A Journal Recommendation Method Based on Contrastive Learning of Multi-Granularity Heterogeneous Attribute Graphs

Abstract

The present invention discloses a journal recommendation method based on multi-granularity heterogeneous attribute graph contrast learning, which relates to the technical field of data analysis. The journal recommendation method includes the following steps: data processing, multi-granularity semantic feature extraction, multi-granularity structural feature extraction based on contrast learning of heterogeneous graph neural networks, and adaptive learning. Based on heterogeneous graph neural networks, contrast learning, convolutional neural networks, and text preprocessing methods, the present invention processes the text into multi-granularities and processes it separately from the semantic and structural directions, uses hierarchical adaptive learning to integrate these processed features into a final optimal text feature, uses softmax to obtain the final classification result; and integrates multiple loss functions to train the model considering the particularities of papers and journals. Finally, a high-quality effect of journal recommendation for a specific research field based only on the content of papers is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and specifically, to a journal recommendation method based on multi-granularity heterogeneous attribute graph contrast learning. Background Art

[0002] In recent years, many emerging interdisciplinary research fields have emerged in academic research. These fields usually span multiple disciplines, with a large content span and a wider range of available journals for publication. This makes it often difficult for researchers, especially those in newly concerned or emerging fields, to choose appropriate journals for publication.

[0003] Current journal recommendation methods can generally be divided into two categories: content-unbased and content-based methods. Content-unbased methods generally recommend journals through information such as publication history and social networks, and are not applicable to new scholars. Most of the existing content-based journal recommendation methods are only for specific academic paper platforms or journal types, and there is currently basically no journal recommendation method for specific research fields, especially emerging interdisciplinary research fields. Content-based journal recommendation methods can essentially be regarded as a text classification problem. Current text classification methods can be divided into the following four categories:

[0004] Text classification based on traditional machine learning methods: Traditional machine learning methods, such as SVM, require a feature engineering step to convert text into embeddings. The classification performance of the method is often directly related to the quality of the embeddings. Since high-quality feature engineering depends on professional domain knowledge, it is difficult for these methods to obtain good performance.

[0005] Text classification based on deep neural networks: Deep learning models represent text as embeddings. Deep neural networks have been quite widely applied in the field of text classification. For example, TextCNN, LSTM, Transformer, etc. These methods often have high requirements for the dataset.

[0006] Text classification based on contrast learning: Contrast learning is mainly applied to self-supervised and unsupervised research tasks. Its effect in supervised research tasks is not very competitive. However, compared with supervised tasks such as text classification that require additional training costs, contrast learning has lower costs and can be used as an auxiliary method.

[0007] Text classification based on graph neural networks (GNN): The biggest difference between graph neural networks and the above methods is that GNN can not only process Euclidean space data, but also process non-Euclidean space data, that is, graph-structured data. GNN has been successfully applied to text classification and achieved good results. However, due to the complexity and diversity of the graph structure, the performance of GNN-based text classification methods varies greatly on different data, but this also means that GNN-based methods often have a high room for improvement. Summary of the Invention

[0008] The purpose of the present invention is to provide a journal recommendation method based on multi-granularity heterogeneous attribute graph contrast learning to solve the problems proposed in the background technology.

[0009] To achieve the above purpose, the present invention provides the following technical solutions:

[0010] A journal recommendation method based on multi-granularity heterogeneous attribute graph contrast learning, comprising the following steps:

[0011] Step 1) Data processing, collecting the abstracts and titles of papers, and extracting information at the document granularity, sentence granularity, word granularity, topic granularity, and title granularity of the papers;

[0012] Step 2) Multi-granularity semantic feature extraction, inputting the information at the document granularity, sentence granularity, and word granularity into the multi-granularity semantic feature extraction module, and obtaining semantic features X word , X abs and X title at the three granularities through pre-training and a feature enhancement network composed of a convolutional layer and a fully connected layer;

