Academic recommendation method and system based on local multi-network community discovery method

Abstract

The invention discloses an academic recommendation method and system based on a local multi-network community discovery method, and the method comprises the steps: S1, collecting common author information and literature reference information in literatures in dblp, constructing an author network G0, a literature network G1, and an interlayer link P between the two networks, and representing the author-literature network G = {G0, G1, P}; s2, according to the author network G0 and the literature network G1 constructed in the step S1, a local multi-network community discovery method based on the node membership degree is gradually expanded outwards from a target author or a target literature to obtain an author community C0 and a literature community C1, and the author community C0 and the literature community C1 are expressed as C = {C0, C1}; and S3, giving any author or literature, setting a node corresponding to the author or literature as a seed node, executing the step S2 to obtain a community C = {C0, C1}, and recommending an author group and literature in closer contact with the target author or literature to the user. According to the method, the author network, the literature network and the relation between the author network and the literature network are fully considered, the quality of community structures in the author network and the literature network can be improved to the maximum extent only by using local information, and meanwhile, the cross-network consistency is ensured; and thus, author groups and literatures which are in closer contact with the target authors or the target literatures are recommended to the user.

Description

Technical Field

[0001] The present invention relates to a recommendation method, in particular to an academic recommendation method based on a local multi-network community discovery method. Background Art

[0002] With the rapid development of academic research, the accumulation of academic resources and the increasing complexity of research fields, the demand for effective information acquisition and personalized recommendation in the academic community is becoming stronger and stronger. As an important branch of information recommendation technology, an academic recommendation system aims to provide relevant academic resources (such as literature, authors, research directions, etc.) according to the interests and needs of users (such as researchers, scholars, etc.). Based on manual research of relevant academic resources, it requires a large amount of manpower and the obtained results are prone to problems such as one-sidedness and subjectivity. Summary of the Invention

[0003] The technical problem to be solved by the present invention is how to recommend to users an accurate group of authors closely related to the target author or target literature and related literature.

[0004] The present invention solves the above technical problem by the following technical means: an academic recommendation method based on a local multi-network community discovery method, comprising the following steps:

[0005] S1. In dblp, collect co-author information and literature citation information in the literature to construct an author network G 0 and a literature network G 1 , and an inter-layer link P between the two networks, denoted as an author-literature network G = {G 0 , G 1 , P};

[0006] S2. According to the author network G 0 and the literature network G 1 constructed in step S1, gradually expand outward from the target author or target literature based on the local multi-network community discovery method based on node membership to obtain an author community C 0 and a literature community C 1 , denoted as C = {C 0 , C 1};

[0007] S3. Given any author or literature, set the corresponding node of the author or literature as a seed node and execute step S2 to obtain a community C = {C 0 , C 1}, and recommend to the user a group of authors and literature that are more closely related to the target author or literature.

[0008] As a further optimized technical solution, step S1 specifically includes the following steps:

[0009] S11. Extract the co - author information of the documents in DBLP to establish the author network G 0 = <V 0 , E 0 >, where any node in the set V 0 represents an author, and E 0 is the adjacency matrix representing the edge relationship. If there is a relationship that authors i and j in the set V 0 co - completed a certain document, then there is an edge relationship between the nodes corresponding to the two authors, that is, in the adjacency matrix E 0

[0010] S12. Extract the document citation information of the documents in DBLP to establish the document network G 1 = <V 1 , E 1 >. Any node in the set V 1 represents a document, and E 1 is the adjacency matrix representing the edge relationship. If document i in the set V 1 cites document j, then there is an edge relationship between the nodes corresponding to the two documents, that is, in the adjacency matrix E 1

[0011] S13. Extract the corresponding relationship between the authors and the network between the author network G 0 and the document network G 1 , that is, if authors A1, A2, A3 in the author network G 0 co - completed the document P in the document network G 1 , then the edges (A1, P), (A2, P), (A3, P) connecting the author network G 0 and the document network G 1 are used as inter - layer edges to construct the inter - layer link P between the two networks;

[0012] S14. Execute steps S11 - S13 to obtain the author - document network G = {G 0 , G 1 , P}.

[0013] As a further optimized technical solution, step S2 specifically includes the following steps:

[0014] S21. Based on the author - document network G = {G 0 , G 1 , P} extracted in step S1, assume that the seed node is the node s in the author network G 0 , and obtain the document network G 1 ​​The set of nodes {s 1 , …, s k} in 0 initializes the author community C 0 in the author network G 1 as C 1 = {s 1 , …, s k}, initializes the local node set M = {M i | i = 0, 1}, where M contains the set of nodes in the two outer layers of the neighborhood of the community. Initialize the weight δ of each network to 0.5.

