Personalized Recommendation Method, System and Medium Based on Graph Neural Network Combined with Curiosity

Abstract

The present invention discloses a personalized recommendation method, system, computer device and storage medium based on graph neural network combined with curiosity. The method includes: obtaining a dataset of user-item interactions, using a latent semantic module to obtain interaction information between users and items and feature vectors of users and items, inputting the obtained data into a graph convolution module and a curiosity module respectively to obtain a predicted relevance score of an item for a specific user and a predicted curiosity score respectively; obtaining a personalized item recommendation list for the user according to the predicted relevance score and the predicted curiosity score; training the latent semantic module and the graph convolution module in the curiosity model using the dataset, inputting the user and item numbers into the trained curiosity model to obtain a personalized item recommendation list for the user. By deeply integrating curiosity with the recommendation system, the present invention effectively improves the diversity of recommendation results on the premise of ensuring a certain degree of recommendation accuracy.

Description

Technical Field

[0001] The present invention relates to the field of recommendation system research, and particularly to a personalized recommendation method, system, computer device and storage medium based on graph neural network combined with curiosity. Background Art

[0002] With the development of technology, modern Internet technology is becoming more and more developed. People can receive more and more information every day, and the types of information received are becoming more and more diverse, which is called "information explosion" or "information overload". Due to the emergence of this problem, tools to assist humans in information processing have gradually become a necessity. Search engines and recommendation systems are two representative technologies to solve the problem of information overload. Traditional search engines can only filter information according to the keywords provided by users, which can solve most of people's needs, but do not achieve the personalization of information filtering and automatically perceive users' preferences. Therefore, the recommendation system is proposed. It can automatically perceive users' preference habits and automatically screen out the information that best meets users' needs from a large amount of information according to users' personalized needs, without the need for users to explicitly provide relevant information about the content they need. Essentially, the recommendation system is to evaluate items that users have never seen, contacted and used on behalf of users, including books, film and television works, news, music, restaurants, tourist attractions, etc. As an important means of information filtering, the recommendation system is one of the most effective methods to solve the problem of information overload at present and is the core technology of Internet products for users.

[0003] However, traditional recommendation systems are often accuracy-oriented, which easily leads to the problem of "information island", that is, repeatedly recommending items that exactly match users' past preferences, making it difficult for them to come into contact with other potentially interesting items. As time goes by, users gradually lose interest in the convergent recommended content and are eager to accept new things to satisfy their own exploration and knowledge-seeking desires. Curiosity, as an important psychological characteristic of humans, can effectively guide people to explore the unknown. Summary of the Invention

[0004] In order to solve the above-mentioned deficiencies of the prior art, the present invention provides a personalized recommendation method, system, computer device and storage medium based on graph neural network combined with curiosity. During the model training process, curiosity is incorporated, so that curiosity can act on the upstream recommendation process, realizing the deep integration of curiosity and the recommendation system. On the premise of ensuring a certain degree of recommendation accuracy, the diversity of the recommendation results is effectively improved.

[0005] The first object of the present invention is to provide a personalized recommendation method based on graph neural network combined with curiosity.

[0006] The second object of the present invention is to provide a personalized recommendation system based on a graph neural network combined with curiosity.

[0007] The third object of the present invention is to provide a computer device.

[0008] The fourth object of the present invention is to provide a storage medium.

[0009] The first object of the present invention can be achieved by adopting the following technical solutions:

[0010] A personalized recommendation method based on a graph neural network combined with curiosity, the method comprising:

[0011] Obtain a data set of user-item interactions;

[0012] According to the data set, use the latent semantic module in the curiosity model to obtain the interaction information between the user and the item and the feature vectors of the user and the item;

[0013] Input the interaction information between the user and the item and the feature vectors of the user and the item into the graph convolution module in the curiosity model to obtain the predicted relevance score of each item for a specific user;

[0014] Input the feature vectors of the user and the item into the curiosity module in the curiosity model to obtain the predicted curiosity score of each item for a specific user;

[0015] According to the predicted relevance score and the predicted curiosity score of each item for a specific user, obtain a personalized item recommendation list for the user;

[0016] Use the data set to train the latent semantic module and the graph convolution module in the curiosity model to obtain a trained curiosity model;

[0017] Input the user and item numbers into the trained curiosity model to obtain a personalized item recommendation list for the user.

[0018] Further, using the latent semantic module in the curiosity model to obtain the interaction information between the user and the item and the feature vectors of the user and the item, including:

[0019] According to the item information in the data set, obtain the item label feature vector;

[0020] Preprocess the data set to obtain the interaction information between the user and the item;

[0021] According to the interaction information between the user and the item and the item label feature vector, obtain the feature vectors of the user and the item.

[0022] Further, obtaining the feature vectors of the user and the item based on the interaction information between the user and the item and the item label feature vector includes:

[0023] Obtaining the user feature vector according to the interaction information between the user and the item;

[0024] Constructing the item feature vector for each item according to the interaction information between the user and the item and the item label feature vector, including:

[0025] Obtaining the item identity feature vector according to the interaction information between the user and the item;

[0026] Obtaining the item feature vector according to the item identity feature vector and the item label feature vector, and the formula is as follows:

[0027]

[0028] Wherein, is the item identity feature vector, represents the transpose of the matrix composed of the item label feature vectors; W ∈ R N×T is the label feature conversion matrix, where each row W n,· ∈ R T represents the proportion of each label in the item corresponding to the row number. The sum of each row is 1; N is the total number of items, and T is the total number of labels; the label feature conversion matrix is a weight matrix recording the proportion of items on each label and is a trainable parameter matrix, which is initialized according to the original information of the item labels in the dataset.

