Business-driven data assetification unified management method and system

Abstract

The application relates to the technical field of computer data processing, and provides a data asset unification management method and system based on business driving, which solves the technical problems of data asset discovery lag and low management efficiency. The method comprises the following steps: acquiring database access logs and business metadata; analyzing the database access logs, and constructing a sparse tensor according to the screened triplets; performing dimension reduction decomposition operation on the sparse tensor by using an alternating least square method to obtain user potential vectors, data potential vectors and time potential vectors; taking the data potential vectors corresponding to the unvisited data in the vectors as input vectors, inputting the input vectors into a trained neural collaborative filtering model, and obtaining the predicted preference values of the unvisited data; arranging the unvisited data in descending order according to the predicted preference values, and identifying potential target data based on the predicted preference values, so as to realize the unified management of data assets, and realize intelligent active discovery and efficient unified management of the data assets.

Description

Technical Field

[0001] This application relates to the field of computer data processing technology, and in particular to a business-driven unified management method and system for data assetization. Background Technology

[0002] As enterprises deepen their digital transformation, data has become a core production factor. Unified management of data assets driven by business needs is crucial for unlocking data value and optimizing resource allocation. Intelligent methods for assessing and efficiently distributing massive data assets can improve the efficiency of data flow and the level of business decision-making within enterprises.

[0003] Existing solutions typically rely on manual maintenance of metadata or simple statistical analysis based on historical access logs, using popularity rankings to aid data discovery.

[0004] However, existing solutions neglect the dynamic evolution of business needs over time and struggle to capture the complex, high-order nonlinear relationships between data assets and user actions. This results in an inability to accurately predict the potential value of unaccessed data when faced with sparse interactive data. Consequently, existing technologies suffer from the inability to effectively extract the potential value of data from dynamic and sparse historical behavior, leading to delayed data asset discovery and low management efficiency. Summary of the Invention

[0005] The purpose of this application is to provide a business-driven unified management method and system for data assetization, in order to solve the technical problems in the existing technology that make it difficult to effectively mine the potential value of data from dynamic and sparse historical behavior, resulting in lagging data asset discovery and low management efficiency.

[0006] Firstly, this application provides a business-driven unified management method for data assetization, including:

[0007] Obtain database access logs and business metadata. Database access logs include user ID, data ID, and access timestamp. Business metadata includes department attributes, business tags, and security levels of accessed and unaccessed data.

[0008] The database access logs are parsed to obtain triples including user identifier, data identifier and access timestamp. The triples are filtered using business metadata, and a sparse tensor is constructed based on the filtered triples. The non-zero element values ​​of the sparse tensor represent the attention of the accessed data.

[0009] The sparse tensor is decomposed by alternating least squares to obtain the user factor matrix, data factor matrix and time factor matrix. Row vectors are extracted from the user factor matrix, data factor matrix and time factor matrix respectively to obtain the user latent vector, data latent vector and time latent vector.

[0010] The data latent vector corresponding to the unvisited data in the user latent vector, time latent vector, and data latent vector is used as the input vector and fed into the trained neural collaborative filtering model to obtain the predicted preference value of the unvisited data.

[0011] Unaccessible data is sorted in descending order based on predicted preference values, and potential target data is identified based on the predicted preference values ​​to achieve unified management of data assets.

[0012] Optionally, before sorting the unaccessed data in descending order based on the predicted preference values ​​and identifying potential target data based on the predicted preference values ​​to achieve unified management of data assets, the method further includes:

[0013] Obtain the job permissions corresponding to the user identifier;

[0014] When the security level of unaccessed data in the business metadata is greater than the user's access permissions, the predicted preference value corresponding to the unaccessed data will be set to zero.

[0015] Optionally, before inputting the data latent vector corresponding to the unvisited data from the user latent vector, temporal latent vector, and data latent vector into the trained neural collaborative filtering model to obtain the predicted preference value for the unvisited data, the method further includes:

[0016] Obtain a training sample set, which includes multiple training samples. Each training sample includes an input feature combination consisting of user latent vector, time latent vector and data latent vector corresponding to the accessed data, and a true preference value extracted from the sparse tensor corresponding to the input feature combination as a label value.

[0017] For each training sample in the training sample set, perform the following steps:

[0018] The input features from each training sample are combined and fed into a pre-defined initial neural collaborative filtering model. Feature cross-operation is performed through the multi-layer network structure inside the model to obtain the predicted preference value.

[0019] The loss function value of the initial neural collaborative filtering model is determined based on the label value and predicted preference value of each training sample.

[0020] If the loss function value does not meet the preset training stopping condition, adjust the model parameters of the initial neural collaborative filtering model to obtain an updated initial neural collaborative filtering model. Then, return to input the combined input features into the updated initial neural collaborative filtering model to obtain the predicted preference value. Continue until the loss function value meets the preset training stopping condition to obtain the trained neural collaborative filtering model.

[0021] Optionally, the database access logs are parsed to obtain a triplet containing a user identifier, a data identifier, and an access timestamp. The triplet is then filtered using business metadata, including:

[0022] The original records of user identifier, data identifier and access timestamp for each access record in the database access log are combined to obtain a triplet corresponding to each access record;

[0023] Extract the security level corresponding to the data identifier from the business metadata;

[0024] Remove triples from all access records whose security level falls within the preset prohibited access range to obtain the filtered triples.

[0025] Optionally, a sparse tensor is constructed based on the filtered triples. The non-zero element values ​​of the sparse tensor represent the attention level of the accessed data, including:

[0026] The access timestamps in the filtered triples are divided into time windows, and access timestamps falling into the same time window are mapped to the same time slice identifier.

[0027] Records with the same user ID, the same data ID, and the same time slice ID in the filtered triples are categorized to obtain multiple triple combinations.

[0028] The frequency of occurrence of each triplet combination is counted. According to the preset weight allocation relationship, a weight coefficient is assigned to the business tag of the accessed data corresponding to each triplet combination, and the product of the frequency of occurrence and the weight coefficient is used as the attention level.

[0029] Establish a three-dimensional coordinate system that includes user, data, and time dimensions, and determine the target position in the three-dimensional coordinate system based on user identifier, data identifier, and time slice identifier;

[0030] Fill the target location with attention and set the unfilled positions in the three-dimensional coordinate system to zero to obtain a sparse tensor.

[0031] Optionally, the sparse tensor is decomposed using alternating least squares to obtain the user factor matrix, data factor matrix, and time factor matrix, including:

[0032] The number of duplicate user identifiers, data identifiers, and time slice identifiers is counted to construct the initial user matrix, initial data matrix, and initial time matrix;

[0033] Each initial matrix is ​​assigned a random initial value based on the preset dimension parameters;

[0034] Tensor reconstruction operations are performed on all initial matrices to obtain reconstructed tensors. An error loss function is constructed based on the Euclidean distance between the sparse tensor and the reconstructed tensor.

