A method for recommending graduate literature resources based on big data technology

Abstract

This invention discloses a method for recommending graduate literature resources based on big data technology, including: a similarity-based pre-recommendation method and a deep learning-based personalized recommendation method. The similarity-based pre-recommendation method initially measures the literature of interest to the user, generates a recommendation table for the user, and collects user-project interaction information through user feedback. The deep learning-based personalized recommendation method first generates personalized user and project representations, uses a graph neural network to construct user-project interaction associations, and uses a learnable hypergraph network to establish global user-user and project-project associations. Combined with the pre-recommendation method, it generates a label-optimized recommendation strategy. This invention introduces a personalized recommendation strategy based on the patterns of graduate students' research-based learning activities. By employing big data information technology, it achieves precise literature resource recommendations driven by the personalized research-based learning needs of graduate students.

Description

Technical Field

[0001] The present invention relates to the field of personalized document resource recommendation, and in particular to a method for recommending graduate document resources based on big data technology. Background Art

[0002] With the rapid development of internet technology, online literature search and learning, using the internet as a medium, have become a common practice for researchers to access academic information and stay abreast of research progress and cutting-edge research. However, the vast amount of literature resources available on academic search engines has also led to serious problems of information overload and asymmetry, as well as learning difficulties. Introducing personalized recommendation strategies is an effective way to address these issues. Implementing personalized literature resource recommendations relies on analyzing and modeling the needs of researchers, the users of these resources. Compared to experienced researchers, graduate students are relatively new to research, with limited research experience and potentially few or no published papers. Traditional modeling and analysis methods that analyze their research interests based on their existing research activities and results are difficult to adapt to the realities of graduate students. Graduate students' research topics primarily originate from their advisors or research teams. The existing research foundation of advisors or research teams is a crucial foundation for their research projects. When conducting literature searches, in addition to focusing on cutting-edge developments in their research fields, they also consider the research progress of teams working on the same or similar topics. Currently, there are no literature resource recommendation methods specifically tailored to graduate students. To address this, the inventors have designed a personalized literature resource recommendation method based on the patterns of graduate students' research learning activities and leveraging big data technology. Summary of the Invention

[0003] The embodiment of the present invention provides a graduate student literature resource recommendation method based on big data technology, which recommends personalized literature of interest to different graduate students based on their basic information and literature reading habits, thereby improving the graduate student literature retrieval efficiency and learning efficiency.

[0004] The embodiment of the present invention provides a method for recommending graduate student literature resources based on big data technology, including:

[0005] Obtain information on research areas of interest to users;

[0006] The similarity-based pre-recommendation method matches the user's research field information with the preset document type to obtain the similarity;

[0007] Arrange the document types in descending order of similarity to obtain a preliminary user document recommendation table;

[0008] Obtain user feedback information;

[0009] Based on user feedback information and preliminary user literature recommendation table, user-item interaction information is obtained;

[0010] A personalized recommendation method based on deep learning recommends literature resources by module based on user-item interaction information;

[0011] The personalized recommendation method based on deep learning: builds a self-supervised deep learning grid, including: a user / item personalized representation module; a user-item interaction representation module; a user-user and item-item global association module; a self-supervised reinforcement learning module and an iterative update module.

[0012] In this solution, the similarity-based pre-recommendation method further includes:

[0013] Obtain user's representation text information;

[0014] Based on the preset vector space model (VSM), the user representation is mapped from text form to a high-dimensional vector space (user);

[0015] Mapping different document types into the same space (project);

[0016] Match the user-item vectors in the corresponding space to obtain the similarity;

[0017] By judging whether the similarity between the user-item vectors in the space is greater than the preset similarity threshold, if so, the corresponding document type is set to the type of document that the corresponding user is interested in; if not, the user is not interested.

[0018] In this solution, the similarity-based pre-recommendation method further includes:

[0019] Sort the documents by similarity from high to low;

[0020] Obtain the binary classification judgment information made by the user on the matching between the recommended literature resources and their current research needs;

[0021] According to the binary classification judgment information of the user terminal on the matching between the recommended literature resources and the current research needs of the user terminal, the next recommended literature resource information and the user-project interaction table are obtained.