[0013] Step 3) Multi-granularity structural feature extraction based on heterogeneous graph neural network contrast learning, inputting the information at the document granularity, sentence granularity, word granularity, topic granularity, and title granularity into the multi-granularity structural feature extraction module based on heterogeneous graph neural network contrast learning, constructing a global heterogeneous attribute graph, constructing a global heterogeneous topological graph according to the global heterogeneous attribute graph, then constructing four local subgraphs, using heterogeneous graph convolution based on a double-layer attention mechanism to pre-train the constructed global heterogeneous attribute graph, global heterogeneous topological graph, and four local subgraphs, using cross-loop graph contrast learning to randomly select positive and negative pairs, and then further enhancing the features of the pre-trained graphs through a fully connected network to obtain structural features X graphs from six perspectives;

[0014] Step 4) Adaptive learning, using hierarchical adaptive learning to fuse the three semantic features and six structural features into an optimal text feature X fin , where multiple loss functions are integrated to train the model;

[0015] Step 5), using softmax to classify X fin to obtain the journal recommendation result.

[0016] Based on the above technical solutions, the present invention also provides the following optional technical solutions:

[0017] In an alternative solution: in step 1), the abstract is decomposed into the following three granularities: the document granularity that treats the entire abstract as a paragraph, the sentence granularity that decomposes the abstract into several sentences, and the word granularity that decomposes the abstract into several words. The BTM method is used to mine the potential topics of the abstract to form the topic granularity, and the title of the paper forms the title granularity.

[0018] In an alternative solution: in step 2), for the word granularity, the initial word embeddings are initially trained using the Word2vec method, and then the initial word embeddings are feature-enhanced through a feature enhancement network module composed of a convolutional layer and a fully connected layer to obtain the semantic features of the word granularity:

[0019] X word = g(C(X ini ) * W fc + b fc )

[0020] where X ini represents the initial word embeddings, C(*) represents the convolution operation, W fc represents the fully connected layer, b fc represents the bias term of the fully connected layer, and g(*) represents the non-linear activation function.

[0021] In an alternative solution: in step 2), for the information of the abstract and title granularities, their initial embeddings are first trained using Doc2vec and BERT, and then the corresponding granularity semantic features X abs and X title .

[0022] In an alternative solution: the specific steps of the multi-granularity structural feature extraction based on the heterogeneous graph neural network contrast learning in step 3) are as follows:

[0023] Step S1: From a structural perspective, the information of the five granularities of the document, sentence, word, title, and topic of the input text is used to construct a global heterogeneous attribute graph G att , G att has four types of nodes: D, S, T, W, and five types of connections: D-W, D-S, D-T, T-W, S-W;

[0024] Step S2: By replacing the attribute vectors of the D and S nodes in the global heterogeneous attribute graph G att with TFIDF vectors without semantic information and replacing the attribute vector of the node W with a one-hot vector without semantic information, a global heterogeneous topological graph G top is constructed;

[0025] Step S3: From the global heterogeneous topological graph G topThree local subgraphs are separated from it, consisting of each type of node except D and node D: D-W, D-T, D-S. The local subgraph contains the global heterogeneous topological graph G top For two types of nodes and their corresponding connection relationships in it, replace the attribute vector of the S node in the D-S local graph with the Sent2vec vector to obtain the fourth local subgraph D-Sv;

[0026] Step S4: Perform pre-training on the six graphs constructed above using heterogeneous graph convolution based on a two-layer attention mechanism to obtain the pre-trained vector X pre ; Extract the global heterogeneous topological graph G top 、the global heterogeneous attribute graph G att and the multi-dimensional node information and structures in the four local subgraphs, and randomly select good positive and negative pairs through cross-loop graph contrast learning;

[0027] Step S5: Input the pre-trained global heterogeneous topological graph G top 、the global heterogeneous attribute graph G att and the attribute vectors of the D nodes in the four local subgraphs into the fully connected layer for further feature enhancement, and take the outputs of these fully connected layers to obtain the structural features X of six perspectives graphs .

