Wallboard outsourcing enterprise recommendation method, system and device based on multi-modal data fusion and storage medium

Abstract

The invention discloses a wallboard outsourcing enterprise recommendation method, system and device based on multi-modal data fusion and a storage medium, and belongs to the technical field of information recommendation. The invention discloses a wallboard outsourcing enterprise recommendation method. The method comprises the following steps: obtaining related data of an outsourcing enterprise and a demand enterprise; potential features are mined according to the related data of the outsourcing enterprise, and mined potential feature vectors are obtained; potential features are captured according to the related data of the demand enterprise, and captured potential feature vectors are obtained; performing score matrix fusion and filling on the mined potential feature vectors and the mined potential feature vectors to obtain a score matrix; after the scoring matrix is corrected, the corrected scoring matrix is obtained, the first N items of the corrected scoring matrix are recommended to the target demand user, and the problem of demand-outsourcing scoring data sparsity is effectively solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of information recommendation, and particularly relates to a method, system, device and storage medium for recommending wall panel external cooperation enterprises based on multi-modal data fusion. Background Art

[0002] The application of network collaborative manufacturing in the industrial field has achieved extensive progress, especially outstanding in dealing with complex and customized production environments. As an industry highly dependent on precision production and resource coordination, especially when it comes to the external cooperation of wall panels, which are core components, introducing network collaborative manufacturing has important strategic significance. In the traditional wall panel external cooperation supply chain model, the selection of external cooperation manufacturers is fixed, and the resource allocation efficiency is low, resulting in insufficient supply chain response ability and flexibility. The production capacity and resource scheduling of external cooperation enterprises have not been fully optimized, thus restricting the overall production efficiency.

[0003] In recent years, although there have been studies proposing to build a network collaborative manufacturing external cooperation service platform for wall panel manufacturing, the existing platforms are still insufficient in the analysis and utilization of interactive data. Specifically, user behavior data (such as ratings, browsing history, collections, etc.) have not been fully mined, resulting in a lack of accurate modeling and efficient matching capabilities for user needs. This not only limits the accuracy and personalization level of the recommendation system, but also makes it difficult to effectively respond to complex and changing market demands, further restricting the resource allocation efficiency of external cooperation enterprises and the optimization effect of the supply chain.

[0004] With the rapid development of technologies such as big data, cloud computing, and artificial intelligence, especially the gradual application of machine learning and deep learning in the industrial field, it provides strong technical support for the recommendation system of external cooperation enterprises. However, the recommendation of wall panel external cooperation enterprises still faces the following several technical challenges:

[0005] 1) The problem of sparse rating data between users and external cooperation enterprises makes it difficult for traditional collaborative filtering-based methods to fully mine effective user preferences;

[0006] 2) The inconsistency between explicit feedback and implicit preferences leads to biases in the recommendation model. Explicit feedback only reflects part of the user's needs, while potential and unexpressed preferences have a greater impact on the recommendation quality;

[0007] 3) The neglect of time-series dynamic change factors makes traditional recommendation algorithms unable to effectively capture the long-term evolution and periodic changes of action attributes. Summary of the Invention

[0008] The purpose of the present invention is to provide a method, system, device and storage medium for recommending wall panel external cooperation enterprises based on multi-modal data fusion, which is used to solve the technical problem of sparse demand-external cooperation rating data in the prior art.

[0009] To achieve the above object, the present invention is implemented by the following technical solutions:

[0010] The present invention discloses a method for recommending external cooperation enterprises of wall panels based on multi-modal data fusion, comprising the following steps:

[0011] Obtain relevant data of external cooperation enterprises and demand enterprises;

[0012] Mine potential features based on the relevant data of external cooperation enterprises to obtain the mined potential feature vector; capture potential features based on the relevant data of demand enterprises to obtain the captured potential feature vector;

[0013] Perform scoring matrix fusion and filling on the mined potential feature vector and the mined potential feature vector to obtain a scoring matrix;

[0014] After correcting the scoring matrix, obtain the corrected scoring matrix, and recommend the top N items recommended by the corrected scoring matrix to the target demand user.

[0015] Further, the relevant data includes attribute data, behavior data and scoring matrix of external cooperation enterprises and demand enterprises;

[0016] The attribute data of the external cooperation enterprise includes the service type and service description of the enterprise; the attribute data of the demand enterprise includes the demand type, price preference, procurement cycle preference and service equipment preference of the enterprise;

[0017] The behavior data includes the browsing times, browsing duration and favorite records of the demand enterprise.

[0018] Further, the mining of potential features is performed using the Attention-LSTM model;

[0019] The mined potential feature vector consists of the weight W1 of the Attention-LSTM model, the auxiliary information Y of the external cooperation enterprise j j , and the Gaussian noise ε j variables;

[0020] where, v j = Attention-LSTM(W1, Y j ) + ε j ;

[0021]

[0022] where: v j is the output vector of each sub-item of the business document, W1 is all the weight parameters in the model, Y j is the sub-item j of the business document of the external cooperation enterprise, and ε jis Gaussian noise that follows a normal distribution, I is the identity matrix, σ v 2 is the output vector V j is the variance of;

[0023] When W1 follows a Gaussian prior distribution, the conditional distributions of W1 and the potential vector V of the external cooperation enterprise are as follows:

[0024]

[0025] where: W1 are all the weight parameters in the Attention-LSTM model, σ w 2 and σ v 2 are the variances of the weight vector W1 and the output vector V respectively, w k is the weight sub-item in W1, ε j is Gaussian noise, I is the identity matrix, v j is the output vector of each document sub-item.

