CN112148876A - Paper classification and recommendation method - Google Patents

Paper classification and recommendation method Download PDF

Info

Publication number
CN112148876A
CN112148876A CN202011009122.9A CN202011009122A CN112148876A CN 112148876 A CN112148876 A CN 112148876A CN 202011009122 A CN202011009122 A CN 202011009122A CN 112148876 A CN112148876 A CN 112148876A
Authority
CN
China
Prior art keywords
representation
attribute
node
neighbor
paper
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202011009122.9A
Other languages
Chinese (zh)
Other versions
CN112148876B (en
Inventor
谭春阳
吴震
何亮
戴新宇
梁华盛
颜强
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University
Original Assignee
Nanjing University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University filed Critical Nanjing University
Priority to CN202011009122.9A priority Critical patent/CN112148876B/en
Publication of CN112148876A publication Critical patent/CN112148876A/en
Application granted granted Critical
Publication of CN112148876B publication Critical patent/CN112148876B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/335—Filtering based on additional data, e.g. user or group profiles
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/335—Filtering based on additional data, e.g. user or group profiles
    • G06F16/337—Profile generation, learning or modification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/10—Text processing
    • G06F40/194—Calculation of difference between files
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Databases & Information Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention provides a paper classification and recommendation method, which comprises the following steps: sampling each thesis node to obtain candidate attribute neighbor nodes, wherein all structure neighbors of each thesis node are used as candidate structure neighbor nodes; generating attribute representation and structure representation of the thesis node and the neighbor node; fusing the structure representation of the paper node and the structure representation of the attribute neighbor to obtain a fused structure representation, and fusing the attribute representation of the paper node and the attribute representation of the final structure neighbor to obtain a fused attribute representation; then training is carried out through a reconstruction loss function in the neural network training process, and the representation is continuously updated; splicing the updated fusion structure representation and fusion attribute representation to obtain a low-dimensional vector representation of the paper node; inputting the low-dimensional representation of part of the thesis nodes with the category labels in the thesis network into a support vector machine, and training to obtain a classification model; and predicting the category of the thesis node through a classification model, and classifying the thesis and recommending the thesis of the same category by using the category.

Description

Paper classification and recommendation method
Technical Field
The invention belongs to the field of data mining, and particularly relates to a paper classification and recommendation method.
Background
At present, network representation learning is a popular research direction in the field of data mining, and has very wide application in real life, such as node classification, node clustering, recommendation systems and the like. The thesis network representation learning is a research field of low-dimensional representation learning of the thesis network containing attribute information on nodes, and low-dimensional representation is learned for each node and applied to offline application. The mainstream method includes a representation learning method using only the structural information and a multisource information fusion representation learning method combining the structural information and the attribute information at the same time. Meanwhile, the multi-source information is merged into the representation learning method combining the structural information and the attribute information, and the learned representation obtains better performance compared with a method only utilizing the structural information.
Although current multi-source information fusion methods have made great progress, these methods usually ignore the problem of outliers in the attribute network. Specifically, a paper node may not belong to the same category as its neighboring nodes or neighboring nodes with similar attributes, but many models are currently based on such similarity assumptions; in addition, the structure and the attribute information have the phenomenon of inconsistent distribution, namely the description of the structure and the attribute information to the node is not uniform. The similarity assumption and consistency assumption based on the current model are not consistent with the network distribution of the paper in reality, so that the representation capability of the model is limited.
Network data is widely available in real life, such as a paper reference network, a social media communication network and the like. How to learn the nodes in the paper network to obtain a low-dimensional representation and apply the low-dimensional representation to offline tasks (such as paper classification, paper recommendation and the like) is a research hotspot. Some methods of network representation learning are mainly described below.
Perozzi et al ([ Perozzi et al, 2014] deep walk: Online learning of social representation) proposed a deep walk model, which walks network data by a random walk method for the first time to obtain a plurality of walk sequences as pseudo sentences. And then training the walking sequences by using a word vector learning method Skip-gram and a hierarchical softmax optimization method so as to obtain a low-dimensional node representation.
Grover et al ([ Grover et al, 2016] node2vec: Scalable feature learning for networks) improved on the basis of a Deepwalk model in a random walk strategy, and proposed that biased random walks exist and the depth and the width of the walk are controlled based on parameters p and q. And after a biased walking sequence is obtained, learning by a Skip-gram algorithm and a negative sampling optimization method to obtain a low-dimensional vector representation of the node.
Tang et al ([ Tang et al, 2015] LINE: Large-scale information network embedding) then proposed a LINE model, and first proposed concepts of first-order similarity and second-order similarity for measuring the similarity between nodes. And training in an Edge sampling optimization mode to obtain low-dimensional vector representation of the nodes.
Yang et al ([ Yang et al, 2015] Network representation learning with rich text information) proposed the first model TADW that incorporates text information into paper Network representation learning, which utilizes both structural information and attribute information by a matrix decomposition method. Because of the simultaneous use of multi-source information, the effectiveness of TADW is generally better than models that use only structural information.