[0028] In an alternative solution: The construction process of the global heterogeneous attribute graph G att is as follows:

[0029] (6) Use the preprocessed paper abstract as the D node, use the sentences and words obtained by splitting and tokenizing the abstract as the S node and W node, and connect each D to the S node obtained by splitting its sentences;

[0030] (7) Use the latent topics of the abstract mined by BTM as the T node, and represent the attribute vector of the T node with the probability distribution of words. For each D node, assign the top 5 topics T with the highest probability to it and connect them;

[0031] (8) Use Doc2vec to train the document to obtain the attribute vector of the D node, use Sent2vec based on BERT to train the sentence to obtain the attribute vector of the S node, and use Word2vec to train the word to obtain the attribute vector of the W node;

[0032] (9) If a document D contains a word W, construct a connection between the corresponding D and W. Similarly, if a sentence S contains a word W, construct a connection between the corresponding S and W. The connection of T-W is constructed based on the word probability distribution of each topic;

[0033] For the title of each paper, considering the particularity of the title, use BERT to train it into an embedding and then splice it after the corresponding D node attribute vector to obtain the global heterogeneous attribute graph G att 。

[0034] In an alternative solution: The specific steps of the adaptive learning in step 4) are as follows: Step D1: Fuse the three semantic features X word 、X abs and X title into a text semantic feature X sem through adaptive learning:

[0035] X sem =∑α i X i

[0036] where X i is X word 、X abs and X title ;

[0037] Step D2: Fuse the six-perspective structural features X graphs into a text structural feature X str through adaptive learning:

[0038] X str =∑β i X i

[0039] where X i is X graphs ;

[0040] Step D3: Respectively pass these two features through the fully connected layer and then fuse them into the final optimal text feature X fin through adaptive learning:

[0041] X fin =∑γ i X i W i

[0042] where X i is X str ,X sem ; W i represents the fully connected layer.

[0043] In an alternative solution: In step 4), the loss function introduces the Asoftmax loss in the field of face recognition and the infoNCE loss in the field of contrastive learning on the basis of cross-entropy. Among them, the loss of the model includes the pre-training loss, semantic loss, structural loss, and prediction loss.

[0044] In an alternative solution: the pre-training loss is cross-entropy, where the graph pre-training loss L HGCN :

[0045]

[0046] uses Asoftmax with text semantic features as the semantic loss, where the semantic loss Ls em :

[0047]

[0048] where x ∈ X sem , m ≥ 1 represents an integer controlling the margin size. When m = 1, Asoftmax degenerates into ordinary Softmax;

[0049] The infoNCE loss L in the field of contrastive learning const :

[0050]

[0051] where H and H′ represent two graphs in the graph pair, d represents a D node in H, d+ represents a homogeneous D node in H′, d- represents a heterogeneous D node in H′, and t is a hyperparameter, called the temperature coefficient in softmax;

[0052] Combine it with the Asoftmax loss L str calculated using the structural feature X graph as the structural loss L str :

[0053] L str = δL const + L graph + η||Θ||

[0054] The final text feature X fin Calculates the prediction loss L through Asoftmax pred :

[0055]

[0056] The overall loss function of the model is:

[0057] L = λ HGCN L HGCN + λ sem L sem + λ graph L graph + λ const L const + λ pred L pred .

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] Based on the content of the paper, the present invention recommends journals. Starting from both multi-granularity and semantic structure directions, various information in the text is fully extracted; through multi-granularity and feature enhancement networks, the semantic information of the text at each granularity is fully extracted. By constructing a global heterogeneous graph and a local heterogeneous graph, combined with the proposed cross-loop graph contrast learning, the structural information of the text from various perspectives is fully extracted and mined; and considering the particularity of using journals as labels, multiple loss functions are integrated to simultaneously mine the differences and commonalities, global and local properties, and problems such as excessive intra-class differences and noisy labels in text information. The work of recommending papers and journals for specific fields is completed; and it has been experimentally proven that the effect of the model is better than existing text classification methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is the overall flowchart of a journal recommendation method based on multi-granularity heterogeneous attribute graph contrast learning of the present invention.