[0015] S22. Measure the sparsity of the nodes inside and outside the community using the difference in membership relationships between nodes. Denote the membership of the local node set M 0 in the author network G 1 or the literature network G i as z i . Measure the compactness between the nodes in the community using the compactness of the membership relationship between nodes. Initialize the membership z 0 of the local node set M 1 in the author network G i or the literature network G i ;

[0016] S23. To maintain the consistency of the communities between the networks, construct a unified membership relationship, denoted as the unified membership U, and achieve consistency by restricting the proximity between the unified membership U and the membership z w of the local node set of the network w;

[0017] S24. Combine the results of steps S22 and S23 to jointly learn the membership z 0 of the author network G 1 or the literature network G i , the unified membership U, and the weights δ = {δ 1 , δ 2} of the two networks. Adopt an alternating iteration strategy, that is, when the other two variables remain unchanged, update the remaining variable;

[0018] S25. Community expansion: After each round of iteration to update the membership z i of each network, the unified membership U, and the weight δ of each network, the nodes with high membership in the author network G 0 or the literature network G 1 are added to the corresponding community C of the network, and update the local node set M = {M 0 | i = 0, 1} of the author network G 1 or the literature network G i ;

[0019] Iteratively execute steps S24 and S25 until the communities C of the author network G 0 and the literature network G 1 tend to be stable, and return = {C 0 , C 1}.

[0020] As a further optimized technical solution, in step S22, initialize the local node set M in the author network G 0 or the literature network G 1 with the membership degree z i corresponding to it i

[0021]

[0022] where is the regularization term, β is the regularization parameter, and ||X|| F is the Frobenius norm;

[0023] In step S23, initialize the unified membership degree U according to formula (2)

[0024]

[0025] where δ w represents the weight value occupied by network w;

[0026] In step S24, combine formula (1) and formula (2) to jointly learn the membership degree z of the author network G 0 or the literature network G 1 , the unified membership degree U, and the weights of the two networks δ = {δ i , δ 1 , δ 2}:

[0027]

[0028] where z w represents the membership degree corresponding to the local node set M in network w i , and V w represents the local node set M in network w i .

[0029] As a further optimized technical solution, in step S3, for the author or literature required by the user, use it as a seed node and execute step S2 to obtain the community C = {C 0 , C 1}, and use the authors in C 0 as a closely related group of authors, and C 1The documents in it are recommended to users as a closely related document collection.

[0030] The present invention also provides an academic recommendation system based on a local multi-network community discovery method, including the following modules:

[0031] An author-document network construction module, which is used to collect co-author information and document citation information in the documents in dblp to construct an author network G 0 and a document network G 1 , as well as an inter-layer link P between the two networks, expressed as an author-document network G = {G 0 , G 1 , P};

[0032] A community acquisition module, which is used to gradually expand outward from a target author or a target document based on the local multi-network community discovery method based on node membership degree according to the constructed author network G 0 and document network G 1 to obtain an author community C 0 and a document community C 1 , expressed as C = {C 0 , C 1};

[0033] A recommendation module, which is used to given any author or document, set the corresponding node of the author or document as a seed node and execute step S2 to obtain a community C = {C 0 , C 1}, and recommend to the user a group of authors and documents that are more closely related to the target author or document.

[0034] As a further optimized technical solution, the author-document network construction module specifically includes:

[0035] An author information extraction unit, which is used to extract co-author information of the documents in dblp to establish an author network G 0 = <V 0 , E 0 >, any node in the set V 0 represents an author, and E 0 is an adjacency matrix representing the edge relationship. If there is a relationship that authors i and j in the set V 0 co-author a certain document, there is an edge relationship between the corresponding nodes of the two authors, that is, in the adjacency matrix E 0 in

[0036] A document information extraction unit, which is used to extract document citation information of the documents in dblp to establish a document network G 1 = <V 1 , E 1>. Set V 1 Any node in represents a document, and E 1 is an adjacency matrix representing an edge relationship. If the set V 1 in the document i cites the document j, then there is an edge relationship between the nodes corresponding to the two documents, that is, the adjacency matrix E 1 in

[0037] Corresponding relationship extraction unit, used to extract the author network G 0 and the corresponding relationship between the author and the network in the document network G 1 , that is, in the author network G 0 the authors A1, A2, and A3 jointly complete the document P in the document network G 1 , then the edges (A1, P), (A2, P), (A3, P) connecting the author network G 0 and the document network G 1 are used as inter-layer edges, thereby constructing an inter-layer link P between the two networks;

[0038] Author-document network acquisition unit, used to obtain the author-document network G = {H 0 , H 1 , P} according to the execution results of the author information extraction unit, the document information extraction unit, and the corresponding relationship extraction unit.