The TriDNR proposed by Pan et al ([ Pan et al, 2016] Tri-Party Deep Network Representation) is a semi-supervised attribute Network Representation learning model based on a simple neural Network, the Deep association among three relations of node-node, node-attribute and label-attribute is considered for the first time by the TriDNR, and two training networks are constructed: a structure training network and an attribute-label training network, which are shared by the two to represent the nodes. And through an alternate training mode, the representation of the node in the structure diagram and the representation in the attribute-label network are continuously and alternately updated, so that the low-dimensional representation can simultaneously fuse the structure information, the attribute information and the supervision information, and the two networks can mutually influence and promote during training.
Yang et al ([ Yang et al, 2017] From properties to links: Deep network embedding on completing graphs) bases on consistency assumptions of structure and properties. For example, in a network of articles, a node (article) belonging to the natural language processing category is more likely to refer to a plurality of articles in the same domain, and at the same time, the contents between the articles referred to each other are more likely to be similar. Therefore, theoretically, a relationship of mutual mapping exists from the structure to the attribute to some extent. By such an assumption, consistency of the two in the depth space is maintained, aiming to enhance the robustness of the representation, thereby obtaining a higher quality representation. However, by analysis, there is a gap between real-world data and such consistency assumptions, limiting the quality of the representation learned by such methods.
Disclosure of Invention
The purpose of the invention is as follows: in the thesis network representation learning method, gaps exist between similarity assumptions and consistency assumptions which are usually based and real data distribution, and the representation quality is limited. The problem to be solved by the invention is how to dig more hidden information under real data distribution, reduce the difference between hypothesis and real data, thus learning to obtain a high-quality low-dimensional thesis node representation, and apply the low-dimensional thesis node representation to offline tasks such as thesis classification and thesis recommendation.
The invention specifically provides a low-dimensional vector representation learning method for a thesis network structure by fusing sampling neighbor information, which comprises the following steps:
step 1, obtaining the candidate of the truncated attribute neighbor of each paper node in the paper network G by truncation neighbor sampling, and using a set symbol CaRepresenting a truncated attribute neighbor candidate set of a paper node p;
step 2, based on the truncated attribute, neighbor candidateAnd (3) collecting, and calculating to obtain a final truncated attribute neighbor: applying cosine similarity metric approach to paper node p and truncated attribute neighbor candidate set of nodes
Figure RE-GDA0002737443240000031
In the above, the attribute of paper node p is expressed as one-hot
Figure RE-GDA0002737443240000032
Truncating the attribute-only hot representation set of the attribute neighbor candidate set to
Figure RE-GDA0002737443240000033
Calculating and reserving K attribute neighbors with highest similarity as final truncation attribute neighbors; where m represents the number of truncated attribute neighbors of paper node p,
Figure RE-GDA0002737443240000034
representing the ith truncated attribute neighbor in the truncated attribute neighbor candidate set for paper node p,
Figure RE-GDA0002737443240000035
representing the attribute one-hot representation of the ith truncation attribute neighbor, wherein the value of i is 1-m;
step 3, reserving T structure neighbors of the thesis node p as final structure neighbors;
step 4, establishing a structural multilayer self-encoder AEsThe structure unique representation of the paper nodes is used as input, and the preliminary structure representation and the structure representation of attribute neighbors of each paper node in the attribute network are obtained through a structure multilayer self-encoder; build Attribute multilayer autoencoder AEaThe attribute unique representation of the paper nodes is used as input, and the preliminary attribute representation and the attribute representation of the structure neighbor of each paper node in the attribute network are obtained through an attribute multilayer self-encoder;
step 5, fusing the preliminary structure representation of the paper node and the structure representation of the attribute neighbor to obtain a fused structure representation of the paper node, and fusing the preliminary attribute representation of the paper node and the attribute representation of the attribute neighbor to obtain a fused attribute representation of the paper node;
step 6, enabling the fusion structure representation and the attribute multilayer self-encoder in the step 5 to interact through reconstruction operation, and enabling the fusion attribute representation and the structure multilayer self-encoder in the step 5 to interact;
step 7, splicing the fusion structure representation obtained through the interaction in the step 6 and the fusion attribute representation obtained through the interaction in the step 6 to obtain a low-dimensional representation of the paper node;
step 8, inputting the low-dimensional representation of the labeled thesis nodes in the thesis network into a support vector machine classifier for training to obtain a trained classifier;
and 9, inputting the thesis nodes without labels in the thesis network into a classifier, obtaining the prediction labels of the thesis nodes as classification results, and performing thesis classification and thesis recommendation of the same category.
In step 1, the thesis network refers to a network in which nodes are thesis and the nodes include attribute information (title text, abstract text, body text, etc.), such as a thesis network Cora ([ Sen et al, 2008)]A collection classification in network data) containing title text and summary text information; the truncated neighbor sampling is: randomly walking each paper node by giving a walking step length L (set to be 10) and a walking number gamma (set to be 80) of each node, obtaining a walking sequence of each paper node by walking, and obtaining a truncated attribute neighbor candidate set C of each paper node by statistically de-duplicating the paper nodes in the sequenceap。
The step 2 comprises the following steps:
step 2-1, setting a truncated attribute neighbor candidate set of the paper node p
Figure RE-GDA0002737443240000041
Representing the ith truncated attribute neighbor in the truncated attribute neighbor candidate set of paper node p, truncated attribute neighbor candidate set CapIs a collection of property-only-hot representations
Figure RE-GDA0002737443240000042
To represent
Figure RE-GDA0002737443240000043
The property of (1) is expressed by single heat, and the value of i is 1-m;
one-hot representation of attributes of paper node p
Figure RE-GDA0002737443240000044
And attribute-only representation set of truncated attribute neighbor candidate set
Figure RE-GDA0002737443240000045
And (3) performing cosine similarity calculation:
Figure RE-GDA0002737443240000046
Figure RE-GDA0002737443240000047
to represent
Figure RE-GDA0002737443240000048
And ith attribute unique hot representation
Figure RE-GDA0002737443240000049
Cosine similarity of (d);
then m sets of cosine similarity results are obtained as
Figure RE-GDA00027374432400000410
Step 2-2, sorting the cosine similarity results of the truncated attribute neighbors of the paper node p in a descending order, and taking the largest first K (for example, the first 20) as the final attribute neighbors of the paper node p; if m is<K, performing a playback random sampling to obtain a set of
Figure RE-GDA00027374432400000411
Wherein
Figure RE-GDA00027374432400000412
Representing the sampled i-th attribute neighbor, CApRepresenting a collection of these attribute neighbors.