[0029] Further, training the latent semantic module and the graph convolution module in the curiosity model using the dataset to obtain the trained curiosity model includes:

[0030] During training, if the number of iterations is less than 5, then: input the data of the interaction between the user and the item in the dataset into the latent semantic module, input the interaction information between the user and the item and the feature vectors of the user and the item obtained in the latent semantic module into the graph convolution module, take the predicted correlation score output by the graph convolution module as the final score, calculate the loss according to the final score, and then update the parameters in the model;

[0031] Otherwise: input the data of the interaction between the user and the item in the dataset into the latent semantic module, input the interaction information between the user and the item and the feature vectors of the user and the item obtained in the latent semantic module into the graph convolution module, and input the feature vectors of the user and the item into the curiosity module; calculate the final score according to the predicted correlation score output by the graph convolution module and the predicted curiosity score obtained by the curiosity module, calculate the loss according to the final score, and then update the parameters in the model;

[0032] The objective function used during training is as follows:

[0033]

[0034] Among them, represents the number of neighbors of user u m on the user-item interaction bipartite graph, which is a topological graph composed of the interaction information between users and items; i p represents the item accessed by the user, i j while represents the item not accessed by the user. represents the final score of the curiosity model for user u m accessing item i p The final score; the second additive term is an L2 regularization term; the parameter λ is used to control the regularization strength; E represents all the feature vectors of users and items. represents the final score of the curiosity model for user u m accessing item i j The final score; the second additive term is an L2 regularization term; the parameter λ is used to control the regularization strength; E represents all the feature vectors of users and items.

[0035] Furthermore, inputting the feature vectors of the user and the item into the curiosity module in the curiosity model to obtain the predicted curiosity scores of each item for a specific user includes:

[0036] Calculating the curiosity stimulation value according to the feature vectors of the user and the item, including:

[0037] Quantifying the complexity of the item on the label according to the feature vector of the item;

[0038] Calculating the user personalization factor according to the feature vectors of the user and the item;

[0039] Calculating the curiosity stimulation value according to the user personalization factor and the complexity;

[0040] Obtaining the predicted curiosity scores of the item for a specific user according to the curiosity stimulation value, including:

[0041] Drawing a frequency distribution histogram of the stimulation value according to the occurrence frequency of the stimulation value in each equal value interval;

[0042] Fitting an inverted U-shaped curiosity curve representing the relationship between the curiosity value and the stimulation value according to the frequency distribution histogram;

[0043] Obtaining the predicted curiosity scores of the item for a specific user by using the inverted U-shaped curiosity curve according to the curiosity stimulation value.

[0044] Furthermore, inputting the interaction information between the user and the item and the feature vectors of the user and the item into the graph convolution module in the curiosity model to obtain the predicted correlation scores of each item for a specific user includes:

[0045] Perform graph convolution operations on the interaction information between users and items and the feature vectors of users and items using multiple convolutional layers;

[0046] Perform layer combination operations on the outputs of graph convolutional layers with different numbers of convolutions to obtain the final output of the graph neural network;

[0047] Calculate the predicted relevance score of each item for a specific user based on the output of the graph neural network.

[0048] Furthermore, the formula for the graph convolution operation is as follows:

[0049]

[0050]

[0051] where the subscript l represents the graph convolutional layer where the vector is located, represents the set of neighbor nodes of user node u m ; represents the set of neighbor nodes of item node i n ; the symmetric normalization term follows the design of the standard GCN; l is a positive integer greater than or equal to 1; represent the feature vectors of users and items respectively; represent the feature vectors of users and items output by the l-th convolutional layer respectively;

[0052] The formula for the layer combination operation is as follows:

[0053]

[0054]

[0055] where are the feature vectors of users and items obtained by the layer combination operation respectively; α l ≥0, is a trainable parameter;

[0056] The calculating the predicted relevance score of each item for a specific user based on the output of the graph neural network includes:

[0057] Using the method of taking the vector inner product of the feature vectors of users and items to calculate the predicted relevance score of each item for a specific user, the formula is as follows:

[0058]

[0059] where is the predicted relevance score of each item for a specific user, represents the transpose of the vector .

[0060] Further, obtaining a personalized item recommendation list for the user according to the predicted relevance score and the predicted curiosity score of each item for a specific user includes:

[0061] Obtaining the final recommendation score for the user according to the predicted relevance score and the predicted curiosity score of each item for a specific user;

[0062] Sorting according to the final recommendation score to obtain a personalized item recommendation list for the user.

[0063] The second object of the present invention can be achieved by adopting the following technical solutions:

[0064] A personalized recommendation system based on a graph neural network combined with curiosity, the system includes:

[0065] A dataset acquisition module for acquiring a dataset of user-item interactions;

[0066] A feature vector generation module for obtaining the interaction information between the user and the item and the feature vectors of the user and the item according to the dataset by using the latent semantic module in the curiosity model;

[0067] A relevance score prediction module for inputting the interaction information between the user and the item and the feature vectors of the user and the item into the graph convolution module in the curiosity model to obtain the predicted relevance score of each item for a specific user;

[0068] A curiosity score prediction module for inputting the feature vectors of the user and the item into the curiosity module in the curiosity model to obtain the predicted curiosity score of each item for a specific user;

[0069] A predicted item recommendation list generation module for obtaining a personalized item recommendation list for the user according to the predicted relevance score and the predicted curiosity score of each item for a specific user;

[0070] A training module for training the latent semantic module and the graph convolution module in the curiosity model by using the dataset to obtain a trained curiosity model;

[0071] An item recommendation list module for inputting the user and item numbers into the trained curiosity model to obtain a personalized item recommendation list for the user.

