Personalized archive recommendation method based on graph contrast learning

Abstract

The invention relates to a personalized archive recommendation method based on graph contrast learning, and belongs to the technical field of information retrieval and recommendation systems. According to the method, the interaction between the user and the archive resources, the social relationship of the user and the multi-source heterogeneous information in the characteristics of the archive are fused, and the graph comparison learning technology is combined to learn the personalized archive recommendation strategy about the user, so that the defects of the existing archive recommendation method are overcome, and more accurate and personalized services are provided. According to the method, the complex relationship between the user preference and the archive characteristics can be more accurately captured, so that a recommendation result which better meets the user demand is provided.

Description

Technical Field

[0001] The present invention belongs to the technical field of information retrieval and recommendation systems, and in particular relates to a personalized archive recommendation method based on graph comparative learning. Background Art

[0002] In the digital age, massive amounts of data have exploded into people's lives in various forms. At the same time, a large amount of information resources have accumulated in various archive management systems (such as e-government platforms, corporate knowledge bases, etc.). In order to improve the efficiency of users in obtaining the required information during the process of searching and using archives, personalized archive recommendations have become one of the key means to improve user experience.

[0003] The accuracy of recommendation results in the real world often stems from the interaction between users and items, such as purchase, evaluation, forwarding, collection and other complex behaviors. Traditional recommendation methods are difficult to effectively capture these highly complex connections. Since graph structures can depict complex interactive relationships, in recent years, methods based on graph neural networks (GNNs) have gradually been introduced into recommendation systems to improve recommendation performance by modeling the relationship between entities. In particular, graph contrastive learning methods, as an emerging learning paradigm, can enhance the robustness and generalization ability of representations by maximizing the mutual information between positive sample pairs, further improving the recommendation effect based on graph structures.

[0004] Existing archive recommendation systems mainly rely on content-based recommendation and collaborative filtering strategies. Although they have achieved certain success, they have certain limitations in mining the deep structure of user interests. Content-based recommendation matches and recommends similar types of archives based on the content characteristics of archives, but ignores the interaction and social relationship between users; while collaborative filtering recommendation predicts preferences based on the user's historical behavior and the behavior of other similar users, but does not handle cold start problems (new users or new archives) well and is difficult to capture deep semantic information. In addition, most existing methods focus on processing a single type of user-archive interaction graph, still ignoring the user's social relationship and the similarity of archive resources themselves, and fail to fully integrate multi-source heterogeneous information, which to a certain extent affects the personalization and accuracy of archive recommendations. Summary of the invention

[0005] 1. Technical issues to be resolved

[0006] The technical problem to be solved by the present invention is how to provide a personalized profile recommendation method based on graph comparative learning to solve the shortcomings of existing profile recommendation methods and provide more accurate and personalized services.

[0007] (II) Technical solution

[0008] In order to solve the above technical problems, the present invention proposes a personalized profile recommendation method based on graph contrast learning, which includes the following steps: data preprocessing, graph construction, graph contrast node representation learning and personalized profile recommendation generation;

[0009] The data preprocessing stage is used to extract user and profile data features by collecting the user's browsing profile history, the user's social relationship data, and analyzing the profile text and other metadata;

[0010] The graph construction phase is used to construct the user-profile interaction graph (UI graph), the user-user social relationship graph (UU graph) and the profile feature graph (II graph) based on the processed data, and normalize the nodes and edges in each graph;

[0011] The graph comparison node representation learning stage is used to construct two sets of comparison view pairs based on the multi-source heterogeneous information in profile recommendation, namely, the user-profile interaction graph and the user-user social relationship graph, and the user-profile interaction graph and the profile feature graph. The graph attention mechanism is introduced to encode the nodes in the graph, and the low-dimensional embedding representation of users and profiles obtained by optimizing the comparison learning task is designed.