[0035] According to the preset update order, while keeping any two initial matrices unchanged, the remaining initial matrices are iteratively calculated and updated with the goal of minimizing the error loss function, until the difference value output by the error loss function is less than the convergence threshold. The updated initial matrices are then determined as the user factor matrix, data factor matrix, and time factor matrix, respectively.

[0036] Secondly, this application provides a business-driven unified management system for data assetization, including:

[0037] The acquisition module is used to acquire database access logs and business metadata. The database access logs include user identifiers, data identifiers, and access timestamps. The business metadata includes departmental attributes, business tags, and security levels of accessed and unaccessed data.

[0038] The parsing module is used to parse the database access logs to obtain triples including user identifier, data identifier and access timestamp. The triples are filtered using business metadata, and a sparse tensor is constructed based on the filtered triples. The non-zero element values ​​of the sparse tensor represent the attention of the accessed data.

[0039] The computation module is used to perform dimensionality reduction decomposition on sparse tensors using the alternating least squares method to obtain user factor matrix, data factor matrix and time factor matrix, and extract row vectors from user factor matrix, data factor matrix and time factor matrix respectively to obtain user latent vector, data latent vector and time latent vector.

[0040] The input module is used to input the data latent vector corresponding to the unvisited data from the user latent vector, time latent vector, and data latent vector into the trained neural collaborative filtering model to obtain the predicted preference value of the unvisited data.

[0041] The sorting module is used to sort unaccessed data in descending order based on the predicted preference value, and to identify potential target data based on the predicted preference value, so as to achieve unified management of data assets.

[0042] Thirdly, this application provides an electronic device, comprising:

[0043] Memory, used to store computer programs;

[0044] A processor is used to execute computer programs to implement the steps of a business-driven unified management method for data assetization as described in the first aspect above.

[0045] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the business-driven unified management method for data assetization described in the first aspect above.

[0046] The beneficial effects of this application are:

[0047] This application presents a business-driven unified management method for data assets. By integrating dynamic database access logs with static business metadata, it constructs a comprehensive asset assessment data foundation, addressing the problem of existing solutions neglecting the dynamic evolution of business needs over time. It also mines the deep correlation features implicit in database access logs, solving the technical problem of existing technologies struggling to extract effective features for value assessment from dynamic and sparse historical behavior. Furthermore, it overcomes the deficiency of traditional linear statistical models in capturing the complex high-order nonlinear relationships between data assets and user operations, enabling accurate prediction of the potential value of unaccessed data even before user interaction. Finally, it solves the problems of delayed data asset discovery and low management efficiency in existing technologies, achieving unified management of data assets based on prediction-driven approaches.

[0048] Furthermore, the access timestamp is divided into time windows and mapped to time slice identifiers. Records with the same dimension are grouped into triplet combinations. The frequency of occurrence and the weight coefficient are calculated as the attention score by combining the preset weight distribution relationship of the business tags. Finally, the attention score is filled into the target position in the three-dimensional coordinate system to construct a sparse tensor, thereby improving the accuracy of value prediction for unaccessed data. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A flowchart illustrating a business-driven unified management method for data assetization provided in this application embodiment;

[0051] Figure 2This application provides a schematic diagram of a process for obtaining sparse tensors in an embodiment of the present application.

[0052] Figure 3 A flowchart illustrating a training neural collaborative filtering model provided in an embodiment of this application;

[0053] Figure 4 A schematic diagram of the structure of a business-driven unified management system for data assetization is provided in this application embodiment;

[0054] Figure 5 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0055] The challenges faced by enterprises in managing data assets during digital transformation include the difficulty in capturing the dynamic evolution of business needs over time from dynamic and sparse historical behaviors, and the inability to mine the complex high-order nonlinear relationships between data assets and user operations through simple statistical analysis. This results in large deviations in the prediction of the potential value of unaccessed data and a lag in asset discovery.

[0056] This application constructs a sparse tensor representing historical interaction attention by parsing database access logs and business metadata, introducing a time dimension to solve the problem of extracting dynamic evolution features. Subsequently, it uses alternating least squares to perform dimensionality reduction decomposition on the sparse tensor to extract user latent vectors, data latent vectors, and time latent vectors to solve the feature extraction problem caused by data sparsity. Then, these linear feature vectors are input into a neural collaborative filtering model for deep feature cross-interaction, focusing on overcoming the difficulty of capturing high-order nonlinear associations. Finally, based on the generated predicted preference values, it accurately identifies potential target data in unaccessed data, realizing the proactive discovery and unified management of data assets.

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] The core of this application is to provide a business-driven unified management method for data assetization, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0059] Step 101: Obtain database access logs and business metadata. Database access logs include user identifiers, data identifiers, and access timestamps. Business metadata includes department attributes, business tags, and security levels of accessed and unaccessed data.

[0060] In this step, database access logs refer to historical log files recording user operations on the database, which may include at least one of SQL query logs, data warehouse audit logs, and API call records. Business metadata refers to structured data describing data asset attributes and business context information, which may include table ownership information, content description information, and management control information. User identifiers are unique codes that distinguish different business personnel in the enterprise's identity authentication system; these can be employee IDs or account IDs. Data identifiers are codes that uniquely distinguish different data tables or data views; these can be table names or metadata IDs.

[0061] Access timestamp refers to the specific moment a user initiates a data operation. Accessed data refers to data assets that exist in the database access log within a defined time window. Unaccessed data refers to data assets that do not appear in the database access log within a defined time window but exist in the business metadata. Department attribute refers to the organizational structure classification to which the user or data asset belongs, which may include department name or department code. Business tag refers to keywords used to identify the business theme of the data asset, which can be at least one of sales, finance, and inventory. Security level refers to the confidentiality or sensitivity level of the data asset.

[0062] In this embodiment, historical operation records within a preset time period are first collected through an enterprise data warehouse or unified log auditing interface. The user identity information, accessed object information, and operation time information are extracted from these records, serving as the user identifier, data identifier, and access timestamp, respectively. Simultaneously, a list of attribute descriptions for the entire data asset is obtained through the API interface of the metadata management platform. This list covers target objects that have appeared in the logs and potential objects that have not appeared in the logs, as shown in Table 1.

[0063] Table 1: List of Attribute Descriptions

[0064]

[0065] For example, in one specific implementation, a record is obtained from the audit log of Company A showing that user U1 queried data table D1 at time T1, and an attribute description list is obtained from the metadata system, thereby completing the preparation of basic data.