[0072] The third object of the present invention can be achieved by adopting the following technical solutions:

[0073] A computer device includes a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the above-mentioned personalized recommendation method is implemented.

[0074] The fourth object of the present invention can be achieved by adopting the following technical solutions:

[0075] A storage medium stores a program, and when the program is executed by a processor, the above-mentioned personalized recommendation method is implemented.

[0076] The present invention has the following beneficial effects compared with the prior art:

[0077] 1. The present invention proposes a personalized recommendation method based on a graph neural network combined with curiosity for modeling complex curiosity. This method can effectively obtain user curiosity that conforms to the psychological definition based on item tags, discover the user's preference components on the tags, and use curiosity to help the recommendation system give a recommendation list that better meets the user's needs.

[0078] 2. The present invention innovatively attempts to incorporate curiosity during the model training process, enabling curiosity to act on the upstream recommendation process rather than re-ranking after obtaining the recommendation results downstream, achieving a deep integration of curiosity and the recommendation system. This method can effectively improve the diversity of recommendation results on the premise of ensuring a certain degree of recommendation accuracy, promote users to explore unknown fields, and help break the "information cocoon" and "information island" dilemmas that may be caused by accuracy-oriented recommendation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on the structures shown in these drawings.

[0080] Figure 1 It is a flowchart of the personalized recommendation method based on a graph neural network combined with curiosity according to Embodiment 1 of the present invention.

[0081] Figure 2 It is a schematic diagram of the principle of the personalized recommendation method based on a graph neural network combined with curiosity according to Embodiment 1 of the present invention.

[0082] Figure 3 It is a schematic diagram of the graph convolution operation and layer combination operation in the graph convolution module according to Embodiment 1 of the present invention.

[0083] Figure 4 It is a flowchart of the curiosity model training according to Embodiment 1 of the present invention.

[0084] Figure 5This is the framework diagram of the personalized recommendation system based on graph neural network combined with curiosity in Embodiment 2 of the present invention.

[0085] Figure 6 This is the structural block diagram of the computer device in Embodiment 3 of the present invention. Detailed implementation manners

[0086] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. It should be understood that the specific embodiments described are only used to explain the present application and not to limit the present application.

[0087] Inspired by the psychological curiosity theory, the present invention provides a personalized recommendation method, system, computer device and storage medium based on graph neural network combined with curiosity. The personalized recommendation method based on graph neural network combined with curiosity is called Cur-GCN (Curiosity-promoted GCN). This method learns high-quality user and item feature information through the graph neural network recommendation method, and at the same time learns the label (Tag) information of the item to model the complex curiosity of users. Different from the existing methods, the present invention incorporates curiosity during the training process, incorporates curiosity into the prediction output of the recommendation system in an appropriate proportion, uses Bayesian personalized ranking (BPR) as the objective function, and uses the method of gradient backpropagation for training, so that the individual curiosity level can effectively guide the parameter update of the model, thereby realizing the deep integration of curiosity and the recommendation system.

[0088] Embodiment 1:

[0089] As Figure 1 、 2 shown, the personalized recommendation method based on graph neural network combined with curiosity provided in this embodiment includes the following steps:

[0090] S101. Obtain a dataset of user-item interactions.

[0091] In this embodiment, the publicly available movie dataset Movielens-1M, book dataset Amazon-Book, and catering dataset Yelp2018 are collected for the training and testing of the present invention.

[0092] S102. According to the dataset, use the latent semantic module in the curiosity model to obtain the interaction information between users and items and the feature vectors of users and items.

[0093] As shown Figure 2 in the figure, in this embodiment, the curiosity model includes a latent semantic module, a graph convolutional module, and a curiosity module.

[0094] Furthermore, step S102 includes:

[0095] (1) Obtain the item label feature vector according to the item information in the dataset.

[0096] Construct an item label dataset according to the item information in the dataset, and further construct a feature vector set of item labels.

[0097] By counting the label information contained in the original information of the items in the dataset, the occurrence frequency of each label is counted, and the labels with a label occurrence frequency greater than 50 are selected for modeling, weighing between the overhead of model parameters and effective modeling for labels.

[0098] According to the statistically obtained label set Tag = {tag1,…,tag t ,…,tag T}, model a feature vector for each label to form a label feature vector set where the total number of labels is T.

[0099] (2) Preprocess the dataset to obtain the interaction information of users and items.

[0100] Since Movielens-1M is relatively small itself, the original dataset is directly used for experiments. For Amazon-Book and Yelp2018, due to the extremely large size of the original dataset and limited experimental conditions, the original dataset was processed with 10-core during data preprocessing. That is, for the Amazon-Book dataset, in the case of duplicate user-item interactions (i.e., a user may access the same item at different times), first, users and items were screened to ensure that each user has at least 30 access records and each item has at least 30 access records; since Cur-GCN does not require time information, the access records were then de-duplicated and screened again to ensure that each user has accessed at least 10 items and each item has been accessed by at least 10 users. The preprocessing operation of ensuring that each user has accessed at least 10 items and each item has been accessed by at least

[0101] 10 users is called the 10-core operation. For Yelp2018, since it is even larger than Amazon-Book, 10-core processing was performed after ensuring that users and items have at least 50 access records without de-duplication.

[0102] The interaction information between users and items will form a user-item interaction graph, and graph convolution operations will be performed based on this interaction graph in the graph convolution module.

[0103] (3) Construct the feature vectors of users and items according to the interaction information between users and items and the item label feature vectors.

