Project recommendation method and device based on attribute-extended dual message propagation graph

Abstract

The present invention belongs to the field of recommendation technology, and specifically relates to a project recommendation method and device based on an attribute-expanded dual message propagation graph. The method includes obtaining user data, project data, and attribute data to construct a user-project-attribute tripartite graph structure; obtaining initial user, project, and attribute embedding vectors through random mapping technology; inputting them into a preference information graph neural network to obtain user and project preference embedding vectors; inputting them into a similarity information graph neural network to obtain user and project similarity embedding vectors; calculating the user's predicted attribute score for the project based on the inner product of the fused user final embedding vector and the project final embedding vector; inputting the predicted score into a user-project-attribute matrix, finding the project and attribute corresponding to the highest predicted score, forming a recommended explanation list, and taking the top-ranked attributes as explanations for the recommended project. The present invention improves the accuracy and interpretability of the recommendation system.

Description

Technical Field

[0001] The present invention belongs to the technical field of recommendation, and in particular relates to a project recommendation method and device based on a dual message propagation graph with attribute extension. Background Art

[0002] In an era of rapid information technology development, everyone generates a rich variety of information anytime, anywhere (photos shared, videos shot, comments left, videos watched, products purchased, etc.). Smart devices can also be used to send and receive a wide variety of information. In terms of generated information, this information mostly includes sharing life insights, product recommendations, news, educational resources, music, online friendships, and live video streaming. It covers everything from historical figures to China's magnificent landscapes. In terms of received information, it also includes a wide variety of film and television works, knowledge books from various countries, and a dazzling array of delicious foods. With such a wide variety of information integrated into a small, portable smart device and displayed to users, the problem of information explosion arises. Therefore, how to use collaborative filtering systems based on user data to provide users with novel and desirable product recommendations has become a hot topic of research.

[0003] Currently, applying Graph Neural Network (GNN) technology to the Collaborative Filtering Recommendation System (CFRS) can better improve the accuracy and interpretability of recommendations. By introducing GNN technology, the interaction behavior between users and items can be better explicitly modeled, and the high-order relationships hidden in the interaction behavior can be captured. This method of modeling high-order relationships will undoubtedly greatly improve the recommendation performance.

[0004] However, the impact of the message propagation mechanism on recommendation performance in GNN-based collaborative filtering recommendation systems has received limited attention, and existing technologies have not considered the crucial issue of interpretability in recommendation performance. For item recommendation systems, high interpretability of item recommendation results is crucial for improving user experience. Interpretability is often considered a key indicator for evaluating user experience and holds significant research value in industrial applications. Summary of the Invention

[0005] This paper combines graph neural network technology with a dual-message propagation mechanism to propose a dual-message propagation graph collaborative filtering recommendation algorithm based on attribute extension. This algorithm aims to improve the interpretability of recommendations by capturing preference and similarity messages in the user-item-attribute tripartite graph through the dual-message propagation mechanism. First, a large amount of irregular data in the dataset is strictly classified and filtered under limited concepts, and the results are input as attributes into the user / item interaction graph. Then, based on the dual-message propagation mechanism, a collaborative filtering recommendation algorithm based on attribute extension is proposed. After obtaining the scores of all user / item / attribute pairs, the algorithm provides a final recommendation list and reasonable explanations for the recommendations.

[0006] In a first aspect of the present invention, the present invention provides an item recommendation method based on a dual message propagation graph with attribute extension, the method comprising:

[0007] Obtain user data, project data, and attribute data, and construct a user-project-attribute tripartite graph structure;

[0008] The user-item-attribute tripartite graph structure is processed by random mapping technology to obtain the user initial embedding vector, the item initial embedding vector and the attribute initial embedding vector;

[0009] Inputting the user initial embedding vector, the item initial embedding vector, and the attribute initial embedding vector into the preference information graph neural network, obtaining the user preference embedding vector through encoding aggregation of high-order relations, and obtaining the item preference embedding vector through encoding aggregation based on attribute extension;

[0010] Inputting the user initial embedding vector, the item initial embedding vector, and the attribute initial embedding vector into the similarity information graph neural network, obtaining the user similarity embedding vector through encoding aggregation based on high-order relations; and obtaining the item similarity embedding vector through encoding aggregation based on attribute extension;

[0011] The user preference embedding vector and the user similarity embedding vector are fused to generate the user final embedding vector; the item preference embedding vector and the item similarity embedding vector are fused to generate the item final embedding vector;

[0012] The predicted attribute score of the user for the item is calculated based on the inner product of the final embedding vector of the user and the final embedding vector of the item;

[0013] The predicted score is input into the user-item-attribute matrix, and the item and attribute corresponding to the highest predicted score are found from the user-item-attribute matrix. A list of recommended explanations is formed in descending order, and the top-ranked attributes are taken as the explanations of the recommended item.