[0026] Furthermore, the capturing of potential features is performed using a VAE model;

[0027] The captured potential feature vector consists of the weight W of the VAE model, the auxiliary information X of the demanding enterprise i i and the Gaussian noise ε i variables;

[0028] where: u i = VAE(W, X i , S i ) + ε i ;

[0029]

[0030] where: u i is the output vector of each document sub-item, W are all the weight parameters in the VAE model, X i is the sub-item i of the preferred document of the demanding enterprise, ε i is Gaussian noise that follows a normal distribution, I is the identity matrix, σ u 2 is the output vector V i is the variance of;

[0031] When the weight W follows a Gaussian prior distribution, the conditional distributions of W and the potential vector U of the demanding enterprise are as follows:

[0032]

[0033] where: σ W 2and σ U 2 are the variances of the weight vector W and the output vector U in the VAE model, respectively. w k is the weight sub-item in W, and u i is the output vector of each document sub-item. S i is the scoring matrix.

[0034] Furthermore, the fusion and filling of the scoring matrix are performed using the PMF technique;

[0035] The conditional distribution of the obtained scoring matrix R is as follows:

[0036]

[0037] where: U is the potential vector of the demanding enterprise, V is the potential vector of the subcontracting enterprise, and R ij is the score of the demanding enterprise i for the subcontracting enterprise j. u i and v j are the potential vectors of the demanding enterprise i and the subcontracting enterprise j, respectively. σ 2 is the variance of the Gaussian distribution.

[0038] Furthermore, the correction of the scoring matrix includes a primary correction and a secondary correction;

[0039] The primary correction is performed using the BERT model for semantic analysis and sentiment calculation, and the expression used is as follows:

[0040] R' = R0 + λ1W R ;

[0041]

[0042] where λ1 is a hyperparameter that controls the influence of comment sentiment on score correction.

[0043] Furthermore, the secondary correction is to apply the pyramid behavior characteristics to perform a secondary correction on the scoring matrix. The specific steps are as follows:

[0044] The behavior of the demanding enterprise is divided into three layers of pyramid behavior characteristics in the time dimension: short-term, medium-term, and long-term. Weighted modeling is performed on the three layers of pyramid behavior characteristics of short-term, medium-term, and long-term. The specific calculation formula is as follows:

[0045] The behavior characteristics are stratified by time window and represented in vector form:

[0046] Short-term behavior vector: B short = {b 1 , b 2 , …, b TS};

[0047] Medium-term behavior vector: B med ={b TS+1 , b TS+2 , …, b Tm};

[0048] Long-term behavior vector: B long ={b Tm+1 , b Tm+2 , …, b Ti};

[0049] Among them, TS, Tm, and Ti respectively represent the length breakpoints of the time windows of short-term, medium-term, and long-term behaviors;

[0050] The behavior characteristics within all time windows can be represented as a three-dimensional vector b t ={b fr , b ti , b co}(t = 1, 2, …, Ti);

[0051] Among them, b fr is the number of times the user browses; b ti is the browsing duration of the user; b co is the user's favorite record;

[0052] Different weights are assigned to the three sub-items of each b t , and the calculation formula is as follows:

[0053]

[0054] Among them, W fr , W ti , W co are the weights of the three behavior characteristics of short-term, medium-term, and long-term; t is the interaction behavior record duration of the demand enterprise, λ is the adjustment parameter for controlling the change trend of each time window, and α, β, γ are the parameters for adjusting the change trend between windows, where α > β > γ;

[0055] Through the following weighted function, the behavior data of each demand enterprise is weighted and adjusted to obtain the complete behavior sequence of the user:

[0056] B i (t) = b fr (t)·W fr (t) + b ti (t)·W ti (t) + b co (t)·W co ;

[0057] Among them: b fr is the number of times the user browses, b ti is the browsing duration of the user, bco For the user's favorite records, W fr ,W ti ,W co are the weights of three behavioral characteristics: short-term, medium-term, and long-term.

[0058] The obtained complete behavior sequence is input into the LSTM network to capture the temporal characteristics of user behavior. Subsequently, the contribution of each time step to the final representation is dynamically adjusted through the Attention mechanism. The specific calculation formula is as follows:

[0059] h t = LSTM(B i (t));

[0060] Where: h t is the output of the LSTM network at time step t, and B i (t) is the complete behavior sequence of each demand enterprise.

[0061] Next, the Attention layer takes the hidden state h t output by the LSTM as input, thereby giving different weights to each time step to obtain the attention score. The calculation formula is as follows:

[0062]

[0063] Where, h t is the hidden state of the LSTM network at time step t; W a is a weight matrix; b a is the bias term; w a is an output weight vector.

[0064] By performing a softmax operation on the score score(h t ) of each time step, the score is converted into an attention weight. The calculation formula is as follows:

[0065]

[0066] Where, α t represents the attention weight at time step t, and score(h t ) is the attention score at time step t;

[0067] Using the calculated attention weight α t to weight the hidden state h t output by the LSTM, the weighted behavior representation B weighted is obtained. The calculation formula is as follows:

[0068]

[0069] Based on the weighted behavioral feature B weighted , perform a secondary correction on the scoring matrix, and the formula is as follows:

[0070] R” = R' + λB weighted ;

[0071] Among them, R’ is the scoring matrix after the first correction, and λ is a hyperparameter that controls the influence of behavioral features on scoring correction.

[0072] The present invention also discloses a wall panel external cooperation enterprise recommendation system based on multi-modal data fusion, which includes

[0073] a data acquisition module for acquiring relevant data of external cooperation enterprises and demand enterprises;

[0074] a sparse scoring matrix completion module for mining potential features based on the relevant data of external cooperation enterprises to obtain the mined potential feature vectors; capturing potential features based on the relevant data of demand enterprises to obtain the captured potential feature vectors; performing scoring matrix fusion and filling on the mined potential feature vectors and the captured potential feature vectors to obtain a scoring matrix;

[0075] a comment sentiment analysis correction module for performing a first correction on the scoring matrix according to the sentiment analysis result to obtain a corrected scoring matrix;

[0076] a pyramid behavioral feature optimization module for performing a secondary correction on the scoring matrix according to behavior and timing information to obtain a corrected scoring matrix;

[0077] an external cooperation enterprise recommendation ranking module for recommending the top N items recommended by the corrected scoring matrix to the target demand user.

