A multi-behavior sequence recommendation method and system based on a graph neural network

Abstract

The application discloses a multi-behavior sequence recommendation method and system based on a graph neural network, relates to the technical field of recommendation systems, and comprises the following steps: constructing a global item graph by using multi-behavior sequence data, obtaining a user interaction graph based on the interaction of a user in the global item graph, constructing a user node feature matrix and an item node feature matrix in the user interaction graph, performing feature learning on the user interaction graph by using a graph neural network, updating the representation of nodes, and generating a personalized recommendation list by using a reinforcement learning algorithm and combining the feedback of the user based on the updated user interaction graph. The application can effectively solve the cold start problem and the problem of a decrease in recommendation accuracy in a traditional recommendation system by using the graph neural network and the reinforcement learning algorithm, provides more targeted and timely recommendation results for users, and thus enhances user experience and satisfaction.

Description

Technical Field

[0001] This invention relates to the field of recommender system technology, and in particular to a multi-behavior sequence recommender method and system based on graph neural networks. Background Technology

[0002] In recent years, with the development of the internet and big data technologies, recommender systems have become an indispensable component of digital platforms. Traditional recommender algorithms mainly rely on users' historical behavioral data, such as click, browsing, and purchase records, typically using collaborative filtering or content-based methods. However, these methods have certain limitations when dealing with diverse user behaviors and complex relationships. Furthermore, a shortcoming of existing technologies is that traditional recommender systems often ignore the multi-behavioral sequence features of users and the relationship networks between items. These features are crucial for accurately understanding user preferences. Specifically, existing methods often fail to effectively integrate multiple user behaviors (such as clicks, purchases, ratings, etc.) to construct a more comprehensive user profile. The user interaction graph lacks dynamism and cannot reflect changes in user preferences in real time, resulting in limited timeliness and accuracy of recommendation results. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, this invention provides a multi-behavior sequence recommendation method and system based on graph neural networks to solve the problem of recommendation accuracy caused by the diversity of user behavior and dynamic interaction changes.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] In a first aspect, embodiments of the present invention provide a multi-behavior sequence recommendation method based on graph neural networks, which includes: constructing a global item graph using multi-behavior sequence data; and obtaining a user interaction graph based on user interactions in the global item graph.

[0007] In the user interaction graph, a user node feature matrix and an item node feature matrix are constructed, and a graph neural network is used to learn features from the user interaction graph and update the node representations.

[0008] Based on the updated user interaction graph, a personalized recommendation list is generated using reinforcement learning algorithms and incorporating user feedback.

[0009] As a preferred embodiment of the multi-behavior sequence recommendation method based on graph neural networks described in this invention, the multi-behavior sequence data includes click count, purchase count, item rating, item category, user ID, and item ID.

[0010] As a preferred embodiment of the multi-behavior sequence recommendation method based on graph neural networks described in this invention, the method involves constructing a global item graph using multi-behavior sequence data, including the following steps.

[0011] Extract all item IDs from multi-line sequence data and use them as item nodes;

[0012] Define the relationships between items, extract item categories from the user's multi-behavior sequence data based on the defined relationships, and establish edges between item nodes;

[0013] Using the constructed item nodes and edges, an initial structure of a global item graph is created using a graph database. The number of times two items are purchased together is used as the weight of the edge, representing the frequency at which two items are purchased together. The similarity of the number of clicks is used to calculate the weight of the edge, representing the degree of user attention to the item.

[0014] Based on changes in user behavior, a dynamic update mechanism is designed to adjust the global item map in real time, and visualization tools are used to display the relationships between items, outputting the completed global item map.

[0015] As a preferred embodiment of the multi-behavior sequence recommendation method based on graph neural networks described in this invention, the method for obtaining a user interaction graph based on user interactions in a global item graph includes the following steps:

[0016] Extract all user IDs from multi-behavior sequence data as user nodes, and item IDs as item nodes. Create edges by combining user and item interaction behaviors, and construct a user interaction graph using a graph database.