[0061] Figure 2 It is the schematic diagram of data preprocessing in the present invention.

[0062] Figure 3 It is the schematic diagram of multi-granularity semantic feature extraction in the present invention.

[0063] Figure 4 It is the schematic diagram of multi-granularity structural feature extraction based on heterogeneous graph neural network contrast learning in the present invention.

[0064] Figure 5 It is the schematic diagram of adaptive learning in the present invention.

[0065] Figure 6 It is the schematic diagram of a recommendation example on the LncRNA dataset of the present invention.

[0066] Figure 7 It is the schematic diagram of a recommendation example on the MicroRNA dataset of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0067] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The embodiments listed in the present invention are only used to illustrate the present invention, and are not used to limit the scope of the present invention. Any obvious modification or change made to the present invention does not depart from the spirit and scope of the present invention.

[0068] In one embodiment, as Figures 1 - 5As shown, a journal recommendation method based on multi-granularity heterogeneous attribute graph contrast learning is provided, and this method includes the following steps:

[0069] I. Data preprocessing

[0070] Extract the abstract and title of the paper, and decompose the abstract into the following three granularities: 1) The document granularity of regarding the entire abstract as a paragraph. 2) The sentence granularity of decomposing the abstract into several sentences. 3) The word granularity of decomposing the abstract into several words.

[0071] Then use the BTM method to mine the latent topics of the abstract, and together with the title of the paper, obtain the information of five granularities of the paper document, sentence, word, topic, and title. Process the preprocessed data from both semantic and structural perspectives.

[0072] II. Multi-granularity semantic feature extraction

[0073] From the semantic perspective, process the information of the document, word, and title granularities of the text to fully mine the semantic information of the text. Specifically, for the word granularity, first use the Word2vec method to initially train the initial word embeddings, and then use a feature enhancement network module composed of a convolutional layer and a fully connected layer to enhance the features of the initial word embeddings to obtain the semantic features X word :

[0074] X word = g(C(X ini ) * W fc + b fc )

[0075] Among them, X ini represents the initial word embeddings, C(*) represents the convolutional operation, W fc represents the fully connected layer, b fc represents the bias term of the fully connected layer, and g(*) represents the non-linear activation function.

[0076] For the information of the abstract and title granularities, first use Doc2vec and BERT to train their initial embeddings respectively, and then obtain the corresponding granularity semantic features X abs and X title .

[0077] III. Multi-granularity structural feature extraction based on heterogeneous graph neural network contrast learning

[0078] From the structural perspective, first, input the information of five granularities of the text document, sentence, word, title, and topic to construct a global heterogeneous attribute graph G att, it has four types of nodes: D, S, T, and W, and five types of connections: DW, DS, DT, TW, and SW. The construction process is described in detail as follows:

[0079] 1) The preprocessed paper abstract is used as the D (Document) node, and the sentences and words obtained by abstract sentence segmentation are used as S (Sentence) nodes and W (Word) nodes, and each D is connected to the S node obtained by its sentence segmentation.

[0080] 2) The potential topics of the abstracts mined by BTM are used as T (Topic) nodes. The attribute vector of the T node is represented by the probability distribution of words. For each D node, the T with the highest probability is assigned to it. o p5 topics T and connect them together, and 200 topics T are mined in the experiment.

[0081] 3) Use Doc2vec to train documents to obtain the attribute vector of the D node, use Sent2vec based on BERT to train sentences to obtain the attribute vector of the S node, and use Word2vec to train words to obtain the attribute vector of the W node.

