generating the privacy compliance report and book recommendation optimization suggestions are as follows. The sensitive features and behavior monitoring logs in the updated local training parameters are matched to generate a preliminary privacy review dataset; By using privacy protection rules, the initial privacy review dataset is minimized, anonymized, differentially private, and access rights verified to identify potential compliance risks and generate a privacy risk flag list. Based on the privacy risk labeling list, the updated local training parameters and behavior monitoring logs are corrected and encrypted to generate compliant corrected parameter and log data. Perform data integrity and compliance checks on the parameters and log data after compliance correction, and generate a privacy compliance report; Based on privacy compliance reports and behavior monitoring logs, the response efficiency, recommendation accuracy, and cache usage of the locally trained model are analyzed to generate book recommendation optimization suggestions.

Images