A user personalized demand prediction method fusing multi-modal review information

Abstract

This invention discloses a method for predicting personalized user needs by integrating multimodal review information. The steps include: 1. Constructing a user-product purchase relationship network based on the user's purchase history; 2. Constructing the attributes of edges in the network graph using the multimodal reviews published by the user; 3. Mining the user's topic distribution across text and image modalities using a multimodal variational autoencoder; 4. Modeling the interaction information between the user and products and the semantic information of the multimodal review content using a graph attention network; 5. Selecting an appropriate loss function to train and optimize the model. This invention combines variational autoencoders and graph attention networks, which can fully and comprehensively mine user preference information for products from both text and image modal data, and accurately represent users and products, thereby achieving a more accurate prediction of personalized user needs.

Description

Technical Field

[0001] This invention relates to the fields of user preference mining and user personalized demand prediction, and in particular to a method for predicting user personalized demand by integrating multimodal comment information. Background Technology

[0002] With the development of the internet and e-commerce platforms, online shopping has gradually become the mainstream shopping method, and users can publish multimodal reviews (e.g., text, images, etc.). These multimodal reviews express cross-modal semantic information, possess information independent of each modality, and contain rich user preference information. However, online shopping leads to a lack of face-to-face communication between merchants and users, making it difficult to accurately capture user preferences and needs. Therefore, it is necessary to organize massive amounts of user purchase history data and multimodal review data to comprehensively mine user preferences and accurately predict personalized user needs.

[0003] In the existing fields of user preference identification and demand prediction, traditional methods neglect the joint modeling of user comment text and comment image data, thus failing to obtain multimodal, fine-grained representations of user preferences. Furthermore, existing methods cannot jointly model users' purchase history and multimodal comment content, resulting in an inability to accurately identify users' fine-grained preferences and a lack of explanation for users' purchase motivations. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a method for predicting personalized user needs by integrating multimodal review information. It aims to jointly model user review text and image data using a variational autoencoder model, thereby learning the distribution of user multimodal preferences. Furthermore, it combines graph attention networks to jointly model user product interaction records and the reasons for user product interaction (i.e., the distribution of user multimodal preferences). This approach enhances the representational power of multimodal data on user preferences, refines the granularity of user preference identification, and improves the accuracy of personalized needs prediction.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] The present invention provides a method for predicting personalized user needs by integrating multimodal comment information, characterized by the following steps:

[0007] Step 1: Construct the dataset;

[0008] Step 1.1: Represent the purchase relationship network between users and goods using a directed graph G = (V, M), where users and goods are nodes in the directed graph G, and V represents the set of all nodes. If user node i purchases goods node j, it means that there is an edge between user node i and goods node j, denoted as e. ij ∈M, where M represents the set of all edges, and edge e ij Corresponding to an attention coefficient α ij and topic distribution λ ij And user node i and product node j are each other's neighbors;

[0009] Step 1.2: Obtain the comment text of user node i to product node j, and after word segmentation and stop word removal, obtain the word list of the comment text, denoted as _____. in, Y represents the list of words ij The n'th word in; n t This indicates the total number of words in the word list;

[0010] Calculate the word frequency representation vector of the comment text of user node i to product node j across all words in the dictionary. in, This indicates the number of times the w-th word in the comment text appears in the dictionary, where W represents the number of all unique words in the comment texts of all user nodes to product nodes.