[0082] 4) If a document D contains a word W, a connection is built between the corresponding D and W. Similarly, if a sentence S contains a word W, a connection is built between the corresponding S and W. The connection of TW is built based on the word probability distribution of each topic.

[0083] 5) In addition to the above information, there is also the title of each paper. Considering the particularity of the title, BERT is used to train it as an embedding and then spliced ​​into the corresponding D node attribute vector. The constructed global heterogeneous attribute graph is defined as G att .

[0084] In order to better capture the topological structure of the graph, a global heterogeneous topological graph G is constructed. top , which is a graph G without semantic attributes att Specifically, the topological version of graph G alt The attribute vectors of nodes D and S are replaced with TFIDF vectors without semantic information, and the attribute vector of node W is replaced with a one-hot vector without semantic information. The other structures remain unchanged. The graph constructed in this way is defined as the global heterogeneous topology graph G top .

[0085] The local structure can better represent the influence of a single type of node on the text than the global graph; to further capture the local structure of the graph, four local graphs are constructed.

[0086] Specifically, from the global heterogeneous topological graph G topThree local subgraphs are isolated from it, consisting of each type of node except D and node D: D-W, D-T, D-S. These local graphs only contain the global heterogeneous topological graph G top Two types of nodes and the corresponding connection relationships in it. Replace the attribute vector of the S node in the D-S local graph with the Sent2vec vector to obtain the fourth local graph D-Sv.

[0087] As is well known, GCN is the most commonly used method for graph neural networks. For any graph G=(V, E), let X be the matrix of nodes and their feature vectors of graph G, let A' = A + I be the adjacency matrix of graph G with self-connections added, and let M be the degree matrix, where M ii = ∑ j A′ ij , then the propagation rule of GCN is as follows:

[0088]

[0089] where W (l) is a trainable transformation matrix, σ(·) is a non-linear activation function H (l) represents the hidden representation of the nodes in the l-th layer, H (0) = X.

[0090] However, due to heterogeneity, GCN cannot be simply applied to heterogeneous graph neural networks. We use heterogeneous graph convolution (HGCN), set a transformation matrix for each type of node in the heterogeneous graph, and project different types of nodes in the graph into the same implicit common space through various transformation matrices. The propagation rule of HGCN is as follows:

[0091]

[0092] where is a submatrix of representing adjacent nodes of type r,

[0093] The heterogeneous graph convolution of the double-layer attention mechanism calculates the importance weights between different types of nodes using type-level and node-level attention. The propagation rule is as follows:

[0094]

[0095] where β r represents the double-layer attention matrix.

[0096] We use the heterogeneous graph convolution based on the double-layer attention mechanism to pre-train the six graphs constructed above to obtain the pre-trained vector Xpre :

[0097] X pre =H (L)

[0098] The double - layer attention mechanism only captures the differences between different types of node information, without considering their commonalities. These commonalities reflect the common dependencies of the text on different types of information and structures, and are also important information that needs to be captured. After extracting multi - dimensional node information and structures through the construction of the above - mentioned graph, contrast learning is introduced next to capture the commonalities of multi - dimensional information and structures.

[0099] Cross - loop graph contrast learning is proposed, which is a contrast learning method that simultaneously includes two structures of attribute - topology, and three connections of global - global, global - local, and local - local.

[0100] Specifically, every two of the above six graphs are connected as a pair. The D nodes of the same type in the graph pair are regarded as positive sample pairs, and the D nodes of different types are regarded as negative sample pairs. At this step, positive and negative sample pairs are randomly selected based on cross - loop graph contrast learning. Then, the attribute vectors of the D nodes of these six pre - trained graphs are respectively input into the fully - connected layer for further feature enhancement, and the outputs of these fully - connected layers are recorded as X graphs .

[0101] IV. Adaptive Learning

[0102] So far, semantic features X word , X abs and X title at three granularities and structural features X graphs from six perspectives are obtained. Hierarchical adaptive learning is used to fuse them into an optimal text feature.