[0066] Step 102: Parse the database access logs to obtain triples including user identifier, data identifier and access timestamp. Filter the triples using business metadata and construct a sparse tensor based on the filtered triples. The non-zero element values ​​of the sparse tensor represent the attention level of the accessed data.

[0067] In this step, a triple refers to a standardized structured data unit that represents discrete user operation records, which can be represented in the form of [user, data, time]. A sparse tensor is a high-dimensional array structure where most elements have zero values, with non-zero values ​​stored only at locations where actual interactions have occurred. The non-zero element values ​​refer to the specific numerical values ​​in the sparse tensor corresponding to the coordinates of the interacting behavior. Attention level is a quantifiable indicator representing the degree to which users recognize the business value of data assets, and can be a comprehensive score calculated based on access frequency and business weight.

[0068] Step 201: Combine the original records of the user identifier, data identifier, and access timestamp for each access record in the database access log to obtain a triplet corresponding to each access record.

[0069] In this embodiment, each log entry in the database access log is traversed, and three core fields are extracted: user identifier, data identifier, and access timestamp. These three fields are then combined sequentially to form a structured triple, ensuring that each triple uniquely corresponds to a historical access behavior in the log. For example, two original log records are obtained: record A[user U1, data D1, 2023-10-01 10:00:00] and record B[user U1, data D1, 2023-10-01 14:30:00]. These are combined to obtain triples A[U1, D1, 2023-10-01 10:00:00] and B[U1, D1, 2023-10-01 14:30:00].

[0070] Step 202: Extract the security level corresponding to the data identifier from the business metadata.

[0071] In this embodiment, the data identifier in the triple is used as the index key to perform a lookup and matching in the pre-loaded business metadata table to obtain the security level attribute value corresponding to the data asset, which is used for subsequent compliance judgment. For example, for data identifier D1, its security level is extracted as L2 from the metadata; assuming there is another record involving data identifier D4, its security level is extracted as L4.

[0072] Step 203: Remove triples whose security level is within the preset prohibited access range from all triples corresponding to all access records to obtain the filtered triples.

[0073] In this step, the prohibited access zone refers to the range of sensitive levels that are not allowed to be analyzed or disclosed according to the company's data security management regulations, such as Level 3 and above. The filtered triples refer to the valid interaction records retained after security compliance cleaning.

[0074] In this embodiment, a preset prohibited access interval parameter is obtained, and it is determined whether the security level falls within this interval. If it falls within the interval, the corresponding triplet is removed from the processing list; if it does not, the triplet is retained. This process is repeated until all triplets are traversed, and the final set is the filtered triplet. For example, the prohibited access interval is set to [L4, L5]. For triplets A and B involving D1, since L2 is not within the prohibited interval, they are retained. For records involving D4, since L4 falls within the prohibited interval, they are discarded. The final filtered triplet set includes triplet A and triplet B.

[0075] like Figure 2 As shown, Figure 2 This is a schematic diagram of a process for obtaining sparse tensors provided in an embodiment of this application.

[0076] Step 211: Divide the access timestamps in the filtered triples into time windows, and map the access timestamps that fall into the same time window to the same time slice identifier.

[0077] In this step, time window segmentation refers to dividing a continuous timeline into discrete segments at fixed time intervals, which can be done by hour, day, or week. A time slice identifier is a code used to uniquely identify a specific time window; it can be a date ID or a time period number.

[0078] In this embodiment, the access timestamps in the filtered triples are read, and the time period to which the timestamp belongs is calculated based on a preset time window size. For example, specific time points are mapped to a unified time slice identifier through rounding or modulo operations, thereby achieving dimensionality reduction and aggregation of the time dimension. For example, if the time window is set to natural days, the timestamps of triple A "2023-10-01 10:00:00" and triple B "2023-10-01 14:30:00" both fall on the same day, and therefore are mapped to the same time slice identifier 20231001.

[0079] Step 212: Classify the records in the filtered triples that have the same user identifier, the same data identifier, and the same time slice identifier to obtain multiple triple combinations.

[0080] In this step, a triplet combination refers to the set of all interaction records of the same user on the same data asset within the same time window.

[0081] In this embodiment, the triple list after time mapping is grouped according to the key combination [user identifier, data identifier, time slice identifier]. Records with identical keys are merged together to form several non-overlapping triple combinations. For example, triples A and B have normalized keys of [U1, D1, 20231001], so these two records are grouped into the same triple combination.

[0082] Step 213: Count the frequency of occurrence of each triplet combination, assign weight coefficients to the business tags of the accessed data corresponding to each triplet combination according to the preset weight allocation relationship, and use the product of the frequency of occurrence and the weight coefficient as the attention level.

[0083] In this step, frequency of occurrence refers to the number of original access records included in a single triplet combination. Weighting relationships are pre-defined based on the business attribute dimensions of data assets. Specific allocation principles include: the degree of data's impact on core decisions, the real-time requirements of the data, and the closeness of the data's correlation within the business chain. For example, revenue data involved in core profit calculations has a significant impact on business decisions and is assigned a higher weight coefficient; material data, which plays a supporting role, has a relatively low value density and is assigned a basic weight coefficient. By establishing this weight mapping based on a data value assessment indicator system, subjective business importance is transformed into calculable quantitative characteristics. The weight coefficient refers to the quantitative multiplier assigned to a specific business label.

[0084] In this embodiment, the frequency of occurrence is first calculated by counting the number of records in each triplet combination. Then, based on the business tags corresponding to the data identifiers in the combination, the preset weight allocation table is queried to obtain the corresponding weight coefficients. Finally, use the formula Calculate attention ,in Indicates level of attention. Indicates frequency of occurrence. This represents the weighting coefficient. For example, statistical analysis shows that the combination [U1,D1,20231001] includes two records, A and B, hence the frequency. Querying the metadata reveals that the business tag for D1 is revenue data. Based on preset rules, this tag corresponds to a weight coefficient. The level of attention was calculated. .

[0085] Step 214: Establish a three-dimensional coordinate system including user dimension, data dimension and time dimension, and determine the target position in the three-dimensional coordinate system based on user identifier, data identifier and time slice identifier.

[0086] In this step, the three-dimensional coordinate system refers to the logical space structure used to construct the sparse tensor, which consists of the user axis, data axis, and time axis. The target location refers to the spatial coordinate point determined by the specific user, data, and time slice index.