[0011] The length n of the comment text from user node i to product node j t After padding with zero vectors to reach length N, the resulting vector is input into the BERT model for processing, yielding the initial text representation vector, denoted as... in, Represents the initial representation vector T of the text ij The representation of the nth word in the text;

[0012] Step 1.3: After obtaining the comment image data of user node i to product node j and performing pixel unification processing, the initial representation vector of the comment image is obtained, denoted as... in, Image P represents the comment. ij The eigenvalues ​​in the s-th dimension, where S represents the dimension of the initial eigenvector;

[0013] Step 2: Construct a multimodal variational autoencoder network, including: a modal encoding module, a modal fusion module, and a modal decoding module;

[0014] Step 2.1: Construct a modality coding module, including a text coding network and an image coding network; wherein, the text coding network includes a Bi-LSTM network and a first fully connected layer; the image coding network includes a pre-trained VGG-19 network and a second fully connected layer;

[0015] Step 2.1.1: Using n as the current time step, process the comment text T. ij Representation of the nth word in The input is given to the Bi-LSTM network, and the forget gate state is obtained by using equations (1)-(5) respectively. Input gate state Output gate status Current cell state at the nth time step and the unit output at the current nth time step

[0016]

[0017]

[0018]

[0019]

[0020]

[0021] In equations (1)-(5), This represents element-wise multiplication, where σ(·) is the sigmoid activation function, tanh is the hyperbolic tangent function, and W... f and U f These represent the two weighting coefficients of the forget gate, b f W represents the bias vector of the forget gate. I and U I These represent the two weighting coefficients of the input gate, b I W represents the bias vector of the input gate. o and U o These represent the two weighting coefficients of the output gate, b o W represents the bias vector of the output gate. c and U c These represent the two weighting coefficients of the cell unit, b c This represents the bias vector of a cell unit. This represents the cell state at time step n-1. This represents the unit output at time step (n-1). When n=1, let...

[0022] Step 2.1.2: The first fully connected layer outputs text latent features using equation (6).

[0023]

[0024] In equation (6), W1 represents the weight matrix of the first fully connected layer; This represents the output of the cell at the Nth time step.

[0025] Step 2.1.3: Use the pre-trained VGG-19 network to process the comment image P. ij Characterization is performed to obtain image feature vectors. Therefore, equation (7) is used to analyze the image feature vector. Encode to obtain latent features of the image

[0026]

[0027] In equation (7), W2 represents the weight matrix of the second fully connected layer;

[0028] Step 2.2: The modal fusion module uses equation (8) to calculate the multimodal shared representation λ after text and image fusion. ij ;

[0029]

[0030] In equation (8), λ ij ~LN(μ ij ,σ ij LN represents the logistic normal distribution. It is the mean of a logistic normal distribution, where, This represents the mean of the multimodal comments from user node i on product node j on the k-th topic. It is the variance of the logistic normal distribution, where, ε represents the variance of the multimodal comment content of user node i on the k-th topic for product node j. ij It is a random variable that follows a normal distribution with a mean of 0 and a variance of 1;

[0031] Step 2.3: Construct the modal decoding module, including: a text decoder network and an image decoder network;

[0032] The text decoder network supports λ ij Decoding and reconstruction are performed to obtain the text reconstruction feature vector. in, Represents the text reconstruction feature vector T ij The reconstructed representation of the nth word in ';

[0033] The image decoder network supports λ ij Decoding and reconstruction are performed to obtain the image reconstruction feature vector. in, Image P represents the comment. ij The eigenvalues ​​after reconstruction in the s-th dimension;

[0034] Step 3: Processing the graph attention network:

[0035] Step 3.1: Define the number of updates as d, and initialize d = 1;

[0036] Step 3.1.1: Calculate the features of user node i updated in the (d-1)th time using equation (9).

[0037]

[0038] In equation (9), N i Let |N| represent the set of all neighboring product nodes of user node i. i | Represents the number of all neighboring product nodes of user node i;

[0039] Step 3.1.2: Calculate the features of the product node j updated in the (d-1)th time using equation (10).

[0040]

[0041] In equation (10), N j Let |N| represent the set of all neighboring user nodes of product node j. j | represents the number of all neighboring user nodes of product node j;

[0042] Step 3.2: Calculate the k-th head attention coefficient between user node i and its neighboring product node j in the (d-1)-th update using equation (11).