[0017] By leveraging the number of new interactions between users and items, a dynamically adaptive graph update mechanism is designed to update edge weights, represented as follows:

[0018] w′=w+α×Δ;

[0019] Where w′ represents the updated edge weight, w represents the initial weight, α represents the adjustment factor, and Δ represents the number of new interactions between the user and the item;

[0020] Based on the edge weight update, the updated user interaction graph is obtained.

[0021] As a preferred embodiment of the multi-behavior sequence recommendation method based on graph neural networks described in this invention, the following steps are included in constructing the user node feature matrix and the item node feature matrix in the user interaction graph:

[0022] Based on the user interaction graph, count the number of historical interactions;

[0023] The most frequently interacted item categories by users are identified and represented using one-hot encoding to obtain preferred item categories.

[0024] Collect the timestamp of the user's most recent interaction to obtain the most recent interaction time;

[0025] By integrating historical interaction counts, preferred item categories, and recent interaction times, a user node feature matrix is ​​obtained;

[0026] Based on the user interaction graph, count the number of times an item is interacted with;

[0027] Obtain the item category from the global item graph, represent it using one-hot encoding, and then obtain the item attributes.

[0028] Calculate the average rating of all users for the same item;

[0029] The number of interactions, item attributes, and average ratings are integrated to obtain the item node feature matrix.

[0030] As a preferred embodiment of the multi-behavior sequence recommendation method based on graph neural networks described in this invention, the method involves: learning features from the user node feature matrix and the item node feature matrix using a graph neural network to update the node representations, including the following steps.

[0031] Define an adjacency matrix to represent the user interaction graph and perform normalization processing;

[0032] In graph convolutional networks, the update formulas for the feature matrices of user and item nodes are expressed as follows:

[0033]

[0034] Among them, H (l+1) H represents the feature matrix of user and item nodes in the (l+1)th layer. (l) W represents the feature matrix of user and item nodes in the l-th layer. (l) The weight matrix represents the learned function, and σ represents the activation function. Represents the normalized adjacency matrix;

[0035] The initial feature matrix is ​​composed of the feature matrices of user and item nodes;

[0036] By aggregating the feature matrices of users and items through multi-layer graph convolution operations, a new representation of each user node and item node is output.

[0037] As a preferred embodiment of the multi-behavior sequence recommendation method based on graph neural networks described in this invention, the method involves generating a personalized recommendation list based on the updated user interaction graph, using a reinforcement learning algorithm, and incorporating user feedback. This includes the following steps:

[0038] Based on the updated user interaction graph, the updated feature vectors of user and item nodes are obtained. The updated feature vectors of user nodes and the user behavior at the current time are used to form the state space. The updated feature vectors of item nodes are used to form a recommendation list to form the action space. Based on user feedback, a reward mechanism is set.

[0039] Using DQN, calculate the Q value of taking an action in a given state, and update the Q value using the Bellman equation;

[0040] The ε-greedy strategy is used to balance exploration and exploitation to determine action selection;

[0041] Based on the learned Q-value, the probability distribution of recommended items is output, and the recommendation list with the largest Q-value is selected. When generating the recommendation list, the user's real-time behavior data is taken into account, and the recommendation content is dynamically adjusted to generate a personalized recommendation list containing the ID and name of the recommended items.

[0042] Secondly, the present invention provides a multi-behavior sequence recommendation system based on graph neural networks, including a user interaction graph construction module, which is responsible for constructing a global item graph using multi-behavior sequence data and obtaining the user interaction graph based on the user's interaction in the global item graph;

[0043] The user interaction graph feature learning module is responsible for constructing user node feature matrices and item node feature matrices in the user interaction graph, and using graph neural networks to learn features from the user interaction graph and update the node representations.

[0044] The personalized recommendation list generation module is responsible for generating personalized recommendation lists based on the updated user interaction graph, using reinforcement learning algorithms, and incorporating user feedback.

[0045] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements any step of the multi-behavior sequence recommendation method based on graph neural networks as described in the first aspect of the present invention.