[0078] The present invention also discloses a terminal device, and the processor implements the steps of the above method when executing the computer program.

[0079] The present invention also discloses a computer-readable storage medium, the computer-readable storage medium stores a computer program, and is characterized in that the computer program implements the steps of the above method when being executed by a processor.

[0080] Compared with the prior art, the present invention has the following beneficial effects:

[0081] The present invention discloses a method for recommending wall panel external cooperation enterprises based on multi-modal data fusion. By combining the Attention-LSTM and VAE models, it deeply mines the auxiliary information of external cooperation enterprises and demand enterprises, effectively solving the problem of sparse demand-outsourcing scoring data. By using the BERT model to combine explicit feedback and implicit preferences, it optimizes the recommendation accuracy and personalization level. At the same time, due to the establishment of pyramid behavior feature correction, the system can dynamically capture the long-term evolution trend and periodic changes of user needs, improving the response ability and adaptability of the recommendation algorithm to complex market demands, thereby optimizing the resource allocation of external cooperation enterprises and the efficiency of supply chain management, and significantly improving the overall production efficiency of printing machine wall panel manufacturing enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 It is the process framework of the method for recommending wall panel external cooperation enterprises based on multi-modal data fusion of the present invention;

[0083] Figure 2 It is the structural principle schematic diagram of the PMF of the present invention that combines Attention-LSTM and VAE models;

[0084] Figure 3 It is the system diagram of the method for recommending wall panel external cooperation enterprises based on multi-modal data fusion of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0085] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0086] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present invention described here can be implemented in an order different from those illustrated or described here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0087] The present invention discloses a method for recommending external cooperation enterprises of wall panels based on multi-modal data fusion, comprising the following steps:

[0088] Step 1: Obtain relevant data of the demand enterprise and the external cooperation enterprise, including attribute data, behavior data, and a scoring matrix;

[0089] Step 2: According to the auxiliary attribute data of the external cooperation enterprise provided in Step 1, use the Attention-LSTM model to mine its potential features;

[0090] Step 3: According to the auxiliary attribute data of the demand enterprise provided in Step 1, use the VAE model to capture its potential features;

[0091] Step 4: According to the potential feature vectors obtained in Step 2 and Step 3, use the PMF technology to perform scoring matrix fusion and filling;

[0092] Step 5: According to the scoring matrix formed in Step 4, use the BERT model to perform semantic analysis and sentiment calculation to correct the scoring matrix;

[0093] Step 6: According to the scoring matrix corrected in Step 5, apply the pyramid behavior features to perform secondary correction on the scoring matrix;

[0094] Step 7: According to the scoring matrix secondarily corrected in Step 6, recommend the top N items to the target demand enterprise.

[0095] Preferably, the attribute data and behavior data include:

[0096] According to the category and characteristics of the object, the attribute data is divided into auxiliary attribute data of the external cooperation enterprise and auxiliary attribute data of the demand enterprise; among them, the auxiliary attribute data of the external cooperation enterprise includes the service type and service description of the enterprise; the auxiliary attribute data of the demand enterprise covers four key factors: demand type, price preference, procurement cycle preference, and service equipment preference of the enterprise;

[0097] The behavior data comes from the interaction activities of the demand enterprise, specifically three types of log data: the number of views, viewing duration, and favorite records, reflecting the behavior patterns and interest preferences shown in the process of the user interacting with the platform.

[0098] Preferably, the potential feature vector obtained by using Attention-LSTM to mine the auxiliary information of the external cooperation enterprise consists of three parts: the weight W1 of Attention-LSTM, the auxiliary information Y of the external cooperation enterprise j j and the Gaussian noise ε j variables; as follows:

[0099] v j = Attention-LSTM(W1, Y j) + ε j ;

[0100]

[0101] where: v j is the output vector of each business document sub - item, W1 is all the weight parameters in the model, Y j is the sub - item j of the business document of the external cooperation enterprise, ε j is Gaussian noise following a normal distribution, I is the identity matrix, σ v 2 is the variance of the output vector V j .

[0102] Assume that the weight W1 also follows a Gaussian prior distribution. The conditional distributions of W1 and the potential vector V of the external cooperation enterprise are as follows:

[0103]

[0104] where: W1 is all the weight parameters in the Attention - LSTM model, σ w 2 and σ v 2 are the variances of the weight vector W1 and the output vector V respectively, w k is the weight sub - item in W1, ε j is Gaussian noise, I is the identity matrix, v j is the output vector of each document sub - item.

[0105] Preferably, the potential feature vector obtained by using VAE to capture the auxiliary information of the demand enterprise consists of three parts: the weight W of VAE, the auxiliary information X i of the demand enterprise i i and the Gaussian noise ε

[0106] u i = VAE(W, X i , S i ) + ε i ;

[0107]

[0108] where: u i is the output vector of each document sub - item, W is all the weight parameters in the VAE model, X i is the sub - item i of the preference document of the demand enterprise, ε i is Gaussian noise following a normal distribution, I is the identity matrix, σ u 2 is the variance of the output vector V i .

[0109] Assume that the weight W also follows a Gaussian prior distribution, and the conditional distributions of W and the potential vector U of the demanding enterprise are as follows:

[0110]

[0111] where: σ W 2 and σ U 2 are the variances of the weight vector W and the output vector V in the VAE model respectively, w k is the weight sub-item in W, u i is the output vector of each document sub-item, and S i is the scoring matrix.

[0112] Preferably, for fusing and filling the scoring matrix by using the PMF technology, the conditional distribution of the scoring matrix R is:

[0113]

[0114] where: U is the potential vector of the demanding enterprise, V is the potential vector of the subcontracting enterprise, R ij is the score of the demanding enterprise i for the subcontracting enterprise j, u i and v j are the potential vectors of the demanding enterprise i and the subcontracting enterprise j respectively, and σ 2 is the variance of the Gaussian distribution.

