A paper recommendation method guided by social relationships
By constructing a connection between cold-start users and papers using a deep learning neural network guided by social relationships, and employing attention mechanisms for weighting and pooling, the recommendation problem for cold-start users is solved, achieving more accurate paper recommendation results.
Patent Information
- Application Number
- CN202211459161.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-17
- Publication Date
- 2025-11-14
- Estimated Expiration
- 2042-11-17
AI Technical Summary
Existing paper recommendation methods have failed to effectively address the recommendation problem for cold-start users, who lack direct interaction history, resulting in poor recommendation performance.
We employ a deep learning neural network guided by social relationships. Through embedding, pooling, and prediction layers, we leverage users' social relationships to build connections between users and papers. We then combine this with attention mechanisms for weighting and pooling to generate more accurate recommendation results.
It enables effective paper recommendations for users in the cold start phase, improves the accuracy and personalization of recommendations, and solves the problem of inconsistent user interests.
Smart Images
Figure CN115757985B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of deep learning, specifically relating to a paper recommendation method based on social relationships. Background Technology
[0002] With the rapid development of information technology, academic papers are now extremely complex, and the number of publications is growing exponentially, making it difficult for researchers to judge the relevance of papers and select the right ones. Simultaneously, due to technological advancements, countries are placing increasing emphasis on education, leading to a year-on-year increase in the number of graduate students. For newly enrolled students or research novices, finding suitable papers is crucial. Therefore, paper recommendation methods have been proposed and developed by researchers. Paper recommendation tools can help improve the efficiency of researchers in retrieving papers and reduce their workload. Generally, paper recommendation models include the following main categories: collaborative filtering-based recommendation methods, content-based recommendation methods, graph-based recommendation methods, and hybrid recommendation methods. Personalized recommendations of papers tailored to the different needs of researchers can be implemented.
[0003] Existing methods all prioritize recommendations based on existing interactions between users and papers. Therefore, current work does not consider how to recommend papers to users during their initial user experience. Summary of the Invention
[0004] The purpose of this invention is to provide a paper recommendation method based on social relationships. By guiding students' social relationships, it constructs a connection between users and papers, explores the information dissemination patterns in social networks, and solves the paper recommendation problem for users in the cold start phase.
[0005] The technical solution to achieve the purpose of this invention is: a paper recommendation method based on social relationships, wherein the method is based on a deep learning neural network, the neural network including an embedding layer, a pooling layer and a prediction layer, and specifically includes the following steps:
[0006] Step (1): Considering users' social relationships, the embedding layer uses word2vec and classification feature embedding methods to process the paper dataset, and obtains user feature embedding vectors, social relationship paper feature embedding vectors and candidate paper feature embedding vectors and outputs them;
[0007] Step (2): Employ an attention mechanism to perform a weighted summation of each dimension of the social relationship paper feature embedding vector in the user feature embedding vector and the social relationship paper feature embedding vector based on social relationship guidance, thereby obtaining the weight w. j An attention mechanism is used to weight each dimension of the social relationship paper feature embedding vector and the candidate paper feature embedding vector based on social relationships, resulting in a weight w. candidateUsing weight w j and weight w candidate The social relationship paper feature embedding vector of each dimension is guided, and the guided social relationship paper feature embedding vector is pooled to obtain the final social relationship paper feature embedding vector; the final social relationship paper feature embedding vector, user feature embedding vector, and candidate paper feature embedding vector are aggregated to obtain the representation vector, that is, the output of the aggregation layer;
[0008] Step (3): In the prediction layer, the representation vector is fully connected through a multilayer perceptron and then dimensionality is reduced to output a one-dimensional value, which represents the score of the candidate paper.
[0009] Furthermore, the paper dataset in step (1) includes user characteristics, social relationship papers, and candidate papers;
[0010] The user characteristics include user paper direction characteristic x0, user school characteristic s, and user paper publication year characteristic t;
[0011] The user's paper direction feature x0 is represented as follows:
[0012] x0={x 0.1 x 0.2 , ..., x 0.n}
[0013] In the above formula, x 0.1 ~x 0.n This represents the n directions of the user's paper, x 0.i This represents the i-th direction of the user's paper.