[0046] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the multi-behavior sequence recommendation method based on graph neural networks as described in the first aspect of the present invention.

[0047] The beneficial effects of this invention are as follows: By utilizing graph neural networks to learn features from user interaction graphs, node representations can better reflect the relationships between users and items. This method not only considers users' historical behavior but also introduces relationships between items, thereby capturing deeper interaction patterns during feature learning. Through reinforcement learning algorithms, combined with real-time user feedback, the system can dynamically adjust recommendation strategies to ensure that recommended content better meets users' real-time needs. Employing the DQN algorithm, the system achieves intelligent decision-making in both the state and action spaces, thereby improving the accuracy and personalization of recommendations. Ultimately, this invention effectively solves the cold start problem and the decline in recommendation accuracy in traditional recommendation systems, providing users with more targeted and timely recommendation results, thus enhancing user experience and satisfaction. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of the multi-behavior sequence recommendation method based on graph neural networks in Example 1.

[0050] Figure 2 This is a flowchart of the multi-behavior sequence recommendation system based on graph neural networks in Example 1. Detailed Implementation

[0051] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0052] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0053] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0054] Example 1, referring to Figure 1 and Figure 2This is the first embodiment of the present invention, which provides a multi-behavior sequence recommendation method based on graph neural networks, including the following steps:

[0055] S1. Construct a global item graph using multi-behavior sequence data. Based on user interactions within the global item graph, obtain a user interaction graph, including the following steps.

[0056] S1.1 Multi-row sequence data includes click count, purchase count, item rating, item category, user ID, and item ID.

[0057] It should be noted that by integrating multi-behavioral sequence data such as user click counts, purchase counts, ratings, item categories, user IDs, and item IDs, a comprehensive capture of user behavior is achieved. This data integration provides a richer feature base for subsequent recommendation systems, enabling the system to more accurately analyze user preferences and ultimately improve the accuracy and relevance of recommendations.

[0058] Preprocess the multi-behavior sequence data, including removing duplicate interaction records, standardizing the timestamp format, and ensuring the temporal order of all behavior records. Extract features from the original multi-behavior sequence data, including user ID, item ID, behavior type, and timestamp.

[0059] Extract all item IDs from the multi-behavior sequence data as item nodes. Each node represents all recommendable items, such as movies, books, and goods, ensuring the completeness of the item nodes and avoiding the omission of newly added items.

[0060] Define the relationships between items, including defining co-purchase relationships based on items purchased by users at the same time, defining co-click relationships based on items clicked by users within similar time periods, defining similar attribute relationships based on item categories, and defining rating relationships based on the similarity of user ratings of items. Based on the defined relationships, extract item categories from the user's multi-behavior sequence data and establish edges between item nodes.

[0061] Using the constructed item nodes and edges, an initial structure for a global item graph is created using a graph database (such as NetworkX). In the graph, each item node is connected to other related item nodes, forming a network. Each edge is assigned a weight to represent the strength of the relationship between items. The number of times items are purchased together is used as the weight of the edge to represent the frequency at which two items are purchased together. The similarity of the number of clicks is used to calculate the weight of the edge to represent the degree of user attention to the item.

[0062] Based on changes in user behavior, a dynamic update mechanism is designed to adjust the global item graph in real time. When a user interacts with a new item, the related co-purchase and co-click relationships are updated. The weights of the relationships between items are recalculated periodically to reflect the latest user behavior data. For items that have not been interacted with for a long time, their weight in the graph can be reduced to avoid affecting the recommendation effect. Visualization tools are used to display the relationships between items, and the completed global item graph is output.

[0063] It should be noted that a global item graph is constructed using multi-behavior sequence data. By extracting item IDs and defining relationships between items, the structure of item nodes and edges is established. This process enables visualization and dynamic monitoring of the relationships between items, allowing the recommendation system to better understand the mutual influence between items. Consequently, the system considers item relevance during recommendations, thereby improving user click-through rates and purchase rates.

