A method for intelligent display and recommendation of enterprise data

Abstract

The present invention relates to the technical field of enterprise big data recommendation, and discloses an enterprise data intelligent display and recommendation method, S1: collecting feedback data of users on the market on enterprise projects, performing user interest modeling based on the feedback data, and obtaining the user's preference potential factors for the projects; S2: applying sparse constraints to non-zero elements in the preference potential factors, and constructing a sparse latent factor model; S3: obtaining similarity vectors between projects, and performing diversity penalty term calculation based on the similarity vector to obtain the diversity penalty term; S4: constructing a final objective function based on the sparse latent factor model and the diversity penalty term to obtain the final objective function, performing recommendation optimization based on the final objective function to obtain the final recommendation result, and balancing the prediction error and different types of regularization terms based on the objective function, which can effectively solve the problem of hot project bias in the recommendation system, thereby improving the diversity and novelty of the recommended content.

Description

Technical Field

[0001] The present invention relates to the technical field of enterprise big data recommendation, and in particular to a method for intelligently displaying and recommending enterprise data. Background Art

[0002] As enterprises’ demand for data utilization increases and market competition intensifies, enterprises must provide more personalized product and project recommendation services, which requires the use of data to better understand market users’ needs and preferences for projects. In the fierce market competition, enterprises need to improve efficiency and reduce costs by refining recommendation data, which requires in-depth analysis of internal data to support the formulation of recommendation data;

[0003] In the existing project data recommendation system, there is a tendency to recommend projects that are already popular, which leads to other projects that may be equally good but less well-known being neglected. Since the recommendation system tends to recommend known popular projects, newly emerging innovative projects are difficult to gain exposure, thereby inhibiting the innovation momentum of enterprises and slowing down the promotion of new products or projects. Therefore, a method for intelligent display and recommendation of enterprise data is proposed herein. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcoming of the conventional recommendation method in the prior art that newly emerging innovative projects are difficult to be exposed, and to propose an enterprise data intelligent display and recommendation method.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for intelligently displaying and recommending enterprise data, the method comprising the steps of:

[0007] S1: Collect feedback data from users in the market on enterprise projects, conduct user interest modeling based on the feedback data, and obtain potential factors of users' preferences for projects;

[0008] S2: impose sparse constraints on the non-zero elements in the preference latent factor and construct a sparse latent factor model;

[0009] S3: Obtain the similarity vector between the items, and calculate the diversity penalty item based on the similarity vector to obtain the diversity penalty item;

[0010] S4: Construct the final objective function based on the sparse latent factor model and the diversity penalty term to obtain the final objective function, optimize the recommendation based on the final objective function, and obtain the final recommendation result.

[0011] Feedback data includes user data, project data and enterprise evaluation data. Feedback data is expressed as D = {(u,i,r ui)}, where u is user data, i is project data, and r ui Ratings for businesses by users.

[0012] Where u is the user identifier, usually a unique ID or user name in the database;

[0013] i is the project identifier, which can be the enterprise project product ID or enterprise project decision plan ID;

[0014] r ui Feedback data of user u on project i, which can be rating, clicking, purchasing, etc. of products or decision-making solutions in enterprise projects;

[0015] Preferably, the steps for obtaining the preference potential factor are:

[0016] Through the user's rating matrix R for the project, the rating matrix R includes R ui , the rating matrix R is an m×n matrix, where m is the number of users and n is the number of items. User u has rated item i, then R ui =r ui , where R ui =r ui Represents the user's rating of the enterprise in the rating matrix;

[0017] Among them, according to the collected user evaluation data on the project ui Construct a rating matrix R. If user u has rated the project, then R ui =r ui ; If there is no evaluation, the position may be empty or represented by a special symbol (special processing may be required in actual calculations, such as filling in default values ​​or using special algorithms to process sparse matrices);

[0018] Since not all users have rated all businesses, this matrix tends to be sparse;

[0019] Matrix decomposition is performed based on the rating matrix R. The implementation of matrix decomposition is as follows:

[0020] Decompose the rating matrix R into two low-rank matrices P and Q, define P as the user potential factor matrix, Q as the enterprise potential factor matrix,