In step 3, obtaining T (for example, set to be 20) final structure neighbors of the paper node through random sampling; if the number of the structure neighbor candidates of the thesis node is less than T, the random sampling is put back to obtain a set of
Figure RE-GDA00027374432400000413
Wherein
Figure RE-GDA00027374432400000414
And i represents the ith structure neighbor of the node p, and the value of i is 1-T.
Step 4 comprises the following steps:
step 4-1, build a structural multi-layer autoencoder ([ Hinton et al, 2006)]A fast learning algorithm for deep belief nets)AEsStructured multilayer autoencoder AEsThe device comprises an encoding module and a decoding module, wherein the encoding module is as follows:
y1＝σ(W1xs+b1)
ym＝σ(Wmym-1+bm)
wherein, yiRepresents the output of the i-th fully-connected layer, WiAnd biThe distribution represents the projection matrix and offset vector in the ith layer,
Figure RE-GDA0002737443240000051
structural one-hot representation of the representation input, dsσ is the sigmoid function for the dimension of the one-hot representation of the structure. WiAnd biIs trained along with the representation learning method, wherein
Figure RE-GDA0002737443240000052
Denotes dsA complex space of dimensions.
The decoding module inverts the coding module to obtain output
Figure RE-GDA0002737443240000053
The parameters are trained along with the learning method through reconstruction, and the loss function of the structural multi-layer self-encoder is
Figure RE-GDA0002737443240000054
Figure RE-GDA0002737443240000055
Wherein n issAs to the number of the nodes,
Figure RE-GDA0002737443240000056
the structure input for the ith node is expressed uniquely,
Figure RE-GDA0002737443240000057
for the structural representation resulting from the decoding by the encoder,
Figure RE-GDA0002737443240000058
to penalize the parameter, e.g., set to 10, more penalties are given to non-zero values in the representation.
Paper node p and attribute neighbor CApStructural independent thermal representation input structure multilayer self-encoder AE ofsObtaining a preliminary structural representation of the nodes of the thesis at the hidden layer
Figure RE-GDA0002737443240000059
And structural representation of attribute neighbors
Figure RE-GDA00027374432400000510
Wherein
Figure RE-GDA00027374432400000511
Obtaining a preliminary structure representation for the ith attribute neighbor from the hidden layer;
step 4-2, establishing an attribute multilayer self-encoder AEaThe method is consistent with the step 4-1, and the paper node p and the structure neighbor C are connectedspAttribute-independent representation of input attributes multilayer autoencoder AEaIn hidden layer theoryPreliminary attribute representation of text nodes
Figure RE-GDA00027374432400000512
And attribute representation of structure neighbors
Figure RE-GDA00027374432400000513
Wherein
Figure RE-GDA00027374432400000514
A preliminary attribute representation derived from the hidden layer for the ith structure neighbor. l is 1-T;
the step 5 comprises the following steps:
step 5-1, a method by point-by-point attention mechanism ([ Luong et al, 2015)]Effective applications to Attention-based Neural Machine Translation), calculating a structural similarity coefficient by point multiplication, and expressing the structural similarity according to the primary structure of a paper node p
Figure RE-GDA00027374432400000515
And structural representation of attribute neighbors
Figure RE-GDA00027374432400000516
Calculating point-by-point similarity:
Figure RE-GDA00027374432400000517
wherein ep,jNormalizing the point-by-point similarity between the thesis node p and the attribute neighbor K to obtain a similarity coefficient alphap,j：
Figure RE-GDA0002737443240000061
Wherein exp represents a natural index; normalization by the inverse mechanism yields a new inverse attention coefficient
Figure RE-GDA0002737443240000062
Figure RE-GDA0002737443240000063
Wherein the reverse attention coefficient indicates the attention degree of the paper node to each neighbor;
neighbor information representation weighted by attention coefficient calculation
Figure RE-GDA0002737443240000064
Figure RE-GDA0002737443240000065
Step 5-2, obtaining the fusion structure representation of the thesis node
Figure RE-GDA0002737443240000066
Step 5-3, calculating to obtain weighted neighbor information representation
Figure RE-GDA0002737443240000067
Step 5-4, obtaining the fusion attribute representation of the paper node
Figure RE-GDA0002737443240000068
Step 5-2 comprises: the importance of the neighbors is controlled using a weighted sum mechanism:
Figure RE-GDA0002737443240000069
obtaining a fused structural representation of a paper node
Figure RE-GDA00027374432400000610
Wherein the parameter lambdasAnd the value is 0-1, and the method is used for controlling the importance degree represented by the neighbor information.