[0014] In the second aspect of the present invention, the present invention also provides a project recommendation device based on an attribute-extended dual message propagation graph, comprising: at least one processor; and at least one memory storing a computer program; wherein, when the computer program is executed by the at least one processor, the project recommendation device performs the steps of a project recommendation method based on an attribute-extended dual message propagation graph according to the first aspect of the present invention.

[0015] Advantages and beneficial effects of the present invention:

[0016] This invention proposes a project recommendation method and device based on an attribute-expanded dual message propagation graph, suitable for recommending project domain scenarios to users. The invention employs a dual message propagation mechanism to capture high-order similarity message propagation patterns and utilizes two graph neural networks to separately process preference messages and similarity messages, which are prevalent in user-item interaction networks. This simplifies the traditional recommendation process of capturing similarity messages by capturing high-order preference relationships. The method also provides reasons for the recommendations along with the list of recommendations, ensuring both the effectiveness of the recommendations and significantly improving their interpretability. In terms of data processing and representation, the invention utilizes data processing techniques to analyze and classify project attributes in a database and employs a graph neural network to explicitly represent user-item interactions, user-neighbor similarity, and project-to-similar similarity. This enhances the explicit representation of the data, improves the model's interpretability, and explicitly demonstrates the rationale of the recommendations. Two graph neural networks are used to handle the two types of message propagation: one for preference message propagation and the other for similar message propagation. This simplifies the model's processing flow, makes the recommendation model more intuitive, ensures the effectiveness of the recommendations, and improves the accuracy of the recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a project recommendation method based on a dual message propagation graph with attribute extension according to an embodiment of the present invention;

[0018] Figure 2 A user-item-attribute tripartite graph structure according to an embodiment of the present invention;

[0019] Figure 3 A user-item-attribute tripartite graph structure according to a preferred embodiment of the present invention;

[0020] Figure 4 The following is a schematic diagram of a recommended explanation process in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] Figure 1 This is a flow chart of a project recommendation method based on a dual message propagation graph with attribute extension according to an embodiment of the present invention; Figure 1 As shown, the method includes:

[0023] 101. Obtain user data, project data, and attribute data, and construct a user-project-attribute tripartite graph structure;

[0024] In an embodiment of the present invention, relevant user data, item data, and attribute data can be obtained from an existing recommendation system database. These data may include the interaction relationship between users and items, and the interaction relationship between items and attributes. For example, it may be that a user watched a certain movie, or dined at a certain restaurant, etc. These user-item interaction behaviors are all generated under certain intentions, and these intentions represent attribute information such as location, survival needs, and character identity.

[0025] Based on the user data, project data and attribute data collected above, user data is mapped to user nodes, project data is mapped to project nodes, and attribute data is mapped to attribute nodes; according to the interactive relationship between user nodes and project nodes, and between project nodes and attribute nodes, a user-project-attribute tripartite graph structure is constructed.