[0021] The goal of user interest modeling is to find a low-dimensional latent factor space so that the vector representations of users and companies in this space can well predict the user's rating of the company. This is achieved through the following steps:

[0022] Among them, the user potential factor matrix P is an m×k matrix, m is the number of users, k is the number of potential factors, and the vector of each row of the user potential factor matrix P is pu , represents the weights of k potential factors of user u;

[0023] The enterprise potential factor matrix Q is an n×k matrix, where n is the number of enterprises and k is the number of potential factors. The vector of each column is q i , represents the weights of k potential factors of enterprise i;

[0024] Define the prediction matrix as The score matrix R and the prediction matrix Minimize the difference between them and get the minimization function L(P,Q). Optimize based on L(P,Q):

[0025] For user u, keep Q fixed and solve for the row vector p in P u Get p u The user latent factor vector Among them, q i T for q i The transpose of q i T The latent factor vector q of item i i Convert to a row vector, q i q i T is the latent factor vector q for item i i The result of multiplying it with itself is the autocovariance matrix of item i in the latent factor space;

[0026] For item i, keeping P fixed, solve for column vector q in Q i Get q i The potential factor vector of the enterprise Among them, p u T For p u The transpose of u p u T is the latent factor vector p of user u u The result of multiplying it with itself is the autocovariance matrix of user u in the latent factor space;

[0027] Get user latent factor vector and the enterprise potential factor vector That is, the latent factor of preference.

[0028] The goal of user interest modeling is to find a low-dimensional latent factor space so that the vector representations of users and companies in this space can well predict the user's rating of the company.

[0029] Preferably, the steps for obtaining the minimization function are:

[0030] The process of obtaining the minimization function is: Among them, (u,i)∈D represents user u and item i in dataset D, Yes u The transpose of q i Multiply, that is, the predicted score of user u for item i, λ is the regularization coefficient, which is used to control the strength of the regularization term, ||p u || 2 and||q i || 2 They are p u and q i The Euclidean norm of is obtained to minimize the loss function.

[0031] The minimization function is the sum of squares of the differences between the predicted and actual ratings plus a regularization term, the purpose of which is to prevent fitting in the subsequent process.

[0032] Preferably, the steps for obtaining the sparse latent factor model are:

[0033] Based on the user potential factor matrix P and the enterprise potential factor matrix Q, a sparse penalty term is added and the sparse penalty term is defined as Ω(p u ) and Ω(q i ),Ω(p u )=λ S ||p u || 0 ,Ω(q i )=λ S ||q i || 0 , where ||·|| 0 represents the number of non-zero elements in the vector, λ s is a hyperparameter that controls the strength of the sparse penalty term, based on the sparse penalty term Ω(p u ) and Ω(q i ), update the user potential factor matrix P and the enterprise potential factor matrix Q. The updating method of the user potential factor matrix P is:

[0034] Fix the enterprise potential factor matrix Q and update p in the user potential factor matrix P u , after updating

[0035] Fix the user potential factor matrix P and update q in the enterprise potential factor matrix Q i , after updating

[0036] Define a thresholding function θ to reduce the user's latent factor vector and the enterprise potential factor vector The number of non-zero elements in the , through the threshold function, to obtain the optimized user potential factor vector and the enterprise potential factor vector

[0037] Among them, the user latent factor vector is defined as and the enterprise potential factor vector The elements in are x. For each element x, it is processed based on the threshold function to obtain the processed element x. When |x|>θ, the value of the element is retained, that is, x remains unchanged. When |x|≤θ, the element is set to 0.

[0038] By adding sparse penalty terms and thresholding function operations, the row vectors in P and Q are finally obtained. and column vector This is the sparse latent factor we obtained. The final P and Q are the sparse latent factor models, expressed as

[0039] Preferably, the steps of obtaining the threshold function are:

[0040] Based on user latent factor vector and the enterprise potential factor vector The formula for obtaining the threshold function is: in, and They are the user latent factors after the tth iteration. and enterprise potential factors The value of and They are the user latent factors after thresholding. and enterprise potential factors The value of determines which elements should be set to zero. and The elements in are processed based on the threshold function to obtain and based on and If the absolute value of is greater than θ, the value of the element is retained, that is, the element remains unchanged; if it is less than θ, the element is set to 0;