[0103] Specifically, first, the semantic features at three granularities are fused into a text semantic feature X sem :

[0104] X sem =∑α i X i

[0105] where X i is X word , X abs and X title .

[0106] At the same time, the structural features X graphs from six perspectives are fused into a text structural feature X str :

[0107] Xstr = ∑β i X i

[0108] where X i is X graphs ;

[0109] Then, these two features are respectively passed through a fully connected layer and then fused through adaptive learning to form the final optimal text feature X str :

[0110] X str = ∑β i X i

[0111] where X i is X graphs ;

[0112] Then, these two features are respectively passed through a fully connected layer and then fused through adaptive learning to form the final optimal text feature X fin :

[0113] X fin = ∑γ i X i Wx

[0114] where X i is X str , X sem ; W i represents the fully connected layer;

[0115] Use softmax to classify X fin to obtain the journal recommendation result.

[0116] V. Loss Function

[0117] It does not rely on the cross-entropy loss function used in previous text classification methods. Instead, the Asoftmax loss in the field of face recognition and the infoNCE loss in the field of contrastive learning are introduced on the basis of cross-entropy. The loss of the model is divided into four parts:

[0118] Graph pre-training loss, semantic loss, structural loss, prediction loss.

[0119] (1) Graph pre-training loss:

[0120] It is found that although Asoftmax is more suitable for the research problem, as the number of parameters of the model increases and the number of layers deepens, the effect of the Asoftmax loss function decreases faster than that of the traditional cross-entropy. Therefore, cross-entropy is still used as the graph pre-training loss L HGCN :

[0121]

[0122] (2) Semantic loss:

[0123] Compared with the traditional cross - entropy loss, Asoftmax can significantly reduce the intra - class gap while expanding the inter - class gap, which is very suitable for this research problem where the intra - class gap of labels is greater than the inter - class gap. Therefore, the Asoftmax using semantic feature X sem is used as the semantic loss L sem :

[0124]

[0125] where x ∈ X sem , m ≥ 1 represents an integer controlling the margin size. When m = 1, Asoftmax degenerates into ordinary Softmax.

[0126] (3) Structural loss:

[0127] From a structural perspective, the positive and negative pairs selected by the cross - loop contrast learning mentioned above are used, and the infoNCE loss in the field of contrast learning is introduced to calculate the contrast learning loss L const :

[0128]

[0129] where H and H′ represent two graphs in the graph pair, d represents a D node in H, d+ represents a D node of the same type in H′, d− represents a D node of a different type in H′, and t is a hyperparameter, which is called the temperature coefficient in softmax.

[0130] Combined with the Asoftmax loss L str calculated using the structural feature X graph as the structural loss L str :

[0131] L str = δL Gonst + L graph + η||Θ||

[0132] (4) Prediction loss:

[0133] After obtaining the final text feature X fin the prediction loss L pred is also calculated using Asoftmax:

[0134]

[0135] The overall loss function of the model is:

[0136] L = λHGCN L HGCN +λ sem L sem +λ graph L graph +λ const L const +λ pred L pred 。

[0137] 1. Experimental results of the LncRNA-related paper dataset

[0138] LncRNA-related articles in the past five years were obtained from Pubmed, which is a typical emerging cross-research hot spot area. After analysis and screening, the top 50 journals with the most published articles and their papers were selected as the dataset, which contained a total of 15,114 LncRNA-related papers.

[0139] The performance of the model is shown in Table 1, and examples of the model's journal recommendations are as Figure 6 shown.

[0140] Table 1. Top1-Top10 classification accuracy results on the LncRNA dataset

[0141]

[0142] 2. Experimental results of the MicroRNA-related paper dataset

[0143] MicroRNA is also an emerging cross-research field in recent years, and its research is also a hot topic among researchers. MicroRNA-related articles in the past five years were obtained from Pubmed. After analysis and screening, the top 50 journals with the most published articles and their papers were selected as the dataset, which contained a total of 29,094 MicroRNA-related papers.