[0022] In this solution, the user / item personalized representation module is specifically:

[0023] Get user information;

[0024] Based on the fusion of multiple user information, the user's personalized representation information is obtained;

[0025] Send the user information to the preset embedding layer to obtain the user information and generate embedded representation information;

[0026] The attention network is used to measure the influence of different features on the recommendation task, and the fully connected network layer is used to fuse different weighted features to obtain the user's primary representation information;

[0027] Generate embedding representations for different document types based on the preset embedding layer to obtain primary representation information of the project.

[0028] In this solution, the user-item interaction representation module is specifically:

[0029] According to the preset message passing mechanism of the graph convolutional network, the updated user / item representation information is obtained;

[0030] Send the item-interaction table information to the preset graph convolutional network to obtain the user-item interaction graph;

[0031] The user-item local interaction relationship is learned by aggregating adjacent node representations to obtain the first comparative learning view.

[0032] In this solution, the user-user and project-project global association modules are specifically:

[0033] Adopting learnable hypergraph structure for association learning;

[0034] The learnable hypergraph structure consists of a set of learnable hyperedges, with different users / items as different nodes. Each learnable hyperedge acts as an information hub, connecting all users / items with different weights. From a global perspective, the information of all users / items is aggregated to update the embedding representation of its own node, thereby obtaining a second comparative learning view.

[0035] Different hyperedges serve as different channels to capture the complex connection relationships between users and users, and items and items from multiple semantic dimensions.

[0036] In this solution, the learnable hypergraph network structure is specifically:

[0037] Generate a learnable hypergraph parameter matrix through low-rank decomposition method;

[0038] According to the user primary representation and the item primary representation, a learnable multi-layer perceptron network is combined to generate a learnable hypergraph structure.

[0039] In this solution, the self-supervised reinforcement learning module is specifically:

[0040] Adopting contrastive learning method, different views are used as contrastive learning objects, and the mutual information maximization mechanism is used to optimize the model gradient;

[0041] The first contrastive learning view and the second contrastive learning view are contrastively learned; the same user / item in different views is used as a positive example pair; different users / items in different views are used as negative example pairs, and the mutual information maximization function is used as the optimization equation to obtain a discriminative user / item representation.

[0042] In this solution, the self-supervised reinforcement learning module also includes:

[0043] Use data enhancement to alleviate data noise in user data;

[0044] The user-item interaction graph is randomly masked and sent to the graph convolutional network to obtain the third contrastive learning view;

[0045] The third contrastive learning view is co-supervised with the first contrastive learning view.

[0046] In this solution, the iterative update module specifically adopts an iterative update strategy to periodically update the user-item interaction table and model parameters.

[0047] The present invention effectively solves the problems of data sparsity, data noise and over-smoothing of traditional graph neural networks in the document recommendation scenario by adopting a self-supervised reinforcement learning method and a learnable hypergraph structure. In particular, for the problem of data sparsity, a local-global feature comparison learning strategy is adopted to allow deep learning models to perform collaborative supervision from a local-global perspective and make up for each other's information loss. At the same time, the information maximization function is used as an optimization strategy to guide the model to learn richer knowledge by constructing additional supervision tasks. For the problem of data noise, a data enhancement strategy is used to mask the user-item interaction edges to exclude target documents that the user is not interested in in the user-item interaction table. For the problem of over-smoothing of traditional graph neural networks, a learnable hypergraph structure is designed to aggregate all user / item information at once from a global perspective, avoiding the problem of node over-smoothing caused by multiple aggregations of graph neural networks.