[0014] Furthermore, the method for determining the social relationship in a paper is as follows:
[0015] Step (11) identifies other students who have the same teacher as user u as user u's social relationships, represented as:
[0016] (u, U1, U2..., U n}∈U t
[0017] In the above formula, U1~U n This refers to other students who have the same teachers as user u and enrolled earlier than user u. t This represents the collection of students (U) of the teacher;
[0018] Step (12): Based on the student set U, integrate the papers that each student is interested in to obtain the paper database J. The relationship between the papers and the students is as follows:
[0019]
[0020] In the above formula, Juj This refers to a collection of papers that student j is interested in. 1j ~j nj This refers to a collection of papers by student J. uj The paper in J u1 ~J un The table indicates that each student's collection of papers is in the paper database;
[0021] Step (13): Determine the social relationship papers of user u: The collection of papers on user u's social relationships includes {J u1 J u2 , ..., J un}; where {U1, U2, ..., U n All of them have the same teachers as user u and enrolled in school earlier than user u. For user u, their social relationships are represented in the following paper:
[0022] {j 11 j 21 , ...j n1 j 12 j 22 , ..., j n2 , ..., j 1m j 2m , ...j nm}
[0023] Furthermore, step (1) also includes the following steps:
[0024] Step (14): Use word2vec to process user features and obtain the user's paper direction feature vector X. i :
[0025]
[0026] In the formula, x i This indicates obtaining the i-th word in the word vector x representing the user's paper's direction. This represents the result of using the word2vec word embedding model for the i-th direction;
[0027] Step (15): Transfer the user's paper direction feature vector X i The user feature embedding vector u is obtained by concatenating the user's school feature s and the user's published paper year feature t. embedding :
[0028] u embedding = [X1, X2, ..., X n ,s,t]
[0029] Among them, X1~X nWord embeddings representing an individual's research direction, s representing the user's school characteristics, and t representing the user's publication year characteristics;
[0030] Step (16): The embedding vector of each paper is defined as the paper title feature, the paper publication time and the paper category feature;
[0031] Step (17): Use word2vec to process the paper title features: Let title represent the paper title, then the word vector function expression of the title is:
[0032]
[0033] Step (18): Concatenate the paper title features, paper category features, and paper publication time to obtain the social relationship paper feature embedding vector j. embedding :
[0034] j embedding = [Title, c, t]
[0035] Step (19): Perform the same feature processing on the candidate papers as in steps (17) and (18) to obtain the candidate paper feature embedding vector candidate. embedding .
[0036] Furthermore, step (2) includes the following steps:
[0037] Step (21): Obtain the weight w j w j The function expression is as follows:
[0038] w j =f(u embedding Title social )
[0039] =Linear(MLP([u embedding Title social u embedding *Title social ]))
[0040] In the formula, Title social An embedding vector representing the title of a paper on social relations;
[0041] Step (22): Obtain the weight w candidate w candidate The function expression is as follows:
[0042] w candidate =f(candidate) embedding j embedding )
[0043] =linear(MLP([candidate]) embedding j embedding candidate embedding *j embedding In the formula, j embedding Represents the complete embedding vector of a paper on social relations;
[0044] Step (23): Use weight w j and weight w candidate The embedding vector of each dimension of social relationship paper features is guided, and the guided embedding vector is e. i The guided social relationship paper feature embedding vector is pooled to obtain the final social relationship paper feature embedding vector e:
[0045] e i =(w j ·w candidate ·j embedding )
[0046] e = pooling(e1 + e2 + ... + e) n )
[0047] In the formula, e1~e n This represents the processed user social relationship paper embedding vector, and pooling() represents the pooling layer, which performs the accumulation operation.
[0048] Step (24) yields the representation vector:
[0049] The user feature vector, the pooled user social relationship paper vector, and the candidate paper vector are concatenated; the concatenated vector is represented by e. concat The formula is as follows:
[0050] e concat =concat(u embedding e, candidate embedding )
[0051] =[u embedding e, candidate embedding ].
[0052] Furthermore, step (3) specifically involves:
[0053] The output vector is then passed through a fully connected multilayer perceptron for dimensionality reduction. The resulting one-dimensional value is the score of the candidate paper, calculated using the following formula:
[0054] output = softmax(ReLU(e))
[0055] The inner multilayer perceptron uses ReLU as the activation function, while the outer multilayer perceptron uses softmax() as the activation function to calculate scores.
[0056] Compared with the prior art, the significant advantages of this invention are:
[0057] Existing methods directly utilize the direct interaction between users and papers to build networks, but this approach is ineffective for cold-start users, who do not interact with the data. This invention addresses cold-start users by proposing a social relationship-guided approach to construct a channel between them and papers, enabling information dissemination and achieving the goal of recommendation between cold-start users and papers.