[0064] S1.2 Extract all user IDs from the multi-behavior sequence data as user nodes and item IDs as item nodes. Create edges by combining the interaction behaviors between users and items, and construct a user interaction graph using a graph database (such as Neo4j, NeoTech).

[0065] Real-time monitoring of user interactions, such as new user actions on the platform (e.g., new clicks, new purchases), triggers a graph update process when a new interaction occurs. When a user frequently interacts with a certain type of item, the item's weight is increased. Based on the number of new interactions between users and items, a dynamically adaptive graph update mechanism is designed to update edge weights, represented as follows:

[0066] w′=w+α×Δ;

[0067] Where w′ represents the updated edge weight, w represents the initial weight, α represents the adjustment factor (which can be set to 0.1), and Δ represents the number of new interactions between the user and the item;

[0068] If a user interacts with a certain type of item more than a set threshold (e.g., 5 times), a new edge is automatically added to represent a strong preference relationship. Conversely, if the number of interactions is less than the threshold (e.g., 1 time), the edge will be considered for deletion. After all edge weight adjustments and edge addition / deletion operations, the graph structure is updated to ensure that the graph always reflects the latest user preferences. Based on the edge weight updates, the updated user interaction graph is obtained.

[0069] It should be noted that, based on the global item graph, user IDs and their interactions with items are extracted to construct a user interaction graph. A dynamic, adaptive graph update mechanism promptly reflects the latest user interactions. This mechanism ensures the timeliness of the user interaction graph, enabling rapid responses to changing user preferences, thereby enhancing user experience and improving the real-time nature and personalization of recommendations.

[0070] S2. In the user interaction graph, construct the user node feature matrix and the item node feature matrix, and use a graph neural network to learn features from the user interaction graph and update the node representations, including the following steps.

[0071] S2.1. Based on the user interaction graph, count the number of historical interactions, represented as follows:

[0072] I(u i )=∑ j I(u i ,p j );

[0073] Among them, I(u i ) represents user u i The total number of interactions with all items, u i Let p represent the i-th user node. j Let I(u) represent the j-th item node. i ,p j ) represents user u i With item p j The number of interactions between them, where i represents the user node index and j represents the item node index;

[0074] The most frequently interacted item categories by users are identified and represented using one-hot encoding to obtain preferred item categories.

[0075] Collect the timestamp of the user's most recent interaction, normalize it to a value between 0 and 1 to represent the user's activity level, and obtain the most recent interaction time;

[0076] By integrating historical interaction counts, preferred item categories, and recent interaction times, a user node feature matrix is ​​obtained;

[0077] Based on the user interaction graph, the number of times an item is interacted with is counted, represented as follows:

[0078] p(p j )=∑ i I(u i ,p j );

[0079] Where p(p j ) represents item p j Number of interactions;

[0080] Obtain the item category from the global item graph, represent it using one-hot encoding, and then obtain the item attributes.

[0081] The average rating of all users for the same item is represented as ,

[0082]

[0083] Among them, A(p j ) represents the average rating of all users for the same item, r(u i ,p j ) represents user u i For item p j The rating, R(p) j ) indicates that for item p j The number of users who provided ratings;

[0084] The number of interactions, item attributes, and average ratings are integrated to obtain the item node feature matrix.

[0085] It should be noted that, based on the user interaction graph, a feature matrix for users and items is constructed. By statistically analyzing information such as the number of historical interactions, preferred item categories, and recent interaction time, a comprehensive feature description of users and items is achieved. The construction of this feature matrix provides a foundation for subsequent graph neural network feature learning, ensuring that the system can fully utilize users' historical behavioral information during the learning process, thereby improving the accuracy of recommendations.