[0026] In an embodiment of the present invention, it is assumed that in a recommendation system, user data, project data, and attribute data are obtained from the recommendation system database, and users, projects, and attributes are mapped to nodes respectively. There is a certain interaction between user nodes, project nodes, and attribute nodes. In daily life, everyone has different interests. It is precisely because of these interests that they interact with a wide variety of goods and items. This data contains user-project interaction information and project-attribute correspondences, which correspond to five project nodes (i1, i2, i3, i4, i5), four user nodes (u1, u2, u3, u4), and five attribute nodes (a1, a2, a3, a4, a5). Figure 2As shown in (a), there are links between four users and five items. These links are explicitly represented as user-item interactions. By explicitly representing these links, we can obtain a user-item interaction graph and an item-attribute bipartite graph. By integrating these interaction graphs and bipartite graphs, we can obtain a user-item-attribute tripartite graph. These interactions in the item-attribute bipartite graph and the user-item-attribute tripartite graph can be understood as users having an undiscovered interest in the items. How to analyze these interactions and fully explore the deeper intentions hidden in these interactions is a pressing issue for every industrial recommender system. Each item also has its own unique attributes, such as type, color, and purpose. Strictly speaking, every specific object has a limited number of attributes. For example, the various attributes of a movie are determined after its release, and there is no need for in-depth exploration. The attributes of this movie may attract specific user groups, such as some user groups who like science fiction films, some who like animation, and some who like romance films. The interactions between users and items are inevitably driven by underlying reasons. These interactions arise because the items possess certain attributes that attract these users, and because these users are interested in these attributes. Industrial recommendation systems utilize vast amounts of user / item interaction data. Based on this vast amount of interaction data, a user profile can be generated for each user. This profile can then be used to predict ratings and make recommendations.

[0027] 102. Process the user-item-attribute tripartite graph structure through random mapping technology to obtain the user initial embedding vector, the item initial embedding vector and the attribute initial embedding vector;

[0028] In the embodiment of the present invention, the interaction between users and items is represented in the user-item-attribute tripartite graph as edges connecting user nodes and item nodes, and the correspondence between items and attributes is also represented in the form of edges connecting item nodes and attribute nodes. Therefore, the initial embedding vector is reflected in the form of edges at the initial moment. The present invention treats the edges in the graph as channels for message propagation, which can update these embedding vectors, thereby realizing item recommendations based on the updated embedding vectors. Figure 3As shown in (a), user u1 obtains attribute messages from similar items i3 and i4 of items i1 and i2. In real life, this message propagation pattern is common. When a user is interested in a certain type of movie, he or she will also be interested in movies similar to this type of movie. This is because similar movies also contain elements that the user is interested in, namely attributes. In order to obtain multi-hop messages from the graph, such as second-order similarity messages, it is generally obtained through the link of user u2 → item i3 → user u3 → item i1 → user u1, as shown in Figure 3 (b) shows that the preference information between items and attributes is directly obtained in the form of attribute a3→item i3, attribute a4→item i4, as shown in Figure 3 As shown in (c); based on the basic single-hop message propagation model, high-order preference messages are propagated through even-hop links, while high-order similarity messages are propagated through odd-hop links.

[0029] In this embodiment of the present invention, assuming that there are M users in the recommendation system database, user u can be encoded as an M-dimensional vector O u =[0,…,1,…,0] T , where only the element at the position where user u is located is 1, while the elements at other positions are all 0. At the same time, since there are usually a large number of users, items and other additional data in the recommendation system, the number of users is very large. In order to reduce O u The dimensions of can be reduced by random mapping techniques to obtain low-dimensional and dense vectors, and these low-dimensional vectors are used as initial embedding inputs to the recommendation model. For example, the mapping process of user u is expressed as:

[0030]

[0031] In the formula, the value of d is much smaller than M, e u Denote the initial embedding vector of user u. The same method can be used to obtain the initial embedding vector e of item i. i The mapping process of project node i is expressed as:

[0032]

[0033] Where item i can be encoded as an N-dimensional vector O i , the value of d is much smaller than N,e i is represented as the initial embedding vector of item i.

[0034] Assume that all project attributes are divided into C categories, and each category is subdivided into countable V C The initial embedding of the attributes is represented as:

[0035]

[0036] Where, (C,V C ) represents C and V C An operation that generally performs a sum operation on all types of attributes. In the process of analyzing the attributes of many items, V C The size of should be limited and reasonable, because having reasonable classification criteria is an important basis for the rigor of the data mining process.

[0037] 103. Input the user initial embedding vector, the item initial embedding vector, and the attribute initial embedding vector into the preference information graph neural network, obtain the user preference embedding vector through encoding aggregation of high-order relations, and obtain the item preference embedding vector through encoding aggregation based on attribute extension;

[0038] In this embodiment of the present invention, the preference message processing process is divided into two stages: encoding and aggregation. In the encoding stage, the preference messages received by the user and the project are encoded. In the aggregation stage, the encoded preference messages are integrated together using an aggregation function to represent the preference embedding of the user and the project.