[0058] In the prediction module guided by social relationships, this invention uses an attention mechanism based on social network relationships to provide combined guidance for both user features and paper features, resulting in a higher-quality representation. This is superior to ordinary deep learning networks and solves the problem of inconsistent user interest in each social relationship paper. This invention uses a two-layer attention mechanism, which can more accurately identify user interests, thereby achieving better recommendation results. Attached Figure Description
[0059] Figure 1 This is a user paper relationship graph guided by social relationships in an example of the present invention.
[0060] Figure 2 This is a schematic diagram of the overall model in an embodiment of the present invention.
[0061] Figure 3 This is a schematic diagram of the attention mechanism in an embodiment of the present invention. Detailed Implementation
[0062] The present invention will now be described in further detail with reference to the accompanying drawings.
[0063] This example integrates a paper recommendation method guided by social relationships, including:
[0064] S1: Representation learning module guided by social relationships: For social relationship paper features, we construct a social network and find the papers that users are associated with on the social network as social relationship paper features; We embed user features, paper features, and candidate paper features into user feature embedding vectors, paper feature embedding vectors, and candidate paper feature embedding vectors through embedding.
[0065] S2: Social Relationship-Guided Prediction Module: The user feature embedding vector and the candidate paper feature embedding vector are weighted by social relationship guidance and attention mechanism to obtain the updated features. The social relationship paper feature embedding vector is accumulated through pooling layer to obtain the final social paper embedding vector. The social paper feature embedding vector, user feature embedding vector and candidate paper embedding vector are connected in the connection layer. The resulting connected embedding vector is predicted by deep learning network to obtain the result.
[0066] In this embodiment, module S1 is a social relationship-guided representation learning module, including:
[0067] The input for this example consists of three parts: user features, paper features, and candidate paper features. These features are used to transform the user features, paper features, and candidate paper features into an initial embedding vector.
[0068] User characteristics include user's research direction, university affiliation, and publication year. User personal characteristics are used as input, where:
[0069] The user's paper direction features are processed into word vectors using word2vec and then embedded. Let x0 represent any one of the direction words in the user's paper. Then x0 can be represented as:
[0070] x0={x 0.1 x 0.2 , ..., x 0n}
[0071] In the above formula, x 0.1 ~x 0.n This represents the n directions of the user's paper, x 0.i This represents the i-th direction of the user's paper. The word embedding uses the word2vec model for output, and the specific function expression for obtaining the word vectors of the user's research direction is:
[0072]
[0073] In the formula, x i This indicates obtaining the i-th word in the word vector x representing the user's paper's direction. This represents the result of using the word2vec word embedding model for the i-th direction.
[0074] User school features are represented by categorical features, and we use s to represent user school features, and
[0075] s∈S
[0076] Where s is the user's school feature, and S is the set of schools in the dataset, that is, s is a school in the set of schools S.
[0077] The year in which a user publishes a paper is also embedded as a classification feature, and we use t to represent the year in which a user publishes a paper.
[0078] From the above three features, we can deduce that the user feature is obtained by embedding and concatenating three features: user's paper direction feature, user's school feature, and user's paper publication year feature. Let u... embedding Representing user characteristics, the function expression for u is:
[0079] u embedding = [X1, X2, ..., X n ,s,t]
[0080] Among them, X1~X n The word embeddings represent an individual's research direction, 's' represent the user's school characteristics, and 't' represent the user's publication year characteristics. Connecting these elements gives the quantity u. embedding The function expression.
[0081] Depend on Figure 1 As can be seen, in this embodiment, the user's papers are obtained from the user's social network relationships. We consider the social relationship between a user and their fellow students as the closest social relationship. Therefore, we believe that the user will definitely be interested in papers that their fellow students are interested in. So we denote the user as u. For user u, the papers that students with the same teachers as user u have read in the past are papers that user u will be interested in.
[0082] The relationship between user u and the other students is as follows:
[0083] {u, U1, U2..., U n}∈U t
[0084] In the above formula, U1~U n This refers to other students who have the same teachers as user u and enrolled earlier than user u. t This refers to the collection of students (U) of the teacher.
[0085] We use J as a database of papers, and the relationship between papers and students is as follows:
[0086]
[0087] In the above formula, J uj This refers to a collection of papers that student j is interested in. 1j ~j nj This refers to a collection of papers by student J. uj The paper in J u1 ~J un The table shows that each student's collection of papers is in the paper database.