[0043]

[0044] In equation (7), LeakeyReLU(·) represents the activation function, and a and W3 are two parameter matrices to be learned. Let || denote the feature of a certain neighboring product node j′ in the set of neighboring product nodes of user node i after the (d-1)th update, excluding neighboring product node j, || denotes the concatenation operation of vectors, and T denotes the matrix transpose operation.

[0045] Step 3.3: Calculate the feature vector of user node i updated in the dth time using equations (12) and (13) respectively. Features of product node j

[0046]

[0047]

[0048] In equations (12)-(13), The edge e between node i and node j ij The probability on the k-th topic; and K represents the number of topics;

[0049] Step 4: Construct a user-personalized demand prediction model consisting of a multimodal variational autoencoder network and a graph attention network, and train and optimize it.

[0050] Step 4.1: Construct the loss function of the graph attention network for the (d-1)th update using equation (14).

[0051]

[0052] In equation (14), Let Q represent the set of non-neighboring nodes of user node i. i Represents the set of non-neighbor nodes of user node i. The number of nodes in the middle; Represents the set of non-neighbor nodes of user node i. The expectation of all nodes in the middle, The set of non-neighbor nodes of user node i in the dth update. Any node j n Features;

[0053] Construct the reconstruction loss for comment image data using equation (15).

[0054]

[0055] In equation (15), E eij～M Let M represent the expected value of all edges in the edge set M;

[0056] Construct the reconstruction loss of text data using equation (16)

[0057]

[0058] In equation (16), Represents the initial representation vector T ij Representation of the nth word in Is it a representation of the w-th word in the dictionary? If so, then let... Set it to 1; otherwise, let it be 1. =0;

[0059] Using equation (17), construct the KL loss L between the true distribution and the variational distribution. kl :

[0060]

[0061] Using equation (18), construct the overall loss function of the user personalized demand prediction model in the dth update.

[0062]

[0063] Step 5.2: Train the user personalized demand prediction model using the Adam algorithm and calculate the overall loss function. When the number of updates d reaches its maximum value d max Time or overall loss function L total Upon convergence, training is stopped, thus obtaining the optimal demand prediction model, which can be used to predict the personalized product demand of any user.

[0064] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the user personalized demand prediction method, and the processor is configured to execute the program stored in the memory.

[0065] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the user personalized demand prediction method.

[0066] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0067] 1. This invention preprocesses user purchase behavior data, filters out nodes with sparse interaction relationships, sums and averages the initial features of the text and images on all edges that each node has interacted with, and uses this sum as the initial feature of each node, thereby completing the preprocessing process of massive behavioral data, which is beneficial to improving the efficiency of subsequent prediction of personalized user needs.

[0068] 2. This invention constructs a multimodal variational autoencoder model to jointly model text and image data in user reviews, and learns a shared representation that can express multimodal semantic information. This is beneficial for refining the granularity of user preference information and expressing users' personalized preference information more comprehensively and completely.

[0069] 3. This invention constructs a graph attention network, building the graph based on the interaction information between users and products, and constructing the attributes of the edges in the graph based on the user's multimodal preference information. During the graph optimization learning process, it ultimately obtains user-product node representations that integrate both the structural information of the graph and the multimodal semantic information of the edges, thereby greatly improving the representation ability of the nodes and the accuracy of user demand prediction.

[0070] 4. This invention can be used in the task of predicting users' personalized needs on e-commerce platforms. It can predict what products users will buy and explain the reasons for their purchases. It has a wide range of applications and practical value. Attached Figure Description

[0071] Figure 1 This is a flowchart of the user personalized needs prediction method of the present invention. Detailed Implementation

[0072] In this embodiment, a method for predicting personalized user needs by integrating multimodal comment information is described, specifically, as follows: Figure 1 As shown, the procedure is as follows:

[0073] Step 1: Construct the dataset;

[0074] Step 1.1: Represent the purchase relationship network between users and goods using a directed graph G = (V, M), where users and goods are nodes in the directed graph G, and V represents the set of all nodes. If user node i purchases goods node j, it means that there is an edge between user node i and goods node j, denoted as e. ij ∈M, where M represents the set of all edges, and edge e ij Corresponding to an attention coefficient α ij and topic distribution λ ij User node i and product node j are considered as neighbors; the final attention coefficient is the sum or average of the multi-head attention coefficients.