[0039] In an embodiment of the present invention, the user preference embedding vector obtained by encoding and aggregating high-order relationships includes encoding according to similar embedding vectors of project nodes at each layer, preference message relationship functions obtained by user nodes from attribute nodes via paths, and unit matrices to obtain preference information propagated by user nodes from attribute nodes; and aggregating the preference information propagated by user nodes from attribute nodes to obtain the user preference embedding vector.

[0040] In this embodiment of the present invention, the calculation formula of the user preference embedding vector includes:

[0041] Through high-order message propagation, user u can obtain the preference message sent by attribute a through the u←i←a message propagation path:

[0042]

[0043] Where, Represents the preference information of user node u propagated from attribute node a in the l-1 layer via the path u←i←a, W1∈R d Indicates e i The weight matrix for feature extraction can be directly set as the unit matrix here. This matrix is ​​the matrix for the path user node u←item node i←attribute node a; Represents the similarity embedding vector of item node i at the l-1 layer. In the first layer, the similarity embedding vector of the item node has not been calculated yet, so the preference embedding vector of the user node at the first order can be directly calculated by Calculate, e i represents the initial embedding vector of item node i; Represents the preference message relationship function obtained by the user node from the attribute node a via the path u←i←a; when obtaining high-order relationship embedding, one thing to note is that is defined as follows:

[0044]

[0045] Where, It represents the proportion of preference messages obtained by the user from attribute a via the path u←i←a. u | represents the number of items interacted by user u, if |I u The larger the value of |, the smaller the proportion of messages received by the user from project i. The smaller the value; |U i | represents the number of users who have interacted with project i. If |U i The larger | is, the less likely it is that project i is attractive to users. The smaller the value; |Attr i | represents the number of attributes of item i. If |Attr i The larger the value of |, the less attribute a can reflect the characteristics of project i. The smaller the value; |I a | represents the number of items with attribute a. If |I a The larger the value of |, the more widespread the attribute a is in the project. It cannot be considered a distinctive attribute and users are unlikely to be attracted by this attribute and interact with the project. The smaller the value. During the iteration process, the user's l-th order preference embedding can be obtained:

[0046]

[0047] Where, represents the l-order preference embedding of user u, that is, the preference embedding vector of user node u at the l-th layer. With the project node as the bridge, the attribute message obtained by user u from the high-order message propagation path, σ(·) is the activation function. Figure 3 As shown in (c), user u1 obtains attribute messages from similar items i3 and i4 of items i1 and i2. In real life, this message propagation pattern is common. When a user is interested in a certain type of movie, he or she will also be interested in movies similar to this type of movie. This is because similar movies also contain elements that the user is interested in, namely attributes.

[0048] After obtaining the user's L-order preference embedding and L-order similarity embedding, the user's high-order embedding representation can be obtained:

[0049]

[0050] The preference embedding of user u and the similarity embedding of user u are aggregated to form the final preference embedding and final similarity embedding of user u:

[0051]

[0052] Where, represents the preference embedding vector of user node u; α l It represents the first weight factor that controls the importance of the l-th order relationship, and its value can be set to 1 / L; L represents the number of layers of high-order relationships.

[0053] In the embodiment of the present invention, high-order similar messages only act on the dual-message propagation stage, and the first-order similar messages, second-order similar messages, ..., N-order similar messages of the target user are obtained in sequence from low order to high order. Similarly, the preferred messages are also obtained in sequence from low order to high order.

[0054] In an embodiment of the present invention, the project preference embedding vector obtained by encoding aggregation based on attribute extension includes encoding according to the classification matrix obtained after the project node is classified, the weight matrix for feature extraction of the classification matrix, and the relationship function of the attribute node propagated to the project node to obtain the preference information propagated between the project node and the attribute nodes it has; aggregating the preference information propagated between the project node and the attribute nodes it has to obtain the project preference embedding vector.

[0055] In this embodiment of the present invention, the calculation formula of the item preference embedding vector includes:

[0056] In the encoding phase, the preference message propagated between a given item i and its attribute a is encoded as follows:

[0057]

[0058] Where, represents the intrinsic attribute message propagated from attribute a to item i, indicating the amount of message item i has on attribute a; It is a Matrix i,a The weight matrix for feature extraction is the matrix for the propagation path from project node i to attribute node a; Matrix i,a is the matrix obtained after classification of item i; Represents the relationship function that propagates the attribute node a to the item node i.