[0086] S2.2 Define an adjacency matrix to represent the user interaction graph and perform normalization processing, where interaction between users and items is represented by 1, and non-interacting nodes are represented by 0;

[0087] In graph convolutional networks, the update formulas for the feature matrices of user and item nodes are expressed as follows:

[0088]

[0089] Among them, H (l+1) H represents the feature matrix of user and item nodes at layer l+1, which is the updated representation of user and item nodes output. (l+1) The dimension is N×F (l+1) N represents the number of nodes, F (l+1) H represents the feature dimension of the (l+1)th layer. (l) Let F represent the feature matrix of user and item nodes in the l-th layer, where each row corresponds to a feature vector of a user and item node. The dimension of the feature matrix is ​​N×F. (l) F (l) W represents the feature dimension of this layer. (l) The weight matrix representing the learning process has dimension F. (l) ×F (l +1) This is used to map the features of the current layer to the feature space of the next layer. σ represents the ReLU (Rectified Linear Unit) activation function, expressed as σ(x) = max(0,x). This represents the normalized adjacency matrix, used to represent the connection relationships between nodes, ensuring that the weights of neighboring nodes are considered when aggregating features.

[0090] Setting the number of layers, usually 2 or 3, is to balance computational complexity and the expressive power of graph convolutional networks. The more layers, the more complex the feature relationships the model can capture, but the computational cost and the risk of overfitting will also increase.

[0091] The initial feature matrix consists of the feature matrices of user and item nodes, denoted as follows:

[0092]

[0093] Among them, H (0) Let U represent the initial feature matrix, U represent the user node feature matrix, and P represent the item node feature matrix.

[0094] By aggregating the feature matrices of users and items through multiple layers of graph convolution operations, a new representation of each user node and item node is output, represented as follows:

[0095]

[0096]

[0097] in, Let represent the updated feature vector of the i-th user node. H represents the updated feature vector of the j-th item node, m represents the number of user nodes, and H represents the number of user nodes. (L) H represents the node feature matrix output by the Lth layer of the graph convolutional network, containing the feature representations of all nodes after the Lth layer of graph convolution operations. [i] indicates the feature matrix H. (L) Extract the i-th row, representing the i-th user node. Since the indices of user nodes in the feature matrix are consecutive, they can be directly accessed via [i]. [m+j] represents the position of the user node in the feature matrix H. (L) Extract the (m+j)th row, which represents the j-th item node.

[0098] It should be noted that feature learning on the user interaction graph using a graph neural network enables the updating of user and item node representations. In this process, adjacency matrix normalization and graph convolution operations allow node features to effectively aggregate information from their neighbors. This feature learning process allows the model to capture richer user preferences and item characteristics, thereby improving recommendation relevance and user satisfaction.

[0099] S3. Based on the updated user interaction graph, using reinforcement learning algorithms and incorporating user feedback, generate a personalized recommendation list, including the following steps:

[0100] Based on the updated user interaction graph, the updated feature vectors of user and item nodes are obtained. The updated feature vectors of user nodes and the user behavior at the current time are used to form the state space. The updated feature vectors of item nodes are used to form a recommendation list to form the action space. Based on user feedback, a reward mechanism is set. If the user clicks or buys a recommended item, the reward is 1; otherwise, it is 0. A discount factor is used to calculate the present value of future rewards to enhance the learning of long-term behavior.

[0101] Using DQN (Deep Q-Network), the Q-value of the action taken in the given state is calculated, and the Q-value is updated using the Bellman equation, expressed as follows:

[0102]

[0103] Where Q(s) t ,a t ) indicates that in state s t Take action a t The current Q-value, α represents the learning rate, which is a hyperparameter between 0 and 1 that controls the degree to which new information influences the current Q-value, s t Representing the state space, a t Represents the action space, r t γ represents the immediate reward a user receives after interacting with the recommended item at time step t, and s represents the discount factor. t+1 This represents the next state after time step t. Indicates that in state s t+1 The maximum Q value among all possible actions 'a';

[0104] The ε-greedy strategy is used to balance exploration and exploitation to determine action selection;

[0105] Based on the learned Q-value, the probability distribution of recommended items is output, and the recommendation list with the largest Q-value is selected. When generating the recommendation list, the user's real-time behavior data (such as recent clicks, browsing history, etc.) is considered, and the recommendation content is dynamically adjusted to generate a personalized recommendation list containing the ID and name of the recommended items.