[0075] Step 1.2: Obtain the comment text of user node i to product node j, and after word segmentation and stop word removal, obtain the word list of the comment text, denoted as _____. in, Y represents the list of words ij The n'th word in; n t This indicates the total number of words in the word list;

[0076] Calculate the word frequency representation vector of the comment text of user node i to product node j across all words in the dictionary. in, This indicates the comment text as the first entry in the dictionary. wThe number of times each word appears, W represents the number of all unique words in the comment text of all user nodes to product nodes;

[0077] The length n of the comment text from user node i to product node j t After padding with zeros to reach length N, for comment texts with fewer than N words, zeros are added; for texts with more than N words, the excess words are deleted. This is then processed in the BERT model to obtain the initial text representation vector, denoted as... in, Represents the initial representation vector T of the text ij The representation of the nth word in the text;

[0078] Step 1.3: After obtaining the comment image data of user node i to product node j and performing pixel unification processing, the initial representation vector of the comment image is obtained, denoted as... in, Image P represents the comment. ij The eigenvalues ​​in the s-th dimension, where S represents the dimension of the initial eigenvector;

[0079] Step 2: Construct a multimodal variational autoencoder network, including: a modal encoding module, a modal fusion module, and a modal decoding module;

[0080] Step 2.1: Construct a modality coding module, including a text coding network and an image coding network; wherein, the text coding network includes a Bi-LSTM network and a first fully connected layer; the image coding network includes a pre-trained VGG-19 network and a second fully connected layer;

[0081] Step 2.1.1: Using n as the current time step, process the comment text T. ij Representation of the nth word in The input is fed into the Bi-LSTM network, and the forget gate state is obtained by using equations (1)-(5) respectively. Input gate state Output gate status Current cell state at the nth time step and the unit output at the current nth time step

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] In equations (1)-(5), This represents element-wise multiplication, where σ(·) is the sigmoid activation function, tanh is the hyperbolic tangent function, and W... f and U f These represent the two weighting coefficients of the forget gate, b f W represents the bias vector of the forget gate. I and U I These represent the two weighting coefficients of the input gate, b I W represents the bias vector of the input gate. o and U o These represent the two weighting coefficients of the output gate, b o W represents the bias vector of the output gate. c and U c These represent the two weighting coefficients of the cell unit, b c This represents the bias vector of a cell unit. This represents the cell state at time step n-1. This represents the unit output at time step (n-1). When n=1, let...

[0088] Step 2.1.2: The first fully connected layer outputs the text latent features using equation (6).

[0089]

[0090] In equation (6), W1 represents the weight matrix of the first fully connected layer; This represents the output of the cell at the Nth time step.

[0091] Step 2.1.3: Use the pre-trained VGG-19 network to process the comment image P. ij In terms of characterization, the VGG-19 network architecture contains 19 hidden layers (including 16 convolutional layers and 3 fully connected layers). Each layer utilizes the output of the previous layer to further extract more complex features from the image. Each layer can be viewed as an extractor of multiple local features, significantly improving the network's feature extraction performance. This yields the image feature vector. Therefore, equation (7) is used to analyze the image feature vector. Encode to obtain latent features of the image And image latent features and textual latent features The vectors have the same dimension;

[0092]

[0093] In equation (7), W2 represents the weight matrix of the second fully connected layer;

[0094] Step 2.2: The modal fusion module uses equation (8) to calculate the multimodal shared representation λ after text and image fusion. ij ;