[0012] The personalized profile recommendation stage is used to take the recommendation task as the main task, multi-source comparative learning as the auxiliary task, jointly optimize the total loss function, and output the final personalized profile recommendation result after training is completed.

[0013] (III) Beneficial effects

[0014] The present invention proposes a personalized profile recommendation method based on graph contrast learning. The present invention not only introduces the advanced graph contrast learning mechanism into the profile recommendation system, but also integrates the multi-source heterogeneous information between users and profile resources, significantly improving the performance of the recommendation system and user experience. Compared with traditional content recommendation and collaborative filtering methods, the present invention can more accurately capture the complex relationship between user preferences and profile features, thereby providing recommendation results that are more in line with user needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A personalized profile recommendation flow chart of the present invention;

[0016] Figure 2 Diagram of the personalized profile recommendation framework based on graph comparative learning. DETAILED DESCRIPTION

[0017] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below in conjunction with the drawings and examples.

[0018] The present invention relates to the technical field of information retrieval and recommendation systems, and in particular to machine learning technology, especially graph neural network technology, and aims to realize a personalized archive recommendation method based on graph comparative learning.

[0019] The purpose of this invention is to propose a personalized profile recommendation method based on graph comparative learning, which learns personalized profile recommendation strategies for users by integrating multi-source heterogeneous information in the interaction between users and profile resources, user social relationships, and profile characteristics themselves, and combining graph comparative learning technology, so as to overcome the shortcomings of existing profile recommendation methods and provide more accurate and personalized services.

[0020] The algorithm flow of the present invention is as follows:

[0021]

[0022] The present invention provides a personalized profile recommendation method based on graph contrast learning. The specific implementation steps are as follows: Figure 1 As shown, it includes four stages: data preprocessing, graph construction, graph comparison node representation learning, and personalized profile recommendation generation.

[0023] The data preprocessing stage is used to extract user and profile data features by collecting the user's browsing profile history, the user's social relationship data, and analyzing the profile text and other metadata;

[0024] The graph construction phase is used to construct the user-profile interaction graph (UI graph), the user-user social relationship graph (UU graph) and the profile feature graph (II graph) based on the processed data, and normalize the nodes and edges in each graph;

[0025] The graph comparison node representation learning stage is used to construct two sets of comparison view pairs based on the multi-source heterogeneous information in profile recommendation, namely, the user-profile interaction graph and the user-user social relationship graph, and the user-profile interaction graph and the profile feature graph. The graph attention mechanism is introduced to encode the nodes in the graph, and the low-dimensional embedding representation of users and profiles obtained by optimizing the comparison learning task is designed.

[0026] The personalized profile recommendation stage is used to take the recommendation task as the main task, multi-source comparative learning as the auxiliary task, jointly optimize the total loss function, and output the final personalized profile recommendation result after training is completed.

[0027] The specific scheme of the present invention is introduced as follows.

[0028] 1. Data preprocessing: Extract user and profile data features by collecting user browsing history, user social relationship data, and analyzing profile text and other metadata.

[0029] First, by collecting users' browsing, downloading, marking and other behavior logs in the archive management system, we can clarify the users' specific operations on different archives and their timestamps, which can be used to extract the interaction relationship between different users and various archives.

[0030] Then, the user's social network information is extracted from the social platform or internal communication tools, including friend lists, follow-up relationships, jointly participated project teams, etc., to characterize the interactive relationship between users.

[0031] Finally, the content of the archives is analyzed, such as text summaries, keywords, classification labels, etc., and combined with metadata, such as creation date, author and other information, to provide a multi-dimensional description of the archives, which is then used to characterize the characteristics of the archival resources themselves.