[0039] As a further optimized technical solution, the community acquisition module specifically includes:

[0040] Initialization unit, used to construct the author-document network G = {G 0 , G 1 , P} extracted by the author-document network construction module. The following assumes that the seed node is the node s in the author network G 0 . According to the inter-layer link P, obtain the node set {s 1 , …, s 1 , …, s k} in the document network G 0 that has a strong inter-layer link with the seed node s. Initialize the author community C 0 = {s} in the author network G 1 , and the document community C 1 = {s 1 , …, s k} in the document network G i . Initialize the local node set M = {M i |i = 0, 1}, where M i contains the outer two-layer neighborhood node set of the community C

[0041] A tightness measurement unit, which is used to measure the sparsity of nodes inside and outside the community by using the difference in membership relationships between nodes, and the author network G 0 or the literature network G 1 The membership of the local node set M i in is represented as z i , and the tightness between nodes in the community is measured by using the tightness of the membership relationship between nodes. Initialize the author network G 0 or the literature network G 1 The membership z i corresponding to the local node set M i ;

[0042] A unified membership construction unit, which is used to construct a unified membership relationship, denoted as the unified membership U, in order to maintain the consistency of communities between networks, and realizes the consistency by restricting the proximity between the unified membership U and the membership z w of the local node set of the network w;

[0043] A learning unit, which is used to jointly learn the membership z 0 of the author network G i or the literature network G i , the unified membership U, and the weights δ = {δ 1 , δ 2} of the two networks, and adopts an alternating iteration strategy, that is, when the other two variables remain unchanged, the remaining variable is updated;

[0044] A community expansion unit, which is used to update the membership z i of each network, the unified membership U, and the weight δ of each network in each round of iteration. After that, the nodes with high membership in the author network G 0 or the literature network G 1 are added to the community C of the corresponding network, and the local node set M = {M 0 |i = 0, 1} of the author network G 1 or the literature network G i is updated;

[0045] An iterative execution unit, which is used to iteratively execute the learning unit and the community expansion unit until the communities C 0 of the author network G 1 and the literature network G 0 , C 1} tend to be stable, and return = {C

[0046] As a further optimized technical solution, in the tightness measurement unit, the author network G 0 or the literature network G 1 is initialized with the formula (1) for the local node set Mi The corresponding membership degree z i

[0047]

[0048] where is the regularization term, β is the regularization parameter, and ||X|| F is the Frobenius norm;

[0049] In the unified membership degree construction unit, the unified membership degree U is initialized according to formula (2)

[0050]

[0051] where δ w represents the weight value occupied by the network w;

[0052] In the learning unit, combining formula (1) and formula (2), jointly learn the membership degree z 0 of the author network G 1 or the literature network G i , the unified membership degree U and the weights of the two networks δ = {δ 1 , δ 2}:

[0053]

[0054] where z w represents the membership degree corresponding to the local node set M i in the network w, and V w represents the local node set M i in the network w

[0055] As a further optimized technical solution, in the recommendation module, for the author or literature required by the user, taking it as a seed node, execute the community acquisition module to obtain the community C = {C 0 , C 1}, and recommend the authors in C 0 as a closely related group of authors and the literature in C 1 as a closely related literature set to the user

[0056] The advantages of the present invention are as follows: The method of the present invention fully considers the author network, the literature network and the connection between the two networks. Only using local information can maximize the quality of the community structure in the author network and the literature network, and at the same time ensure cross-network consistency, so as to recommend a group of authors and literature with closer connections to the target author or target literature for the user

[0057] The method of the present invention measures the possibility that a node belongs to the same community as a seed node through the membership relationship of the nodes in step S22. The membership relationship of nodes within the same community is high, while that of nodes outside the community is low.

[0058] To maintain the consistency of communities between networks, a unified membership relationship, denoted as U, is constructed in step S23, and the consistency is achieved by restricting the proximity of U and the membership degree z w of the local node set of network w.

[0059] The global detection method requires global network information, while the method of the present invention only uses local information and iteratively executes steps S24 and S25 through a local expansion method. In step S24, the weights, membership degrees, and unified membership degrees of the author network and the literature network are jointly learned.