Step 5-3 comprises: calculated by the following formulaWeighted neighbor information representation
Figure RE-GDA00027374432400000611
Figure RE-GDA00027374432400000612
Step 5-4 comprises: the importance of the neighbors is controlled using a weighted sum mechanism:
Figure RE-GDA00027374432400000613
obtaining a fused attribute representation of a paper node
Figure RE-GDA00027374432400000614
Wherein the parameter lambdaaAnd the value is 0-1, and the method is used for controlling the importance degree represented by the neighbor information.
The step 6 comprises the following steps:
step 6-1, build structure-to-attribute decoder, let the fused structure representation in step 5-2
Figure RE-GDA0002737443240000071
Deriving a reconstructed property representation by a structure-to-property decoder
Figure RE-GDA0002737443240000072
Interacting through reconstruction operations to obtain a loss function for a structure-to-attribute decoder
Figure RE-GDA0002737443240000073
Figure RE-GDA0002737443240000074
Wherein n isaIn order to be the number of attributes,
Figure RE-GDA0002737443240000075
for the property-one-hot representation of node p,
Figure RE-GDA0002737443240000076
is a penalty coefficient;
step 6-2, build attribute to structure decoder, let step 5-4 fuse attribute representation
Figure RE-GDA0002737443240000077
Deriving a reconstructed structure representation by an attribute-to-structure decoder
Figure RE-GDA0002737443240000078
Interacting through reconstruction operations to obtain a loss function for an attribute-to-structure decoder
Figure RE-GDA0002737443240000079
Figure RE-GDA00027374432400000710
Wherein n issIn order to be the number of attributes,
Figure RE-GDA00027374432400000711
for the property-one-hot representation of node p,
Figure RE-GDA00027374432400000712
is a penalty factor. The parameters are trained with the method.
And 7, splicing the fusion structure representation and the fusion attribute representation after the interaction of the step 6-1 and the step 6-2 to be used as the final representation output of the nodes of the thesis.
According to the method provided by the invention, the information of the nodes is further supplemented by sampling the structure neighbors and the attribute neighbors of the thesis nodes, so that the problem that the theoretical assumption is inconsistent with the real data distribution is solved, and the node representation quality in the thesis network representation learning is improved. And the method of the invention provides a truncation sampling method aiming at the problem of sampling efficiency, relieves the problem of complexity and is easier to expand into a large-scale thesis network.
The invention has the following beneficial effects:
the technical level is as follows:
1 original thesis network representation learning method is basically based on two basic assumptions: similarity assumptions (nodes that are similar in structure or similar in attributes are more similar) and consistency assumptions (predicted preferences for structure information and attribute information are consistent). However, data analysis finds that a gap exists between the hypothesis and the real data distribution, and the phenomenon limits the representation quality learned by the current model. In order to alleviate the problem, the method of the invention excavates and fuses additional attribute information and structure information, reduces the difference between the assumed data distribution and the real data distribution, and improves the representation quality.
The truncation neighbor sampling method provided by the invention avoids high-complexity calculation for each thesis node, reduces the complexity of the method, and is suitable for large-scale attribute network data sets.
3 the reverse attention mechanism provided by the invention enables each thesis node to fuse more diverse neighbor information, and the reverse attention mechanism has a better effect through experimental verification.
The application level is as follows: the low-dimensional vector representation learning method for the paper network structure by fusing the sampled neighbor information can be applied to the paper network with any attribute, is not limited to the network with the text attribute, can be conveniently expanded to a large-scale paper network representation learning task, and performs paper classification and paper recommendation on the basis.
Drawings
The foregoing and/or other advantages of the invention will become further apparent from the following detailed description of the invention when taken in conjunction with the accompanying drawings.
Fig. 1 is a flowchart of a paper network representation learning method based on a depth self-encoder according to an embodiment.
Fig. 2 is a flow chart of truncating neighbor samples.
FIG. 3 is a flow chart of a reverse attention mechanism.
FIG. 4 is a flow chart of the method of the present invention in conjunction with a deep auto-encoder based paper network representation learning method.
FIG. 5 is a graph of the effect of the invention and other methods on node classification on the Cora dataset and the DBLP dataset.