[0115] Preferably, for semantic analysis and sentiment calculation by using the BERT model, the calculation expression is:

[0116] R' = R0 + λ1W R ;

[0117]

[0118] Preferably, the steps for secondarily correcting the scoring matrix by applying the pyramid behavior characteristics are:

[0119] Divide the behavior of the demanding enterprise into three layers of pyramid behavior characteristics of short-term, medium-term, and long-term in the time dimension, and perform weighted modeling on various behavior characteristics. The specific calculation formula is as follows:

[0120] Short-term behavior vector: B short = {b 1 , b 2 , …, b TS};

[0121] Medium-term behavior vector: B med = {b TS+1 , b TS+2 , …, b Tm};

[0122] Long - term behavior vector: B long ={b Tm+1 ,b Tm+2 ,…,b Ti};

[0123] Wherein, TS, Tm, and Ti respectively represent the length breakpoints of the time windows for short - term, medium - term, and long - term behaviors;

[0124] The behavior characteristics within all time windows can be represented as a three - dimensional vector b t ={b fr ,b ti ,b co}(t = 1, 2…Ti);

[0125] Wherein, b fr is the number of views, b ti is the browsing duration, b co is whether it is favorited;

[0126] Assign different weights to the three sub - items of each b t , and the calculation formula is as follows:

[0127]

[0128] Wherein, W fr ,W ti ,W co are the weights of the three behavior characteristics, t is the interaction behavior record duration of the demanding enterprise, λ is the adjustment parameter for controlling the change trend of each time window, and α, β, γ are the parameters for adjusting the change trend between windows, with α>β>γ; The recommended parameter values are: λ1 = 0.4, λ2 = 0.15, λ3 = 0.07, α = 1, β = 0.5, β = 0.2;

[0129] Through the above - mentioned weighting function, the behavior data of each demanding enterprise is weighted and adjusted. Specifically, the behavior data of each demanding enterprise in the past t days will be multiplied by the corresponding time - weighted value, and the weighted behavior data is summarized to obtain the complete behavior sequence of the user:

[0130] B i (t)=b fr (t)·W fr (t)+b ti (t)·W ti (t)+b co (t)·W co ;

[0131] Where: b fr is the number of views of the user, b ti is the browsing duration of the user, bco For the user's favorite records, W fr ,W ti ,W co are the weights of the short-term, medium-term, and long-term behavioral characteristics.

[0132] Input the weighted complete behavior sequence into the LSTM network to capture the temporal characteristics of user behavior. Dynamically adjust the contribution of each time step to the final representation through the Attention mechanism. The specific calculation formula is as follows:

[0133] Weighted behavior sequence B i (t) is processed by the LSTM model to obtain the hidden state h at each time step t :

[0134] h t = LSTM(B i (t));

[0135] Among them: h t is the output of the LSTM network at time step t, and B i (t) is the complete behavior sequence of each demand enterprise.

[0136] Next, the Attention layer takes the hidden state h output by the LSTM t as input, so as to give different weights to each time step and obtain the attention score. The calculation formula is as follows:

[0137]

[0138] Among them, h t is the hidden state of the LSTM at time step t, W a is a weight matrix used to map the hidden state to a higher-dimensional space, b a is the bias term used to adjust the mapping result, w a is an output weight vector used to calculate the final score;

[0139] By performing a softmax operation on the score score(h t ) at each time step, convert the score into an attention weight; the calculation of the attention weight is as follows:

[0140]

[0141] Among them, the denominator is the sum of the exponents of the scores of all time steps, used for normalization, and α t is the Attention weight for time step t;

[0142] Use the calculated attention weight α tPerform weighting on the hidden state h output by the LSTM t to obtain the weighted behavior representation B weighted :

[0143]

[0144] Based on the weighted behavior feature B weighted , perform a secondary correction on the scoring matrix:

[0145] R” = R' + λB weighted ;

[0146] where R’ is the scoring matrix after the first correction, and λ is a hyperparameter that controls the influence of the behavior feature on the scoring correction.

[0147] The present invention also discloses a wall panel external cooperation enterprise recommendation system that integrates review text analysis and pyramid features, including:

[0148] A wall panel external cooperation data collection module, responsible for collecting the attribute data and behavior data of enterprises from the platform;

[0149] A sparse scoring matrix completion module, which supplements and optimizes the missing data in the matrix based on the attribute information;

[0150] A comment sentiment analysis correction module, which adjusts the scoring matrix by analyzing the sentiment tendency of user comments;

[0151] A pyramid behavior feature optimization module, used to combine the behavior data with the time series change to perform a secondary correction on the matrix;

[0152] An external cooperation enterprise recommendation ranking module, which generates a priority recommendation list of the top N enterprises based on the scoring results.

[0153] The following further describes the present invention in detail with reference to the accompanying drawings:

[0154] See Figures 1 to 3 As shown, the embodiment of the present invention discloses a wall panel external cooperation enterprise recommendation method and system based on multi-modal data fusion. The method includes the following steps:

[0155] Step 1, obtain the relevant data of the demand enterprise and the external cooperation enterprise, including attribute data, behavior data, and scoring matrix;

[0156] This step specifically includes the following steps:

[0157] Data will be retrieved from the external cooperation platform database, including attribute data and behavioral data. The auxiliary attribute data of external cooperation enterprises include service type and service description; the auxiliary attribute data of demand enterprises cover four key factors: demand type, price preference, procurement cycle preference, and service equipment preference. Behavioral data comes from the interaction activities of demand enterprises, such as the number of views, view duration, and favorite records. After data cleaning and outlier handling, all data will be organized into a structured format and stored in the form of a CSV table;

[0158] Step 2: According to the auxiliary attribute data of external cooperation enterprises provided in Step 1, use the Attention-LSTM model to mine its potential features;

[0159] This step specifically includes the following steps:

[0160] The potential feature vector obtained by mining the auxiliary information of external cooperation enterprises using Attention-LSTM consists of three parts: (1) the weight W1 of Attention-LSTM; (2) the auxiliary information Y of external cooperation enterprise j j ; (3) Gaussian noise ε j Variable;

[0161] As follows:

[0162] v j = Attention-LSTM(W1, Y j ) + ε j ;

[0163]

[0164] Where: v j is the output vector of each sub-item of the business document, W1 is all the weight parameters in the model, Y j is the sub-item j of the business document of the external cooperation enterprise, ε j is Gaussian noise that follows a normal distribution, I is the identity matrix, and σ v 2 is the variance of the output vector V j ;

[0165] Assume that the weight W1 also follows a Gaussian prior distribution, and the conditional distributions of W1 and the potential vector V of the external cooperation enterprise are as follows:

[0166]

[0167] Where: W1 is all the weight parameters in the Attention-LSTM model, and σ w 2 and σ v 2 are the variances of the weight vector W1 and the output vector V respectively, wk is a weight sub-item in W1, ε j is Gaussian noise, I is the identity matrix, v j is the output vector of each document sub-item.