[0088] Papers on determining user u's social relationships: The collection of papers on user u's social relationships includes {J} u1 J u2 , ..., J un}; where {U1, U2, ..., U n All of them have the same teachers as user u and enrolled in school earlier than user u. For user u, their social relationships are represented in the following paper:
[0089] {j 11 j 21 , ...j n1 j 12 j 22 , ..., j n2 , ..., j 1m j 2m , ...j nm}
[0090] For each paper, its embedding vector is divided into three parts: paper title features, paper publication time and paper category features.
[0091] An example is: "A collaborative filtering recommendation algorithm based on temporal behavior, 2013, Journal of Software", where "A collaborative filtering recommendation algorithm based on temporal behavior" is the title feature, "2013" is the publication date, and "Journal of Software" is the category feature.
[0092] We embed the features of the paper title using the word2vec word embedding model, and let 'title' represent the paper title. The word vector function expression for the title is then:
[0093]
[0094] In the above formula, Title uj This refers to the j-th paper among the relevant papers for user uj. This represents the result of using the word2vec word embedding model for the j-th paper among the user's relevant papers.
[0095] Paper category features are represented by categorical features. We use c to represent the paper category feature, and...
[0096] c∈C
[0097] Where c is the paper category feature, and C is the set of paper categories in the dataset (including Journal of Software, Chinese Journal of Computers, Computer Science, Pattern Recognition and Artificial Intelligence, etc.), that is, c is a certain category in the set of paper categories C.
[0098] The publication date of the paper is also embedded as a classification feature, and we use t to represent the year in which the user published the paper.
[0099] From the above three features, we can conclude that the paper features are obtained by embedding and concatenating three features: paper title feature, paper category feature, and paper publication time feature. Let j embedding Let j represent the paper embedding vector. embedding The function expression is as follows:
[0100] j embedding = [Title, c, t]
[0101] Here, Title represents the word embedding of the paper's title, c represents the paper's category feature, and t represents the paper's publication time feature. Connecting these elements together gives the quantity j. embedding The function expression.
[0102] Candidate papers are a collection of papers recommended to students, so their characteristics are the same as those of the papers themselves, including title, publication date, and category. We use "Candidate" to represent the candidate paper database.
[0103] For each candidate paper, its embedding vector is also divided into three parts: paper title features, paper publication time and paper category features.
[0104] We let the candidate embedding Representing the embedding vector of the candidate paper, we obtain the candidate. embedding The function expression is:
[0105] candidate embedding = [Title, c, t]
[0106] Where Title represents the word embedding of the candidate paper's title, c represents the candidate paper's category feature, and t represents the candidate paper's publication time feature. Connecting these elements gives the candidate... embedding The function expression.
[0107] In this embodiment, module S2 is a prediction module for social relationship guidance, including:
[0108] The aggregation layer is guided by social relationships. The aggregation layer uses an attention mechanism to weight and connect user features, paper features, and candidate paper features together. The pooling layer, connection layer, and prediction layer are also guided by social relationships.
[0109] Since the user's social relationship paper database consists of papers that are of interest to the user's senior classmates, but the user's interest in each paper in the social relationship database varies, this example uses an attention mechanism. By calculating the relevance between user characteristics and papers in the social relationship paper database, the degree of user interest in the papers is calculated, and a weighted coefficient is obtained.
[0110] Attention mechanisms such as Figure 3 As shown, the steps are as follows:
[0111] The input consists of two vector embeddings, denoted as embedding1 and embedding2, and the output is the weight, denoted as w.
[0112]
[0113] Where w is the weight, representing the correlation between embedding1 and embedding2. A larger w indicates a stronger correlation between embedding1 and embedding2; a smaller w indicates a weaker correlation. `Linear()` normalizes the multidimensional vector output by `MLP()`. `MLP()` is a multilayer perceptron, and this layer is fully connected, using ReLU as the activation function. We use... This represents the product of two vectors, embedding1 and embedding2.
[0114] This means connecting these vectors together.
[0115] We can obtain the correlation weights between two vectors through the attention mechanism.