[0087] In this embodiment, the unique value sets of user identifier, data identifier, and time slice identifier are extracted from all filtered triples to construct a three-dimensional index mapping table. Based on the user identifier, data identifier, and time slice identifier in the currently processed triple combination, the corresponding integer index value is searched in the index mapping table. The triple This determines the target position in the three-dimensional coordinate system. For example, assuming an index mapping is established: U1 is mapped to user axis index 1, D1 is mapped to data axis index 1, and time slice 20231001 is mapped to time axis index 5, then for this triplet combination, its target position in the three-dimensional coordinate system is determined to be coordinates (1,1,5). Finally, the attention score of 4.0 is filled into this position, completing the data cleaning, aggregation, and value quantification process.

[0088] Step 215: Fill the target position with attention and set the unfilled positions in the three-dimensional coordinate system to zero to obtain a sparse tensor.

[0089] In this step, a sparse tensor refers to a high-dimensional array structure in which most elements have a value of zero, and non-zero values ​​are stored only in the locations where there is actual interaction.

[0090] In this embodiment, all triplet combinations for which attention has been calculated are iterated, and the attention value corresponding to each combination is... Write to target location In the 3D coordinate system, all other coordinate points besides these target locations are assigned a default value of 0, indicating that the user has not generated any explicit interaction with the corresponding data assets or has had extremely low attention during these time windows. The resulting 3D data structure, consisting of a small number of non-zero values ​​and a large number of zero values, is a sparse tensor. For example, filling the tensor with an attention level of 4.0. Assume the total size of the coordinate system is... Except for position Apart from other recorded locations, the values ​​of all other coordinate points are 0.

[0091] Step 103: Use the alternating least squares method to perform dimensionality reduction decomposition on the sparse tensor to obtain the user factor matrix, data factor matrix and time factor matrix, and extract row vectors from the user factor matrix, data factor matrix and time factor matrix respectively to obtain the user latent vector, data latent vector and time latent vector.

[0092] In this step, alternating least squares refers to a mathematical algorithm that solves a matrix or tensor decomposition model through iterative optimization. The user factor matrix is ​​a set matrix representing all user features in a low-dimensional space. The data factor matrix is ​​a set matrix representing all data asset features in a low-dimensional space. The time factor matrix is ​​a set matrix representing temporal context features in a low-dimensional space. A row vector is a row of values ​​in the horizontal direction of a matrix, used to represent the features of a single entity. The user latent vector is the feature vector corresponding to a specific user in the user factor matrix. The data latent vector is the feature vector corresponding to a specific data asset in the data factor matrix. The time latent vector is the feature vector corresponding to a specific time segment in the time factor matrix.

[0093] In this embodiment, the number of unique elements in each dimension of the sparse tensor is first counted to determine the number of rows in each factor matrix, and a latent feature number much smaller than the original dimension is set as the number of columns to initialize the three factor matrices. Then, a fixed-variable strategy is adopted, in each iteration, the values ​​of two matrices are fixed in sequence, and the third matrix is ​​solved and updated using the least squares principle. The purpose is to make the tensor reconstructed by the product of the three matrices approximate the original sparse tensor as closely as possible. This process is repeated until the error converges, resulting in the final stable user factor matrix, data factor matrix, and time factor matrix. Finally, based on the specific identifier index, the corresponding row vectors are extracted from these three matrices respectively.

[0094] For example, in one specific embodiment, the hidden feature dimension is set to 3. The aforementioned sparse tensor X is decomposed to obtain the user factor matrix U, data factor matrix D, and time factor matrix T as shown below:

[0095] Example of user factor matrix U:

[0096] ;

[0097] Example of data factor matrix D:

[0098] ;

[0099] Example of time factor matrix T:

[0100] ;

[0101] For user U1, the user latent vector u1=[0.1,0.3,0.5] is extracted from the first row of the U matrix; for data table D1, the data latent vector d1=[0.7,0.1,0.2] is extracted from the first row of the D matrix; for data table D2, the data latent vector d2=[0.5,0.5,0.5] is extracted from the second row of the D matrix; and for time T1, the time latent vector t1=[0.9,0.1,0.0] is extracted from the first row of the T matrix. These vectors condense the implicit attributes of their respective entities in the business interaction network.

[0102] Step 301: Count the number of duplicate user identifiers, data identifiers, and time slice identifiers to construct the initial user matrix, initial data matrix, and initial time matrix.

[0103] In this step, the deduplication count refers to the total number of unique elements across all dimensions involved in the sparse tensor. The initial user matrix is ​​a matrix variable used to temporarily store user features before the decomposition iteration begins. The initial data matrix is ​​a matrix variable used to temporarily store data asset features before the decomposition iteration begins. The initial time matrix is ​​a matrix variable used to temporarily store time features before the decomposition iteration begins.

[0104] In this embodiment, the number of unique values ​​N for all data identifiers is counted based on business metadata, where N includes accessed data and unaccessed data; the indices corresponding to all non-zero elements in the sparse tensor are traversed, and the number of unique values ​​for user identifiers is counted for each. The number of unique values ​​for time slice identifiers Based on the pre-defined implicit feature dimensions The construction dimension is The initial user matrix has the following dimensions: The initial data matrix and its dimensions are The initial time matrix. For example, the number of unique users is obtained by counting. Data deduplication count Number of duplicates in time slices Define the dimensions of hidden features. The dimensions to be constructed are as follows: , , The three initial matrices.

[0105] Step 302: Set random initial values ​​for each initial matrix according to the preset dimension parameters.

[0106] In this step, the dimension parameter refers to the hyperparameter that determines the number of columns in the factor matrix, i.e., the latent feature dimension. A random initial value refers to a decimal that conforms to a specific probability distribution, typically a normally distributed value with a mean of 0 and a small variance.

[0107] In this embodiment, a Gaussian or uniform distribution random number generator is used to sequentially fill each element position of the three constructed initial matrices. This ensures that the absolute values ​​of the initial values ​​are small to prevent gradient explosion or numerical instability in the early stages of iteration. For example, filling the aforementioned three initial matrices... Random numbers within an interval.

[0108] Step 303: Perform tensor reconstruction operation on all initial matrices to obtain reconstructed tensors. Based on the Euclidean distance between the sparse tensor and the reconstructed tensor, construct the error loss function.

[0109] In this step, the reconstructed tensor refers to the approximate tensor restored through computation. The Euclidean distance is the square root of the sum of the squares of the differences between corresponding elements of two tensors, used to measure the degree of approximation. The error loss function is the objective function used to guide the optimization direction; a smaller value indicates a better decomposition effect.

[0110] In this embodiment of the application, formula (1) is used:

[0111] (1);

[0112] Compute the reconstructed tensor , where · represents the cross product of vectors. Denotes the first user matrix of the initial user matrix. Column vector, Represents the first element of the initial data matrix. Column vector, Represents the first time matrix of the initial time matrix. Column vectors. This involves fitting the original tensor to the sum of the outer products of the corresponding column vectors of the three matrices. Constructing the error loss function. ,in For the original sparse tensor, It is the Frobenius norm. The regularization coefficient is used to prevent overfitting. This function quantifies the reconstruction error under the current matrix parameters.