[0104] (3-1) Construct user feature vectors for each user according to the interaction information between users and items.

[0105] Construct a set of user feature vectors for each user according to the interaction information between users and items Set of user identity feature vectors Directly form the set of user feature vectors M is the total number of users.

[0106] (3-2) Construct item feature vectors for each item according to the interaction information between users and items and the item label feature vectors.

[0107] Construct a set of item feature vectors E for each item according to the interaction information between users and items and the item label feature vectors I ={E Iid ,E tag}, and the item feature vector set contains the set of item identity feature vectors and the set of item label feature vectors where the total number of items is N.

[0108] Integrate the label feature vectors into the item feature vectors according to the proportion of each label in the items in the original dataset, specifically including:

[0109] The calculation formula for integrating the label feature vectors into the item feature vectors is as follows:

[0110]

[0111] where, is the identity feature vector of the item, represents the transpose of the matrix composed of label feature vectors, W ∈ R N×T is a label feature transformation matrix, where each row W n,· ∈ R TIt represents the proportion of each tag in the item corresponding to the line number. Therefore, the sum of each line should be 1, and a normalization function softmax(·) is used here to achieve this. This feature transformation matrix is a weight matrix that records the proportion of items on each tag and is a trainable parameter matrix, which will be initialized according to the original information of item tags in the dataset.

[0112] S103. Input the interaction information between the user and the item and the feature vectors of the user and the item into the graph convolution module in the curiosity model to obtain the predicted relevance score of each item for a specific user.

[0113] The predicted relevance score of each item for a specific user means that each user will generate a predicted score for all items.

[0114] Furthermore, as Figure 3 shown, step S103 includes:

[0115] (1) Perform graph convolution operations on the interaction information between the user and the item and the feature vectors of the user and the item using multiple convolutional layers.

[0116] The graph convolution module in this embodiment includes three convolutional layers, and the graph convolution operations are performed using the three convolutional layers.

[0117] According to the information shown on the user-item interaction graph (if there is a link between the user and the item, it means they are neighbors to each other), for all nodes in the interaction graph, aggregate their neighbor information to achieve the learning of graph local features;

[0118] The calculation formula for the graph convolution operation is as follows:

[0119]

[0120]

[0121] Among them, the superscript (l) in the node feature vector represents the label of the graph convolution layer where the vector is located, represents the set of neighbor nodes of the user node u m , represents the set of neighbor nodes of the item node i n . The symmetric normalization term follows the design of the standard GCN. The normalization operation can prevent the scale of the feature vector from becoming too large as the number of graph convolution operations increases. Existing research has pointed out that using such a normalization term has the best effect. l is a positive integer greater than or equal to 1, represents the feature vector of the user and the item output by the latent semantic module.

[0122] (2) Perform a layer combination operation on the outputs of graph convolutional layers with different numbers of convolutions to obtain the final output of the graph neural network.

[0123] The calculation formula for the layer combination operation is as follows:

[0124]

[0125]

[0126] Among them, are the feature vectors of the user and the item obtained by the layer combination operation, and α l ≥0 represents the importance of the features of the l-th layer when constructing the output feature vector of the GCN, and it is a trainable parameter; the sigmoid function ensures that the weight is between 0 and 1 to prevent the scale of the output feature vector from being too large; L refers to the total number of graph convolutional layers.

[0127] (3) Calculate the predicted relevance score of each item for a specific user according to the output of the graph neural network.

[0128] Use the method of taking the vector inner product of the feature vectors of the user and the item to calculate the predicted relevance score of each item for a specific user. The calculation formula is as follows:

[0129]

[0130] Among them, is the feature vector of user u m , is the feature vector of item i n , represents the transpose of the vector . Take the vector inner product of the feature vectors of the user and the item, and then use the sigmoid(·) operation to scale it to between 0 and 1 to form the relevance score of the item for a specific user.

[0131] S104. Input the feature vectors of the user and the item into the curiosity module in the curiosity model to obtain the predicted curiosity scores of each item for a specific user.

[0132] Furthermore, step S104 includes:

[0133] (1) Calculate the curiosity stimulation value according to the feature vectors of the user and the item.

[0134] (1-1) Quantify the complexity of the item on the label using the information entropy according to the feature vector of the item.

[0135] First, use the information entropy to quantify the complexity of the item on the label.

[0136] The calculation formula is as follows:

[0137] dist(i n ) = softmax(W n,· ),

[0138]

[0139] Among them, the result dist(i n ) of the proportion of an item on each label calculated on the trainable weight matrix is an array of length T. By calculating the information entropy of the item on the label, the unified stimulation value of each item for all users according to its own complexity is obtained. However, since the tolerance for unknown things varies among different users, the degree of stimulation of the same item to the curiosity of users is also different.

[0140] (1-2) Calculate the user personalization factor according to the feature vectors of the user and the item.

[0141] Calculate the user personalization factor to express the personalized complexity curiosity stimulation value of the item for different users.

[0142] (1-2-1) First, calculate the proportion of the user on each label, and the calculation formula is as follows:

[0143]

[0144] Among them, the proportion dist(u m ) of the user component on each label obtained is an array of length T, expressing the user's familiarity with each label.

[0145] (1-2-2) Then calculate the JS divergence to represent the difference degree between the user label distribution and the item label distribution, that is, the user's familiarity with the item component. The calculation formula of the JS divergence is as follows:

[0146]

[0147]

[0148] Among them, p i is an element in the probability distribution P, q i is an element in the probability distribution Q. The JS divergence is to calculate the relative entropy between the probability distributions P and Q to represent the difference between the two distributions. Its value range is between 0 and 1. The more similar P and Q are, the smaller the JS divergence, and the value is 0 when they are exactly the same. Substitute the user's distribution on the label and the item's distribution on the label to get D JS (dist(u m )||dist(i n)) can represent the difference degree between the user label distribution and the item label distribution.