[0059] Since the proportions of attribute objects in projects are different, the attribute weights still rely on manual design and have little impact on the interpretability of the recommendation. Therefore, in this embodiment of the present invention, W3 can be set to the unit matrix. Expressed as Where |Attr i | represents the number of attributes that project i has in the attribute matrix. The more attributes a project contains, the wider the audience of this project is, or at least it is not unpopular. Then the proportion of this attribute in this project is smaller, that is, The smaller the value; if the project contains fewer attributes, it means that the project is targeted at a certain type of user, then the proportion of messages of this attribute in this project will be greater, that is, The larger the value, the greater the value of |I a | represents the number of items with attribute a in the dataset. If |I a The larger the value, the more common the attribute a is in the project. The smaller; if |I a The smaller the value, the greater the proportion of messages it accounts for. The bigger.

[0060] In the aggregation phase, after project i encodes all received preference messages, these encoded preference messages are aggregated together to form the attribute distribution inherent to project i:

[0061]

[0062] Where p i,Attr represents the attribute information inherent in item i, that is, the preference embedding vector of item node i, and σ(·) is the activation function. Since there is no similarity between attribute categories and they are divided completely according to different standards, σ(·) is set to the unit function here.

[0063] 104. Input the user initial embedding vector, the item initial embedding vector, and the attribute initial embedding vector into a similar information graph neural network, obtain a user similarity embedding vector through encoding aggregation based on high-order relations, and obtain a item similarity embedding vector through encoding aggregation based on attribute extension;

[0064] In an embodiment of the present invention, the method of obtaining a user similarity embedding vector by encoding and aggregating based on high-order relations includes encoding according to a relationship function propagated from a user node to other user nodes, similarity embedding vectors of user nodes at each layer, and a transformation matrix to obtain similar messages transmitted from a user node to other user nodes; and aggregating similar messages transmitted from a user node to other user nodes to obtain a user similarity embedding vector.

[0065] In this embodiment of the present invention, the method for calculating the user similarity embedding vector includes:

[0066] Similar to the process of processing preference messages, the process of processing similar messages also consists of two stages: encoding and aggregation. In the encoding stage, taking a given similar link (u, v) as an example, the similar message transmitted from user v to user u is encoded at layer l as:

[0067]

[0068] Where, represents the similarity information propagated between user node i and user node v at layer l, W2 is the transformation matrix; Indicates the similar embedding vector of user node v in the l-1 layer. In the first layer, the similar embedding vector of the user node has not been calculated at this time, so the similar embedding vector of the user node in the first order can be directly calculated by Calculate, e v Represents the initial embedding vector of user node v. Unlike the preference message encoding process, W2 does not need to consider the different contributions of the message sources, because the link process between similar neighbor users cannot reflect the importance of different users. Therefore, W2 is set to the identity matrix, which is the matrix for the propagation path from user node u to user node v and from item node i to item node j. The same as the preference message encoding process, Controls the similarity between user u and user v:

[0069]

[0070] According to life experience, when the number of items that user u and user v interact with each other is greater (i.e., I u ∩I v The more ), the more similar the interests of two users are (i.e. Of course, considering the popularity of the project, the more popular an item is, the less it reflects the similarity between the two users (i.e. |U i |The bigger, smaller), because in this case it is difficult to tell whether they are interacting with these projects because of similar interests or because these projects are too popular. Reflects the activity level of user u and user v in the dataset, |I u | represents the number of interactive items generated by user u, |I v | represents the number of interactive items generated by user v. When user u and user v are more active, the corresponding |I u |and|I vThe larger the value of | becomes, the more likely there is a coincidence in their interactions with certain items, which in turn reduces the persuasiveness. The smaller the value, the more likely they are not interacting with these items out of similar interests.

[0071] In the aggregation stage, the encoded similar messages are integrated and represented as the target user’s l-th order similarity embedding through the aggregation function:

[0072]

[0073] Where, N represents the similarity embedding vector of user node u at layer l, which describes the common features of user u and its neighbors at layer l. u Denotes other users with similar links to user u. Similar to the preference message processing, the activation function σ(·) is set to the unit function.