[0032] In particular, for the collected search data, incomplete, erroneous or duplicate data entries are removed by combining data cleaning methods to ensure data quality and consistency. At the same time, missing values ​​are processed by filling default values, predictive completion, etc. to ensure the integrity and value of the data used for subsequent recommendations. For the text data involved, especially the archive's own features and user features, text features are extracted through natural language processing operations such as word segmentation, stop word removal, and stem extraction. In order to ensure that the subsequent graph representation learning process can fully capture the characteristic information of users and archives, the word embedding technology Word2Vec is introduced to convert such text features into low-dimensional dense vector features as the input data for initialization. Therefore, each user and each archive will be described by a feature vector to reflect the real semantic features.

[0033] 2. Graph construction: Based on the processed data, we construct the user-profile interaction graph (UI graph), user-user social relationship graph (UU graph) and profile feature graph (II graph). At the same time, we normalize the nodes and edges in each graph to ensure that the weights at different scales are comparable, thereby fully exploring the potential connections between multi-source heterogeneous information.

[0034] In the personalized profile recommendation method based on graph contrast learning, the construction of graph structure is a crucial step. First, a user-profile interaction graph (UI graph) is constructed based on the processed user-profile interaction data, in which each user and each profile is represented as a node. If there is any association between the user and the profile (such as browsing, downloading, favorites, etc.), there is an edge between the two nodes in the graph. Specifically, the UI graph is represented by an association matrix A. u_iIt is characterized by a matrix, where each row represents a user node, each column represents a profile node, and the matrix elements represent the intensity or frequency of the user's behavior on a specific profile, that is, the interaction edge with weight.

[0035] Then, based on the extracted user social information, a user-user social relationship graph (UU graph) is established, where nodes represent users and edges represent social connections between two users, which can be quantified by factors such as the number of common friends and interaction frequency. Specifically, the UU graph is constructed through an adjacency matrix A u_u Each row and column represents a different user, and the matrix elements are the social association strengths, which can intuitively capture the social network structure of users. At the same time, each user node is associated with its own attribute feature vector, which is represented by a separate feature matrix X u_u Each row represents a user, each column represents a different feature (such as occupation, position, education, etc.), and the elements in the matrix are the initialized word embedding values.

[0036] Finally, based on the archive resource content information, by measuring the similarity between archives, an archive feature graph (II graph) is constructed to reflect the association between archive contents. Specifically, each archive is associated with its own attribute feature vector obtained in step 1, which is represented by a separate feature matrix X i_i Characterization, each row represents a file, each column is a different feature dimension of the file, and the elements in the matrix are the initialized word embedding values. At the same time, based on the initialized file's own feature vector, the cosine similarity between any vectors is calculated. When the similarity value is greater than the set threshold (such as 0.6), it is considered that there is an intrinsic correlation between the two files. Therefore, the nodes in the II graph are files, and the similarity between files is reflected as weighted edges. The II graph is also composed of the adjacency matrix A i_i Formal description, each row and column represents a different file, and the elements in the matrix are similarity values.

[0037] In particular, in order to ensure the comparability of node features and edge weights at different scales and improve the stability of subsequent graph neural network model training, the matrices of each type of graph structure (UI graph, UU graph, II graph) are normalized during the implementation process. The graph construction stage further extracts and mines the potential connections between multi-source heterogeneous information in the archive recommendation process, which helps to achieve more accurate and personalized archive recommendation services in the future.

[0038] 3. Graph contrast node representation learning: Based on the multi-source heterogeneous information in profile recommendation, the user-profile interaction graph and the user-user social relationship graph, the user-profile interaction graph and the profile feature graph are constructed into two sets of contrast view pairs, and the graph attention mechanism is introduced to encode the nodes in the graph. The low-dimensional embedding representation of users and profiles obtained by optimizing the contrastive learning task is designed.

[0039] (1) Contrast view construction

[0040] An inherent challenge facing the graph contrastive learning framework is how to construct effective contrastive views to learn node representations. Based on the multi-source heterogeneous information in profile recommendation, the user-profile interaction graph and the user-user social relationship graph both start from the user himself and characterize the user's behavior. In order to learn discriminative user representations, the user-profile interaction graph and the user-user social relationship graph are regarded as a pair of contrastive views. Similarly, the user-profile interaction graph and the profile feature graph are both characterized around the profile's own features. Therefore, the two are used as a pair of contrastive views for learning profile representations.