[0041] Preferably, the process of obtaining the diversity penalty term is as follows:

[0042] Define item data i to consist of multiple individual items, represented by i = {z 1 , z2 , ....z n},

[0043] Any two separate items in the definition are represented by z j and z m , for each individual item z in item i j , we calculate its similarity s with all other items mj ;

[0044] The similarity vector is denoted as s jm , in, Indicates z j The latent factor vector of Indicates z m The latent factor vector of for The norm of for The norm of express The transpose of

[0045] Define the diversity penalty term as Ω(L i ), where L i represents the total similarity of the items and, Get the diversity penalty term.

[0046] Among them, the items in the recommendation list are very similar, so the similarity between them is s jm will be relatively high, resulting in Ω(L i ) is large. On the contrary, if the items in the recommendation list are very diverse, the similarity between them is jm will be relatively low, resulting in Ω(L i ) has a smaller value;

[0047] Preferably, the steps for obtaining the objective function are:

[0048] Based on the sparse latent factor model and diversity penalty term, an objective function is constructed, which includes an error term, a regularization term, a combination term and a diversity penalty control term.

[0049] Based on the sparse latent factor model, the error term is obtained. The steps for obtaining the error term are:

[0050] Among them, it means User latent factor vector and the item latent factor vector The dot product of , we get the error term;

[0051] Based on the sparse latent factor model, the regularization term is obtained based on the formula:R (∑ u ||pu|| 2 +∑ i ||qi|| 2 ), where λ R is the regularization parameter, which controls the model complexity and obtains the regularization term;

[0052] Specifically, the regularization term constrains the complexity of the model by summing the Euclidean norms of the user latent factor vector and the enterprise latent factor vector and multiplying them by the regularization parameter;

[0053] Based on the sparse penalty term Ω(p u ) and Ω(q i ) to obtain the combined term, combined with the sparse penalty term Ω(p u ) and Ω(q i ), expressed as: S (||p u || 0 +||q i || 0 ), and obtain the sparse penalty term Ω(p u ) and Ω(q i )

[0054] Based on the diversity penalty term, the diversity penalty term parameter is introduced to obtain the diversity penalty control term: D Ω(L i ), where λ D To control the parameters of the diversity penalty term;

[0055] Based on the error term, regularization term, combination term and diversity penalty control term, the objective function is obtained: Among them, J(p u ,q i ,L i ) represents the objective function;

[0056] Among them, based on the sparse penalty term Ω(p u ) and Ω(q i ) obtains the combination term, combining the sparse penalty term and , which reflects the constraint on the sparsity of the latent factor vector. The significance of the combination term is the constraint on the sparsity of the latent factor vector: the combination term is formed by adding the sparse penalty term. Its purpose is to consider the sparsity of both the user's latent factor and the item's latent factor during the optimization process. Sparsity is of great significance in the recommendation system because it allows the model to focus on the most important and representative features, reduce the impact of noise and irrelevant factors, and thus improve the accuracy and relevance of the recommendation.

[0057] The present invention has the following beneficial effects:

[0058] In the present invention, by performing sparse constraints on latent factors, the recommendation system can focus more on the most important and representative features, thereby improving the relevance and accuracy of the recommendations. At the same time, by introducing diversity penalty items, the system can increase the diversity of the recommendation list while maintaining relevance. By establishing a sparse latent factor model and considering the similarity between items, the system can reduce over-recommendation of popular items and increase the exposure rate of long-tail items. The sparse latent factor model and diversity penalty items are combined to construct the final objective function. The prediction error and different types of regularization items are balanced based on the objective function, which can effectively solve the problem of bias in popular items in the recommendation system, thereby improving the diversity and novelty of recommended content. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A method step diagram of an enterprise data intelligent display and recommendation method proposed by the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] Embodiment 1

[0062] like Figure 1 As shown, the present invention proposes a method for intelligently displaying and recommending enterprise data, the steps of which include:

[0063] S1: Collect feedback data from users in the market on enterprise projects, conduct user interest modeling based on the feedback data, and obtain potential factors of users' preferences for projects;

[0064] Collect feedback data, which includes user data, project data, and enterprise evaluation data. Feedback data is represented by D = {(u,i,r ui )}, where u is user data, i is project data, and r ui Ratings for businesses by users.