[0149] (1 - 2 - 3) Finally, calculate the similarity and dissimilarity between the user and the item in terms of the label component distribution, and then calculate the user personalization factor.

[0150] Calculate the similarity and dissimilarity between the user and the item in terms of the label component distribution, and then calculate the user personalization factor The calculation formula is as follows:

[0151]

[0152]

[0153]

[0154] Among them, the hyperparameter β ∈ [0, 1] is used to control the ratio of dissimilarity and similarity in the personalization factor. Experiments show that setting β = 0.3 can obtain good results.

[0155] (1 - 3) Calculate the curiosity stimulation value according to the user personalization factor and the complexity of the item in terms of labels.

[0156] After calculating the user personalization factor and the information entropy of the distribution of the item in terms of label components, the complexity curiosity stimulation value of the item for a specific user can be calculated. The calculation formula is as follows:

[0157]

[0158] (2) Obtain the predicted curiosity score of the item for a specific user according to the curiosity stimulation value.

[0159] (2 - 1) Draw a frequency distribution histogram of the stimulation value according to the occurrence frequency of the stimulation value in each equal value range.

[0160] Draw a frequency distribution histogram of the stimulation value according to the occurrence frequency of the stimulation value in each equal value range.

[0161] Among them, generally, it is statistically divided into 100 value ranges according to the distance between the maximum and minimum values of the stimulation value.

[0162] (2 - 2) According to the frequency distribution histogram, fit an inverted U-shaped curiosity curve representing the relationship between the curiosity value and the stimulation value.

[0163] Design a mathematical curve according to the Wundt curve theory and fit an inverted U-shaped curiosity curve representing the relationship between the curiosity value and the stimulation value.

[0164] The mathematical curve is as follows:

[0165] y = y1 + bias,

[0166]

[0167] The trainable parameters of the curve are α1, α2, β, γ, and bias, a total of five. Setting these parameters all has a purpose. Among them, the two parameters α1 and α2 are used to control the steepness of the left and right U-shaped edges of the inverted U shape. β is used to control the abscissa position of the highest point of the curve. γ is used to control the maximum value of the curve. Bias is used to adjust the vertical position of the entire curve when necessary. It can be clearly seen that the inverted U shape of the curve is actually composed of two semi-U-shaped curves spliced together at x = β. At the position of x = β of the curve, the left limit and the right limit are equal. According to the definition of a continuous curve in mathematics, the curve is continuously differentiable at x = β. Therefore, the entire curve is continuously differentiable, and the gradient backpropagation method can be used to update the parameters for fitting. When actually conducting experiments, it is found that the initialization method has a greater impact on the curve fitting effect. Therefore, some prior knowledge about the curve parameters will be obtained from the frequency distribution histogram of the stimulus values during fitting. For example, it can be inferred that the abscissa position of the highest point of the curve is approximately near the abscissa of the midpoint of the top edge of the highest bar in the frequency distribution histogram, which can be used to initialize the parameter β. The maximum value of the curve may also be near the ordinate of the midpoint of its top edge, which can be used to initialize the parameters γ and bias. Through such analysis, a suitable initial value can be assigned to the curve parameters, which helps to learn a high-quality and personalized curiosity curve. The method used during fitting is the least non-linear least squares method.

[0168] (2 - 3) According to the curiosity stimulus value, using the inverted U-shaped curiosity curve, obtain the predicted curiosity score of the item for a specific user.

[0169] Normalize the obtained curiosity value, and then obtain the predicted curiosity score of the item for a specific user. The formula is as follows:

[0170]

[0171] Among them, the curiosity value normalization operation is performed using the softmax(·) function, and the normalization operation is performed within the curiosity array calculated for different items by the same user.

[0172] S105. According to the predicted relevance score and the predicted curiosity score of each item for a specific user, obtain the personalized item recommendation list for the user.

[0173] Furthermore, step S105 includes:

[0174] (1) Obtain the final recommendation score of the user according to the predicted relevance score and the predicted curiosity score of each item for a specific user.

[0175] The calculation formula of the final recommendation score is as follows:

[0176]

[0177] Among them, the curiosity integration ratio coefficient η is used to control the curiosity integration level in the recommendation score, so as to ensure that the final recommendation score can take into account both the relevance between the recommended items and the user's preferences and the user's curiosity performance.

[0178] (2) Sort according to the final recommendation score to obtain the personalized item recommendation list of the user.

[0179] Perform recommendation sorting according to the final recommendation score obtained in the above steps to obtain the personalized recommendation list of each user.

[0180] S106. Use the dataset to train the latent semantic module and the graph convolutional module in the curiosity model to obtain a trained curiosity model.

[0181] Furthermore, as Figure 4 shown, step S106 includes:

[0182] (1) In the first five iterations of training, use the predicted relevance score of each item for a specific user as the final score, calculate the loss according to the final score, and then update the parameters in the model.

[0183] This step is mainly for training the feature vector for accuracy.