[0060] Meanwhile, the method of the present invention can optionally select an author in the corresponding author network or a literature in the literature network, use it as a seed node and execute step 2 to obtain a consistent literature community and author community for recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 is a flowchart of an academic recommendation method based on the local multi-network community discovery method of the present invention;

[0062] Figure 2 is a model structure diagram of the author-literature network in the present invention;

[0063] Figure 3 shows the author community from the author network and the literature community from the literature network. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0065] Embodiment 1

[0066] Referring to Figure 1 shown, the academic recommendation method based on the local multi-network community discovery method of the present invention includes the following steps:

[0067] S1. In dblp (Digital Bibliography & Library Project, a computer science English literature retrieval database system), collect the co - author information and literature citation information in the literature, and construct the author network G 0 and the literature network G 1 , as well as the inter - layer link P between the two networks, denoted as the author - literature network G = {G 0 ,G 1 ,P}.

[0068] S2. According to the author network G 0 and the literature network G 1 constructed in step S1, based on the local multi - network community discovery method of node membership, gradually expand outward from the target author or target literature to obtain the author community G 0 and the literature community C 1 , denoted as C = {C 0 ,C 1};

[0069] S3. Given any author or literature, set the corresponding node of this author or literature as the seed node and execute step S2 to obtain the community C = {C 0 ,C 1}, and recommend to the user the author group and literature that are more closely related to the target author or literature.

[0070] The specific implementation steps of each step are as follows.

[0071] Step S1 specifically includes the following steps:

[0072] S11. Extract the co - author information in the dblp literature to establish the author network G 0 = <V 0 ,E 0 >. Any node in the set V 0 represents an author, and E 0 is the adjacency matrix representing the edge relationship. If there is a relationship between author i and author j in the set V 0 who have jointly completed a certain literature, there is an edge relationship between the corresponding nodes of the two authors, that is, in the adjacency matrix E 0

[0073] S12. Extract the literature citation information in the dblp literature to establish the literature network G 1 = <V 1 ,E 1 >. Any node in the set V 1 represents a literature, and E 1 is the adjacency matrix representing the edge relationship. If the set V​1 If Document i cites Document j, there is an edge relationship between the nodes corresponding to the two documents, that is, the adjacency matrix E 1 In

[0074] S13. Extract the author network G 0 and the document network G 1 The corresponding relationship between the authors and the network, that is, the author network G 0 The authors A1, A2, and A3 in jointly complete the documents in the document network G 1 If the document P in, then connect the author network G 0 and the document network G 1 The edges (A1, P), (A2, P), and (A3, P) as inter-layer edges, thus constructing the inter-layer link P between the two networks;

[0075] S14. Execute steps S11 - S13 to obtain the author-document network G = {G 0 , G 1 , P}. The model structure is as Figure 2 shown.

[0076] Step S2 specifically includes the following steps:

[0077] S21. Based on the author-document network G = {G 0 , G 1 , P} extracted in step S1, assume the seed node is the node s in the author network G 0 , obtain the node set {s 1 , …, s 1 , …, s k} in the document network G 0 with a strong inter-layer link to the seed node s according to the inter-layer link P, initialize the author community C 0 = {s} in the author network G 1 , and the document community C 1 = {s 1 , …, s k} in the document network G i |i = 0, 1}, where M contains the two-layer neighborhood node sets outside the community. Initialize the weight δ of each network to 0.5.

[0078] S22. Use the difference in membership relationship between nodes to measure the sparsity of nodes inside and outside the community, and represent the membership of the local node set M 0 in the author network G 1 or the document network G i as z i, the tightness between nodes within a community is measured by the tightness of the membership relationship between nodes, and the author network G is initialized using formula (1) 0 or the literature network G 1 and the membership degree z corresponding to the local node set M i in i

[0079]

[0080] where is the regularization term, β is the regularization parameter, and ||X|| F is the Frobenius norm.

[0081] S23. To maintain the consistency of communities between networks, a unified membership relationship is constructed, denoted as the unified membership degree U, and the consistency is achieved by restricting the proximity between the unified membership degree U and the membership degree z of the local node set of network w. The unified membership degree U is initialized according to formula (2) w

[0082]

[0083] where δ w represents the weight value occupied by network w.

[0084] S24. Combining formula (1) and formula (2), jointly learn the membership degree z 0 of the author network G 1 or the literature network G i , the unified membership degree U, and the weights δ of the two networks = {δ 1 , δ 2}:

[0085]

[0086] where z w represents the membership degree corresponding to the local node set M i in network w, and V w represents the local node set M i in network w.