[0065] Where u is the user identifier, usually a unique ID or user name in the database;

[0066] i is the project identifier, which can be the enterprise project product ID or enterprise project decision plan ID;

[0067] r uiFeedback data of user u on project i, which can be rating, clicking, purchasing, etc. of products or decision-making solutions in enterprise projects;

[0068] Specifically, the feedback data includes user data u and item data i, which is expressed as D = {(u,i,r ui )}, user data u can be the unique ID or user name in the database, etc., which can identify the user's identity; project data i can be the enterprise project product ID, enterprise project decision plan ID, etc.; enterprise evaluation data r ui It can be users' ratings of products or decision-making plans in enterprise projects (such as 1-5 points), click behaviors (0 / 1 indicates whether clicked), purchase behaviors (0 / 1 indicates whether purchased), etc.; these data can be collected from multiple data sources such as the enterprise's user database, sales records, website logs, user feedback platforms, etc.

[0069] S2: impose sparse constraints on the non-zero elements in the preference latent factor and construct a sparse latent factor model;

[0070] The steps to obtain the preference latent factors are:

[0071] Through the user's rating matrix R for the project, the rating matrix R includes R ui , the rating matrix R is an m×n matrix, where m is the number of users and n is the number of items. User u has rated item i, then R ui =r ui , where R ui =r ui Represents the user's rating of the enterprise in the rating matrix;

[0072] Among them, according to the collected user evaluation data on the project ui Construct a rating matrix R. If user u has rated the project, then R ui =r ui ; If there is no evaluation, the position may be empty or represented by a special symbol (special processing may be required in actual calculations, such as filling in default values ​​or using special algorithms to process sparse matrices);

[0073] Since not all users have rated all businesses, this matrix tends to be sparse;

[0074] Matrix decomposition is performed based on the rating matrix R. The implementation of matrix decomposition is as follows:

[0075] Decompose the rating matrix R into two low-rank matrices P and Q, define P as the user potential factor matrix, Q as the enterprise potential factor matrix,

[0076] The goal of user interest modeling is to find a low-dimensional latent factor space so that the vector representations of users and companies in this space can well predict the user's rating of the company. This is achieved through the following steps:

[0077] Among them, the user potential factor matrix P is an m×k matrix, m is the number of users, k is the number of potential factors, and the vector of each row of the user potential factor matrix P is p u , represents the weights of k potential factors of user u;

[0078] The enterprise potential factor matrix Q is an n×k matrix, where n is the number of enterprises and k is the number of potential factors. The vector of each column is q i , represents the weights of k potential factors of enterprise i;

[0079] Define the prediction matrix as The score matrix R and the prediction matrix Minimize the difference between them and get the minimization function L(P,Q). Optimize based on L(P,Q):

[0080] For user u, keep Q fixed and solve for the row vector p in P u Get p u The user latent factor vector Among them, q i T for q i The transpose of q i T The latent factor vector q of item i i Convert to a row vector, q i q i T is the latent factor vector q for item i i The result of multiplying it with itself is the autocovariance matrix of item i in the latent factor space;

[0081] For item i, keeping P fixed, solve for column vector q in Q i Get q i The potential factor vector of the enterprise Among them, p u T For p u The transpose of u p u T is the latent factor vector p of user u u The result of multiplying it with itself is the autocovariance matrix of user u in the latent factor space;

[0082] Get user latent factor vector and the enterprise potential factor vector That is, the latent factor of preference.

[0083] The goal of user interest modeling is to find a low-dimensional latent factor space so that the vector representations of users and companies in this space can well predict the user's rating of the company.

[0084] Specifically, in the context of user interest modeling, high-dimensional data often contains a large amount of information, which is redundant or noisy. It is necessary to use certain mathematical methods (such as matrix decomposition technology) to map the high-dimensional data into a low-dimensional space. Then, in the low-dimensional space, the data is more representative and concise, and can better reveal the essential characteristics behind the data.