[0041] (2) Graph Representation Coding

[0042] The graph representation encoding process mainly uses graph convolution operations to learn the embedded representation of users and profiles based on the input contrast view pairs. For user-side encoding, a two-layer graph attention network is used as the encoder f θ To effectively aggregate the information of neighbors in a given graph, the weights between each neighbor sample are adaptively learned to perform feature aggregation by considering the importance of different neighbor nodes, thereby better capturing the interaction between nodes. After L-layer network embedding propagation, the user node representation in the user-user social relationship graph The learning process is as follows:

[0043]

[0044] Where W is the network learning weight parameter matrix, σ is the nonlinear activation function Sigmoid, represents the set of users that have social connections with user u, is the embedding representation of any neighbor v of user u in the UU graph in the previous layer of the network, α uv It is the weight coefficient between user node u and its adjacent user v, emphasizing the importance of the link relationship between different users. It is calculated by the Softmax function and its formula is expressed as:

[0045]

[0046] in(·) T and || represent transposition and concatenation operations respectively, is a weight vector, and LeakyRELU(·) is a nonlinear activation function. The comparison view UU graph and UI graph are implemented using a graph attention network with shared parameters. Therefore, the process of obtaining user node embedding based on the user-profile interaction graph is similar and can be expressed as follows:

[0047]

[0048] where α ujis the weight coefficient of the interaction between user node u and its adjacent profile j, It is the embedding representation of the adjacent profile j of user u in the UI graph obtained in the previous layer of the network.

[0049] For the archive encoding process, the graph attention network learning is also used. Its embedding propagation process is consistent with the user representation learning process, but the input view is the user-archive interaction graph and the archive feature graph. The corresponding learned archive low-dimensional embedding is expressed as:

[0050]

[0051]

[0052] where α iv ,α ij They represent the weight coefficients of the interaction between profile i and user v and the link between profile i and profile j, is the embedding representation of profile i and adjacent user v in the UI graph obtained in the previous layer of the network, It is the embedded representation of the adjacent file j of file i in Figure II obtained in the previous layer of the network.

[0053] (3) Multi-source contrastive representation learning

[0054] Based on the user and profile embedding representations obtained by graph encoding modeling, the final user representation and profile representation are learned using the contrastive representation learning mechanism. The profile interaction target behavior and social relationship auxiliary behavior of the same user are constructed as positive example pairs. The interactive target behaviors and social auxiliary behaviors of different users u' are constructed as negative example pairs. The learning objectives of constructing user profile comparison using positive and negative pairs are:

[0055]

[0056] Where sim(·) represents the cosine similarity function.

[0057] Similarly, the positive example pairs are constructed by using the user interaction target behavior of the same profile and the profile similarity relationship auxiliary behavior Construct negative example pairs based on interactive target behaviors and similar auxiliary behaviors of different profiles i' The comparison learning objectives of constructing profile images using positive and negative pairs are:

[0058]

[0059] 4. Personalized profile recommendation generation: Taking the recommendation task as the main task and multi-source comparative learning as the auxiliary task, the total loss function is jointly optimized. After the training is completed, the final personalized profile recommendation result is output.