[0113] Step 304: According to the preset update order, while keeping any two initial matrices unchanged, with the goal of minimizing the error loss function, iteratively calculate and update the remaining initial matrices until the difference value output by the error loss function is less than the convergence threshold. The updated initial matrices are then determined as the user factor matrix, data factor matrix, and time factor matrix, respectively.

[0114] In this embodiment, an iterative loop is entered. The first step is to fix... and Unchanged, for the loss function about Find the partial derivative and set it to 0, then solve for the new... Matrix. Second step: Fix the updated matrix. and old Unchanged, for about Find the partial derivative and solve for the new... Matrix. Third step: Fix the updated matrix. and Unchanged, for about Find the partial derivative and solve for the new... Matrix. Calculate the loss function value for the current round. Compared to the previous round The difference is the variance value. If the difference is greater than the convergence threshold, the next iteration continues; if it is less than the threshold, the iteration stops, and the final result is obtained. The matrices are defined as user factor matrix, data factor matrix, and time factor matrix, respectively. For example, regularization coefficients are set. Convergence threshold After multiple iterations, the difference value decreased to 0.0005, satisfying the stopping condition. At this point, the three initial matrices in memory converged to the following: user factor matrix U, data factor matrix D, and time factor matrix T:

[0115] Example of user factor matrix U:

[0116] ;

[0117] Example of data factor matrix D:

[0118] ;

[0119] Example of time factor matrix T:

[0120] ;

[0121] This is the final solution result.

[0122] like Figure 3 As shown, Figure 3 This is a flowchart illustrating a training neural collaborative filtering model provided in an embodiment of this application.

[0123] Step 1041: Obtain a training sample set, which includes multiple training samples. Each training sample includes an input feature combination consisting of user latent vector, time latent vector and data latent vector corresponding to the accessed data, and a true preference value extracted from the sparse tensor corresponding to the input feature combination as a label value.

[0124] In this step, the training sample set refers to the dataset used for supervised learning, including the model's input features and expected output. The label value refers to the true numerical value used as the standard answer during supervised training, corresponding here to the historical actual attention recorded in the sparse tensor.

[0125] In this embodiment, all non-zero elements in the sparse tensor are traversed to obtain the user dimension index, data dimension index, and time dimension index corresponding to each non-zero element. Based on these three indices, the corresponding row vectors, namely the user latent vector, data latent vector, and time latent vector, are extracted from the user factor matrix, data factor matrix, and time factor matrix generated in the previous steps, respectively. These three vectors are concatenated in a preset order to form an input feature combination. Simultaneously, the specific value of the non-zero element in the sparse tensor is read and used as the label value corresponding to the input feature combination. Each input feature combination is paired with its corresponding label value to construct a training sample, and all training samples are collected into a training sample set. For example, for user U1, data D1, and time T1, their latent vectors are extracted as follows: , , The input feature combination is obtained by concatenation. Simultaneously, the true attention level of 4.0 at that location is read from the sparse tensor as the label value. .

[0126] Step 1042: For each training sample in the training sample set, perform the following steps:

[0127] The input features from each training sample are combined and fed into a pre-defined initial neural collaborative filtering model. Feature cross-operation is performed through the multi-layer network structure inside the model to obtain the predicted preference value.

[0128] In this step, the initial neural collaborative filtering model refers to a deep neural network model whose network structure has been built but whose weight parameters have not yet been optimized. Specifically, the neural collaborative filtering model adopts a two-branch fusion architecture, including a generalized matrix factorization layer and a multilayer perceptron layer. The generalized matrix factorization layer uses the element-wise product of the user latent vector, the data latent vector, and the time latent vector to capture linear feature associations; the multilayer perceptron layer concatenates the above three latent vectors as input and captures high-order nonlinear feature interactions through a multilayer fully connected network.

[0129] In this embodiment, a neural network comprising multiple fully connected layers is constructed, and the weight matrices and bias terms of each layer are initialized to obtain an initial neural collaborative filtering model. The input features are combined and input into the input layer of this model. Data is passed layer by layer through the hidden layers, and each layer is calculated using formula (2):

[0130] (2);

[0131] Perform calculations, where As weight, This is the output of the previous layer. For bias, The activation function is ReLU. After multiple nonlinear transformations, the output layer outputs a scalar value, which is the predicted preference value. For example, combining the above input features... The input is an initial 3-layer fully connected network model. After forward propagation, the output layer obtains the predicted preference values. .

[0132] Step 1043: Determine the loss function value of the initial neural collaborative filtering model based on the label value and predicted preference value of each training sample.

[0133] In this step, the loss function value is a numerical indicator that quantifies the degree of difference between the model's predicted output and the true label.

[0134] In this embodiment, mean squared error or binary cross-entropy is selected as the loss function. The label values... Compared with predicted preference values Substitute the values ​​into the loss function formula for calculation. For example, use the mean squared error formula. This allows us to calculate the loss function value that reflects the prediction bias of the current model. For example, the true label... Compared with the predicted value Substitute into the formula to calculate the single-sample loss. .

[0135] Step 1044: If the loss function value does not meet the preset training stopping condition, adjust the model parameters of the initial neural collaborative filtering model to obtain an updated initial neural collaborative filtering model, and return to input the input feature combination into the updated initial neural collaborative filtering model to obtain the predicted preference value, until the loss function value meets the preset training stopping condition, and obtain the trained neural collaborative filtering model.

[0136] In this step, the training stopping condition refers to the rule for determining whether the model has finished training, which may include the loss function value falling below a specific threshold or the number of iterations reaching an upper limit. Model parameters refer to the variables in the neural network that can be learned and adjusted, mainly including weights and biases.

[0137] In this embodiment, it is determined whether the calculated loss function value is less than a preset convergence threshold. If it is greater than the threshold, the gradient of the loss function with respect to each model parameter is calculated using the backpropagation algorithm. The gradient descent optimizer is used to update the model's weights and biases based on the calculated gradient direction. The forward propagation calculation is re-executed using the updated parameters, and the above process is repeated until the loss function value converges or the preset number of iterations is reached. Finally, the model with fixed parameters is retained as the trained neural collaborative filtering model. For example, the training stopping condition is set as follows: Therefore, the gradient is calculated through backpropagation and the network weights are adjusted. The updated model is then input again. To obtain new predicted values Calculate the new loss. The stopping condition is met. Training is complete, and the model is output.