[0087] An alternating iteration strategy is adopted, that is, when the other two variables remain unchanged, the remaining variable is updated.

[0088] S25. Community expansion: After each round of iteration, updating the membership degree z i of each network, the unified membership degree U, and the weights δ of each network, those with high membership degrees in the author network G 0 or the literature network G 1 are added to the corresponding network community C, and the author network G 0 or the literature network G is updated​1 The local node set M = {M i | i = 0, 1}.

[0089] S26. Iteratively execute steps S24 and S25 until the communities C of the author network G 0 and the literature network G 1 tend to be stable, and return = {C 0 , C 1}.

[0090] S3. Given any author or literature, execute step S2 to obtain the community C = {C 0 , C 1}, and recommend to the user the group of authors and the literature that are more closely related to the target author or literature.

[0091] For the author or literature required by the user, use it as a seed node and execute step S2 to obtain the community C = {C 0 , C 1}. Take the authors in C 0 as the group of closely related authors, and the literature in C 1 as the set of closely related literature and recommend them to the user.

[0092] The following gives an example to prove the feasibility of the method:

[0093] We select the dataset collected by dblp in July 2016, and use Professor Jie Tang from Tsinghua University as the seed node. He is a researcher whose main research interest is data mining. The quality of the community is judged by observing whether the detected community nodes are related to Jie Tang's research direction. Figure 3 shows the author communities from the author network and the literature communities from the literature network. Due to space limitations, the detailed information is not shown in the author communities and the literature communities. Some authors and the places where these literatures are published are listed. For example, "KDD(17)" means that there are 17 literatures published on KDD in the literature community. Specifically, the detected authors mainly come from the Jie Tang group (Zhichun Wang, Juanzi Li, Jamal Yousaf, etc.) and the authors who cooperate with the Jie Tang group (Xiaowen Dai, Martin A. Ferman, etc.). The literatures in the literature community are mainly published in the conferences related to data mining. The results show that the method successfully identifies the local communities with practical significance from dblp and can recommend accurate results to the users.

[0094] Embodiment 2

[0095] The present invention also provides an academic recommendation system corresponding to Embodiment 1 based on the local multi-network community discovery method, including the following modules:

[0096] The author - literature network construction module is used to collect co - author information and literature citation information in dblp, and construct the author network G 0 and the literature network G 1 , as well as the inter - layer link P between the two networks, denoted as the author - literature network G = {G 0 ,G 1}, P};

[0097] The community acquisition module is used to gradually expand outward from the target author or target literature based on the local multi - network community discovery method based on node membership degree according to the constructed author network G 0 and the literature network G 1 , to obtain the author community C 0 and the literature community C 1 , denoted as C = {C 0 ,C 1};

[0098] The recommendation module is used to, given any author or literature, set the corresponding node of the author or literature as the seed node and execute step S2 to obtain the community C = {C 0 ,C 1}, and recommend to the user the group of authors and literatures that are more closely related to the target author or literature.

[0099] Among them, the author - literature network construction module specifically includes:

[0100] The author information extraction unit is used to extract co - author information in dblp literature to establish the author network G 0 = <V 0 , E 0 >. Any node in the set V 0 represents an author, and E 0 is the adjacency matrix representing the edge relationship. If there is a relationship that authors i and j in the set V 0 co - complete a certain literature, then there is an edge relationship between the corresponding nodes of the two authors, that is, in the adjacency matrix E 0 in

[0101] The literature information extraction unit is used to extract literature citation information in dblp literature to establish the literature network G 1 = <V 1 , E 1 >. Any node in the set V 1 represents a literature, and E 1 is the adjacency matrix representing the edge relationship. If literature i in the set V 1 cites literature j, then there is an edge relationship between the corresponding nodes of the two literatures, that is, the adjacency matrix E1 China

[0102] A corresponding relationship extraction unit for extracting the author network G 0 and the literature network G 1 The corresponding relationship between the author and the network, that is, in the author network G 0 Authors A1, A2, and A3 jointly complete the literature P in the literature network G 1 Then connect the edges (A1, P), (A2, P), and (A3, P) of the author network G 0 and the literature network G 1 As the inter-layer edges, so as to construct the inter-layer link P between the two networks;

[0103] An author-literature network acquisition unit for obtaining the author-literature network G = {G 0 , G 1 , P} according to the execution results of the author information extraction unit, the literature information extraction unit, and the corresponding relationship extraction unit.