[0168] The model structure for mining potential feature vectors of auxiliary information of external cooperation enterprises using Attention-LSTM includes: (1) Input layer; (2) Embedding layer; (3) LSTM layer; (4) Attention mechanism; (5) Feature fusion layer; (6) Linear layer; (7) Dropout layer; (8) Output layer;

[0169] Input layer: Receives the business descriptive document of the external cooperation enterprise. Let the input document be:

[0170] Y = {y1, y2,.., y l}, Y ∈ R l×p ;

[0171] where l is the length of the document, and p is a vector of a fixed dimension to which each word is transformed by Word2Vec;

[0172] Embedding layer: Maps the input words to a low-dimensional embedding vector space. The input Y is converted into an embedding matrix, which is used as the input for the lower LSTM layer;

[0173] E = Embedding(Y);

[0174] LSTM layer: Extracts the context features of the document through LSTM. The input E passes through the LSTM layer to obtain a hidden state sequence:

[0175] H = LSTM(E);

[0176] Attention mechanism: Fuses the context vector v output by the Attention mechanism with the auxiliary information vector y of the external cooperation enterprise, expressed as:

[0177]

[0178] where v is the output context feature of the Attention mechanism, and ε is Gaussian noise;

[0179] Feature fusion layer: Fuses the context vector v output by the Attention mechanism with the auxiliary information vector y of the external cooperation enterprise, expressed as:

[0180] Z = Concat(v, y);

[0181] Linear layer: The fused feature Z undergoes a non-linear transformation through the linear layer:

[0182] z linear =tanh(W z Z+b);

[0183] Among them, W z The weight matrix of the linear layer, b is the bias vector;

[0184] Dropout layer: To prevent overfitting, a Dropout operation is introduced after the output of the linear layer:

[0185] z dropout = Dropout(z linear ,p);

[0186] Among them, p is the probability hyperparameter of Dropout;

[0187] Output layer: The model generates the potential feature vector of the outsourced enterprise through the output layer and transforms it using the tanh activation function:

[0188] S = tanh(W out z dropout +b out );

[0189] Among them, W out is the weight matrix of the output layer, b out is the bias vector;

[0190] Through the above processing, the latent vector output by the LSTM structure can be expressed as:

[0191] S j =Attention-LSTM(W1,Y j );

[0192] Among them, W1 represents all weights, Y j is the child j of the document;

[0193] Step 3: Based on the auxiliary attribute data of the demand enterprise provided in step 1, the VAE model is used to capture its potential characteristics;

[0194] The potential feature vector obtained by using VAE to capture the auxiliary information of demand enterprises consists of three parts: (1) the weight W of VAE; (2) the auxiliary information X of demand enterprise i i ; (3) Gaussian noise ε i Variable; where:

[0195] u i =VAE(W,X i ,S i )+ε i ;

[0196]

[0197] where: u i is the output vector of each document sub-item, W is all the weight parameters in the VAE model, X i is the sub-item i of the document preferred by the demanding enterprise, ε i is Gaussian noise that follows a normal distribution, I is the identity matrix, σ u 2 is the variance of the output vector V i of.

[0198] Assume that the weight W also follows a Gaussian prior distribution, and the conditional distributions of W and the latent vector U of the demanding enterprise are as follows:

[0199]

[0200] where: σ W 2 and σ U 2 are the variances of the weight vector W and the output vector U in the VAE model respectively, w k is the weight sub-item in W, u i is the output vector of each document sub-item, S i is the scoring matrix.

[0201] The model structure for mining the latent feature vectors of the auxiliary information of the demanding enterprise using VAE includes: (1) input layer; (2) Embedding layer; (3) feature encoder; (4) latent variable sampling layer; (5) feature decoder; (6) linear layer; (7) Dropout layer; (8) output layer;

[0202] Input layer: Receive the preference attributes of the demanding enterprise and the scoring matrix S i , and assume the input preference attributes are:

[0203] X = {x1, x2,..., x T}, X ∈ R T×p ;

[0204] where, T is the length of the document, and p is a vector with a fixed dimension to which each word is transformed by Word2Vec;

[0205] Embedding layer: Map the input words to a low-dimensional embedding vector space. The input X is converted into an embedding matrix, which is used as the input to the encoder;

[0206] E = Embedding(X);

[0207] Feature encoder: Compress the input feature E into the distribution parameters of the latent space through the encoder network:

[0208] μ, log(σ 2 ) = Encoder(E);

[0209] Among them, μ represents the mean of the latent distribution, and log(σ 2 ) represents the logarithm of the variance of the latent distribution;

[0210] Latent variable sampling layer: Sampling is performed from the latent space using the reparameterization trick. The sampling process is as follows:

[0211] z = μ + σ ⊙ ∈, ∈ ~ N(0, I);

[0212] Among them, σ = exp(0.5 · log(σ 2 )) represents the standard deviation, ⊙ represents element-wise multiplication, and ∈ is a noise variable sampled from the standard normal distribution;

[0213] Feature decoder: Decodes the sampled latent variable z and the demand enterprise score S i together into a feature representation:

[0214] h = Decoder(z, S i );