[0184] The objective function used for training is as follows:

[0185]

[0186] Among them, represents the number of the neighbors of user u m on the user-item interaction bipartite graph, that is, the items represented by other nodes with which it has a link. The user-item interaction bipartite graph is a topological graph composed of the interaction information of users and items; i p is the item accessed by the user, and i j is the item not accessed by the user. represents the predicted score of the model for user u m to access item i p , represents the predicted score of the model for user u m to access item i jThe predicted score. The second additive term of the objective function is an L2 regularization term, which is used to prevent overfitting during the model learning process. Among them, the parameter λ is used to control the regularization strength, and E represents all the feature vectors of users and items. The objective of this BPR objective function is to make the predicted score of the model for user u m accessing item i p be as large as possible compared to the predicted score of the model for user u m accessing item j j of the predicted score.

[0187] (2) Starting from the fifth iteration, calculate the final score based on the predicted relevance score and curiosity score of each item for a specific user, calculate the loss based on the final score, and then update the parameters in the model.

[0188] The training objective function remains unchanged and is still the objective function Loss in step (1) BPR .

[0189] During the entire training process, use the methods of loss backpropagation and gradient descent to update and optimize the trainable parameters (feature vectors and layer combination weights in the graph convolution module).

[0190] In this embodiment, the training is terminated after 200 iterations are completed.

[0191] S107. Input the user and item numbers into the trained curiosity model to obtain the item recommendation list for the user.

[0192] Input the user and item numbers into the latent semantic module in the trained curiosity model, input the data output by the latent semantic module into the graph convolution module and the curiosity module. The graph convolution module outputs the relevance score of the user for each item, and the curiosity module outputs the curiosity score of the user for each item. The relevance score and curiosity score of each item are weighted and summed to obtain the final recommendation score, and then each item is sorted to obtain the item recommendation list for this user.

[0193] Those skilled in the art can understand that all or part of the steps in the method for implementing the above embodiments can be completed by instructing relevant hardware through a program, and the corresponding program can be stored in a computer-readable storage medium.

[0194] It should be noted that although the method operations of the above embodiments are described in a specific order in the drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the described steps can be changed in the execution order. Additionally or alternatively, some steps can be omitted, multiple steps can be combined into one step for execution, and / or one step can be decomposed into multiple steps for execution.

[0195] Example 2:

[0196] As Figure 5 shown, this embodiment provides a personalized recommendation system based on a graph neural network combined with curiosity. The system includes a dataset acquisition module 501, a feature vector generation module 502, a relevance score prediction module 503, a curiosity score prediction module 504, a predicted item recommendation list generation module 505, a training module 506, and an item recommendation list module 507, where:

[0197] The dataset acquisition module 501 is used to acquire a dataset of user-item interactions;

[0198] The feature vector generation module 502 is used to obtain the interaction information between users and items and the feature vectors of users and items according to the dataset by using the latent semantic module in the curiosity model;

[0199] The relevance score prediction module 503 is used to input the interaction information between users and items and the feature vectors of users and items into the graph convolution module in the curiosity model to obtain the predicted relevance score of each item for a specific user;

[0200] The curiosity score prediction module 504 is used to input the feature vectors of users and items into the curiosity module in the curiosity model to obtain the predicted curiosity score of each item for a specific user;

[0201] The predicted item recommendation list generation module 505 is used to obtain a personalized item recommendation list for the user according to the predicted relevance score and the predicted curiosity score of each item for a specific user;

[0202] The training module 506 is used to train the latent semantic module and the graph convolution module in the curiosity model by using the dataset to obtain a trained curiosity model;

[0203] The item recommendation list module 507 is used to input the user and item numbers into the trained curiosity model to obtain a personalized item recommendation list for the user.

[0204] For the specific implementation of each module in this embodiment, reference may be made to the above-mentioned Embodiment 1, which will not be elaborated here one by one; it should be noted that the system provided in this embodiment only takes the above-mentioned division of each functional module as an example. In actual applications, the above functions can be allocated to different functional modules according to needs, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above.

[0205] Example 3:

[0206] This embodiment provides a computer device, which can be a computer, such asFigure 6 As shown in Figure 6 , it includes a processor 602, a memory, an input device 603, a display 604, and a network interface 605 connected through a system bus 601. The processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium 606 and an internal memory 607. The non-volatile storage medium 606 stores an operating system, a computer program, and a database. The internal memory 607 provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. When the processor 602 executes the computer program stored in the memory, the personalized recommendation method of the above-mentioned Embodiment 1 is implemented as follows:

[0207] Obtain a dataset of user-item interactions;

[0208] According to the dataset, use the latent semantic module in the curiosity model to obtain the interaction information between the user and the item and the feature vectors of the user and the item;

[0209] Input the interaction information between the user and the item and the feature vectors of the user and the item into the graph convolutional module in the curiosity model to obtain the predicted relevance score of each item for a specific user;

[0210] Input the feature vectors of the user and the item into the curiosity module in the curiosity model to obtain the predicted curiosity score of each item for a specific user;

[0211] According to the predicted relevance score and the predicted curiosity score of each item for a specific user, obtain a personalized item recommendation list for the user;

[0212] Use the dataset to train the latent semantic module and the graph convolutional module in the curiosity model to obtain a trained curiosity model;

[0213] Input the user and item numbers into the trained curiosity model to obtain a personalized item recommendation list for the user.

[0214] Embodiment 4:

[0215] This embodiment provides a storage medium, which is a computer-readable storage medium and stores a computer program. When the computer program is executed by a processor, the personalized recommendation method of the above-mentioned Embodiment 1 is implemented as follows:

[0216] Obtain a dataset of user-item interactions;

[0217] According to the dataset, use the latent semantic module in the curiosity model to obtain the interaction information between the user and the item and the feature vectors of the user and the item;

[0218] Input the interaction information between the user and the item, as well as the feature vectors of the user and the item, into the graph convolution module in the curiosity model to obtain the predicted relevance scores of each item for a specific user;

[0219] Input the feature vectors of the user and the item into the curiosity module in the curiosity model to obtain the predicted curiosity scores of each item for a specific user;

[0220] According to the predicted relevance scores and predicted curiosity scores of each item for a specific user, obtain the personalized item recommendation list for the user;

[0221] Use the dataset to train the latent semantic module and graph convolution module in the curiosity model to obtain the trained curiosity model;

[0222] Input the user and item numbers into the trained curiosity model to obtain the personalized item recommendation list for the user.