[0116] like Figure 2 As shown, the input to the convergence layer consists of three parts: user feature embedding vector, social relationship paper feature embedding vector, and candidate paper feature embedding vector. We use u embedding Represents the user feature embedding vector, Title social The embedding vector representing the title of a paper on social relations, j embedding Represents the complete embedding vector of a social relations paper, Title candidatel The embedding vector representing the title of a candidate paper. embedding This represents the complete embedding vector of the candidate paper.
[0117] Since users u have varying levels of interest in their social relationship papers, this step uses an attention mechanism to adaptively calculate the user's social relationship representation vector, taking into account the correlation between users and social relationship papers. We use wj If w represents the weight of the paper title indicating user characteristics and social relationships, then w j The function expression is as follows: w j The function expression is as follows:
[0118] w j =f(u embedding Title social )
[0119] =Linear(MLP([u embedding Title social u embedding *Title social ]))
[0120] Since each social relationship paper plays a different role and has varying importance among different candidate papers, we use w to consider the correlation between user social relationship papers and candidate papers. candidate w represents the weight of papers and candidate papers related to user social relationships. candidate The function expression is as follows:
[0121] w candidate =f(candidate) embedding j embedding )
[0122] =linear(MLP([candidatee mbedding j embedding candidate embedding *j embedding ]))
[0123] Since different users exhibit different numbers of behaviors, we need to process this set of behaviors into vectors of equal length and then sum them. Therefore, for each user's social relationship paper, we use e to represent the processed embedding vector. i It means, e i The formula is as follows:
[0124] e i =(w j ·w candidate ·j embedding )
[0125] Pooling layer: This layer accumulates the processed social relationship embedding vectors. The pooling layer formula is as follows, and we denote the output of the pooling layer as e:
[0126] e = pooling(e1 + e2 + ... + e) n )
[0127] In the formula, e1~e n This represents the processed user social relationship paper embedding vector, and pooling() represents the pooling layer, which performs the accumulation operation.
[0128] The connection layer concatenates the user feature vector, the pooled user social relationship paper vector, and the candidate paper vector. The concatenated vector is represented by e. concat The formula is as follows:
[0129] e concat =concat(u embedding e, candidate mbedding )
[0130] =[u embedding e, candidate embedding ].
[0131] The output vector is then passed through a fully connected multilayer perceptron for dimensionality reduction. The resulting one-dimensional value is the score of the candidate paper, calculated using the following formula:
[0132] output = softmax(ReLU(e))
[0133] The inner multilayer perceptron uses ReLU as the activation function, while the outer multilayer perceptron uses softmax() as the activation function to calculate scores.
[0134] Finally, the output scores are sorted. The higher the output score, the more interested the user is in the candidate paper; the lower the output score, the less interested the user is in the candidate paper.
[0135] The model is trained using the standard gradient algorithm with cross-entropy and L2 regularization:
[0136]
[0137] Where x is the input, y is the label of the sample (1 for positive samples and 0 for negative samples), and p(x) is the probability that the sample is positive when the input is x.
[0138] We employ L2 regularization to mitigate overfitting caused by an excessive number of features. The formula is as follows:
[0139]
[0140] Where W represents the parameters of the entire embedding, w j This represents the j-th embedding vector.
Claims
1. A paper recommendation method based on social relationships, the method being based on a deep learning neural network, the neural network comprising an embedding layer, a pooling layer, and a prediction layer, characterized in that, Specifically, the steps include the following: Step (1): Considering users' social relationships, the embedding layer uses word2vec and classification feature embedding methods to process the paper dataset, and obtains user feature embedding vectors, social relationship paper feature embedding vectors and candidate paper feature embedding vectors and outputs them; Step (2): Employ an attention mechanism to perform a weighted summation of each dimension of the social relationship paper feature embedding vector in the user feature embedding vector and the social relationship paper feature embedding vector based on social relationship guidance, thereby obtaining the weight w. j An attention mechanism is used to weight each dimension of the social relationship paper feature embedding vector and the candidate paper feature embedding vector based on social relationships, resulting in a weight w. candidate Using weight w j and weight w candidate The embedding vector of social relationship paper features in each dimension is guided, and the guided embedding vector of social relationship paper features is pooled to obtain the final embedding vector of social relationship paper features. The final embedding vector of social relationship paper features, the embedding vector of user features, and the embedding vector of candidate paper features are aggregated to obtain the representation