[0074] After obtaining the user's L-order preference embedding and L-order similarity embedding:

[0075]

[0076] The preference embedding and similarity embedding are aggregated separately to form the final similarity embedding of user u:

[0077]

[0078] in, represents the similarity embedding vector of user node i, β l It represents the second weight factor that controls the importance of the l-th order relationship, and its value can be set to 1 / L; L represents the number of layers of high-order relationships.

[0079] In the present invention, high-order similar messages only act on the dual-message propagation stage, and the first-order similar messages, second-order similar messages, ..., N-order similar messages of the target user are obtained in sequence from low order to high order. Similarly, the preferred messages are also obtained in sequence from low order to high order.

[0080] In an embodiment of the present invention, the method of obtaining a project similarity embedding vector by encoding aggregation based on attribute extension includes encoding according to a relationship function between project nodes, an initial project embedding vector, and a transformation matrix to obtain similar information propagated between project nodes; and aggregating the similarity information propagated between project nodes to obtain a project similarity embedding vector.

[0081] In this embodiment of the present invention, the method for calculating the item similarity embedding vector includes:

[0082] Like the process of processing preference information, the process of processing similarity information between items also consists of two stages: encoding stage and aggregation stage.

[0083] In the encoding phase, taking a given similar link (i, j) as an example, similar to the calculation process of the user similarity embedding vector, the similarity message propagated from item i to item j is encoded as:

[0084]

[0085] Among them, e j represents the initial embedding vector of item j, is defined as follows:

[0086]

[0087] When judging the similarity between items, based on daily experience, we usually judge the type of their attributes. The more attributes the two items have in common (i.e., a∈Attr i ∩Attr j The more ), the more similar the two items are (i.e. Considering that some common attributes cannot express the particularity of the item, such as color, shape, etc., that is, |I a |The bigger, The smaller it is, the harder it is to use attribute a to separate different items, because attribute a exists in almost all items. i ||Attr j | reflects the popularity of item i and item j in the data, |Attr i | represents the number of attributes that item i has, |Attr j | represents the number of attributes of item j. The more frequently items i and j appear in the dataset, the more they are needed by users, such as some basic daily necessities. This will reduce the accuracy of the recommendation. The smaller the value.

[0088] In the aggregation phase, the encoded similar messages are integrated to represent the similar embedding of the target item through the aggregation function:

[0089]

[0090] Where s ia N represents the similarity embedding of item i, which describes the features shared by item i and its similar items. i represents other items that have similar links to item i. Similar to the preference message processing, the activation function σ(·) is set to the unit function.

[0091] 105. Fuse the user preference embedding vector and the user similarity embedding vector to generate the user final embedding vector; fuse the item preference embedding vector and the item similarity embedding vector to generate the item final embedding vector;

[0092] In this embodiment of the present invention, after multiple layers of dual message propagation, each user (or item) will obtain a final preference embedding representation and a final similarity embedding representation. These two embedding representations are combined to obtain the final embedding of the user (or item):

[0093]

[0094]

[0095] 106. Calculate the predicted attribute score of the user for the item based on the inner product of the final embedding vector of the user and the final embedding vector of the item;

[0096] Given a user-item-attribute pair (u,i,a), its preference score is obtained by the inner product of the final embedding representation of user u and item i, as follows:

[0097]

[0098] Where, represents the predicted rating of user u on item i based on attribute a. represents the general preference of user u on attribute a, represents the common characteristics of project i and its similar projects, then It can be expressed as how the characteristics of item i on attribute a match the preferences of user u; represents the common features of user u and similar users, Indicates which users item i tends to attract based on attribute a, then It can be expressed as the degree to which user u is attracted by the attribute a of item i. The predicted score of the user-item-attribute pair (u,i,a) can be comprehensively measured.

[0099] 107. Input the predicted score into the user-item-attribute matrix, find the item and attribute corresponding to the highest predicted score from the user-item-attribute matrix, form a list of recommended explanations in descending order, and take the top-ranked attributes as the explanations of the recommended item.

[0100] like Figure 4As shown in the figure, for the target user, after all the user-item-attribute pairs are predicted and scored in the score prediction stage, the prediction results are saved in the user-item-attribute matrix of the target user, and then all the item-attribute pairs are sorted in descending order, and the items corresponding to the top-ranked predicted scores are selected to form a recommendation list and displayed to the target user, and the attributes corresponding to the recommended items are also displayed to the target user as the recommendation explanation, thus realizing the recommendation process and the explanation process.