[0223] It should be noted that the computer-readable storage medium in this embodiment can be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of the computer-readable storage medium can include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0224] In summary, the personalized recommendation method based on graph neural network combined with curiosity provided by the present invention constructs a set of feature vectors of users and items according to the interaction data between users and items in the dataset; constructs a label dataset of items according to the original information of the items in the dataset, and further constructs a set of feature vectors of item labels; according to the proportion of each label in the item clarified by the original information of the item in the dataset, integrates the label feature vectors into the item feature vectors proportionally to form the final set of item feature vectors; inputs the user and item feature vectors into the graph convolution module and the curiosity module respectively; among them, the graph convolution module conducts learning for recommendation accuracy, performs graph convolution operations on the topological graph called "user-item interaction bipartite graph" composed of the interaction data between users and items, that is, conducts neighbor information aggregation, aggregates the information of the neighbor nodes of the user node or item node to the own node to realize the learning of local graph structure features; then conducts layer combination operations, combines the outputs of the convolution layers with different convolution times together as the final feature vector output; calculates the predicted relevance score of the item for a specific user by taking the inner product of the two finally output vectors, that is, the user and item feature vectors; while the curiosity module conducts learning for recommendation diversity, the input is the user and item feature vectors, calculates the complexity curiosity stimulation value according to the definition of complexity curiosity; draws a frequency distribution histogram of the stimulation value according to the occurrence frequency of the stimulation value in each equal value interval; then according to the Wundt curve theory, uses the designed mathematical curve to fit to obtain a curiosity curve showing the inverted U-shaped distribution relationship between the curiosity value and the stimulation value, and normalizes the obtained curiosity value, and further obtains the predicted curiosity score of the item for a specific user; uses the curiosity proportion coefficient to add the curiosity score to the relevance score in a certain proportion to obtain the final recommendation score of each item for a specific user; sorts the final recommendation scores to obtain the personalized recommendation list for each user. During the training process, since the calculation of the curiosity module requires the feature vectors of users and items to be able to reflect the features of users and items to a certain extent, therefore, the curiosity is not calculated first, only the relevance score is calculated for recommendation ranking, and the model is trained for recommendation accuracy. After five iterations of training, the curiosity score is calculated, and thereafter, the final recommendation score including the relevance score and the curiosity score is used for recommendation ranking to train the entire model. Input the specified user and item numbers into the trained model, and according to the predicted relevance score and the predicted curiosity score, the item recommendation list for the specified user can be obtained. By modeling the complexity curiosity, the present invention can discover the labels that users may like but are not reflected in the interaction data on the user's curiosity, expand the label components included in the recommended content obtained by users, thereby alleviating the problem of insufficient diversity in the recommendation method for accuracy and improving the recommendation effect.

[0225] As described above, it is only a preferred embodiment of the present invention patent, but the protection scope of the present invention patent is not limited thereto. Any person skilled in the art within the scope disclosed by the present invention patent, according to the technical solution and inventive concept of the present invention patent, makes equivalent substitutions or changes, all of which belong to the protection scope of the present invention patent.

Claims

1. A personalized recommendation method based on graph neural network combined with curiosity, characterized in that, The method includes: Obtaining a dataset of user-item interactions; Based on the dataset, using the latent semantic module in the curiosity model to obtain the interaction information between the user and the item and the feature vectors of the user and the item; Inputting the interaction information between the user and the item and the feature vectors of the user and the item into the graph convolutional module in the curiosity model to obtain the predicted relevance score of each item for a specific user; Inputting the feature vectors of the user and the item into the curiosity module in the curiosity model to obtain the predicted curiosity score of each item for a specific user; Based on the predicted relevance score and the predicted curiosity score of each item for a specific user, obtaining a personalized item recommendation list for the user; Using the dataset to train the latent semantic module and the graph convolutional module in the curiosity model to obtain a trained curiosity model; Inputting the user and item numbers into the trained curiosity model to obtain a personalized item recommendation list for the user.

2. The personalized recommendation method according to claim 1, wherein The step of, based on the dataset, using the latent semantic module in the curiosity model to obtain the interaction information between the user and the item and the feature vectors of the user and the item includes: Based on the item information in the dataset, obtaining the item label feature vector; Preprocessing the dataset to obtain the interaction information between the user and the item; Based on the interaction information between the user and the item and the item label feature vector, obtaining the feature vectors of the user and the item.

3. The personalized recommendation method according to claim 2, wherein The step of, based on the interaction information between the user and the item and the item label feature vector, obtaining the feature vectors of the user and the item includes: Based on the interaction information between the user and the item, obtaining the user feature vector; Based on the interaction information between the user and the item and the item label feature vector, constructing an item feature vector for each item, including: Based on the interaction information between the user and the item, obtaining the item identity feature vector; Based on the item identity feature vector and the item label feature vector, obtaining the item feature vector, and the formula is as follows: Among them, is the item identity feature vector, represents the transpose of the matrix composed of item label feature vectors; W ∈ R N×T is the label feature conversion matrix, where each row represents the proportion of each label in the items corresponding to the row number. The sum of each row is 1; N is the total number of items, and T is the total number of labels; the label feature conversion matrix is a weight matrix that records the proportion of items on each label and is a trainable parameter matrix, which is initialized according to the original information of item labels in the dataset.