[0215] Linear layer: The decoded feature h undergoes a linear transformation to generate the final latent vector:

[0216] z linear1 = tanh(W z1 h + b1);

[0217] Among them, W z1 is the weight matrix of the linear layer, and b1 is the bias vector;

[0218] Dropout layer: To prevent overfitting, a Dropout operation is introduced after the output of the linear layer:

[0219] z dropout1 = Dropout(z linear , p);

[0220] Among them, p is the probability hyperparameter of Dropout;

[0221] Output layer: The model generates the latent feature vector of the demand enterprise through the output layer, and uses the tanh activation function for transformation:

[0222] S = tanh(W out1 z dropout1 + b out1 );

[0223] Among them, W out1 is the weight matrix of the output layer, and b out1is the bias vector;

[0224] Through the above processing, the latent vector output by the VAE structure can be expressed as:

[0225] S i = VAE(W2, X i , S);

[0226] where W2 represents all weights, and X i is the sub-item i of the attribute;

[0227] Step 4: According to the latent feature vectors obtained in Step 2 and Step 3, use the PMF technology to perform scoring matrix fusion and filling;

[0228] When using the PMF technology to perform scoring matrix fusion and filling, the conditional distribution of the scoring matrix is:

[0229]

[0230] where: U is the latent vector of the demanding enterprise, V is the latent vector of the subcontracting enterprise, and R ij is the score of the demanding enterprise i for the subcontracting enterprise j, u i and v j are the latent vectors of the demanding enterprise i and the subcontracting enterprise j respectively, and σ 2 is the variance of the Gaussian distribution.

[0231] When using the PMF technology to perform scoring matrix fusion and filling, the model performs parameter learning to predict the scoring matrix, which can be expressed as:

[0232] Using the maximum a posteriori estimation, the problem is transformed into minimizing the negative log posterior:

[0233]

[0234] For each known score R, its predicted value R0 can be given by the inner product of the latent feature vectors of the demanding enterprise and the subcontracting enterprise and the linear combination of the auxiliary information:

[0235]

[0236] The loss function includes the observation error and the regularization term, and the form is as follows:

[0237]

[0238] where λ U , λ V , λ W , λ W1 are the regularization coefficients. The above loss function can be optimized by the alternating minimization algorithm, and U, V, W, and W1 are updated sequentially until convergence;

[0239] Step 5: According to the scoring matrix formed in Step 4, use the BERT model for semantic analysis and sentiment calculation to correct the scoring matrix.

[0240] The semantic analysis and sentiment calculation using the BERT model are calculated by the following expression:

[0241] R' = R + λ1W R ;

[0242]

[0243] where λ1 is a hyperparameter that controls the impact of comment sentiment on score correction.

[0244] The steps for semantic analysis and sentiment calculation using the BERT model are as follows:

[0245] (1) Use regular expressions to extract comment entries from the platform logs and filter out valid user evaluation data for subsequent analysis.

[0246] (2) Conduct sentiment classification on each comment entry and label it as a negative review (0), a positive review (1), or a neutral review (2) to construct a sentiment label dataset.

[0247] (3) Use the Tokenizer of BERT to convert non-numeric characters in the text into integer IDs, add [CLS] and [SEP] symbols at both ends of the text, and perform zero-padding at the same time.

[0248] (4) Input the preprocessed labeled ID sequence into the BERT model, perform context modeling on the text through a multi-layer Transformer architecture, and output the representation vector of the Pooler Output layer.

[0249] (5) Use a fully connected classifier to classify the vectors in the Pooler Output layer and predict the sentiment label (negative review, positive review, or neutral review) of each comment.

[0250] (6) Use the cross-entropy loss function to evaluate the error between the prediction result and the true label, thereby measuring the training effect of the model.

[0251] (7) Apply the Adam optimizer to train the model, minimize the cross-entropy loss, and optimize the model parameters of BERT by adaptively adjusting the learning rate.

[0252] (8) Correct the scoring matrix according to the sentiment classification results, combine the sentiment tendency and scoring history of the demand enterprise, and fill in the missing items in the scoring matrix.

[0253] Step 6: According to the scoring matrix corrected in Step 5, apply the pyramid behavior characteristics to perform a secondary correction on the scoring matrix;

[0254] The steps of applying the pyramid behavior characteristics to perform a secondary correction on the scoring matrix are as follows:

[0255] (1) Divide the behavior of the demand enterprise into three layers of pyramid behavior characteristics in the time dimension: short-term, medium-term, and long-term. Perform weighted modeling on various behavior characteristics. The specific calculation formula is as follows:

[0256] Short-term behavior vector: B short ={b 1 ,b 2 ,…,b TS};

[0257] Medium-term behavior vector: B med ={b TS+1 ,b TS+2 ,…,b Tm};

[0258] Long-term behavior vector: B long ={b Tm+1 ,b Tm+2 ,…,b Ti};

[0259] Among them, TS, Tm, and Ti respectively represent the length breakpoints of the time windows of short-term, medium-term, and long-term behaviors;

[0260] The behavior characteristics within all time windows can be represented as a three-dimensional vector b t ={b fr ,b ti ,b co}(t = 1, 2…Ti);

[0261] Among them, b fr is the number of views, b ti is the browsing duration, b co is whether it is collected;

[0262] Assign different weights to the three sub-items of each b t . The calculation formula is as follows:

[0263]

[0264] Among them, W fr ,W ti ,W cow1, w2, w3 are the weights of three behavioral characteristics, t is the interaction behavior record duration of the demand enterprise, λ is the adjustment parameter that controls the change trend of each time window, and α, β, γ are the parameters that adjust the change trend between windows; the recommended parameter values are: λ1 = 0.4, λ2 = 0.15, λ3 = 0.07, α = 1, β = 0.5, β = 0.2

[0265] Through the above weighted function, the behavioral data of each demand enterprise is weighted and adjusted; specifically, the behavioral data of each demand enterprise in the past t days will be multiplied by the corresponding time-weighted value, and the weighted behavioral data will be aggregated to obtain the complete behavioral sequence of the demand enterprise:

[0266] B i (t) = b fr (t) · W fr (t) + b ti (t) · W ti (t) + b co (t) · W co (t);

[0267] Where: b fr is the number of views of the user, b ti is the browsing duration of the user, b co is the favorite record of the user, W fr , W ti , W co are the weights of three behavioral characteristics in the short term, medium term and long term.