[0104] Among them, the community acquisition module specifically includes:

[0105] An initialization unit for constructing the author-literature network G = {G 0 , G 1 , P} extracted by the author-literature network construction module. Here, assume that the seed node is the node s in the author network G 0 Obtain the node set {s 1 , …, s 1 , …, s k} in the literature network G 0 with a strong inter-layer link with the seed node s according to the inter-layer link P, initialize the author community C 0 = {s} in the author network G 1 , and the literature community C 1 = {s 1 , …, s k} in the literature network G i Initialize the local node set M = {M

[0106] |i = 0, 1}, where M contains the two-layer neighborhood node sets outside the community; 0 A compactness measurement unit for measuring the sparsity of the nodes inside and outside the community by using the difference in membership relationships between nodes, and representing the membership of the local node set M 1 in the author network G i as z i , and using the compactness of the membership relationship between nodes to measure the compactness between nodes in the community, and initializing the author network G 0 or the literature network G1 Local node set M i Corresponding membership degree z i ;

[0107] A unified membership construction unit is used to construct a unified membership relationship, denoted as unified membership U, in order to maintain the consistency of communities between networks, and the consistency is achieved by restricting the proximity between the unified membership U and the membership degree z of the local node set of network w w ;

[0108] A learning unit is used to jointly learn the membership degree z of the author network G 0 or the literature network G 1 by combining the execution results of the compactness measurement unit and the unified membership construction unit i , the unified membership U, and the weights δ = {δ 1 , δ 2} of the two networks, and an alternating iteration strategy is adopted, that is, when the other two variables remain unchanged, the remaining variable is updated

[0109] A community expansion unit is used to update the membership degree z of each network, i the unified membership U, and the weight δ of each network in each round of iteration. After updating the membership degree z of each network, i the unified membership U, and the weight δ of each network, the nodes with high membership degree in the author network G 0 or the literature network G 1 are added to the corresponding network community C, and the local node set M = {M 0 |i = 0, 1} of the author network G 1 or the literature network G i is updated;

[0110] An iterative execution unit is used to iteratively execute the learning unit and the community expansion unit until the community C of the author network G 0 and the literature network G 1 tends to be stable, and return = {C 0 , C 1};

[0111] Among them, in the compactness measurement unit, the membership degree z corresponding to the local node set M in the author network G 0 or the literature network G 1 is initialized with formula (1) i ; i

[0112]

[0113] where is the regularization term, β is the regularization parameter, ||X||F is the Frobenius norm;

[0114] In the unified membership construction unit, the unified membership U is initialized according to formula (2)

[0115]

[0116] where δ w represents the weight value occupied by network w;

[0117] In the learning unit, combining formula (1) and formula (2), jointly learn the membership degree z of the author network G 0 or the literature network G 1 of; i , the unified membership U and the weights of the two networks δ = {δ 1 , δ 2}:

[0118]

[0119] where z w represents the membership degree corresponding to the local node set M in network w i , V w represents the local node set M in network w i .

[0120] In the recommendation module, for the author or literature required by the user, use it as a seed node, execute the community acquisition module, and obtain the community C = {C 0 , C 1}, use the authors in C 0 as a closely related group of authors, and the literature in C 1 as a closely related literature set to recommend to the user.

[0121] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An academic recommendation method based on a local multi-network community discovery method, characterized by: The following steps are involved: S1. In dblp, collect the co-author information and citation information in the literature, construct the author network G0 and the literature network G1, and the inter-layer link P between the two networks, expressed as the author-literature network G = {G0, G1, P}; S2. Based on the author network G0 and document network G1 constructed in step S1, the local multi-network community discovery method based on node membership gradually expands outward from the target author or target document to obtain the author community C0 and document community C1, which are expressed as C = {C0, C1}; S3. Given any author or document, set the node corresponding to the author or document as a seed node and execute step S2 to obtain community C = {C0, C1}, and recommend to the user the author group and documents that are more closely related to the target author or document.