[0085] Specifically, the rating matrix R is decomposed into two low-rank matrices P and Q. Each row vector of the user potential factor matrix P represents the weight of a potential factor of the user, and each column vector of the enterprise potential factor matrix Q represents the weight of a potential factor of the enterprise. Through the prediction matrix, the difference between the rating matrix and the prediction matrix is ​​minimized to obtain the minimization function. A series of optimization processes are performed through the minimization function, that is, for user u, Q is kept fixed and for item i, P is kept fixed, and finally the user potential factor vector is obtained. and the enterprise potential factor vector That is, the latent factor of preference.

[0086] The steps to obtain the minimization function are:

[0087] The process of obtaining the minimization function is: Among them, (u,i)∈D represents user u and item i in dataset D, Yes u The transpose of q i multiplication, that is, the predicted score of user u for item i, λ is the regularization coefficient, which is used to control the strength of the regularization term. Its value is obtained by presetting a grid containing multiple regularization coefficient values, evaluating each λ value in the grid, and selecting the λ value that makes the evaluation index reach the optimal (maximum value), ||p u || 2 and||q i || 2 They are p u and q i The Euclidean norm of is obtained to minimize the loss function.

[0088] The minimization function is the sum of squares of the differences between the predicted and actual ratings plus a regularization term, the purpose of which is to prevent fitting in the subsequent process.

[0089] The steps to obtain the sparse latent factor model are:

[0090] Based on the user potential factor matrix P and the enterprise potential factor matrix Q, a sparse penalty term is added and the sparse penalty term is defined as Ω(p u ) and Ω(q i ),Ω(p u )=λ S ||p u || 0 ,Ω(q i )=λ S ||q i || 0 , where ||·|| 0 represents the number of non-zero elements in the vector, λ s is a hyperparameter that controls the strength of the sparse penalty term, based on the sparse penalty term Ω(p u ) and Ω(q i ), update the user potential factor matrix P and the enterprise potential factor matrix Q. The updating method of the user potential factor matrix P is:

[0091] Fix the enterprise potential factor matrix Q and update p in the user potential factor matrix P u , after updating

[0092] Fix the user potential factor matrix P and update q in the enterprise potential factor matrix Q i , after updating

[0093] Define a thresholding function θ to reduce the user's latent factor vector and the enterprise potential factor vector The number of non-zero elements in the , through the threshold function, to obtain the optimized user potential factor vector and the enterprise potential factor vector

[0094] Among them, the user latent factor vector is defined as and the enterprise potential factor vector The elements in are x. For each element x, it is processed based on the threshold function to obtain the processed element x. When |x|>θ, the value of the element is retained, that is, x remains unchanged. When |x|≤θ, the element is set to 0.

[0095] By adding sparse penalty terms and thresholding function operations, the row vectors in P and Q are finally obtained. and column vector This is the sparse latent factor we obtained. The final P and Q are the sparse latent factor models, expressed as

[0096] Specifically, based on the user potential factor matrix P and the enterprise potential factor matrix Q, a sparse penalty term Ω(p u ) and Ω(q u ), by introducing a sparse penalty term, the latent factor vector is made more sparse, thereby focusing on the most important and representative features and improving the relevance and accuracy of recommendations.

[0097] The steps to obtain the threshold function are:

[0098] Based on user latent factor vector and the enterprise potential factor vector The formula for obtaining the threshold function is: in, and They are the user latent factors after the tth iteration. and enterprise potential factors The value of and They are the user latent factors after thresholding. and enterprise potential factors The value of determines which elements should be set to zero. and The elements in are processed based on the threshold function to obtain and based on and If the absolute value of is greater than θ, the value of the element is retained, that is, the element remains unchanged; if it is less than θ, the element is set to 0;

[0099] Specifically, we set and based on the threshold function θ, since and are the values ​​of user latent factor and enterprise latent factor after the first iteration, and are the values ​​of user latent factors and enterprise latent factors after threshold processing. For the user latent factor vector and the enterprise potential factor vector When the element in is greater than 0, the value of the element is retained, and when it is less than or equal to 0, the element is set to 0. In this way, through the thresholding operation, the number of non-zero elements in the latent factor vector is further reduced to obtain a sparse latent factor, and finally a sparse latent factor model is obtained based on and.