[0048] Furthermore, this invention uses a pre-recommendation strategy to collect user-item interaction information, generating supervised labels for subsequent personalized recommendation algorithms. Traditional unsupervised methods make rough recommendations based solely on single pieces of information, such as a graduate student's research field, failing to capture their reading patterns. By designing a personalized recommendation algorithm, we can deeply explore graduate students' reading preferences and interests, recommending more valuable literature, reducing literature search time and improving learning efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flowchart of a method for recommending graduate student literature resources based on big data technology is shown in the present invention;

[0050] Figure 2 The present invention shows a method for recommending graduate student literature resources based on big data technology.

[0051] Figure 3 The figure shows an architecture diagram of a personalized recommendation method based on deep learning provided by the present invention. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0053] In particular, a method for recommending graduate literature resources based on big data technology can be regarded as a graduate literature recommendation system. In the description of this invention, according to the description convention of general recommendation systems, graduate students are described as users and literature types are described as projects. Let the user set U = {u1,u2,u3,…,u j ,…,u J}, item set I={i1,i2,i3,…,i k ,…,i K}.

[0054] Figure 1 A flowchart of a method for recommending graduate student literature resources based on big data technology is shown in the present invention.

[0055] like Figure 1 As shown, the present invention discloses a method for recommending graduate literature resources based on big data technology, comprising:

[0056] S102, obtaining information on research areas of interest to the user;

[0057] S104, a similarity-based pre-recommendation method matches the user's research field information with the preset document type to obtain similarity;

[0058] S106, arranging the document types in descending order of similarity to obtain a preliminary user document recommendation table;

[0059] S108, obtaining user feedback information;

[0060] S110, obtaining user-item interaction information based on user feedback information and the preliminary user document recommendation table;

[0061] S112, a personalized recommendation method based on deep learning, recommends literature resources according to modules based on user-item interaction information.

[0062] It should be noted that keywords are extracted from the information on the research field in which the user is interested, and the keywords are set as user "tags". The documents with the highest similarity to the information on the research field in which the user is interested are prioritized, and a preliminary user document recommendation table is obtained in descending order of similarity and fed back to the user end to obtain user feedback on the preliminary user document recommendation table, for example: dissatisfied with document A, but satisfied with document B. Based on the feedback information and the preliminary user document recommendation table, user-project interaction information is obtained. The personalized recommendation method based on deep learning requires the construction of a self-supervised deep learning grid, including: a user / project personalized representation module; a user-project interaction representation module; a user-user, project-project global association module; a self-supervised reinforcement learning module, and an iterative update module.

[0063] According to an embodiment of the present invention, the similarity-based pre-recommendation method further includes:

[0064] Obtain user's representation text information;

[0065] Based on the preset vector space model (VSM), the user representation is mapped from text form to a high-dimensional vector space (user);

[0066] Mapping different document types into the same space (project);

[0067] Match the user-item vectors in the corresponding space to obtain the similarity;

[0068] By judging whether the similarity between the user-item vectors in the space is greater than the preset similarity threshold, if so, the corresponding document type is set to the type of document that the corresponding user is interested in; if not, the user is not interested.

[0069] It's important to note that the similarity-based pre-recommendation method uses a vector space model (VSM) to map each user and project from textual form into a high-dimensional vector space (user). Similarly, the VSM is used again to map different document types into the same space (project). By determining the similarity between the user-project vectors in this space, it initially recommends documents of potential interest to the user. For example, if the preset similarity threshold is 80, then when the similarity between the user-project vectors in the corresponding space is greater than 80, the corresponding document type is considered to be of potential interest to the user.

[0070] According to an embodiment of the present invention, the similarity-based pre-recommendation method further includes:

[0071] Sort the documents by similarity from high to low;

[0072] Obtain the binary classification judgment information made by the user on the matching between the recommended literature resources and their current research needs;

[0073] According to the binary classification judgment information of the user terminal on the matching between the recommended literature resources and the current research needs of the user terminal, the next recommended literature resource information and the user-project interaction table are obtained.

[0074] It should be noted that the similarity-based pre-recommendation method sorts documents by similarity from high to low, recommending the n most relevant documents to the user each time, where n represents a natural integer greater than 0, for example, n = 20. The user can make a binary judgment on whether the recommended document resources match their current research needs (satisfactory or not), and the next recommendation is automatically generated after the user completes the judgment.