[0060] In the process of contrastive view construction and contrastive learning, both the user-side embedding representation and the profile-side embedding representation cover multi-source information from multiple views and are learned based on a self-supervisory mechanism without relying on external label information. First, the average pooling operation AVGPOOL(·) is introduced to fuse multi-source information, and the final embedding representation of users and profiles is:

[0061]

[0062] Then, based on the known supervisory signals of user-profile interactions in the recommendation task, by using the goal that the prediction score of observed interaction behaviors is higher than that of unobserved interaction behaviors, the main BPR loss target of the recommendation task is constructed as:

[0063]

[0064] Where σ is the nonlinear activation function Sigmoid, o = {u,i,j|R u,i =1,R u,j =0} represents paired training data, (u,i), (u,j) represents the interaction between user u and profile i,j, R u,i =1 indicates the observed interactive positive sample, R u,j = 0 indicates an unobserved interaction negative sample. ui Then it represents the probability of user u interacting with profile i, that is, the possibility, which is obtained through a layer of multi-layer perceptron mapping and expressed as:

[0065] y ui =f(h u ,h i ).

[0066] When the prediction score of the positive sample is higher than that of the negative sample, the personalized recommendation profile i for user u is more credible and can better guide the recommendation effect.

[0067] In order to make full use of the interaction between users and archives and model the association behavior between the two, the multi-source comparison learning goal is used as an auxiliary task. The multi-task learning strategy is combined with the main BPR (Bayesian Personalized Ranking) recommendation goal of the recommendation task to obtain the final overall loss target as follows:

[0068] L=λ*l BPR +l u +l i ,

[0069] Where λ is a hyperparameter that controls the weight of the recommendation main task.

[0070] After joint optimization and training, not only can we learn user and profile embeddings with strong representation capabilities, but we can also predict the probability of recommending different profiles for different users, achieve personalized profile recommendations, and improve the accuracy and experience of profile recommendations.

[0071] The present invention not only introduces advanced graph contrast learning mechanism into the archive recommendation system, but also integrates multi-source heterogeneous information between users and archive resources, significantly improving the performance of the recommendation system and user experience. Compared with traditional content recommendation and collaborative filtering methods, the present invention can more accurately capture the complex relationship between user preferences and archive features, thereby providing recommendation results that better meet user needs.

[0072] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A personalized profile recommendation method based on graph contrast learning, characterized in that: The method includes the following steps: data preprocessing, graph construction, graph comparison node representation learning, and personalized profile recommendation generation; The data preprocessing stage is used to extract user and profile data features by collecting the user's browsing profile history, the user's social relationship data, and analyzing the profile text and other metadata; The graph construction phase is used to construct the user-profile interaction graph (UI graph), the user-user social relationship graph (UU graph) and the profile feature graph (II graph) based on the processed data, and normalize the nodes and edges in each graph; The graph comparison node representation learning stage is used to construct two sets of comparison view pairs based on the multi-source heterogeneous information in profile recommendation, namely, the user-profile interaction graph and the user-user social relationship graph, and the user-profile interaction graph and the profile feature graph. The graph attention mechanism is introduced to encode the nodes in the graph, and the low-dimensional embedding representation of users and profiles obtained by optimizing the comparison learning task is designed. The personalized profile recommendation stage is used to take the recommendation task as the main task, multi-source comparative learning as the auxiliary task, jointly optimize the total loss function, and output the final personalized profile recommendation result after training is completed.

2. The personalized profile recommendation method based on graph contrast learning according to claim 1, characterized in that: The data preprocessing stage specifically includes: First, by collecting the user's browsing, downloading, and marking behavior logs in the archive management system, the user's specific operations on different archives and their timestamps are clarified to extract the interaction relationship between different users and various archives; Then, extract the user's social network information from the social platform or internal communication tools, including friend lists, follow-up relationships, and co-participated project teams, to characterize the interactive relationship between users; Finally, the content of the archives is analyzed and combined with metadata to provide a multi-dimensional description of the archives, which is then used to characterize the characteristics of the archival resources themselves.

3. The personalized profile recommendation method based on graph contrast learning as claimed in claim 2, characterized in that: For the collected search data, we use data cleaning methods to remove incomplete, erroneous or duplicate data entries to ensure data quality and consistency. At the same time, we handle missing values ​​by filling in default values ​​and predictive completion. For the text data involved, especially the archive’s own characteristics and user characteristics, we extract text features through word segmentation, stop word removal, and stemming operations.