[0100] S3: Obtain the similarity vector between the items, and calculate the diversity penalty item based on the similarity vector to obtain the diversity penalty item;

[0101] The process of obtaining the diversity penalty term is as follows:

[0102] Define item data i to consist of multiple individual items, represented by i = {z 1 , z 2 , ....z n},

[0103] Any two separate items in the definition are represented by z j and z m , for each individual item z in item i j , we calculate its similarity s with all other items mj ;

[0104] The similarity vector is denoted as s jm , in, Indicates z j The latent factor vector of Indicates z m The latent factor vector of for The norm of for The norm of express The transpose of

[0105] Define the diversity penalty term as Ω(L i ), where L u represents the total similarity of the items and, Get the diversity penalty term.

[0106] Among them, the items in the recommendation list are very similar, so the similarity between them is s jm will be relatively high, resulting in Ω(L u ) is large. On the contrary, if the items in the recommendation list are very diverse, the similarity between them is jm will be relatively low, resulting in Ω(L i ) has a smaller value;

[0107] Specifically, through the formula The diversity penalty item obtained, if the items in the recommendation list are very similar, then the similarity between them is s jm will be relatively high, resulting in Ω(L i ) value is large; on the contrary, if the items in the recommendation list are very diverse, the similarity between them is s jm will be relatively low, resulting in Ω(Li ) is small. By introducing the diversity penalty term, the system can increase the diversity of the recommendation list while maintaining relevance, reduce over-recommendations for popular items, and increase the exposure of long-tail items.

[0108] S4: Construct the final objective function based on the sparse latent factor model and the diversity penalty term to obtain the final objective function, optimize the recommendation based on the final objective function, and obtain the final recommendation result;

[0109] The steps to obtain the objective function are:

[0110] Based on the sparse latent factor model and diversity penalty term, an objective function is constructed, which includes an error term, a regularization term, a combination term and a diversity penalty control term.

[0111] Based on the sparse latent factor model, the error term is obtained. The steps for obtaining the error term are:

[0112] Among them, it means User latent factor vector and the item latent factor vector The dot product of , we get the error term;

[0113] Specifically, the error term formula is The sum of squares of the differences between the dot product of the user latent factor vector and the item latent factor vector and the actual rating, reflecting the error between the predicted rating and the actual rating;

[0114] Based on the sparse latent factor model, the regularization term is obtained based on the formula: R (∑ u ||pu|| 2 +∑ i ||qi|| 2 ), where λ R is the regularization parameter, which controls the model complexity and obtains the regularization term;

[0115] Specifically, the regularization term constrains the complexity of the model by summing the Euclidean norms of the user latent factor vector and the enterprise latent factor vector and multiplying them by the regularization parameter;

[0116] Based on the sparse penalty term Ω(p u ) and Ω(q i ) to obtain the combined term, combined with the sparse penalty term Ω(p u ) and Ω(q i ), expressed as: S (||p u || 0 +||q i || 0), and obtain the sparse penalty term Ω(p u ) and Ω(q i )

[0117] Based on the diversity penalty term, the diversity penalty term parameter is introduced to obtain the diversity penalty control term: D Ω(L i ), where λ D To control the parameters of the diversity penalty term;

[0118] Based on the error term, regularization term, combination term and diversity penalty control term, the objective function is obtained: Among them, J(p u ,q i ,L i ) represents the objective function;

[0119] Among them, based on the sparse penalty term Ω(p u ) and Ω(q i ) obtains the combination term, combines the sparse penalty term and , reflects the constraint on the sparsity of the latent factor vector. The significance of the combination term is to constrain the sparsity of the latent factor vector: the combination term is formed by adding the sparse penalty term. Its purpose is to consider the sparsity of the user's latent factor and the item's latent factor at the same time during the optimization process. Sparsity is of great significance in the recommendation system because it allows the model to focus on the most important and representative features, reduce the impact of noise and irrelevant factors, and thus improve the accuracy and relevance of the recommendation.