2. The academic recommendation method based on the local multi-network community discovery method according to claim 1, characterized in that: Step S1 specifically includes the following steps: S11. Extract the co-author information of the literature in dblp to build the author network G0= <V 0 ,E 0 >, set V 0 Any node in E represents the author. 0 is the adjacency matrix representing the edge relationship. If the set V 0 If there is a relationship between author i and author j that they have completed a document together, then there is an edge relationship between the nodes corresponding to the two authors, that is, the adjacency matrix E 0 middle S12, extract the literature citation information of the literature in dblp, and use it to establish the literature network G1= <V 1 ,E 1 >, set V 1 Any node in E represents a document. 1 is the adjacency matrix representing the edge relationship. If the set V 1 If document i cites document j, there is an edge relationship between the nodes corresponding to the two documents, that is, the adjacency matrix E 1 middle S13, extracting the correspondence between the author network G0 and the document network G1, that is, the authors A1, A2, and A3 in the author network G0 jointly complete the document P in the document network G1, and the edges (A1, P), (A2, P), and (A3, P) connecting the author network G0 and the document network G1 are used as inter-layer edges, thereby constructing an inter-layer link P between the two networks; S14. Execute steps S11-S13 to obtain the author-document network G = {G0, G1, P}.

3. The academic recommendation method based on the local multi-network community discovery method as claimed in claim 1, characterized in that: Step S2 specifically includes the following steps: S21. Based on the author-document network G = {G0, G1, P} extracted in step S1, it is assumed that the seed node is the node s in the author network G0. According to the inter-layer link P, the node set {s1, ..., s1} in the document network G1 that has a strong inter-layer link with the seed node s is obtained. k }, initialize the author community C0 in the author network G0 = {s}, and the document community C1 in the document network G1 = {s1,…,s k }, initialize the local node set M = {M i |i=0,1}, where M contains the set of two layers of neighboring nodes in the outer layer of the community. Initialize the weight of each network to δ=0.5; S22, using the difference in the membership relationship between nodes to measure the sparsity of nodes inside and outside the community, the local node set M in the author network G0 or the document network G1 i The corresponding membership degree is expressed as z i , use the closeness of the affiliation between nodes to measure the closeness between nodes in the community, initialize the local node set M in the author network G0 or the literature network G1 i The corresponding membership degree z i ; S23. In order to maintain the consistency of the communities between networks, a unified membership relationship is constructed, represented by the unified membership degree U, and the membership degree z of the local node set of network w is restricted. w proximity to achieve consistency; S24, combining the results of step S22 and step S23, jointly learning the membership degree z of the author network G0 or the document network G1 i , unify the membership degree U and the weights of the two networks δ={δ1,δ2}, and adopt an alternating iteration strategy, that is, when the other two variables remain unchanged, update the remaining variables; S25, community expansion: update the membership z of each network in each iteration i After unifying the membership U and the weight δ of each network, the nodes with high membership in the author network G0 or the literature network G1 are added to the community C of the corresponding network, and the local node set M of the author network G0 or the literature network G1 is updated to {M i |i=0,1}; S26. Iterate steps S24 and S25 until the community C of the author network G0 and the document network G1 tends to be stable, and return = {C0, C1}.

4. The academic recommendation method based on the local multi-network community discovery method as claimed in claim 3, characterized in that: In step S22, the local node set M in the author network G0 or the document network G1 is initialized using formula (1): i The corresponding membership degree z i in is the regularization term, β is the regularization parameter, ||X|| F is the Frobenius norm; In step S23, the unified membership degree U is initialized according to formula (2): where δ w Represents the weight value occupied by the network w; In step S24, the membership degree z of the author network G0 or the document network G1 is learned by combining formula (1) and formula (2). i , unified membership U and weights of the two networks δ = {δ1, δ2}: where z w Represents the local node set M in the network w i The corresponding membership degree, V w Represents the local node set M in the network w i .

5. The academic recommendation method based on the local multi-network community discovery method as claimed in claim 1, characterized in that: In step S3, for the author or document required by the user, take it as the seed node, execute step S2, and obtain community C = {C0, C1}, and recommend the authors in C0 as a closely connected author group and the documents in C1 as a closely connected document collection to the user.

6. An academic recommendation system based on a local multi-network community discovery method, characterized by: Includes the following modules: The author-document network construction module is used in dblp to collect co-author information and document citation information in the document, and to construct the author network G0 and document network G1, as well as the inter-layer link P between the two networks, which is expressed as the author-document network G = {G0, G1, P}; The community acquisition module is used to gradually expand outward from the target author or target document based on the constructed author network G0 and document network G1, and obtain the author community C0 and document community C1, which are expressed as C = {C0, C1}; The recommendation module is used to set the node corresponding to any author or document as a seed node and execute step S2 to obtain community C = {C0, C1}, and recommend author groups and documents that are more closely related to the target author or document to the user.