[0144] The performance of the model is shown in Table 2, and examples of the model's journal recommendations are as Figure 7 shown.

[0145] Table 2. Top1-Top10 classification accuracy results on the MicroRNA dataset

[0146]

[0147] As mentioned above, it is only the specific implementation manner of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present disclosure can easily think of changes or substitutions, which should be covered within the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. A journal recommendation method based on multi-granularity heterogeneous attribute graph contrast learning, characterized in that It includes the following steps: Step 1) Data processing, collecting the abstracts and titles of papers, and extracting information at the document granularity, sentence granularity, word granularity, topic granularity, and title granularity of the papers; Step 2) Multi-granularity semantic feature extraction. Input the information at the document granularity, sentence granularity, and word granularity into the multi-granularity semantic feature extraction module, and obtain the semantic features at the three granularities through pre-training and a feature enhancement network composed of convolutional layers and fully connected layers , and ; Step 3) Multi-granularity structural feature extraction based on contrastive learning of heterogeneous graph neural networks. Input the information of document granularity, sentence granularity, word granularity, topic granularity, and title granularity into the multi-granularity structural feature extraction module based on contrastive learning of heterogeneous graph neural networks to construct a global heterogeneous attribute graph. Construct a global heterogeneous topological graph according to the global heterogeneous attribute graph, and then construct four local subgraphs. Use heterogeneous graph convolution based on a double-layer attention mechanism to pre-train the constructed global heterogeneous attribute graph, global heterogeneous topological graph, and four local subgraphs. Use cross-loop graph contrastive learning to randomly select positive and negative pairs, and then pass the pre-trained graphs through a fully connected network respectively to further enhance the features, obtaining structural features from six perspectives ; The specific steps for multi-granularity structural feature extraction based on contrastive learning of heterogeneous graph neural networks are as follows: Step S1: From the structural perspective, information at five granularities of the document, sentence, word, title, and theme of the input text is used to construct a global heterogeneous attribute graph , It has four types of nodes: D, S, T, and W, and five types of connections: D-W, D-S, D-T, T-W, and S-W; Step S2: By replacing the attribute vectors of D and S nodes in the global heterogeneous attribute graph with TFIDF vectors without semantic information and replacing the attribute vector of node W with a one-hot vector without semantic information, a global heterogeneous topological graph is constructed; Step S3: Separate three local subgraphs from the global heterogeneous topological graph by each type of node except D and node D Composition: D-W, D-T, D-S, the local subgraph contains the global heterogeneous topological graph For the two types of nodes and the corresponding connection relationships in , replace the attribute vector of the S node in the D-S local graph with the Sent2vec vector to obtain the fourth local subgraph D-Sv; Step S4: Use heterogeneous graph convolution based on a double-layer attention mechanism to pre-train the six graphs constructed above to obtain pre-trained vectors ; Extract the global heterogeneous topological graph , the global heterogeneous attribute graph and the multi-dimensional node information and structures in the four local subgraphs, and use cross-loop graph contrast learning to randomly select good positive and negative pairs; Step S5: Input the pre-trained global heterogeneous topological graph , the global heterogeneous attribute graph and the attribute vectors of the D nodes of the four local subgraphs into the fully connected layer for further feature enhancement, and obtain the structural features of six perspectives from the outputs of these fully connected layers ; Step 4) Adaptive learning, using hierarchical adaptive learning to fuse three semantic features and six structural features into an optimal text feature , where multiple loss functions are integrated to train the model; Step 5), use softmax to perform classification to obtain the journal recommendation results.

2. The journal recommendation method based on multi-granularity heterogeneous attribute graph contrast learning according to claim 1, characterized in that In the said Step 1), the paper is decomposed into the following five granularities: the document granularity with the entire abstract regarded as a paragraph, the sentence granularity with the abstract decomposed into several sentences, the word granularity with the abstract decomposed into several words, then the BTM method is used to mine the potential topics of the abstract to form the topic granularity, and in addition, the title granularity formed by the title of the paper.