[0120] Specifically, the objective function The output of is a single real value, which represents the overall performance of the entire recommendation system. The smaller the value, the higher the prediction accuracy of the recommendation system, the simpler the model, the sparser the latent factor vector, the better the diversity of the recommendation list, and the significance of the output value. The value indicates that the recommendation list performs well in predicting the score. At the same time, the model is simple, the latent factor vector is sparse, and the recommendation list has high diversity. The diversity of items included in the recommendation list is good, which helps to discover and recommend novel items, improve the diversity and novelty of the recommendations, and give recommendations. The larger the value, the better. A value of 0 indicates that the recommendation system is deficient in some aspects, such as poor prediction accuracy, overfitting of the model, insufficient sparseness of the latent factor vector, or lack of diversity in the recommendation list. The diversity of items included in the recommendation list is poor, resulting in the recommendation list being too concentrated on popular items and ignoring other potential high-quality items.

[0121] In summary, in the objective function of the entire recommendation system, the combination term works together with other terms (error term, regularization term, and diversity penalty term). When optimizing the objective function, the algorithm will try to find a balance between reducing prediction errors, controlling model complexity, increasing the sparsity of latent factors, and improving recommendation diversity. The existence of the combination term prompts the model to tend to choose solutions with sparse latent factors when looking for the optimal solution, so as to better explore the essential characteristics of users and projects and achieve more accurate and diverse recommendations.

[0122] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent display and recommendation of enterprise data, characterized in that: The steps of this method are as follows: S1: Collect feedback data from users on the market about enterprise projects, conduct user interest modeling based on the feedback data, and obtain potential factors of users' preferences for enterprise projects; S2: impose sparse constraints on the non-zero elements in the preference latent factor and construct a sparse latent factor model; S3: Obtain the similarity vectors between enterprise projects, and calculate the diversity penalty item based on the similarity vector to obtain the diversity penalty item; S4: Construct the final objective function based on the sparse latent factor model and the diversity penalty term to obtain the final objective function, optimize the recommendation based on the final objective function, and obtain the final recommendation result.

2. The method for intelligently displaying and recommending enterprise data according to claim 1, characterized in that: The feedback data includes user data, project data and enterprise evaluation data. The feedback data is expressed as D={(u,i,r ui )}, where u is user data, i is project data, and r ui Evaluate data for businesses.

3. The method for intelligently displaying and recommending enterprise data according to claim 1, characterized in that: The steps for obtaining the preference potential factor are: Through the user's rating matrix R for the project, where the rating matrix R includes R ui , and is an m×n matrix, where m is the number of users, n is the number of items, and user u has a rating for item i, then R ui =r ui , where R ui =r ui Represents the user's rating of the enterprise in the rating matrix; Matrix decomposition is performed based on the score matrix R. The implementation steps of the matrix decomposition are as follows: Decompose the rating matrix R into two low-rank matrices P and Q, where P is the user potential factor matrix and Q is the enterprise potential factor matrix. The user potential factor matrix P is an m×k matrix, k is the number of potential factors, and the vector of each row of the user potential factor matrix P is p u ; The enterprise potential factor matrix Q is an n×k matrix, where k is the number of potential factors and each column is a vector of q i , represents the weights of k potential factors of enterprise i; Through a prediction matrix The score matrix R and the prediction matrix Minimize the difference between them and get the minimization function L(P,Q). Optimize based on L(P,Q): For user u, keep Q fixed and solve for the row vector p in P u Get p u The user latent factor vector Among them, q i T Q i The transpose of q i T The latent factor vector q of item i i Convert to a row vector, q i q i T is the latent factor vector q for item i i The result of multiplying it with itself is the autocovariance matrix of item i in the latent factor space; For item i, keeping P fixed, solve for column vector q in Q i Get q i The potential factor vector of the enterprise Among them, p u T For p u The transpose of u p u T is the latent factor vector p of user u u The result of multiplying it with itself is the autocovariance matrix of user u in the latent factor space; Get user latent factor vector and the enterprise potential factor vector That is, the latent factor of preference.

4. The method for intelligently displaying and recommending enterprise data according to claim 3, characterized in that: The steps for obtaining the minimization function are: The process of obtaining the minimization function is: Among them, (u,i)∈D represents user u and item i in dataset D, Yes u The transpose of q i Multiply, that is, user u's predicted score for item i, λ is the regularization coefficient, ||p u || 2 and||q i || 2 They are p u and q i The Euclidean norm of is obtained to minimize the loss function.