Detailed Description
The flow chart is shown in fig. 1. As shown in the figure, the invention provides a paper classification and recommendation method, firstly establishing a baseline model based on a depth self-encoder, which comprises the following steps:
step 1, establishing a depth structure multilayer self-encoder AEsComprising an encoder module and a decoder module, the input being a structural, one-hot representation x of a paper nodesThe encoder output is a low dimensional hidden layer vector
Figure RE-GDA0002737443240000081
Decoder output is reconstructed structure representation x's。
Step 2, establishing a depth attribute multilayer self-encoder AEaComprising an encoder module and a decoder module, the input being a property-independent representation x of a paper nodeaThe encoder output is a low dimensional hidden layer vector
Figure RE-GDA0002737443240000082
Decoder output is reconstructed attribute representation x'a。
Step 3, establishing a decoder with attributes to a structure, and inputting the decoder as attribute hidden layer vectors of the paper nodes
Figure RE-GDA0002737443240000083
Output as a reconstructed structure representation of attributes to structures
Figure RE-GDA0002737443240000084
Step 4, a structure-to-attribute decoder is established, and the structure hidden layer vector of the thesis node is input
Figure RE-GDA0002737443240000085
Output as a reconstructed attribute representation of structure to attribute
Figure RE-GDA0002737443240000086
In this flow, the self-encoder in step 1 and step 2 is implemented by extending the depth self-encoder method ([ Hinton et al, 2006] a fast learning algorithm for deep belief nets), and the encoder module can be expressed as:
y1＝σ(W1x+b1)
ym＝σ(Wmym-1+bm)
wherein, WiAnd biThe distribution represents the projection matrix and offset vector in the i-th layer, x represents the input structure/property one-hot representation, and σ is the sigmoid function. WiAnd biTrained along with the representation learning method. The input of the encoder is the structure one-hot representation and the attribute one-hot representation of the paper nodes, the structure information and the attribute information of the paper network are summarized, and the output dimension of the encoder is set to be 64 dimensions.
The decoder module in steps 1 to 4 can be represented as the inversion process of the encoder module, the input dimension of the decoder module is 64 dimensions, the output dimension is consistent with the input dimension of the encoder, and the decoding output x' is obtained, and the parameters are obtained along with network training.
The training baseline model adopts a reconstruction loss function, and adds all reconstruction losses in steps 1 to 4, for example, the reconstruction loss in step 1 is:
Figure RE-GDA0002737443240000091
wherein
Figure RE-GDA0002737443240000092
Representing the output representation from the decoder module of the encoder module,
Figure RE-GDA0002737443240000093
representing the input from the encoder module(s),
Figure RE-GDA0002737443240000094
to penalize the parameter, e.g., set to 10, more penalties are given to non-zero values in the representation.
Next, a truncated neighbor sampling method proposed in the present invention is described. Flow diagram As shown in FIG. 2, the following notation is used, including with the set notation CapRepresenting a truncated attribute neighbor candidate set for node p. The method comprises the following steps:
step 1, random walk is carried out on each paper node through a given walk step length L (set to be 10 for example) and the walk times gamma (set to be 80 for example) of each paper node, a walk sequence of each paper node is obtained through the walk, paper nodes in the sequence are subjected to statistical deduplication to obtain a truncated attribute neighbor candidate set C of each paper node pap。
Step 2, the attribute of the thesis node p is expressed in a single hot way
Figure RE-GDA0002737443240000095
And attribute-only representation set of truncated attribute neighbor candidate set
Figure RE-GDA0002737443240000096
The cosine similarity calculation is carried out to obtain m cosine similarity result sets as
Figure RE-GDA0002737443240000097
Step 3, sorting the cosine similarity results of the truncated attribute neighbors in a descending order, and taking the largest top K (for example, set to be 20) as the final attribute neighbors of the thesis node p; if m is<K, performing a playback random sampling to obtain a set of
Figure RE-GDA0002737443240000101
In this flow, the cosine similarity calculation formula in step 2 is:
Figure RE-GDA0002737443240000102
the reverse attention mechanism proposed by the present invention is described next. Flow diagram As shown in FIG. 3, the notation is used to include a preliminary structure/attribute representation of a paper node
Figure RE-GDA0002737443240000103
Preliminary structure/attribute representation of attributes/structure neighbors
Figure RE-GDA0002737443240000104
The method comprises the following steps:
step 1, a method of point-by-point attention mechanism ([ Luong et al, 2015)]Effective applications to Attention-based Neural Machine Translation), calculating a structural similarity coefficient by adopting point multiplication, and expressing according to the primary structure of the nodes of the paper
Figure RE-GDA0002737443240000105
And preliminary structural representation of attribute neighbors
Figure RE-GDA0002737443240000106
And calculating the point-by-point similarity, and normalizing the point-by-point similarity of K (set to 20, for example) to obtain a similarity coefficient.
And 2, obtaining a new reverse attention coefficient through the normalization of a reverse mechanism.
In this flow, the point-by-point similarity calculation formula in step 1 is:
Figure RE-GDA0002737443240000107
the normalized formula is:
Figure RE-GDA0002737443240000108
where exp denotes the natural index.