[0106] It should be noted that, based on the updated user interaction graph and combined with user feedback, a personalized recommendation list is generated using reinforcement learning algorithms. By applying DQN and the Bellman equation, the system can strike a balance between exploration and exploitation, adjusting the recommendation strategy in real time. This process achieves dynamic adaptation to user preferences, ensuring that recommended content accurately matches user needs, thereby increasing click-through rates and purchase rates, and ultimately enhancing business conversion.

[0107] This embodiment also provides a multi-behavior sequence recommendation system based on graph neural networks, including: a user interaction graph construction module, which is responsible for constructing a global item graph using multi-behavior sequence data, and obtaining the user interaction graph based on the user's interaction in the global item graph;

[0108] The user interaction graph feature learning module is responsible for constructing user node feature matrices and item node feature matrices in the user interaction graph, and using graph neural networks to learn features from the user interaction graph and update the node representations.

[0109] The personalized recommendation list generation module is responsible for generating personalized recommendation lists based on the updated user interaction graph, using reinforcement learning algorithms, and incorporating user feedback.

[0110] The user interaction graph construction module is responsible for constructing a global item graph using multi-behavior sequence data and obtaining the user interaction graph based on the user's interaction in the global item graph.

[0111] The user interaction graph feature learning module is responsible for constructing user node feature matrices and item node feature matrices in the user interaction graph, and using graph neural networks to learn features from the user interaction graph and update the node representations.

[0112] The personalized recommendation list generation module is responsible for generating personalized recommendation lists based on the updated user interaction graph, using reinforcement learning algorithms, and incorporating user feedback.

[0113] This embodiment also provides a computer device applicable to the multi-behavior sequence recommendation method based on graph neural networks, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the multi-behavior sequence recommendation method based on graph neural networks as proposed in the above embodiment.

[0114] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0115] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the multi-behavior sequence recommendation method based on graph neural networks as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0116] In summary, this invention utilizes graph neural networks to learn features from user interaction graphs, enabling node representations to better reflect the relationships between users and items. This method not only considers users' historical behavior but also incorporates relationships between items, thereby capturing deeper interaction patterns during feature learning. Through reinforcement learning algorithms combined with real-time user feedback, the system can dynamically adjust its recommendation strategy to ensure that recommended content better meets users' real-time needs. Employing the DQN algorithm, the system achieves intelligent decision-making in both the state and action spaces, thereby improving the accuracy and personalization of recommendations. Ultimately, this invention effectively solves the cold start problem and declining recommendation accuracy issues in traditional recommendation systems, providing users with more targeted and timely recommendation results, thus enhancing user experience and satisfaction.

Claims

1. A multi-behavior sequence recommendation method based on graph neural networks, characterized in that: include, Using multi-behavior sequence data, a global item graph is constructed, and based on user interactions within the global item graph, a user interaction graph is obtained. In the user interaction graph, constructing the user node feature matrix and the item node feature matrix includes the following steps. Based on the user interaction graph, count the number of historical interactions; The most frequently interacted item categories by users are identified and represented using one-hot encoding to obtain preferred item categories. Collect the timestamp of the user's most recent interaction to obtain the most recent interaction time; By integrating historical interaction counts, preferred item categories, and recent interaction times, a user node feature matrix is ​​obtained; Based on the user interaction graph, count the number of times an item is interacted with; Obtain the item category from the global item graph, represent it using one-hot encoding, and then obtain the item attributes. Calculate the average rating of all users for the same item; The number of interactions, item attributes, and average ratings are integrated to obtain the item node feature matrix; Graph neural networks are used to learn features from user interaction graphs and update node representations. Based on the updated user interaction graph, a personalized recommendation list is generated using reinforcement learning algorithms and incorporating user feedback, including the following steps: Based on the updated user interaction graph, the updated feature vectors of user and item nodes are obtained. The updated feature vectors of user nodes and the user behavior at the current time are used to form the state space. The updated feature vectors of item nodes are used to form a recommendation list to form the action space. Based on user feedback, a reward mechanism is set. Using DQN, calculate the Q value of taking an action in a given state, and update the Q value using the Bellman equation; The ε-greedy strategy is used to balance exploration and exploitation to determine action selection; Based on the learned Q-value, the probability distribution of recommended items is output, and the recommendation list with the largest Q-value is selected. When generating the recommendation list, the user's real-time behavior data is taken into account, and the recommendation content is dynamically adjusted to generate a personalized recommendation list containing the ID and name of the recommended items.