7. The academic recommendation system based on the local multi-network community discovery method as claimed in claim 6, characterized in that: The author-document network building module specifically includes: Author information extraction unit, used to extract the co-author information of the document in dblp, used to establish the author network G0= <V 0 ,E 0 >, set V 0 Any node in E represents the author. 0 is the adjacency matrix representing the edge relationship. If the set V 0 If there is a relationship between author i and author j that they have completed a document together, then there is an edge relationship between the nodes corresponding to the two authors, that is, the adjacency matrix E 0 middle Document information extraction unit, used to extract document citation information from documents in dblp, used to establish document network G1= <V 1 ,E 1 >, set V 1 Any node in E represents a document. 1 is the adjacency matrix representing the edge relationship. If the set V 1 If document i cites document j, there is an edge relationship between the nodes corresponding to the two documents, that is, the adjacency matrix E 1 middle The corresponding relationship extraction unit is used to extract the corresponding relationship between the author network G0 and the document network G1. That is, if the authors A1, A2, and A3 in the author network G0 jointly complete the document P in the document network G1, the edges (A1, P), (A2, P), and (A3, P) connecting the author network G0 and the document network G1 are used as inter-layer edges, thereby constructing an inter-layer link P between the two networks. The author-document network acquisition unit is used to obtain the author-document network G={G0, G1, P} according to the execution results of the author information extraction unit, the document information extraction unit and the corresponding relationship extraction unit.

8. The academic recommendation method based on the local multi-network community discovery method as claimed in claim 7, characterized in that: The community acquisition module specifically includes: Initialization unit, used for extracting the author-document network G = {G0, G1, P} based on the author-document network construction module. In the following, it is assumed that the seed node is the node s in the author network G0, and the node set {s1,…,s1} in the document network G1 that has a strong inter-layer link with the seed node s is obtained according to the inter-layer link P. k }, initialize the author community C0 in the author network G0 = {s}, and the document community C1 in the document network G1 = {s1,…,s k }, initialize the local node set M = {M i |i=0,1}, where M contains the set of two layers of neighboring nodes in the outer layer of the community, and the weight of each network is initialized to δ=0.5; The compactness measurement unit is used to measure the sparsity of nodes inside and outside the community by using the difference in the membership relationship between nodes. The local node set M in the author network G0 or the document network G1 i The corresponding membership degree is expressed as z i , use the closeness of the affiliation between nodes to measure the closeness between nodes in the community, initialize the local node set M in the author network G0 or the literature network G1 i The corresponding membership degree z i ; The unified membership construction unit is used to maintain the consistency of the communities between networks. It constructs a unified membership relationship, expressed as a unified membership U, and constrains the unified membership U and the membership z of the local node set of network w. w proximity to achieve consistency; The learning unit is used to combine the execution results of the closeness measurement unit and the unified membership construction unit to jointly learn the membership z of the author network G0 or the document network G1 i , unify the membership degree U and the weights of the two networks δ={δ1,δ2}, and adopt an alternating iteration strategy, that is, when the other two variables remain unchanged, update the remaining variables; Community expansion unit, used to update the membership z of each network in each iteration i After unifying the membership U and the weight δ of each network, the nodes with high membership in the author network G0 or the literature network G1 are added to the community C of the corresponding network, and the local node set M of the author network G0 or the literature network G1 is updated to {M i |i=0,1}; The iterative execution unit is used to iteratively execute the learning unit and the community expansion unit until the community C of the author network G0 and the document network G1 tends to be stable, and returns = {C0, C1}.

9. The academic recommendation method based on the local multi-network community discovery method as claimed in claim 8, characterized in that: In the closeness measurement unit, the local node set M in the author network G0 or the document network G1 is initialized using formula (1): i The corresponding membership degree z i in is the regularization term, β is the regularization parameter, ||X|| F is the Frobenius norm; In the unified membership construction unit, the unified membership U is initialized according to formula (2): where δ w Represents the weight value occupied by the network w; In the learning unit, we combine formula (1) and formula (2) to learn the membership degree z of the author network G0 or the document network G1. i , unified membership U and weights of the two networks δ = {δ1, δ2}: where z w Represents the local node set M in the network w i The corresponding membership degree, V w Represents the local node set M in the network w i .

10. The academic recommendation system based on the local multi-network community discovery method according to claim 6, characterized in that: In the recommendation module, for the authors or documents required by the user, they are taken as seed nodes, and the community acquisition module is executed to obtain the community C = {C0, C1}. The authors in C0 are regarded as a closely connected group of authors, and the documents in C1 are regarded as a closely connected collection of documents for recommendation to the user.

Images