The reverse attention coefficient calculation formula in step 2 is as follows:
Figure RE-GDA0002737443240000109
wherein the reverse attention coefficient indicates the degree of attention of the paper node to each neighbor. The weighted neighbor information representation is calculated according to the attention coefficient:
Figure RE-GDA00027374432400001010
how the low-dimensional vector representation learning method for the paper network structure by fusing the sampled neighbor information provided by the invention is applied to the baseline model based on the depth self-encoder is described next. Flow diagram As shown in FIG. 4, the following notation is used, including with the set notation CapRepresenting a truncated attribute neighbor candidate set. The method comprises the following steps:
step 1, obtaining the candidate of the truncated attribute neighbor of each node in the thesis network G by truncation neighbor sampling, and using a set symbol CaRepresenting a truncated attribute neighbor candidate set;
step 2, applying cosine similarity measurement mode to paper node p and node truncation attribute neighbor candidate set
Figure RE-GDA0002737443240000111
In the above, the attribute of paper node p is expressed as one-hot
Figure RE-GDA0002737443240000112
Truncating the attribute-only hot representation set of the attribute neighbor candidate set to
Figure RE-GDA0002737443240000113
Calculating and reserving K attribute neighbors with highest similarity as final truncation attribute neighbors; where m represents the number of truncated attribute neighbors of paper node pThe number of the first and second groups is,
Figure RE-GDA0002737443240000114
representing the ith truncated attribute neighbor in the truncated attribute neighbor candidate set for paper node p,
Figure RE-GDA0002737443240000115
representing the attribute one-hot representation of the ith truncation attribute neighbor, wherein the value of i is 1-m;
step 3, reserving T (for example, set to be 20) structure neighbors of the node p of the node paper as final structure neighbors through random sampling, wherein the set is represented as
Figure RE-GDA0002737443240000116
Wherein
Figure RE-GDA0002737443240000117
Representing the ith structure neighbor of the node p, wherein the value of i is 1-T;
step 4, establishing a structural multilayer self-encoder AEsThe structure unique representation of the paper nodes is used as input, and the preliminary structure representation of each paper node in the paper network and the preliminary structure representation of the attribute neighbor are obtained through a structure multilayer self-encoder; build Attribute multilayer autoencoder AEaThe attribute unique representation of the paper nodes is used as input, and the preliminary attribute representation of each paper node in the attribute network and the preliminary attribute representation of the structure neighbor are obtained through an attribute multilayer self-encoder;
step 5, a reverse attention mechanism and a weighting and method are provided, the preliminary structure representation of the paper node and the preliminary structure representation of the attribute neighbor are fused to obtain a fusion structure representation of the paper node, and the attribute representation of the paper node and the preliminary attribute representation of the structure neighbor are fused to obtain a fusion attribute representation of the paper node through the reverse attention mechanism and the weighting and method;
step 6, enabling the fusion structure representation of the paper node in the step 5 to interact with the attribute multilayer self-encoder through reconstruction operation, and enabling the fusion attribute representation of the paper node in the step 5 to interact with the structure multilayer self-encoder;
and 7, splicing the fusion structure representation obtained through the interaction in the step 6 and the fusion attribute representation obtained through the interaction in the step 6 to obtain a final representation of the thesis node.
In this flow, step 1 is the same as step 1 of the truncated neighbor sampling method, the default walking step length L is set to 80 steps, and the number of walks γ of each paper node is set to 10 times by default.
Step 2 adopts the same method as step 2 and step 3 of the truncated neighbor sampling method.
Step 3, obtaining T final structure neighbors of the paper node through random sampling; if the number of the structure neighbor candidates of the thesis node is less than T, the random sampling is put back to obtain a set of
Figure RE-GDA0002737443240000121
In particular, T is 20.
The step 4 comprises the following steps:
step 4-1 is consistent with the step of establishing the structural multi-layer self-encoder in the baseline model based on the depth self-encoder in the foregoing, and the structural multi-layer self-encoder AE is establisheds. Structural one-hot representation with input as paper node
Figure RE-GDA0002737443240000122
Obtaining an output by a decoding module
Figure RE-GDA0002737443240000123
Paper node p and attribute neighbor CapStructural independent thermal representation input structure multilayer self-encoder AE ofsObtaining a preliminary structural representation of the nodes of the thesis at the hidden layer
Figure RE-GDA0002737443240000124
And preliminary structural representation of attribute neighbors
Figure RE-GDA0002737443240000125
Step 4-2, withIn the foregoing, the steps of establishing the attribute multi-layer self-encoder in the baseline model based on the depth self-encoder are consistent, and the attribute multi-layer self-encoder AE is establisheda. Input as an attribute-only representation of a paper node
Figure RE-GDA0002737443240000126
Obtaining an output by a decoding module
Figure RE-GDA0002737443240000127
Paper node p and structure neighbor CsAttribute-independent representation of input attributes multilayer autoencoder AEaObtaining a preliminary attribute representation of a paper node at a hidden layer
Figure RE-GDA0002737443240000128
And preliminary attribute representation of structural neighbors
Figure RE-GDA0002737443240000129
The step 5 comprises the following steps:
step 5-1, a method by point-by-point attention mechanism ([ Luong et al, 2015)]Effective applications to Attention-based Neural Machine Translation), by calculating a structural similarity coefficient, according to the preliminary structural representation of the nodes of the paper
Figure RE-GDA00027374432400001210
And preliminary structural representation of attribute neighbors
Figure RE-GDA00027374432400001211
Calculating point-by-point similarity:
Figure RE-GDA00027374432400001212
and normalizing the K point multiplied similarity to obtain a similarity coefficient:
Figure RE-GDA00027374432400001213
where exp denotes the natural index. The new inverse attention coefficient is obtained by normalization of the inverse mechanism:
Figure RE-GDA00027374432400001214
wherein the reverse attention coefficient indicates the degree of attention of the paper node to each neighbor. The weighted neighbor information representation is calculated according to the attention coefficient:
Figure RE-GDA0002737443240000131
and 5-2, controlling the importance of the neighbor by adopting a weighting sum mechanism:
Figure RE-GDA0002737443240000132
and obtaining the fusion structure representation of the nodes of the thesis. Wherein λsThe value is 0-1 and is used for controlling the importance degree represented by the neighbor information.