[0095]

[0096] In equation (8), λ ij ~LN(μ ij ,σ ij LN represents the logistic normal distribution. It is the mean of a logistic normal distribution, where, This represents the mean of the multimodal comments from user node i on product node j on the k-th topic. It is the variance of the logistic normal distribution, where, ε represents the variance of the multimodal comment content of user node i on the k-th topic for product node j. ij It is a random variable that follows a normal distribution with a mean of 0 and a variance of 1;

[0097] Step 2.3: Construct the modal decoding module, including: a text decoder network and an image decoder network;

[0098] Text decoder network for λ ij Decoding and reconstruction are performed to obtain the text reconstruction feature vector. in, Represents the text reconstruction feature vector T ij The reconstructed representation of the nth word in ';

[0099] Image decoder network for λ ij Decoding and reconstruction are performed to obtain the image reconstruction feature vector. in, Image P represents the comment. ij The eigenvalues ​​after reconstruction in the s-th dimension;

[0100] The text decoder first divides the multimodal shared representation λ ij The input to the Bi-LSTM model is created through a fully connected layer, which then reconstructs the probability of each word appearing in the original text content; the image decoder uses the multimodal shared representation λ. ij The VGG-19 features of the original image are reconstructed using multiple fully connected layers.

[0101] Step 3: Processing the graph attention network:

[0102] Step 3.1: Define the number of updates as d, and initialize d = 1;

[0103] Step 3.1.1: Calculate the features of user node i updated in the (d-1)th time using equation (9).

[0104]

[0105] In equation (9), N i Let |N| represent the set of all neighboring product nodes of user node i. i | Represents the number of all neighboring product nodes of user node i;

[0106] Step 3.1.2: Calculate the features of the product node j updated in the (d-1)th time using equation (10).

[0107]

[0108] In equation (10), N j Let |N| represent the set of all neighboring user nodes of product node j. j | represents the number of all neighboring user nodes of product node j;

[0109] Step 3.2: Calculate the k-th head attention coefficient between user node i and its neighboring product node j in the (d-1)-th update using equation (11).

[0110]

[0111] In equation (7), LeakeyReLU(·) represents the activation function, and a and W3 are two parameter matrices to be learned. Let || denote the feature of a certain neighboring product node j′ in the set of neighboring product nodes of user node i after the (d-1)th update, excluding neighboring product node j, || denotes the concatenation operation of vectors, and T denotes the matrix transpose operation.

[0112] Step 3.3: Calculate the feature vector h of user node i updated for the dth time using equations (12) and (13) respectively. i d Features of product node j

[0113]

[0114]

[0115] In equations (12)-(13), The edge e between node i and node j ij The probability on the k-th topic; and K represents the number of topics; on the one hand, the influence weight of each neighbor node is calculated to aggregate the features of the neighbor nodes based on the attention coefficient; on the other hand, the topic preference features of the neighbor nodes are aggregated on this basis.

[0116] Step 4: Construct a user-personalized demand prediction model consisting of a multimodal variational autoencoder network and a graph attention network, and train and optimize it.

[0117] Step 4.1: Construct the loss function of the graph attention network for the (d-1)th update using equation (14).

[0118]

[0119] In equation (14), Let Q represent the set of non-neighboring nodes of user node i. i N represents the set of non-neighbor nodes of user node i. i The number of nodes in the middle; Represents the set of non-neighbor nodes of user node i. The expectation of all nodes in the middle, The set of non-neighbor nodes of user node i in the dth update. Any node j n Features;

[0120] The reconstruction loss L for the comment image data is constructed using equation (15). recvgg :

[0121]

[0122] In equation (15), E eij～M Let M represent the expected value of all edges in the edge set M;

[0123] The reconstruction loss L of text data is constructed using equation (16). reclstm :

[0124]

[0125] In equation (16), Represents the initial representation vector T ij Representation of the nth word in Is it a representation of the w-th word in the dictionary? If so, then let... Set it to 1; otherwise, let it be 1. =0;

[0126] Using equation (17), construct the KL loss L between the true distribution and the variational distribution. kl :

[0127]

[0128] Using equation (18), construct the overall loss function of the user personalized demand prediction model in the dth update.