[0138] Step 104: Input the data latent vector corresponding to the unvisited data from the user latent vector, time latent vector, and data latent vector into the trained neural collaborative filtering model to obtain the predicted preference value of the unvisited data.

[0139] In this step, the input vector refers to the combination of features input to the input layer of the neural network model, which can be a long vector concatenated from multiple potential vectors. A trained neural collaborative filtering model refers to a deep neural network model that has been trained under supervised guidance with historical data and whose parameters have been optimized and converged, possessing the ability to capture non-linear relationships between features. The predicted preference value refers to the probability or score output by the model used to estimate a user's level of interest in a specific data asset.

[0140] In this embodiment, target data assets that the current user has not yet accessed but exist in the business metadata are first identified. Then, the user's latent vector, the current context's temporal latent vector, and the corresponding data latent vector of the target data asset are obtained. Next, these three vectors are concatenated or fused to construct a complete input vector. Finally, the input vector is fed into a pre-trained neural collaborative filtering model. Within the model, the data undergoes nonlinear transformations and feature crossings through multiple fully connected networks and activation functions, ultimately generating a numerical value from the output layer. This value represents the user's predicted preference for the unaccessed data in the given time context.

[0141] For example, in one specific embodiment, for data table D2 that user U1 has never accessed, the previously extracted user latent vector u1=[0.1,0.3,0.5], time latent vector t1=[0.9,0.1,0.0], and data latent vector d2=[0.5,0.5,0.5] are combined into an input vector V=[0.1,0.3,0.5,0.9,0.1,0.0,0.5,0.5,0.5]. Vector V is then input into a trained neural collaborative filtering model. The model calculates and outputs a value between 0 and 1, 0.85, which is the predicted preference value, indicating that user U1 has a high latent interest in material data D2 at time T1.

[0142] Step 1051: Obtain the job permissions corresponding to the user identifier.

[0143] In this embodiment, the maximum data access level corresponding to the user's position is obtained by querying the user's identifier based on the enterprise identity authentication management interface or internal permission configuration table. For example, the permission level for a regular salesperson might be set to L2; for a department manager, it might be set to L3. As shown in Table 2:

[0144] Table 2: Job Authority Configuration Table

[0145]

[0146] According to the job permission configuration table, for the current user U1, the job permission is found to be L2 level.

[0147] Step 1052: When the security level of unaccessed data in the business metadata is greater than the role's permissions, set the predicted preference value corresponding to the unaccessed data to zero.

[0148] In this embodiment, all inaccessible data for which predicted preference values ​​have been calculated are traversed. For each data point, its security level in the business metadata is read and compared numerically with the acquired job permissions. If the data's security level is higher than the user's job permissions, it means the data is not visible to the current user. Regardless of how high the model's predicted interest level is, its predicted preference value is forcibly corrected to zero to prevent recommending sensitive data that the user is not authorized to access. If the security level is less than or equal to the job permissions, the original predicted preference value is retained.

[0149] For example, user U1's access level is L2. The model predicts that user U1's preference value for unaccessed data D2 (security level L3) is 0.85, for unaccessed data D4 (security level L4) is 0.20, and for unaccessed data D5 (security level L2) is 0.70. The system performs permission checks: for D2, because L3 is greater than L2, a zeroing operation is triggered, and the preference value is corrected to 0; for D4, because L4 is greater than L2, a zeroing operation is triggered, and the preference value is corrected to 0; for D5, because L2 is equal to L2, the preference value of 0.70 is retained. In the final corrected recommendation list, D5 is ranked first, while D2 and D4 are moved to the end due to compliance filtering, ensuring the security and compliance of data asset management.

[0150] Step 105: Sort the unaccessed data in descending order according to the predicted preference value, and identify potential target data based on the predicted preference value to achieve unified management of data assets.

[0151] In this step, potential target data refers to data assets with high prediction preference values, potential business promotion value, or those that require key management.

[0152] Step 501: Sort the unaccessed data in descending order according to the magnitude of the predicted preference values ​​to obtain the asset recommendation list corresponding to each user identifier, and generate a recommendation display instruction to send the asset recommendation list to the front-end interface.

[0153] In this step, the asset recommendation list refers to a collection of data asset items and their related information, generated for a specific user and sorted by interest. The front-end interface refers to the visual interactive terminal for users to perform data queries and analysis; it can be a web data portal or a BI tool dashboard.

[0154] In this embodiment, predicted preference values ​​for all unaccessed data of a single user are collected, using a single user identifier as the dimension. A quicksort algorithm is used to sort these data from largest to smallest preference value, and the top N data items are extracted to generate a personalized asset recommendation list. A recommendation display instruction in JSON or XML format, including this list data, is constructed and pushed to the user's currently logged-in front-end interface via a RESTful API or WebSocket, where it is rendered and displayed in the "You May Also Like" or "Popular Recommendations" areas. For example, for user U1, their asset recommendation list is shown in Table 3:

[0155] Table 3: Asset Recommendation List

[0156]

[0157] The system will push the instruction containing the above content to the portal homepage.

[0158] Step 502: Sum the predicted preference values ​​of the same unaccessed data to obtain the popularity value.

[0159] In this step, the popularity value refers to a quantitative indicator that reflects the potential demand intensity of a certain data asset among the entire user group, and is a future popularity trend calculated based on the predicted value.

[0160] In this embodiment, the aggregation dimension is transformed, and the predicted preference results for all users are traversed using the data identifier as the key. The predicted preference values ​​for all users who have not accessed the same data are summed using the following formula: ,in For data Popularity value Total number of users For users Data The predicted preference value. The higher the value, the greater the likelihood that the data will be accessed in the future. For example, for unaccessed data D2, user U1's predicted value is 0, user U2's predicted value is 0.8, and user U3's predicted value is 0.9 due to zero permissions. Therefore, the popularity value of D2 is... .

[0161] Step 503: When the heat value is greater than the preset heat judgment threshold, the corresponding unaccessed data is marked as potential target data.

[0162] In this step, the popularity threshold is a critical value used to distinguish between ordinary data and potentially popular data. It can be a fixed value set based on historical statistics or a dynamically calculated percentile value.

[0163] In this embodiment, the popularity threshold is dynamically determined based on the statistical characteristics of the popularity distribution of all data assets within a historical period. Specifically, the calculation principle is as follows: extract the average popularity value of all accessed data within a preset historical period. with standard deviation Set the threshold to ,in This is the sensitivity coefficient, which typically ranges from 1 to 3; or all data can be sorted in descending order of popularity, and the value at the top 10 percentile can be used as the popularity threshold.