Step 5-3, similar to step 5-1, the weighted neighbor information representation is calculated:
Figure RE-GDA0002737443240000133
step 5-4, similar to step 5-2, uses a weighted sum mechanism to control the importance of the neighbors:
Figure RE-GDA0002737443240000134
obtaining a fused attribute representation of the nodes of the paper, wherein lambdaaThe value is 0-1 and is used for controlling the importance degree represented by the neighbor information.
The step 6 comprises the following steps:
step 6-1, build structure-to-attribute decoder, let the fused structure representation in step 5-2
Figure RE-GDA0002737443240000135
Deriving a reconstructed property representation by a structure-to-property decoder
Figure RE-GDA0002737443240000136
Interacting through reconstruction operations to obtain a loss function for a structure-to-attribute decoder
Figure RE-GDA0002737443240000137
Figure RE-GDA0002737443240000138
Wherein n isaIn order to be the number of attributes,
Figure RE-GDA0002737443240000139
for the property-one-hot representation of node p,
Figure RE-GDA00027374432400001310
is a penalty factor.
Step 6-2, build attribute to structure decoder, let step 5-4 fuse attribute representation
Figure RE-GDA00027374432400001311
Deriving a reconstructed structure representation by an attribute-to-structure decoder
Figure RE-GDA00027374432400001312
Interacting through reconstruction operations to obtain a loss function for an attribute-to-structure decoder
Figure RE-GDA00027374432400001313
Figure RE-GDA0002737443240000141
Wherein n issIn order to be the number of attributes,
Figure RE-GDA0002737443240000142
for the property-one-hot representation of node p,
Figure RE-GDA0002737443240000143
is a penalty factor. The decoder parameters are trained along with the method.
In step 7, the fusion structure of the thesis obtained by the interaction in step 6 is represented
Figure RE-GDA0002737443240000144
And step 6, the fusion attribute representation of the thesis node is obtained by interaction
Figure RE-GDA0002737443240000145
Is spliced to obtain
Figure RE-GDA0002737443240000146
As the final output representation.
And 8, inputting the low-dimensional representation of the labeled thesis nodes in the thesis network into a support vector machine classifier for training to obtain the trained classifier.
And 9, inputting the thesis nodes without labels in the thesis network into a classifier, obtaining the prediction labels of the thesis nodes as classification results, and performing thesis classification and thesis recommendation of the same category.
Example 1
This example uses the Cora paper citation data set and the DBLP paper citation data set for experimental and validation work.
Cora is a paper reference dataset containing 2708 paper nodes and 5429 edges expressing reference relationships, and the attribute information of each node is a unique representation extracted from the title and abstract text, with 1433 dimensions. Each node belongs to a paper domain category, with a total of 7 categories.
DBLP is a paper reference dataset containing 18448 paper nodes and 45526 edges expressing reference relationships, and the attribute information of each node is a unique hot representation extracted from the paper text, having 2476 dimensions. Each node belongs to a category of the paper domain, with a total of 4 categories.
In order to verify the effectiveness of the invention, the invention mainly compares the performances of different methods in the node classification task:
deepwalk ([ Perozzi et al, 2014] Deepwalk: Online learning of social representations): a method based on random walk generates a paper node sequence through the walk and learns to obtain a low-dimensional expression vector of a paper node by using a skip-gram algorithm and a hierarchical softmax optimization method in a word2vec tool.
Node2vec ([ Grover et al, 2016] Node2vec: Scalable feature learning for networks): a method based on biased random walk controls the breadth and depth of the walk through parameters p and q, and after a biased walk sequence is obtained, low-dimensional vector representation of a node is obtained through learning through a Skip-gram algorithm and a negative sampling optimization method.
TADW ([ Yang et al, 2015] Network representation learning with rich information): a thesis network representation learning method capable of fusing structure information and attribute information simultaneously utilizes the structure information and the attribute information simultaneously through a matrix decomposition method. Because of the simultaneous use of multi-source information, the effectiveness of TADW is generally better than models that use only structural information.
TriDNR ([ Pan et al, 2016] Tri-Party Deep Network replication): the model is a semi-supervised attribute network representation learning model based on a simple neural network, and two training networks are constructed: a structure training network and an attribute-label training network, which are shared by the two to represent the nodes. The method can simultaneously fuse structure information, attribute information and supervision information. The overhead information used by the method is removed from the comparison for fairness.
MVC _ DNE ([ Yang et al, 2017] From properties to links: Deep network embedding on completing graphs): the method is a depth learning method based on a multilayer self-encoder, and the method assumes that a mutual mapping relationship exists from a structure to an attribute to a certain extent, and the structure and the attribute are kept consistent in a depth space during training.
The method of node classification is selected for comparing the expression effects of the paper node vectors. In the experiment, a Cross-validation (Cross-validation) method is adopted, and an SVM classifier is used for evaluating the classification effect of each data set and the method.
The invention adopts two measures of Micro-average (Micro-F1) and Macro-average (Macro-F1).
The calculation mode of the Micro-F1 is as follows:
Figure RE-GDA0002737443240000151
wherein the Macro-F1 is calculated in the following way:
Figure RE-GDA0002737443240000152
wherein P and R represent precision and recall, respectively.