[0075] Specifically, the document uses the vector space model to map the document title and keywords to the high-dimensional vector space, and performs relevance matching with the project (document type). The most relevant documents are first selected from the first relevant project with the highest relevance of n documents. If the user is satisfied with the recommended documents, then the n+1-2n documents are selected from the most relevant project based on relevance. If the user is dissatisfied, then the first n most relevant documents are selected from the second relevant project for recommendation, and so on.

[0076] By iteratively performing the above operations, a user-item interaction table is collected to generate label data for the deep learning-based personalized recommendation method. That is, if the user is satisfied with the recommended document, the user's interaction count for the item is increased by 1; otherwise, no action is taken.

[0077] According to an embodiment of the present invention, the user / item personalized representation module is specifically:

[0078] Get user information;

[0079] Based on the fusion of multiple user information, the user's personalized representation information is obtained;

[0080] Send the user information to the preset embedding layer to obtain the user information and generate embedded representation information;

[0081] The attention network is used to measure the influence of different features on the recommendation task, and the fully connected network layer is used to fuse different weighted features to obtain the user's primary representation information;

[0082] Generate embedding representations for different document types based on the preset embedding layer to obtain primary representation information of the project.

[0083] It should be noted that the user information includes: user registration information, such as which college, what major, etc. The user / project personalized representation module generates a user personalized representation by integrating multiple user information. Multiple embedding layers are used to generate a personalized representation for each user u. j Generate an embedding representation for each basic information f (such as major, research interests, etc.) of k Generate embedded representation as shown in formula (1):

[0084]

[0085] in Represents user u j and Project i k The embedded representation of the embedded representation of the f-th basic information; Represent the learnable parameters of the embedding layer of users and items respectively, and d represents the number of embedding dimensions. For the embedding representation of multiple user information, vector splicing is used for fusion. like Where F represents the number of basic user information. On top of this, the attention network is used to measure the influence of different features on the recommendation task. In particular, the self-attention network is used to represent the primary user Generate three matrices, namely query key Sum Q, K and V are used in the self-attention mechanism operation. Mapped into three different dimensions of latent representation. Specifically, Q, K and V are represented by three learnable weights and As shown in formula (2):

[0086]

[0087] Among them, Q and K are used to calculate the attention weights. The calculation results are multiplied by V to calculate the influence of different information features on the recommendation task and generate the user primary representation z (u) , as shown in formula (3):

[0088]

[0089] Among them, Attention represents the calculation method of attention weight, softmax represents the nonlinear activation function, and similarly, the embedding layer is used to generate embedding representations for different document types to generate the primary representation z of the project. (i) .

[0090] According to an embodiment of the present invention, the user-item interaction representation module is specifically:

[0091] According to the preset message passing mechanism of the graph convolutional network, the updated user / item representation information is obtained;

[0092] Send the item-interaction table information to the preset graph convolutional network to obtain the user-item interaction graph;

[0093] The user-item local interaction relationship is learned by aggregating adjacent node representations to obtain the first comparative learning view.

[0094] It should be noted that the user-item interaction representation module uses the message passing mechanism of the graph convolutional network to update the user / item representation. The graph convolutional network generates a user-item interaction graph based on the user-item interaction table, learns the local user-item interaction relationship by aggregating adjacent node representations, and generates a first comparative learning view;

[0095] The user-item interaction table records the interaction between users and items, and the user-item interaction graph is represented as If user u j Worked with Project I k If an interaction occurs, then A j,k =1; otherwise, A j,k = 0. The graph convolutional network performs message passing in the form of formula (4):

[0096]

[0097] in, and Represents the distance from adjacent items / users to the central node u j and i k Aggregate information, σ represents the LeakyRelu activation function, which is used to generate nonlinear representation. In order to solve the scale change problem caused by the inconsistent number of interactions, a standardized adjacency matrix is ​​used to generate In terms of form, it can be updated as formula (5):

[0098]

[0099] According to an embodiment of the present invention, the user-user and project-project global association module is specifically:

[0100] Adopting learnable hypergraph structure for association learning;

[0101] The learnable hypergraph structure consists of a set of learnable hyperedges, with different users / items as different nodes. Each learnable hyperedge acts as an information hub, connecting all users / items with different weights. From a global perspective, the information of all users / items is aggregated to update the embedding representation of its own node, thereby obtaining a second comparative learning view.