[0129]

[0130] Step 5.2: Train the user personalized demand prediction model using the Adam algorithm and calculate the overall loss function. When the number of updates d reaches its maximum value d max Time or overall loss function L total Upon convergence, training is stopped, thus obtaining the optimal demand prediction model, which can be used to predict the personalized product demand of any user.

[0131] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the aforementioned user personalized demand prediction method, and the processor is configured to execute the program stored in the memory.

[0132] In this embodiment, a computer-readable storage medium stores a computer program, which, when run by a processor, executes the steps of the above-described user personalized demand prediction method.

Claims

1. A method for predicting personalized user needs by integrating multimodal review information, characterized in that, The procedure is as follows: Step 1: Construct the dataset; Step 1.1: Using a directed graph Represents the purchase relationship network between users and products, where users and products are directed graphs. Each node in, Represents the set of all nodes, if the user node Purchase Product Node This indicates a user node. and product nodes There exists an edge between them, denoted as . , Let the set of all edges be denoted as , and let the edges be denoted as . Corresponding to an attention coefficient and topic distribution ; and user nodes and product nodes They act as neighbors to each other; Step 1.2: Obtain User Nodes For product nodes The comment text is processed by word segmentation and stop word removal to obtain a word list of the comment text, denoted as . ,in, Represents a list of words The first in One word; This indicates the total number of words in the word list; Compute user nodes For product nodes The word frequency representation vector of the comment text over all words in the dictionary ,in, This indicates the comment text as the first entry in the dictionary. The number of times each word appears This represents the number of all unique words in the comment text of all user nodes on the product node; User nodes For product nodes Length of comment text After padding with zero vectors to reach length N, the resulting vector is input into the BERT model for processing, yielding the initial text representation vector, denoted as... ,in, Represents the initial representation vector of the text. The first in The representation of a word; Step 1.3: Obtain User Nodes For product nodes After processing the comment image data and performing pixel unification, the initial representation vector of the comment image is obtained, denoted as... ,in, Image representing a comment In the Eigenvalues ​​in dimension, Indicates the dimension of the initial feature vector; Step 2: Construct a multimodal variational autoencoder network, including: a modal encoding module, a modal fusion module, and a modal decoding module; Step 2.1: Construct a modality coding module, including a text coding network and an image coding network; wherein, the text coding network includes a Bi-LSTM network and a first fully connected layer; the image coding network includes a pre-trained VGG-19 network and a second fully connected layer; Step 2.1.1, with At the current time step, the initial representation vector of the text is... The first in The representation of a word The input is given to the Bi-LSTM network, and the forget gate state is obtained by using equations (1)-(5) respectively. Input gate status Output gate status The current number Cell state at each time step and the current number Unit output at each time step : (1) (2) (3) (4) (5) In equations (1)-(5), This indicates element-wise multiplication. It is the sigmoid activation function. It is the hyperbolic tangent function. and These represent the two weighting coefficients of the forget gate. This represents the bias vector of the forget gate. and These represent the two weight coefficients of the input gate. This represents the bias vector of the input gate. and These represent the two weighting coefficients of the output gate. This represents the bias vector of the output gate. and These represent the two weighting coefficients of the cell unit. This represents the bias vector of a cell unit. Indicates the first The cell state at each time step. Indicates the first The unit output at each time step, when n=1, let , ; Step 2.1.2: The first fully connected layer outputs text latent features using equation (6). ; (6) In equation (6), This represents the weight matrix of the first fully connected layer; Indicates the first The unit output at each time step; Step 2.1.3: Use the pre-trained VGG-19 network to process the comment images. Characterization is performed to obtain image feature vectors. Thus, equation (7) is used to analyze