4. The personalized profile recommendation method based on graph contrast learning as claimed in claim 2, characterized in that: The word embedding technology Word2Vec is introduced to convert text features into low-dimensional dense vector features as the initial input data. Each user and each profile will be described by a feature vector to reflect the real semantic features.

5. The personalized profile recommendation method based on graph contrast learning according to any one of claims 1 to 4, characterized in that: During the graph construction phase, First, a user-profile interaction graph (UI graph) is constructed based on the processed user-profile interaction data, in which each user and each profile is represented as a node. If there is any association between a user and a profile, there is an edge between the two nodes in the graph. Specifically, the UI graph is represented by an association matrix A. u_i It is characterized by each row representing a user node, each column representing a profile node, and the matrix elements represent the intensity or frequency of the user's behavior on a specific profile, that is, the interaction edge with weight; Then, based on the extracted user social information, a user-user social relationship graph (UU graph) is established, in which nodes represent users and edges represent social connections between two users, which are quantified by the number of common friends and interaction frequency factors. Specifically, the UU graph is constructed through an adjacency matrix A u_u Each row and column represents a different user, and the matrix elements are the social association strengths, which can intuitively capture the social network structure of users; at the same time, each user node is associated with its own attribute feature vector, which is represented by a separate feature matrix X u_u Characterization, where each row represents a user, each column is a different feature, and the elements in the matrix are the initialized word embedding values; Finally, based on the content information of archive resources, by measuring the similarity between archives, an archive feature graph (II graph) is constructed to reflect the association between archive contents; specifically, each archive is associated with its own attribute feature vector, which is represented by a separate feature matrix X i_i Characterization, each row represents a file, each column is a different feature dimension of the file, and the elements in the matrix are the initialized word embedding values; at the same time, based on the initialized feature vector of the file itself, the cosine similarity between any vectors is calculated. When the similarity value is greater than the set threshold, it is considered that there is an intrinsic correlation between the two files; therefore, the nodes in the II graph are files, and the similarity between files is reflected as weighted edges; the II graph is also composed of the adjacency matrix A i_i Formal description, each row and column represents a different file, and the elements in the matrix are similarity values.

6. The personalized profile recommendation method based on graph contrast learning according to claim 5, characterized in that: In the graph construction stage, the matrices of the UI graph, UU graph, and II graph are all normalized. The graph construction stage further extracts and mines the potential connections between multi-source heterogeneous information in the archive recommendation process.