[0268] (2) Input the weighted complete behavioral sequence into the LSTM network to capture the temporal characteristics of the demand enterprise's behavior. Dynamically adjust the contribution of each time step to the final representation through the Attention mechanism. The specific calculation formula is as follows:

[0269] The weighted behavioral sequence B i (t) is processed through the LSTM model to obtain the hidden state h t at each time step. The calculation formula of LSTM is as follows:

[0270] h t = LSTM(B i (t));

[0271] Where: h t is the output of the LSTM network at the t time step, and B i (t) is the complete behavioral sequence of each demand enterprise.

[0272] Then the Attention layer takes the hidden state h t output by the LSTM as the input, so as to give different weights to each time step and obtain the attention score. The calculation formula is as follows:

[0273]

[0274] Among them, h t is the hidden state of the LSTM at time step t, W a is a weight matrix used to map the hidden state to a higher-dimensional space, b a is the bias term used to adjust the mapping result, w a is an output weight vector used to calculate the final score;

[0275] By performing a softmax operation on the score score(h t ) for each time step, the score is converted into an attention weight; the calculation of the attention weight is as follows:

[0276]

[0277] Among them, the denominator is the sum of the exponents of the scores of all time steps, used for normalization, α t is the Attention weight for time step t;

[0278] Using the calculated attention weight α t to weight the hidden state h t output by the LSTM to obtain the weighted behavior representation B weighted :

[0279]

[0280] (3) Based on the weighted behavior feature B weighted , perform a quadratic correction on the scoring matrix:

[0281] R” = R' + λB weighted ;

[0282] Among them, R’ is the scoring matrix after the first correction, and λ is a hyperparameter that controls the influence of the behavior feature on the scoring correction;

[0283] Step seven, according to the scoring matrix quadratic-corrected in step six, recommend the top N items to the target demand user.

[0284] This embodiment also discloses a wall panel external cooperation enterprise recommendation system based on multi-modal data fusion, including:

[0285] A wall panel external cooperation data collection module, responsible for collecting the attribute data and behavior data of enterprises from the platform;

[0286] A sparse scoring matrix completion module, which supplements and optimizes the missing data in the matrix based on the attribute information;

[0287] The comment sentiment analysis correction module adjusts the scoring matrix by analyzing the sentiment tendency of user comments;

[0288] The pyramid behavior feature optimization module is used to secondarily correct the matrix by combining the behavior data with the time series change.

[0289] The outsourcing enterprise recommendation ranking module generates a priority recommendation list of the top N enterprises in combination with the scoring results.

[0290] A schematic diagram of a terminal device provided by an embodiment of the present invention. The terminal device of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps in the above-mentioned method embodiments are implemented. Alternatively, when the processor executes the computer program, the functions of each module / unit in the above-mentioned device embodiments are implemented.

[0291] The computer program can be divided into one or more modules / units, and the one or more modules / units are stored in the memory and executed by the processor to complete the present invention.

[0292] The terminal device can be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The terminal device may include, but is not limited to, a processor and a memory.

[0293] The processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0294] The memory can be used to store the computer program and / or module, and the processor realizes various functions of the terminal device by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory.

[0295] If the modules / units integrated in the terminal device are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, to implement all or part of the processes in the above-described embodiment methods of the present invention, it can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0296] The above content is only for explaining the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution according to the technical idea proposed by the present invention fall within the protection scope of the claims of the present invention.

Claims

1. A wall panel outsourcing enterprise recommendation method based on multimodal data fusion, characterized in that: The following steps are involved: Obtain relevant data of outsourcing enterprises and demand enterprises; Mining potential features based on relevant data of outsourced enterprises to obtain the mined potential feature vectors; Capture potential features based on relevant data of the demand enterprise to obtain captured potential feature vectors; The mined latent feature vectors and the mined latent feature vectors are fused and filled in the scoring matrix to obtain a scoring matrix; After the rating matrix is ​​modified, a modified rating matrix is ​​obtained, and the top N items of the modified rating matrix are recommended to the target users.

2. According to the method for recommending wall panel outsourcing companies based on multimodal data fusion according to claim 1, it is characterized in that: The relevant data include attribute data, behavior data and scoring matrix of the outsourcing enterprise and the demand enterprise; The attribute data of the outsourcing enterprise includes the enterprise's service type and service description; the attribute data of the demand enterprise includes the enterprise's demand type, price preference, purchase cycle preference and service equipment preference; The behavior data includes the number of views, browsing time and collection records of the demand enterprise.

3. According to the method for recommending wall panel outsourcing companies based on multimodal data fusion according to claim 1, it is characterized in that: The mining of potential features is performed using an Attention-LSTM model; The potential feature vector mined is composed of the weight W1 of the Attention-LSTM model and the auxiliary information Y j , and Gaussian noise ε j Variable composition; Among them, v j =Attention-LSTM(W1,Y j )+ε j ; Where: v j is the output vector of each business document sub-item, W1 is all the weight parameters in the model, and Y j is the sub-item j of the outsourced enterprise business document, ε j is Gaussian noise that follows a normal distribution, I is the identity matrix, σ v 2 is the output vector V j The variance of When W1 follows a Gaussian prior distribution, the conditional distributions of W1 and the outsourced enterprise potential vector V are as follows: Where: W1 is all the weight parameters in the Attention-LSTM model, σ w 2 and σ v 2 are the variances of the weight vector W1 and the output vector V, respectively, w k is the weight sub-item in W1, ε j is Gaussian noise, I is the identity matrix, v j is the output vector for each document child.