vector, which is the output of the aggregation layer; Step (3): In the prediction layer, the representation vector is fully connected through a multilayer perceptron and then dimensionality is reduced to output a one-dimensional value, which represents the score of the candidate paper. The paper dataset in step (1) includes user characteristics, social relationship papers, and candidate papers; The user characteristics include user paper direction characteristic x0, user school characteristic s, and user paper publication year characteristic t; Where the user's paper direction feature x0 is represented as: x0={x 0.1 ,x 0.2 ,…,x 0.n } In the above formula, x 0.1 ~x 0.n This represents the n directions of the user's paper, x 0.i This represents the i-th direction of the user's paper; The method for determining the social relationship in a paper is as follows: Step (11) identifies other students who have the same teacher as user u as user u's social relationships, represented as: {u,U1,U2…,U n }∈U t In the above formula, U1~U n This refers to other students who have the same teachers as user u and enrolled earlier than user u. t This represents the collection of students (U) of the teacher; Step (12): Based on the student set U, integrate the papers that each student is interested in to obtain the paper database J. The relationship between the papers and the students is as follows: In the above formula, J uj This refers to a collection of papers that student j is interested in. 1j ~j nj This refers to a collection of papers by student J. uj The paper in J u1 ~J un The table indicates that each student's collection of papers is in the paper database; Step (13): Determine the social relationship papers of user u: The collection of papers on user u's social relationships includes {J u1 J u2 ,…,J un }; where {U1,U2,…,U n All of them have the same teachers as user u and enrolled in school earlier than user u. For user u, their social relationships are represented by the paper {j}. 11 ,j 21 ,…j n1 ,j 12 ,j 22 ,…,j n2 ,…,j 1m ,j 2m ,…j nm }; Step (1) also includes the following steps: Step (14): Use word2vec to process user features and obtain the user's paper direction feature vector X. i : In the formula, x i This indicates obtaining the i-th word in the word vector x representing the user's paper's direction. This represents the result of using the word2vec word embedding model for the i-th direction; Step (15): Transfer the user's paper direction feature vector X i The user feature embedding vector u is obtained by concatenating the user's school feature s and the user's published paper year feature t. embedding : u embedding <[X1,X2,…,X n ,st] Among them, X1~X n Word embeddings representing an individual's research direction, s representing the user's school characteristics, and t representing the user's publication year characteristics; Step (16): The embedding vector of each paper is defined as the paper title feature, the paper publication time and the paper category feature; Step (17): Use word2vec to process the paper title features: Let title represent the paper title, then the word vector function expression of the title is: Step (18): Concatenate the paper title features, paper category features, and paper publication time to obtain the social relationship paper feature embedding vector j. embedding : j embedding =[Title,c,t] Step (19): Perform the same feature processing on the candidate papers as in steps (17) and (18) to obtain the candidate paper feature embedding vector candidate. embedding .
2. The method according to claim 1, characterized in that, Step (2) includes the following steps: Step (21): Obtain the weight w j ,w j The function expression is as follows: w j =f(u embedding ,Title social ) =Linear(MLP([u embedding ,Title social ,u embedding *Title social ])) In the formula, Title social An embedding vector representing the title of a paper on social relations; Step (22): Obtain the weight w candidate w candidate The function expression is as follows: w candidate =f(candidate embedding ,j embedding ) =linear(MLP([candidate embedding ,j embedding ,candidate embedding *j embedding ])) In the formula, j embedding Represents the complete embedding vector of a paper on social relations; Step (23): Use weight w j and weight w candidate The embedding vector of each dimension of social relationship paper features is guided, and the guided embedding vector is e. i The guided social relationship paper feature embedding vector is pooled to obtain the final social relationship paper feature embedding vector e: have been i =(w j w candidate ·j embedding ) e=pooling(e1+e2+…+e n ) In the formula, e1~e n This represents the processed user social relationship paper embedding vector, and pooling() represents the pooling layer, which performs the accumulation operation. Step (24) yields the representation vector: The user feature vector, the pooled user social relationship paper vector, and the candidate paper vector are concatenated; the concatenated vector is represented by e. concat The formula is as follows: it is concat =concat(u embedding ,e,candidate embedding ) =[u embedding ,e,candidate embedding ]。 3. The method according to claim 2, characterized in that, Step (3) specifically involves: The output vector is then passed through a fully connected multilayer perceptron for dimensionality reduction. The resulting one-dimensional value is the score of the candidate paper, calculated using the following formula: output = softmax(ReLU(e)) The inner multilayer perceptron uses ReLU as the activation function, while the outer multilayer perceptron uses softmax() as the activation function to calculate scores.