7. The personalized profile recommendation method based on graph contrast learning according to claim 5, characterized in that: The graph comparison node representation learning phase includes: comparison view construction, graph representation encoding and multi-source comparison representation learning; The construction of comparative views includes: the user-profile interaction graph and the user-user social relationship graph both start from the user himself and characterize the user's behavior. In order to learn discriminative user representation, the user-profile interaction graph and the user-user social relationship graph are regarded as a pair of comparative views; similarly, the user-profile interaction graph and the profile feature graph are both characterized around the profile's own features; therefore, the two are regarded as a comparative view pair for learning profile representation; Graph representation encoding includes: learning the embedding representation of users and profiles by using graph convolution operations based on the input contrast view pairs; for user-side encoding, a two-layer graph attention network is used as the encoder f θ To effectively aggregate the information of neighbors in a given graph, the weights between each neighbor sample are adaptively learned to perform feature aggregation by considering the importance of different neighbor nodes, so as to better capture the interaction between nodes; after L-layer network embedding propagation, the user node representation in the user-user social relationship graph The learning process is as follows: Where W is the network learning weight parameter matrix, σ is the nonlinear activation function Sigmoid, represents the set of users that have social connections with user u, is the embedding representation of any neighbor v of user u in the UU graph in the previous layer of the network, α uv It is the weight coefficient between user node u and its adjacent user v, emphasizing the importance of the link relationship between different users. It is calculated by the Softmax function and its formula is expressed as: in,(·) T and || represent transposition and concatenation operations respectively. is a weight vector, LeakyRELU(·) is a nonlinear activation function; the comparison view UU graph and UI graph are implemented using a graph attention network with shared parameters. Therefore, the process of obtaining user node embedding based on the user-profile interaction graph is similar, as shown below: where α uj is the weight coefficient of the interaction between user node u and its adjacent profile j, It is the embedding representation of the adjacent profile j of user u in the UI graph obtained in the previous layer of the network; For the archive encoding process, the graph attention network learning is also used. Its embedding propagation process is consistent with the user representation learning process, but the input view is the user-archive interaction graph and the archive feature graph. The corresponding learned archive low-dimensional embedding is expressed as: where α iv ,α ij They represent the weight coefficients of the interaction between profile i and user v and the link between profile i and profile j, is the embedding representation of profile i and adjacent user v in the UI graph obtained in the previous layer of the network, It is the embedding representation of the adjacent file j of file i in Figure II obtained in the previous layer of network; Multi-source contrastive representation learning includes: learning the final user representation and profile representation based on the user and profile embedding representation obtained by graph encoding modeling using contrastive representation learning mechanism; constructing the profile interaction target behavior and social relationship auxiliary behavior of the same user as positive example pairs The interactive target behaviors and social auxiliary behaviors of different users u' are constructed as negative example pairs. The learning objectives of constructing user profile comparison using positive and negative pairs are: Where sim(·) represents the cosine similarity function; The positive example pairs are constructed by using the user interaction target behavior and the auxiliary behavior of the profile similarity relationship of the same profile. Construct negative example pairs based on interactive target behaviors and similar auxiliary behaviors of different profiles i' The comparison learning objectives of constructing profile images using positive and negative pairs are:

8. The personalized profile recommendation method based on graph contrast learning according to claim 7, characterized in that: The personalized profile recommendation generation stage includes: First, the average pooling operation AVGPOOL(·) is introduced to fuse multi-source information, and the final embedding of users and profiles is expressed as: Then, based on the known supervisory signals of user-profile interactions in the recommendation task, by using the goal that the prediction score of observed interaction behaviors is higher than that of unobserved interaction behaviors, the main BPR loss target of the recommendation task is constructed as: Where σ is the nonlinear activation function Sigmoid, o = {u,i,j|R u,i =1,R u,j =0} represents paired training data, (u,i), (u,j) represents the interaction between user u and profile i,j, R u,i =1 indicates the observed interactive positive sample, R u,j = 0 indicates an unobserved interaction negative sample; y ui Then it represents the probability of user u interacting with profile i, that is, the possibility, which is obtained through a layer of multi-layer perceptron mapping and expressed as: y ui =f(h u ,h i ). When the prediction score of the positive sample is higher than that of the negative sample, the personalized recommendation profile i for user u is more credible and can better guide the recommendation effect.

9. The personalized profile recommendation method based on graph contrast learning according to claim 8, characterized in that: In order to make full use of the interaction between users and archives and model the association behavior between the two, the multi-source comparison learning objective is used as an auxiliary task. The multi-task learning strategy is combined with the main BPR recommendation objective of the recommendation task to obtain the final overall loss target as follows: L=λ*l BPR +l u +l i , Where λ is a hyperparameter that controls the weight of the recommended main task; After joint optimization and training, it not only learns user and profile embeddings with strong representation capabilities, but also can predict the probability of recommending different profiles for different users, thus achieving personalized profile recommendation.

10. The personalized profile recommendation method based on graph contrast learning according to claim 8, characterized in that: In the process of comparative view construction and comparative learning, both the user-side embedding representation and the archive-side embedding representation cover multi-source information of multiple views and are learned based on a self-supervisory mechanism without relying on external label information.

Images