5. The method for intelligently displaying and recommending enterprise data according to claim 1, characterized in that: The steps for obtaining the sparse latent factor model are: Based on the user potential factor matrix P and the enterprise potential factor matrix Q, a sparse penalty term is added, and the sparse penalty term is Ω(p u ) and Ω(q i ),Ω(p u )=λ S ∥p u ∥0,Ω(q i )=λ S ∥q i ∥0, where ∥·∥0 represents the number of non-zero elements in the vector, and λ s is a hyperparameter that controls the strength of the sparse penalty term, based on the sparse penalty term Ω(p u ) and Ω(q i ), update the user potential factor matrix P and the enterprise potential factor matrix Q. The updating method of the user potential factor matrix P is: Fix the enterprise potential factor matrix Q and update p in the user potential factor matrix P u , after updating Fix the user potential factor matrix P and update q in the enterprise potential factor matrix Q i , after updating Set and based on the threshold function θ, obtain the optimized user potential factor vector and the enterprise potential factor vector Through the user latent factor vector and the enterprise potential factor vector For each element x in , process it based on the threshold function θ to obtain the processed element x. When |x|>θ, keep the value of the element, that is, x remains unchanged. When |x|≤θ, set the element to 0. Based on the sparse penalty term and thresholding function operation, the row vectors in P and Q are obtained. and column vector is a sparse latent factor. Based on P and Q, a sparse latent factor model is obtained, which is expressed as 6. The method for intelligently displaying and recommending enterprise data according to claim 5, characterized in that: Based on the threshold function θ, the optimized user potential factor vector is obtained and the enterprise potential factor vector The steps to obtain are: Based on user latent factor vector and the enterprise potential factor vector The formula for obtaining the threshold function is: in, and They are the user latent factors after the tth iteration. and enterprise potential factors The value of and They are the user latent factors after thresholding. and enterprise potential factors For the value of and The elements in are processed based on the threshold function to obtain and based on and The absolute value of and If the absolute value of is greater than θ, the value of the element is retained. and If the absolute value of is less than θ, the corresponding element is set to 0, and finally the optimized user potential factor vector is obtained. and the enterprise potential factor vector 7. The method for intelligently displaying and recommending enterprise data according to claim 1, characterized in that: The process of obtaining the diversity penalty term is as follows: Suppose that the project data i consists of multiple individual projects, represented by i = {z1, z2, .... z n }, Any two separate items in item data i are represented by z j and z m , for each individual item z in item i j , calculate its similarity s with all other items m ; The similarity vector is denoted as s jm , in, Indicates z j The latent factor vector of Indicates z m The latent factor vector of for The norm of for The norm of express The transpose of Get the diversity penalty term through the formula Among them, L i Represents the total similarity of the items.

8. The method for intelligently displaying and recommending enterprise data according to claim 1, characterized in that: The steps for obtaining the objective function are: Based on the sparse latent factor model and diversity penalty term, an objective function is constructed, which includes an error term, a regularization term, a combination term, and a diversity penalty control term. The error term is obtained based on the sparse latent factor model. The steps are: Based on the formula: The formula represents User latent factor vector and the item latent factor vector The dot product of , we get the error term; Based on the sparse latent factor model, the regularization term is obtained based on the formula: R (∑ u ∥pu∥ 2 +∑ i ∥qi∥ 2 ), where λ R is the regularization parameter; Based on the sparse penalty term Ω(p u ) and Ω(q i ) to obtain the combined term, combined with the sparse penalty term Ω(p u ) and Ω(q i ), expressed as: S (∥p u ∥0+∥q i ∥0), and obtain the sparse penalty term Ω(p u ) and Ω(q i ) Based on the diversity penalty term, the diversity penalty term parameter is introduced to obtain the diversity penalty control term: D Ω(L i ), where λ D To control the parameters of the diversity penalty term; Based on the error term, regularization term, combination term and diversity penalty control term, the objective function is obtained: Among them, J(p u ,q i ,L i ) represents the objective function.

Images