[0102] Different hyperedges serve as different channels to capture the complex connection relationships between users and users, and items and items from multiple semantic dimensions.

[0103] It should be noted that the user-user and project-project global association modules use a learnable hypergraph structure for association learning. The learnable hypergraph structure consists of a set of learnable hyperedges. The hyperedges for global user association and global project association are represented as and Where h represents the number of hyperedges. The learnable hyperedge treats different users / projects as different nodes. Each learnable hyperedge acts as an information hub, connecting all users / projects with different weights. It aggregates the information of all users / projects from a global perspective to update the embedding representation of its own node, generating a second comparative learning view. This can be formally expressed as Equation (6):

[0104] Ψ (u) =σ(H (u) ·H (u)T ·g (u) ),Ψ (i) =σ(H (i) ·H (i)T ·g (i) ) (6)

[0105] in, For the first contrastive learning view, the user representation g is generated. (u) , first use the transposed learnable hyperedge H (u)T Will g (u) The information of each user in is passed to the hyperedge, at this time, the vector size dimension is Updated to Then again with H (u) Performing matrix multiplication, each hyperedge passes the aggregated global user information back to the representation of each user, and the vector size dimension is determined by Updated to In particular, different hyperedges serve as different channels to capture the complex connection relationships between users and users and items and items from multiple semantic dimensions.

[0106] According to an embodiment of the present invention, the learnable hypergraph network structure is specifically:

[0107] Generate a learnable hypergraph parameter matrix through low-rank decomposition method;

[0108] According to the user primary representation and the item primary representation, a learnable multi-layer perceptron network is combined to generate a learnable hypergraph structure.

[0109] It should be noted that the learnable hypergraph network structure uses a low-rank decomposition method to generate a learnable hypergraph parameter matrix in order to reduce the number of parameters in the neural network model. Based on the user primary representation and the item primary representation, combined with the multi-layer perceptron network, a learnable hypergraph structure is generated, which is formally expressed as formula (7):

[0110] H (u) =MLP u (z (u) W (uh) ),H (i) =MLP i (z (i) W (ih) ) (7)

[0111] in is a learnable parameter. The hidden layer dimension is much smaller than the number of users (i.e., d<<J). Constructing hyperedges through low-rank decomposition can significantly reduce the number of parameters, shorten the neural network training time, and improve recommendation efficiency. The multi-layer perceptron network (MLP) contains multiple linear layers and nonlinear activation functions, which can be formalized as shown in formula (8):

[0112] MLP(x)=w2·σ(w1·x+b1)+b2 (8)

[0113] in, is the trainable weight parameter, is a trainable bias parameter.

[0114] According to an embodiment of the present invention, the self-supervised reinforcement learning module is specifically:

[0115] Adopting contrastive learning method, different views are used as contrastive learning objects, and the mutual information maximization mechanism is used to optimize the model gradient;

[0116] The first contrastive learning view and the second contrastive learning view are contrastively learned; the same user / item in different views is used as a positive example pair; different users / items in different views are used as negative example pairs, and the mutual information maximization function is used as the optimization equation to obtain a discriminative user / item representation.