the image feature vector. Encode to obtain latent features of the image ; (7) In equation (7), This represents the weight matrix of the second fully connected layer; Step 2.2: The modal fusion module uses equation (8) to calculate the multimodal shared representation after text and image fusion. ; (8) In equation (8), , Represents a logistic normal distribution. It is the mean of a logistic normal distribution, where, Represents user node For product nodes Multimodal commentary content in the first The mean across the topics It is the variance of the logistic normal distribution, where, Represents user node For product nodes Multimodal commentary content in the first Variance across topics It is a random variable that follows a normal distribution with a mean of 0 and a variance of 1; Step 2.3: Construct the modal decoding module, including: a text decoder network and an image decoder network; The text decoder network Decoding and reconstruction are performed to obtain the text reconstruction feature vector. ,in, Represents the text reconstruction feature vector The Middle The representation of each word after reconstruction; The image decoder network Decoding and reconstruction are performed to obtain the image reconstruction feature vector. ,in, Image representing a comment In the The eigenvalues ​​after dimensional reconstruction; Step 3: Processing the graph attention network: Step 3.1: Define the number of updates as d, and initialize d=1; Step 3.1.1: Calculate the user node updated in the (d-1)th time using equation (9). Features ; (9) In equation (9), Represents user node The set of all neighboring product nodes, Represents user node The number of all neighboring product nodes; Step 3.1.2: Calculate the product node updated for the (d-1)th time using equation (10). Features ; (10) In equation (10), Represents product nodes The set of all neighboring user nodes, Represents product nodes The number of all neighboring user nodes; Step 3.2: Calculate the user node updated in the (d-1)th time using equation (11). and its neighboring product nodes The k-th head attention coefficient ; (11) In equation (7), This represents the activation function. and These are two parameter matrices to be learned. This represents the user node that is updated for the (d-1)th time. The set of neighboring product nodes excluding neighboring product nodes A neighboring product node Features This represents the concatenation operation on vectors. This represents the matrix transpose operation; Step 3.3: Calculate the user node updated for the dth time using equations (12) and (13) respectively. eigenvectors and product nodes Features ; (12) (13) In equations (12)-(13), Represents a node and nodes The edge between In the Probability on each topic; and , Indicates the number of topics; Step 4: Construct a user-personalized demand prediction model consisting of a multimodal variational autoencoder network and a graph attention network, and train and optimize it. Step 4.1: Construct the loss function of the graph attention network for the (d-1)th update using equation (14). : (14) In equation (14), Represents user node The set of non-neighboring nodes Represents user node non-neighbor node set The number of nodes in the middle; Represents user node non-neighbor node set The expectation of all nodes in the middle, The user nodes in the table are updated at the dth time. non-neighbor node set any node Features; Construct the reconstruction loss for comment image data using equation (15). : (15) In equation (15), Represents the set of edges The expectation of all edges in the equation; Construct the reconstruction loss of text data using equation (16) : (16) In equation (16), Represents the initial representation vector The first in The representation of a word Is it the first in the dictionary? The representation of a word, if so, then let Set it to 1; otherwise, let it be 1. =0; Construct the KL loss between the true distribution and the variational distribution using equation (17). : (17) Using equation (18), construct the overall loss function of the user personalized demand prediction model in the dth update. : (18) Step 5.2: Train the user personalized demand prediction model using the Adam algorithm and calculate the overall loss function. When the number of updates d reaches its maximum value d max Time or overall loss function Upon convergence, training is stopped, thus obtaining the optimal demand prediction model, which can be used to predict the personalized product demand of any user.

2. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the user personalized demand prediction method of claim 1, wherein the processor is configured to execute the program stored in the memory.

3. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to perform the steps of the user personalized demand prediction method of claim 1.

Images