[0164] The popularity score of each unaccessed data point is compared with a preset popularity threshold. If the popularity score of a data point is significantly greater than the threshold, it is marked as a potential target data in the metadata management database, indicating that it is a high-value asset to be optimized; if it is less than or equal to the threshold, it remains in a normal state. For example, the popularity threshold for the current period is dynamically calculated to be 1.5 based on the quantile algorithm described above. Data point D2 with a popularity score of 1.7 is marked as potential target data; data point D4 with a popularity score of 1.2 is not marked because it does not meet the threshold.

[0165] Step 504: Generate cache preloading instructions for potential target data, and migrate the potential target data from the low-speed storage area to the high-speed cache area according to the cache preloading instructions to complete the unified management of data assets.

[0166] In this step, the cache preloading instruction refers to the control command that triggers the underlying storage system to perform data migration operations. The low-speed storage area refers to persistent storage media with large capacity but slow read / write speeds, such as hard disk drives (HDDs) or cold data object storage. The high-speed cache area refers to temporary storage media with smaller capacity but extremely fast read / write speeds, such as in-memory databases or SSD caching layers.

[0167] In this embodiment, all data identifiers marked as potential target data are scanned. For each potential target data, a cache preloading instruction including the source address, target address, and data block ID is generated. This instruction is executed by calling the cache management interface of a distributed file system or database. Before peak business periods, these data blocks are proactively read from low-speed storage areas and written to high-speed cache areas. When users subsequently initiate actual access requests, the response is directly from the cache, thereby significantly reducing access latency and achieving prediction-driven resource optimization and unified management.

[0168] This application's embodiments construct a comprehensive asset valuation data foundation by integrating dynamic database access logs with static business metadata. It addresses the problem of existing solutions ignoring the dynamic evolution of business needs over time; it mines the deep correlation features hidden in database access logs, solving the technical problem of existing technologies struggling to extract effective features from dynamic and sparse historical behavior for value assessment. It overcomes the deficiency of traditional linear statistical models in capturing the complex high-order nonlinear relationships between data assets and user operations, enabling accurate prediction of the potential value of unaccessed data even before user interaction. It solves the problems of lagging data asset discovery and low management efficiency in existing technologies, achieving unified management of data assets based on prediction-driven approaches.

[0169] Figure 4This application provides a schematic diagram of the structure of a specific implementation of a business-driven unified management system for data assetization, with reference to... Figure 4 The system may include:

[0170] Module 21 is used to obtain database access logs and business metadata. The database access logs include user identifiers, data identifiers, and access timestamps. The business metadata includes department attributes, business tags, and security levels of accessed and unaccessed data.

[0171] The parsing module 22 is used to parse the database access logs to obtain a triplet including user identifier, data identifier and access timestamp. The triplet is filtered using business metadata, and a sparse tensor is constructed based on the filtered triplet. The non-zero element values ​​of the sparse tensor represent the attention of the accessed data.

[0172] The operation module 23 is used to perform dimensionality reduction decomposition on the sparse tensor using the alternating least squares method to obtain the user factor matrix, data factor matrix and time factor matrix, and extract row vectors from the user factor matrix, data factor matrix and time factor matrix respectively to obtain the user latent vector, data latent vector and time latent vector.

[0173] Input module 24 is used to input the data latent vector corresponding to the unvisited data in the user latent vector, time latent vector and data latent vector as input vector into the trained neural collaborative filtering model to obtain the predicted preference value of the unvisited data.

[0174] The sorting module 25 is used to sort unaccessed data in descending order based on the predicted preference value, and to identify potential target data based on the predicted preference value, so as to achieve unified management of data assets.

[0175] This application provides a business-driven unified management system for data assetization, which is used to implement the aforementioned business-driven unified management method for data assetization. Therefore, the specific implementation of the business-driven unified management system for data assetization can be found in the embodiment section of the business-driven unified management method for data assetization described above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0176] Figure 5 A schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application is shown.

[0177] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of a business-driven unified management method for data assetization.

[0178] The electronic device may include a processor 510 and a memory 520 storing computer program instructions.

[0179] Specifically, the processor 510 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0180] Memory 520 may include mass storage for data or instructions. For example, and not limitingly, memory 520 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 520 may include removable or non-removable (or fixed) media. Where appropriate, memory 520 may be internal or external to the integrated gateway disaster recovery device. In a particular embodiment, memory 520 is non-volatile solid-state memory.

[0181] Memory may include read-only memory (ROM), random access memory (RAM), disk storage media devices, optical storage media devices, flash memory devices, and electrical, optical, or other physical / tangible memory storage devices. Therefore, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to the first aspect of this disclosure.

[0182] The processor 510 reads and executes computer program instructions stored in the memory 520 to implement any of the business-driven unified management methods for data assets in the above embodiments.

[0183] In one example, the electronic device may also include a communication interface 530 and a bus 540. Wherein, such as Figure 5 As shown, the processor 510, memory 520, and communication interface 530 are connected through bus 540 and complete communication with each other.

[0184] The communication interface 530 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0185] Bus 540 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 540 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, any suitable bus or interconnect is contemplated herein.

[0186] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described business-driven unified management methods for data assetization.

[0187] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0188] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the business-driven unified management method for data assetization.

[0189] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0190] The foregoing has provided a detailed description of a business-driven unified management method and system for data assetization provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A business-driven unified management method for data assets, characterized in that, include: Obtain database access logs and business metadata. The database access logs include user identifiers, data identifiers, and access timestamps. The business metadata includes department attributes, business tags, and security levels of accessed and unaccessed data. The database access logs are parsed to obtain triples including user identifier, data identifier and access timestamp. The triples are filtered using the business metadata, and a sparse tensor is constructed based on the filtered triples. The non-zero element values ​​of the sparse tensor represent the attention of the accessed data. The sparse tensor is decomposed using alternating least squares to obtain a user factor matrix, a data factor matrix, and a time factor matrix. Row vectors are then extracted from these matrices to obtain user latent vectors, data latent vectors, and time latent vectors, respectively. The user factor matrix represents the set of features of all users, the data factor matrix represents the set of features of all data assets, and the time factor matrix represents the set of features of the time context. A row vector is a row of values ​​in the horizontal direction of the matrix. The user latent vector is the feature vector corresponding to a specific user in the user factor matrix, the data latent vector is the feature vector corresponding to a specific data asset in the data factor matrix, and the time latent vector is the feature vector corresponding to a specific time segment in the time factor matrix. The user latent vector, the time latent vector, and the data latent vector corresponding to the unvisited data in the data latent vector are used as input vectors and input into the trained neural collaborative filtering model to obtain the predicted preference value of the unvisited data. The unaccessed data is sorted in descending order based on the predicted preference value, and potential target data is identified based on the predicted preference value, so as to achieve unified management of data assets.