2. The multi-behavior sequence recommendation method based on graph neural networks as described in claim 1, characterized in that: The multi-behavior sequence data includes the number of clicks, the number of purchases, the rating of the item, the item category, the user ID, and the item ID.

3. The multi-behavior sequence recommendation method based on graph neural networks as described in claim 2, characterized in that: Constructing a global item graph using multi-row sequence data includes the following steps: Extract all item IDs from multi-line sequence data and use them as item nodes; Define the relationships between items, extract item categories from the user's multi-behavior sequence data based on the defined relationships, and establish edges between item nodes; Using the constructed item nodes and edges, an initial structure of a global item graph is created using a graph database. The number of times two items are purchased together is used as the weight of the edge, representing the frequency at which two items are purchased together. The similarity of the number of clicks is used to calculate the weight of the edge, representing the degree of user attention to the item. Based on changes in user behavior, a dynamic update mechanism is designed to adjust the global item map in real time, and visualization tools are used to display the relationships between items, outputting the completed global item map.

4. The multi-behavior sequence recommendation method based on graph neural networks as described in claim 3, characterized in that: Based on user interactions in the global item map, a user interaction map is obtained, including the following steps: Extract all user IDs from multi-behavior sequence data as user nodes, and item IDs as item nodes. Create edges by combining user and item interaction behaviors, and construct a user interaction graph using a graph database. By leveraging the number of new interactions between users and items, a dynamically adaptive graph update mechanism is designed to update edge weights, represented as follows: ； in, This represents the updated edge weight. Indicates the initial weights. Indicates the regulating factor. Indicates the number of new interactions between the user and the item; Based on the edge weight update, the updated user interaction graph is obtained.

5. The multi-behavior sequence recommendation method based on graph neural networks as described in claim 1, characterized in that: Based on the user node feature matrix and the item node feature matrix, a graph neural network is used to learn features from the user interaction graph and update the node representations. This includes the following steps: Define an adjacency matrix to represent the user interaction graph and perform normalization processing; In graph convolutional networks, the update formulas for the feature matrices of user and item nodes are expressed as follows: ； in, Indicates the first The feature matrix of user and item nodes in the layer, Indicates the first The feature matrix of user and item nodes in the layer, The weight matrix representing the learning process. This represents the activation function. Represents the normalized adjacency matrix; The initial feature matrix is ​​composed of the feature matrices of user and item nodes; By aggregating the feature matrices of users and items through multi-layer graph convolution operations, a new representation of each user node and item node is output.

6. A multi-behavior sequence recommendation system based on graph neural networks, based on the multi-behavior sequence recommendation method based on graph neural networks according to any one of claims 1 to 5, characterized in that: include, The user interaction graph construction module is responsible for constructing a global item graph using multi-behavior sequence data and obtaining the user interaction graph based on the user's interaction in the global item graph. The user interaction graph feature learning module is responsible for constructing user node feature matrices and item node feature matrices in the user interaction graph, and using graph neural networks to learn features from the user interaction graph and update the node representations. The personalized recommendation list generation module is responsible for generating personalized recommendation lists based on the updated user interaction graph, using reinforcement learning algorithms, and incorporating user feedback.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the multi-behavior sequence recommendation method based on graph neural networks as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the multi-behavior sequence recommendation method based on graph neural networks as described in any one of claims 1 to 5.

Images