[0117] It should be noted that the self-supervised reinforcement learning module adopts a contrastive learning method, takes different views as contrastive learning objects, and optimizes the model gradient using the mutual information maximization mechanism. In particular, the first contrastive learning view g generated by the graph convolutional network is (u) The second comparative learning view Ψ generated by the learnable hypergraph network (u) Perform contrastive learning, taking the same user / item in different views as positive pairs and different users / items in different views as negative pairs, and using the mutual information maximization function as the optimization equation to enable local features and global features to collaborate and supervise each other, generating discriminative user / item representations. The user mutual information maximization function is described as formula (9):

[0118]

[0119] where cos(·) represents the cosine similarity, which is used to measure the similarity of embeddings of different contrastive learning views; is the temperature coefficient used to adjust the gradient scale. By using the mutual information maximization function as the loss function, the distance between the representations of the same user in the first and second comparative learning views is shortened in high-dimensional space, while the distance between the representations of different users in the first and second comparative learning views is expanded. This allows the algorithm to collaboratively supervise local and global features under self-supervision, further enhancing user representation. Similarly, the item mutual information maximization function is set in a similar manner.

[0120] According to an embodiment of the present invention, the self-supervised reinforcement learning module further includes:

[0121] Use data enhancement to alleviate data noise in user data;

[0122] The user-item interaction graph is randomly masked and sent to the graph convolutional network to obtain the third contrastive learning view;

[0123] The third contrastive learning view is co-supervised with the first contrastive learning view.

[0124] It should be noted that the self-supervised reinforcement learning module uses data enhancement to alleviate the data noise problem. By randomly masking the user-item interaction graph A, a data-enhanced user-item interaction graph is generated. The data is input into the graph convolutional network to generate the third contrastive learning view. The data enhancement method is described as formula (10):

[0125]

[0126] Among them, A′ j,k Represents user u after data enhancement j and Project Il The connection relationship, Represents a random number in the range (0,1), M j,l Represents a random number in the jth row and kth column, corresponding to A′ j,k ∈ represents the preset threshold. If the random number M j,k If it is less than ∈, the user-item interaction graph is masked, otherwise the original graph is retained. The third contrastive learning view replaces A in formula (5) with A′ and performs graph convolution operation to generate the third contrastive learning view embedding representation g′ (u) , g′ (i) Subsequently, the contrastive learning strategy described in formula (9) is adopted to collaboratively supervise the third contrastive learning view with the first contrastive learning view;

[0127] In particular, the personalized recommendation method based on deep learning, after obtaining the user-item relevance representation, makes recommendations according to the literature recommendation method of the similarity-based pre-recommendation method.

[0128] According to an embodiment of the present invention, the iterative updating module specifically adopts an iterative updating strategy to periodically update the user-item interaction table and model parameters.

[0129] It should be noted that the iterative update strategy primarily includes 1) updating the user-item interaction table: Based on graduate students' assessment of the compatibility of recommended literature resources with their current research needs, the user-item interaction table is periodically updated. This serves as the foundation for constructing the graph convolutional network user-item interaction graph and provides more labels for the personalized recommendation algorithm model. 2) updating model parameters: Over time, as the user population changes and the number of literature increases, both user and literature representations need to be updated. This strategy employs the principle of pre-training followed by fine-tuning, retraining the original model parameters to update them.

[0130] Figure 2 The figure shows an architecture diagram of a graduate student literature resource recommendation method based on big data technology provided by the present invention.

[0131] As shown in the figure, the present invention provides a method for recommending graduate literature resources based on big data technology, including: a similarity-based pre-recommendation method and a deep learning-based personalized recommendation method. The similarity-based pre-recommendation method uses user information and user interaction data as a basis to determine a preliminary user literature recommendation table based on similarity. The deep learning-based personalized recommendation method includes: a user and project personalized representation module; a user-project interaction representation module; a user-user and project-project global association module; a self-supervised reinforcement learning module; and an iterative update module.

[0132] Figure 3 The figure shows an architecture diagram of a personalized recommendation method based on deep learning provided by the present invention.

[0133] As shown in the figure, the modules in the personalized recommendation method based on deep learning are displayed independently but closely connected, among which: the first comparative learning view generated by the user-item interaction representation module, the second comparative learning view generated by the user-user, item-item global association module, and the third comparative learning view generated by the self-supervised reinforcement learning module are associated with each other by comparing them.