[0101] When giving a list of recommended items, the basis for giving the recommendation explanation is the list of several attributes with the highest attribute scores in the recommended items, as shown in Table 1.

[0102] Table 1 User-item-attribute pair (u,i,a) matrix calculated by the prediction scoring function

[0103]

[0104] After obtaining the predicted scores for the user-item-attribute pair using the above formula, the results are entered into the matrix shown in Table 1. From the (u, i, a) matrix, it's clear that the highest predicted score, 4.2, is located in row I5 and column A3. Similarly, the second-highest predicted score, 3.4, is located in row I3 and column A2. By sorting all the ratings in the user / item / attribute matrix from highest to lowest, a list of recommended explanations is formed, and the top-ranked attributes are used as explanations for the recommended item. This way, when the system makes a recommendation, we can understand why the item was recommended to the user. For example, the highest score of 4.2 in Table 1 is the first in the final recommendation list, indicating that attribute a in item i best matches user u's preferences, meaning that user u is most interested in item i containing attribute a. This approach provides the rationale for the recommendation along with the system's recommendations, directly improving the interpretability of the recommendation system.

[0105] In an embodiment of the present invention, the present invention also provides a project recommendation device based on an attribute-extended dual message propagation graph, comprising: at least one processor; and at least one memory storing a computer program; wherein, when the computer program is executed by the at least one processor, the project recommendation device performs the steps of a project recommendation method based on an attribute-extended dual message propagation graph according to the first aspect of the present invention.

[0106] Compared with traditional recommendation systems, the project recommendation method proposed in this invention, which is based on a dual message propagation graph with attribute extension, optimizes the computing performance in the graph neural network while ensuring the superiority of recommendation performance, can obtain better recommendation accuracy, and gives reasons for the recommendation while recommending the project, thereby improving the interpretability of the recommendation and thus improving the user's product experience in the recommendation system.

[0107] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium, which may include: ROM, RAM, disk or CD, etc.

[0108] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A project recommendation method based on attribute-extended dual message propagation graph, characterized in that: The method comprises: Obtain user data, project data, and attribute data, and construct a user-project-attribute tripartite graph structure; The user-item-attribute tripartite graph structure is processed by random mapping technology to obtain the user initial embedding vector, the item initial embedding vector and the attribute initial embedding vector; Inputting the user initial embedding vector, the item initial embedding vector, and the attribute initial embedding vector into the preference information graph neural network, obtaining the user preference embedding vector through encoding aggregation of high-order relations, and obtaining the item preference embedding vector through encoding aggregation based on attribute extension; Inputting the user initial embedding vector, the item initial embedding vector, and the attribute initial embedding vector into the similarity information graph neural network, obtaining the user similarity embedding vector through encoding aggregation based on high-order relations; obtaining the item similarity embedding vector through encoding aggregation based on attribute extension; The user preference embedding vector and the user similarity embedding vector are fused to generate the user final embedding vector; the item preference embedding vector and the item similarity embedding vector are fused to generate the item final embedding vector; The predicted attribute score of the user for the item is calculated based on the inner product of the final embedding vector of the user and the final embedding vector of the item; The predicted score is input into the user-item-attribute matrix, and the item and attribute corresponding to the highest predicted score are found from the user-item-attribute matrix. A list of recommended explanations is formed in descending order, and the top-ranked attributes are taken as the explanations of the recommended item.

2. The project recommendation method based on attribute-expanded dual message propagation graph according to claim 1, characterized in that: The method of obtaining the user preference embedding vector by encoding and aggregating high-order relationships includes encoding the preference information propagated by the user node from the attribute node via the path according to the similarity embedding vectors of the project node at each layer, the preference message relationship function obtained by the user node from the attribute node via the path, and the identity matrix; The preference information of user nodes propagated from attribute nodes is aggregated to obtain the user preference embedding vector.