4. According to the method for recommending wall panel outsourcing companies based on multimodal data fusion according to claim 1, it is characterized in that: The capturing of potential features is performed by adopting a VAE model; The captured potential feature vector is composed of the weight W of the VAE model, the auxiliary information X of the demand enterprise i i and Gaussian noise ε i Variable composition; Among them: u i = VAE(W,X i ,S i )+ε i ; Where: u i is the output vector of each document sub-item, W is all the weight parameters in the VAE model, X i is the item i of the preference document of the demanding enterprise, ε i is Gaussian noise that follows a normal distribution, I is the identity matrix, σ u 2 is the output vector u i The variance of When the weight W follows a Gaussian prior distribution, the conditional distributions of W and the demand enterprise potential vector U are as follows: Where: W 2 and σ U 2 are the variances of the weight vector W and output vector U in the VAE model, w k is the weight sub-item in W, u i is the output vector for each document sub-item, S i is the rating matrix.

5. The wall panel outsourcing enterprise recommendation method based on multimodal data fusion according to claim 1 is characterized in that: The scoring matrix fusion and filling are performed using PMF technology; The conditional distribution of the obtained rating matrix R is: Among them: U is the potential vector of the demand enterprise, V is the potential vector of the outsourcing enterprise, R ij is the score of demand enterprise i on outsourcing enterprise j, u i and v j are the potential vectors of demand enterprise i and outsourcing enterprise j, σ 2 is the variance of the Gaussian distribution.

6. The wall panel outsourcing enterprise recommendation method based on multimodal data fusion according to claim 1 is characterized in that: The correction of the scoring matrix includes primary correction and secondary correction; The first correction is performed by using the BERT model for semantic analysis and sentiment calculation, and the expression used is as follows: R'=R0+λ1W R ; Among them, λ1 is a hyperparameter that controls the impact of review sentiment on score correction.

7. The wall panel outsourcing enterprise recommendation method based on multimodal data fusion according to claim 6 is characterized in that: The secondary correction is to apply the pyramid behavior characteristics to perform secondary correction on the scoring matrix, and the specific steps are as follows: According to the time dimension, the demand enterprise behavior is divided into short-term, medium-term and long-term three-layer pyramid behavior characteristics, and the short-term, medium-term and long-term three-layer pyramid behavior characteristics are weighted modeled. The specific calculation formula is as follows: The behavioral features are layered by time window and expressed in vector form: Short-term behavior vector: B short = {b 1 ,b 2 ,…,b TS }; Medium-term behavior vector: B med = {b TS+1 ,b TS+2 ,…,b Tm }; Long-term behavior vector: B long = {b Tm+1 ,b Tm+2 ,…,b Ti }; Among them, TS, Tm and Ti represent the length breakpoints of the time windows of short-term, medium-term and long-term behaviors, respectively; The behavioral characteristics within all time windows can be represented as a three-dimensional vector b t = {b fr ,b ti ,b co }; where t = 1, 2, ... Ti; Among them, b fr is the number of times the user has browsed; b ti b is the browsing time of the user; co Collect records for users; For each b t Different weights are assigned to the three sub-items of , and the calculation formula is as follows: Among them, W fr ,W ti ,W co are the weights of the three behavioral characteristics of short-term, medium-term and long-term; t is the duration of the interactive behavior record of the demand enterprise; λ is the adjustment parameter for controlling the change trend of each time window; α, β, γ are the parameters for adjusting the change trend between windows, α>β>γ; Through the weighting function below, the behavior data of each demand enterprise is weighted and adjusted to obtain the user's complete behavior sequence: B i (t)=b fr (t)·W fr (t)+b ti (t)·W ti (t)+b co (t)·W co (t); Where: b fr is the number of views by the user, b ti is the browsing time of the user, b co is the user's favorite record, W fr ,W ti ,W co It is the weight of three behavioral characteristics: short-term, medium-term and long-term. The obtained complete behavior sequence is input into the LSTM network to capture the temporal characteristics of user behavior. Then, the contribution of each time step to the final representation is dynamically adjusted through the Attention mechanism. The specific calculation formula used is as follows: h t =LSTM(B i (t)); Where: h t is the output of the LSTM network at time step t, B i (t) A complete sequence of actions for each demanding enterprise; Then the Attention layer uses the hidden state h output by LSTM t As input, different weights are given to each time step to obtain the attention score. The calculation formula is as follows: Among them, h t is the hidden state of the LSTM network at time step t; W a is a weight matrix; b a is the bias term; w a is an output weight vector; By calculating the score(h) of each time step t ) performs a softmax operation to convert the score into an attention weight. The calculation formula is as follows: Among them, α t represents the attention weight at time step t, score(h t ) is the attention score at time step t; Use the calculated attention weight α t The hidden state h of the LSTM output t Weighted, the weighted behavior representation B weighted , the calculation formula is as follows: Based on the weighted behavioral features B weighted , perform secondary correction on the scoring matrix, the formula is as follows: R”=R'+λB weighted ; Among them, R' is the rating matrix after the first correction, and λ is a hyperparameter that controls the impact of behavioral characteristics on rating correction.

8. A wall panel outsourcing enterprise recommendation system based on multimodal data fusion, characterized in that: It includes a data acquisition module, which is used to obtain relevant data of outsourcing enterprises and demand enterprises; The sparse scoring matrix completion module is used to mine potential features based on the relevant data of outsourced enterprises and obtain the mined potential feature vectors; Capture potential features based on relevant data of the demand enterprise to obtain captured potential feature vectors; fuse and fill the mined potential feature vectors and the captured potential feature vectors into a scoring matrix to obtain a scoring matrix; The comment sentiment analysis and correction module is used to correct the rating matrix based on the sentiment analysis results to obtain a corrected rating matrix; The pyramid behavior feature optimization module is used to perform secondary correction on the rating matrix based on the behavior and timing information to obtain a corrected rating matrix; Outsourcing enterprise recommendation and ranking module; Used to recommend the top N items of the modified rating matrix to target users.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images