3. The journal recommendation method based on multi-granularity heterogeneous attribute graph contrast learning according to claim 1, characterized in that In step 2), for the word granularity, the initial word embeddings are initially trained using the Word2vec method, and then a feature enhancement network module composed of a convolutional layer and a fully connected layer is used to enhance the features of the initial word embeddings to obtain semantic features at the word granularity: ; Among them, represents the initial word embedding, represents the convolution operation, represents the fully connected layer, represents the bias term of the fully connected layer, represents the non-linear activation function.

4. The journal recommendation method based on multi-granularity heterogeneous attribute graph contrast learning according to claim 3, characterized in that In step 2), for the information at the abstract and title granularities, first use Doc2vec and BERT to train their initial embeddings, and then obtain the semantic features at the corresponding granularities through the feature enhancement network module respectively. and .

5. For the journal recommendation method based on multi-granularity heterogeneous attribute graph contrast learning according to claim 1, the construction process of the global heterogeneous attribute graph is as follows: (1) Take the preprocessed paper abstract as the D node, take the sentences and words obtained by sentence segmentation and word segmentation of the abstract as the S node and the W node, and connect each D with the S nodes obtained by its sentence segmentation; (2) Take the latent topics mined from the abstract using BTM as the T node, and represent the attribute vector of the T node with the probability distribution of words. For each D node, assign the top 5 topics T with the highest probability to it and connect them; (3) Use Doc2vec to train the document to obtain the attribute vector of the D node, use Sent2vec based on BERT to train the sentence to obtain the attribute vector of the S node, and use Word2vec to train the word to obtain the attribute vector of the W node; (4) If a document D contains a word W, a connection is built between the corresponding D and W. Similarly, if a sentence S contains a word W, a connection is built between the corresponding S and W. The T-W connection is built based on the word probability distribution for each topic; (5) For the title of each paper, considering the particularity of the title, use BERT to train it into an embedding and then splice it after the attribute vector of the corresponding D node to obtain the global heterogeneous attribute graph .

6. The journal recommendation method based on contrastive learning of multi-granularity heterogeneous attribute graphs according to claim 1, wherein The specific steps of the above step 4) of adaptive learning are as follows: Step D1: Fuse the three semantic features , and into a text semantic feature through adaptive learning; ； Among them, is , and ; Step D2: Fuse the structural features from six perspectives into a text structural feature through adaptive learning ; ； Among them, is ; Step D3: Respectively pass these two features through the fully connected layer and then fuse them through adaptive learning to form the final optimal text feature ; ； Among them, is , ; represents a fully connected layer.

7. The journal recommendation method based on contrastive learning of multi-granularity heterogeneous attribute graphs according to claim 6, wherein In Step 4), the loss function introduces the Asoftmax loss in the field of face recognition and the infoNCE loss in the field of contrastive learning on the basis of cross-entropy. Among them, the loss of the model includes pre-training loss, semantic loss, structural loss, and prediction loss.

8. The journal recommendation method based on contrastive learning of multi-granularity heterogeneous attribute graphs according to claim 7, wherein The pre-training loss is cross-entropy, where the graph pre-training loss : ； Use Asoftmax with text semantic features as the semantic loss, where the semantic loss : ； Among them, , represents an integer for controlling the size of the corner margin, when , Asoftmax degenerates into ordinary Softmax; InfoNCE Loss in the Field of Contrastive Learning ; ； Among them, H and H' represent two graphs in the graph pair, d represents a D node in H, d+ represents a D node of the same type in H', d- represents a D node of a different type in H', and t is a hyperparameter, called the temperature coefficient in softmax; Combine it with the Asoftmax loss calculated using the structural feature as the structural loss ; ； Final text features Calculate the prediction loss through Asoftmax ; ； The overall loss function of the model is: 。

Images