[0134] In the embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0135] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0136] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0137] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0138] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

Claims

1. A method for recommending graduate literature resources based on big data technology, characterized in that: include: Obtain information on research areas of interest to users; The similarity-based pre-recommendation method matches the user's research field information with the preset document type to obtain the similarity; Arrange the document types in descending order of similarity to obtain a preliminary user document recommendation table; Obtain user feedback information; Based on user feedback information and preliminary user literature recommendation table, user-item interaction information is obtained; A personalized recommendation method based on deep learning recommends literature resources by module based on user-item interaction information; The personalized recommendation method based on deep learning: build a self-supervised deep learning grid, including: user / item personalized representation module; user-item interaction representation module; user-user and item-item global association module; self-supervised reinforcement learning module and iterative update module; The user / item personalized representation module is specifically: Get user information; Based on the fusion of multiple user information, the user's personalized representation information is obtained; Send the user information to the preset embedding layer to obtain the user information and generate embedded representation information; The attention network is used to measure the influence of different features on the recommendation task, and the fully connected network layer is used to fuse different weighted features to obtain the user's primary representation information; Generate embedding representations for different document types based on the preset embedding layer to obtain primary representation information of the project; The user-item interaction representation module is specifically: According to the preset message passing mechanism of the graph convolutional network, the updated user / item representation information is obtained; Send the item-interaction table information to the preset graph convolutional network to obtain the user-item interaction graph; The first comparative learning view is obtained by aggregating adjacent node representations to learn the local user-item interaction relationship; The user-user and project-project global association modules are specifically: Adopting learnable hypergraph structure for association learning; The learnable hypergraph structure consists of a set of learnable hyperedges, with different users / items as different nodes. Each learnable hyperedge acts as an information hub, connecting all users / items with different weights. From a global perspective, the information of all users / items is aggregated to update the embedding representation of its own node, thereby obtaining a second comparative learning view. Different hyperedges serve as different channels to capture the complex connection relationships between users and items from multiple semantic dimensions; The learnable hypergraph network structure is specifically: Generate a learnable hypergraph parameter matrix through low-rank decomposition method; Generate a learnable hypergraph structure based on the user primary representation and the item primary representation in combination with a learnable multi-layer perceptron network; The self-supervised reinforcement learning module is specifically: Adopting contrastive learning method, different views are used as contrastive learning objects, and the mutual information maximization mechanism is used to optimize the model gradient; Performing comparative learning on the first contrastive learning view and the second contrastive learning view; using the same user / item in different views as positive example pairs; using different users / items in different views as negative example pairs, and using the mutual information maximization function as the optimization equation to obtain discriminative user / item representations; The iterative update module specifically adopts an iterative update strategy to periodically update the user-item interaction table and model parameters.

2. The method for recommending graduate student literature resources based on big data technology according to claim 1, characterized in that: The similarity-based pre-recommendation method further includes: Obtain user's representation text information; Based on the preset vector space model, the user representation is mapped from text form to a high-dimensional vector space; Mapping different document types into the same space; Match the user-item vectors in the corresponding space to obtain the similarity; By judging whether the similarity between the user-item vectors in the space is greater than the preset similarity threshold, if so, the corresponding document type is set to the type of document that the corresponding user is interested in; if not, the user is not interested.

3. The method for recommending graduate student literature resources based on big data technology according to claim 1, characterized in that: The similarity-based pre-recommendation method further includes: Sort the documents by similarity from high to low; Obtain the binary classification judgment information made by the user on the matching between the recommended literature resources and their current research needs; According to the binary classification judgment information of the user terminal on the matching between the recommended literature resources and the current research needs of the user terminal, the next recommended literature resource information and the user-project interaction table are obtained.

4. The method for recommending graduate student literature resources based on big data technology according to claim 1, characterized in that: The self-supervised reinforcement learning module further includes: Use data enhancement to alleviate data noise in user data; The user-item interaction graph is randomly masked and sent to the graph convolutional network to obtain the third contrastive learning view; The third contrastive learning view is co-supervised with the first contrastive learning view.

Images