2. The business-driven unified management method for data assets according to claim 1, characterized in that, Before sorting the unaccessed data in descending order according to the predicted preference value and identifying potential target data based on the predicted preference value to achieve unified management of data assets, the method further includes: Obtain the job permissions corresponding to the user identifier; When the security level of unaccessed data in the business metadata is greater than the job permission, the predicted preference value corresponding to the unaccessed data is set to zero.

3. The business-driven unified management method for data assets according to claim 1, characterized in that, Before inputting the data latent vector corresponding to the unvisited data from the user latent vector, the time latent vector, and the data latent vector into the trained neural collaborative filtering model to obtain the predicted preference value for the unvisited data, the method further includes: Obtain a training sample set, which includes multiple training samples. Each training sample includes an input feature combination consisting of a user latent vector, a time latent vector, and a data latent vector corresponding to the accessed data, and a true preference value extracted from the sparse tensor corresponding to the input feature combination as a label value. For each training sample in the training sample set, perform the following steps: The input features from each training sample are combined and fed into a pre-defined initial neural collaborative filtering model. Feature cross-operation is performed through the multi-layer network structure inside the model to obtain the predicted preference value. The loss function value of the initial neural collaborative filtering model is determined based on the label value and predicted preference value of each training sample. If the loss function value does not meet the preset training stopping condition, the model parameters of the initial neural collaborative filtering model are adjusted to obtain an updated initial neural collaborative filtering model. The input feature combination is then input into the updated initial neural collaborative filtering model to obtain the predicted preference value. This process continues until the loss function value meets the preset training stopping condition, resulting in a trained neural collaborative filtering model.

4. The business-driven unified management method for data assets according to claim 1, characterized in that, The database access logs are parsed to obtain a triplet containing a user identifier, a data identifier, and an access timestamp. The business metadata is then used to filter the triplet, including: The original records of the user identifier, data identifier, and access timestamp for each access record in the database access log are combined to obtain a triplet corresponding to each access record; Extract the security level corresponding to the data identifier from the business metadata; Remove the triplets whose security level falls within the preset prohibited access range from all the triplets corresponding to the access records to obtain the filtered triplets.

5. The business-driven unified management method for data assets according to claim 1, characterized in that, A sparse tensor is constructed based on the selected triples. The non-zero element values ​​of the sparse tensor represent the attention level of the accessed data, including: The access timestamps in the filtered triples are divided into time windows, and access timestamps falling into the same time window are mapped to the same time slice identifier. The records in the filtered triplet that have the same user identifier, the same data identifier, and the same time slice identifier are classified to obtain multiple triplet combinations. The frequency of occurrence of each triplet combination is counted. According to the preset weight allocation relationship, a weight coefficient is assigned to the business tag of the accessed data corresponding to each triplet combination. The product of the frequency of occurrence and the weight coefficient is used as the attention level. Establish a three-dimensional coordinate system including user dimension, data dimension and time dimension, and determine the target position in the three-dimensional coordinate system based on the user identifier, the data identifier and the time slice identifier; The attention is filled into the target position, and the unfilled positions in the three-dimensional coordinate system are set to zero to obtain a sparse tensor.

6. The business-driven unified management method for data assets according to claim 5, characterized in that, The sparse tensor is decomposed using alternating least squares to obtain a user factor matrix, a data factor matrix, and a time factor matrix, including: The number of duplicates of the user identifier, the data identifier, and the time slice identifier is counted to construct an initial user matrix, an initial data matrix, and an initial time matrix; Each initial matrix is ​​assigned a random initial value based on the preset dimension parameters; Tensor reconstruction operations are performed on all initial matrices to obtain reconstructed tensors. An error loss function is constructed based on the Euclidean distance between the sparse tensor and the reconstructed tensor. According to the preset update order, while keeping any two initial matrices unchanged, the remaining initial matrices are iteratively calculated and updated with the goal of minimizing the error loss function, until the difference value output by the error loss function is less than the convergence threshold. The updated initial matrices are then determined as user factor matrix, data factor matrix and time factor matrix respectively.

7. The business-driven unified management method for data assets according to claim 1, characterized in that, The unaccessed data is sorted in descending order based on the predicted preference values, and potential target data is identified based on the predicted preference values ​​to achieve unified management of data assets, including: The unaccessed data are sorted in descending order according to the magnitude of the predicted preference values ​​to obtain an asset recommendation list corresponding to each user identifier, and a recommendation display instruction is generated to send the asset recommendation list to the front-end interface; The predicted preference values ​​of the same unaccessed data are summed to obtain the popularity value; When the heat value is greater than the preset heat determination threshold, the corresponding unaccessed data is marked as potential target data; Generate a cache preloading instruction for the potential target data, and migrate the potential target data from the low-speed storage area to the high-speed cache area according to the cache preloading instruction to complete the unified management of data assets.

8. A business-driven unified management system for data assetization, characterized in that, include: The acquisition module is used to acquire database access logs and business metadata. The database access logs include user identifiers, data identifiers, and access timestamps. The business metadata includes department attributes, business tags, and security levels of accessed and unaccessed data. The parsing module is used to parse the database access logs to obtain a triplet including user identifier, data identifier and access timestamp. The triplet is filtered using the business metadata, and a sparse tensor is constructed based on the filtered triplet. The non-zero element values ​​of the sparse tensor represent the attention of the accessed data. The computation module is used to perform dimensionality reduction decomposition on the sparse tensor using the alternating least squares method to obtain a user factor matrix, a data factor matrix, and a time factor matrix. Row vectors are extracted from these matrices to obtain user latent vectors, data latent vectors, and time latent vectors, respectively. The user factor matrix represents a set of features for all users, the data factor matrix represents a set of features for all data assets, and the time factor matrix represents a set of features for the time context. A row vector is a row of values ​​in the horizontal direction of the matrix. A user latent vector is a feature vector corresponding to a specific user in the user factor matrix, a data latent vector is a feature vector corresponding to a specific data asset in the data factor matrix, and a time latent vector is a feature vector corresponding to a specific time segment in the time factor matrix. The input module is used to input the data latent vector corresponding to the unvisited data in the user latent vector, the time latent vector and the data latent vector as input vectors into the trained neural collaborative filtering model to obtain the predicted preference value of the unvisited data. The sorting module is used to sort the unaccessed data in descending order according to the predicted preference value, and to identify potential target data based on the predicted preference value, so as to achieve unified management of data assets.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of a business-driven unified management method for data assetization as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables a business-driven unified management method for data assetization as described in any one of claims 1 to 7.

Images