3. The project recommendation method based on attribute-expanded dual message propagation graph according to claim 2, characterized in that: The calculation formula of the user preference embedding vector includes: Where, represents the preference embedding vector of user node u; α l represents the first weight factor that controls the importance of the l-th order relationship; L represents the number of layers of high-order relationships; represents the preference embedding vector of user node u at layer l; σ(·) is the activation function, represents the preference information of user node u propagated from attribute node a in the l-1 layer via the path u←i←a; W1 represents the identity matrix; Represents the similarity embedding vector of item node i at layer l-1; Represents the preference message relationship function that the user node obtains from the attribute node a via the path u←i←a; |I u | represents the number of project nodes interacted by user node u, |U i | represents the number of user nodes that have interacted with project node i, |Attr i | represents the number of attributes that item node i has in the attribute matrix; |I a |Indicates the number of project nodes with attribute node a.

4. The project recommendation method based on attribute-extended dual message propagation graph according to claim 1, characterized in that: The method of obtaining the item preference embedding vector by encoding aggregation based on attribute extension includes encoding the classification matrix obtained after the item node is classified, the weight matrix for feature extraction of the classification matrix, and the relationship function of the attribute node propagating to the item node, thereby obtaining the preference information propagated between the item node and the attribute node it possesses; The preference information propagated between the project node and its attribute nodes is aggregated to obtain the project preference embedding vector.

5. The project recommendation method based on attribute-extended dual message propagation graph according to claim 4, characterized in that: The calculation formula of the item preference embedding vector includes: Where p i,Attr represents the preference embedding vector of item node i, σ(·) is the activation function, Attr i Represents the attribute set of project node i in the attribute matrix, It represents the preference information propagated between project node i and the attribute node a it possesses, indicating the amount of messages project node i has on attribute node a; Is the matrix Matrix i,a Weight matrix for feature extraction; Matrix i,a is the classification matrix obtained after classification of project node i; Represents the relationship function of propagating attribute node a to item node i, |Attr i | represents the number of attributes that item node i has in the attribute matrix; |I a |Indicates the number of project nodes with attribute node a.

6. The project recommendation method based on attribute-expanded dual message propagation graph according to claim 1, characterized in that: The method of obtaining the user similarity embedding vector by encoding aggregation based on high-order relations includes encoding according to the relationship function propagated from the user node to other user nodes, the similarity embedding vector of the user node at each layer, and the transformation matrix to obtain similar messages transmitted from the user node to other user nodes; Similar messages transmitted from a user node to other user nodes are aggregated to obtain user similarity embedding vectors.

7. The project recommendation method based on attribute-expanded dual message propagation graph according to claim 6, characterized in that: The calculation method of the user similarity embedding vector includes: Where, represents the similarity embedding vector of user node i, β l represents the second weight factor that controls the importance of the l-th order relationship, represents the similarity embedding vector of user node u in layer l, σ(·) is the activation function, Attr i Represents the attribute set of project node i in the attribute matrix, represents the similar information propagated between user node i and user node v at layer l; W2 is the transformation matrix; Represents the similarity embedding vector of user node v at layer l-1; Represents the relationship function of user node u propagating to user node v; |I u | represents the number of project nodes that user node u interacts with, |I v | represents the number of project nodes interacted by user node v, |U i | represents the number of user nodes that have interacted with project node i.

8. The project recommendation method based on attribute-expanded dual message propagation graph according to claim 1, characterized in that: The obtaining of the project similarity embedding vector by encoding aggregation based on attribute extension includes encoding according to the relationship function between project nodes, the project initial embedding vector and the transformation matrix to obtain similarity information propagated between project nodes; The similar information propagated between project nodes is aggregated to obtain the project similarity embedding vector.

9. The project recommendation method based on attribute-expanded dual message propagation graph according to claim 8, characterized in that: The project similarity embedding vector is calculated as follows: Among them, s ia represents the similarity embedding vector of item node i, σ(·) is the activation function, N represents the similarity message propagated from item node i to item node j, i represents the set of neighboring project nodes of project node i, Represents the relationship function between project node i and project node j, e j represents the initial embedding vector of the item, W2 represents the change matrix; |Attr i | represents the number of attributes of item node i, |Attr j | represents the number of attributes of item node j, a represents an attribute node, |I a | represents the number of attribute nodes a.

10. A project recommendation device based on a dual message propagation graph with attribute extension, comprising: at least one processor; and at least one memory storing a computer program; wherein, when the computer program is executed by the at least one processor, the project recommendation device performs the steps of a project recommendation method based on a dual message propagation graph with attribute extension according to any one of claims 1 to 9.

Images