4. The personalized recommendation method according to claim 2, wherein The step of using the dataset to train the latent semantic module and the graph convolutional module in the curiosity model to obtain a trained curiosity model includes: During training, if the number of iterations is less than 5, then: Inputting the data of user-item interactions in the dataset into the latent semantic module, inputting the interaction information between the user and the item and the feature vectors of the user and the item obtained in the latent semantic module into the graph convolutional module, using the predicted relevance score output by the graph convolutional module as the final score, calculating the loss based on the final score, and then updating the parameters in the model; Otherwise: Inputting the data of user-item interactions in the dataset into the latent semantic module, inputting the interaction information between the user and the item and the feature vectors of the user and the item obtained in the latent semantic module into the graph convolutional module, and inputting the feature vectors of the user and the item into the curiosity module; Calculating the final score based on the predicted relevance score output by the graph convolutional module and the predicted curiosity score obtained by the curiosity module, calculating the loss based on the final score, and then updating the parameters in the model; The objective function used during training is as follows: where M is the total number of users, represents the neighbor number of user u m on the user-item interaction bipartite graph, which is a topological graph composed of the interaction information between users and items; i p represents the item accessed by the user, i j while the item not accessed by the user represents the final score of the curiosity model for user u m to access item i p The final score represents the final score of the curiosity model for user u m to access item i j The final score; the second summand is an L2 regularization term; the parameter λ is used to control the regularization strength; E represents all the feature vectors of users and items.

5. The personalized recommendation method according to claim 1, wherein The step of inputting the feature vectors of the user and the item into the curiosity module in the curiosity model to obtain the predicted curiosity score of each item for a specific user includes: Based on the feature vectors of the user and the item, calculating the curiosity stimulation value, including: Quantify the complexity of an item on a label using information entropy based on the feature vector of the item; Calculate the user personalization factor based on the feature vectors of the user and the item; Calculate the curiosity stimulation value based on the user personalization factor and the complexity; Obtain the predicted curiosity score of the item for a specific user based on the curiosity stimulation value, including: Draw a frequency distribution histogram of the stimulation values according to the occurrence frequencies of the stimulation values in each equal value range; Fit an inverted U-shaped curiosity curve showing the relationship between the curiosity value and the stimulation value based on the frequency distribution histogram; Obtain the predicted curiosity score of the item for a specific user using the inverted U-shaped curiosity curve based on the curiosity stimulation value.

6. The personalized recommendation method according to claim 1, wherein The inputting the interaction information of the user and the item and the feature vectors of the user and the item into the graph convolution module in the curiosity model to obtain the predicted relevance score of each item for a specific user includes: Perform graph convolution operations on the interaction information of the user and the item and the feature vectors of the user and the item using multiple convolutional layers; Perform a layer combination operation on the outputs of the graph convolution layers with different numbers of convolutions to obtain the final output of the graph neural network; Calculate the predicted relevance score of each item for a specific user based on the output of the graph neural network.

7. The personalized recommendation method according to claim 6, wherein The formula for the graph convolution operation is as follows: Among them, the subscript l indicates the graph convolutional layer where the vector is located. denotes the set of neighbor nodes of user node u m . denotes the set of neighbor nodes of item node i n ; The symmetric normalization term follows the design of the standard GCN; l is a positive integer greater than or equal to 1. denotes the feature vectors of users and items. respectively denote the feature vectors of users and items output by the l-th convolutional layer. The formula for the layer combination operation is as follows: Among them, are the feature vectors of the user and the item obtained by the layer combination operation respectively; α l ≥ 0, which is a trainable parameter; The calculating the predicted relevance score of each item for a specific user based on the output of the graph neural network includes: Calculate the predicted relevance score of each item for a specific user using the method of taking the vector inner product of the feature vectors of the user and the item. The formula is as follows: wherein, is the predicted relevance score for each item for a specific user, represents the vector transpose of.

8. The personalized recommendation method according to claim 1, wherein The obtaining the personalized item recommendation list for the user based on the predicted relevance score and the predicted curiosity score of each item for a specific user includes: Obtain the final recommendation score for the user based on the predicted relevance score and the predicted curiosity score of each item for a specific user; Sort according to the final recommendation score to obtain the personalized item recommendation list for the user.

9. A personalized recommendation system based on graph neural network combined with curiosity, characterized in that, The system includes: A dataset acquisition module for acquiring a dataset of user-item interactions; A feature vector generation module for obtaining the interaction information of the user and the item and the feature vectors of the user and the item based on the dataset using the latent semantic module in the curiosity model; A relevance score prediction module for inputting the interaction information of the user and the item and the feature vectors of the user and the item into the graph convolution module in the curiosity model to obtain the predicted relevance score of each item for a specific user; A curiosity score prediction module for inputting the feature vectors of the user and the item into the curiosity module in the curiosity model to obtain the predicted curiosity score of each item for a specific user; A predicted item recommendation list generation module for obtaining the personalized item recommendation list for the user based on the predicted relevance score and the predicted curiosity score of each item for a specific user; A training module for training the latent semantic module and the graph convolution module in the curiosity model using the dataset to obtain a trained curiosity model; An item recommendation list module for inputting the user and item numbers into the trained curiosity model to obtain the personalized item recommendation list for the user.

10. A storage medium stores a program, characterized in that, When the described program is executed by a processor, it implements the personalized recommendation method described in any one of claims 1-8.

Images