The effect of the node classification on the Cora dataset and the DBLP dataset of the present invention and other methods is shown in FIG. 5, where 10% and 50% of the nodes of the paper are taken as training data, respectively, and the rest are taken as test data. This process was repeated 30 times, and the average of Micro-F1 (Mi-F1 in the figure) and Macro-F1 (Ma-F1 in the figure) was used as a final index for comparison. Compared with the models Deepwalk and node2vec which do not utilize the attribute information of the paper network, the method utilizing the attribute information is obviously superior in performance. Compared with the methods TADW, TriDNR and MVC-DNE which simultaneously utilize the thesis network structure information and the attribute information, the method of the invention simultaneously combines the depth self-encoder and the truncation neighbor sampling, and can realize better performance. On two data sets and two measurement indexes, the method provided by the invention can be improved by about 2 percent relatively to a second best method, and a great improvement is obtained.
While the invention has been described in terms of a number of papers and methods for their identification and recommendation, which are presently preferred embodiments of the invention, it should be understood that various modifications and adaptations can be made by those skilled in the art without departing from the principles of the invention, and are intended to be within the scope of the invention. All the components not specified in the present embodiment can be realized by the prior art.

Claims (10)

1. A method for paper classification and recommendation, comprising the steps of:
step 1, obtaining the candidate of the truncated attribute neighbor of each paper node in the paper network G by truncation neighbor sampling, and using a set symbol CapRepresenting a truncated attribute neighbor candidate set of a paper node p;
step 2, calculating to obtain a final truncated attribute neighbor based on the truncated attribute neighbor candidate set;
step 3, reserving T structure neighbors of the thesis node p as final structure neighbors;
step 4, establishing a structural multilayer self-encoder AEsThe structure unique representation of the paper nodes is used as input, and the preliminary structure representation and the structure representation of attribute neighbors of each paper node in the attribute network are obtained through a structure multilayer self-encoder; build Attribute multilayer autoencoder AEaThe attribute unique representation of the paper nodes is used as input, and the preliminary attribute representation and the attribute representation of the structure neighbor of each paper node in the attribute network are obtained through an attribute multilayer self-encoder;
step 5, fusing the preliminary structure representation of the paper node and the structure representation of the attribute neighbor to obtain a fused structure representation of the paper node, and fusing the preliminary attribute representation of the paper node and the attribute representation of the attribute neighbor to obtain a fused attribute representation of the paper node;
step 6, enabling the fusion structure representation and the attribute multilayer self-encoder in the step 5 to interact through reconstruction operation, and enabling the fusion attribute representation and the structure multilayer self-encoder in the step 5 to interact;
step 7, splicing the fusion structure representation obtained through the interaction in the step 6 and the fusion attribute representation obtained through the interaction in the step 6 to obtain a low-dimensional representation of the paper node;
step 8, inputting the low-dimensional representation of the labeled thesis nodes in the thesis network into a support vector machine classifier for training to obtain a trained classifier;
and 9, inputting the thesis nodes without labels in the thesis network into a classifier, obtaining the prediction labels of the thesis nodes as classification results, and performing thesis classification and thesis recommendation of the same category.
2. The method of claim 1, wherein the truncated neighbor samples in step 1 are: randomly walking each paper node by a given walking step length L and the walking times gamma of each node, obtaining a walking sequence of each paper node by walking, and obtaining a truncated attribute neighbor candidate set C of each paper node by counting and de-duplicating the paper nodes in the sequencea。
3. The method of claim 2, wherein step 2 comprises:
step 2-1, setting a truncated attribute neighbor candidate set of the paper node p
Figure FDA0002696978610000011
Figure FDA0002696978610000012
Representing the ith truncated attribute neighbor in the truncated attribute neighbor candidate set of paper node p, truncated attribute neighbor candidate set CapIs a collection of property-only-hot representations
Figure FDA0002696978610000013
Figure FDA0002696978610000014
To represent
Figure FDA0002696978610000015
The property of (1) is expressed by single heat, and the value of i is 1-m;
one-hot representation of attributes of paper node p
Figure FDA0002696978610000021
And attribute-only representation set of truncated attribute neighbor candidate set
Figure FDA0002696978610000022
And (3) performing cosine similarity calculation:
Figure FDA0002696978610000023
Figure FDA0002696978610000024
to represent
Figure FDA0002696978610000025
And ith attribute unique hot representation
Figure FDA0002696978610000026
Cosine similarity of (d);
then m sets of cosine similarity results are obtained as
Figure FDA0002696978610000027
Step 2-2, sorting the cosine similarity results of the truncation attribute neighbors of the paper node p in a descending order, and taking the top K maximum neighbors as the final attribute neighbors of the paper node p; if m is less than K, the random sampling is put back to obtain a set of
Figure FDA0002696978610000028
Wherein
Figure FDA0002696978610000029
Representing the ith property of the sampled node pNeighbors, CApRepresenting a collection of these attribute neighbors.
4. The method according to claim 3, wherein in step 3, T final structure neighbors of a paper node are obtained by random sampling; if the number of the structure neighbor candidates of the thesis node is less than T, the random sampling is put back to obtain a set of
Figure FDA00026969786100000210
Wherein
Figure FDA00026969786100000211
And i represents the ith structure neighbor of the node p, and the value of i is 1-T.
5. The method of claim 4, wherein step 4 comprises:
step 4-1, building a structural multilayer self-encoder AEsStructured multilayer autoencoder AEsThe device comprises an encoding module and a decoding module, wherein the encoding module is as follows:
y1＝σ(W1xs+b1)
ym＝σ(Wmym-1+bm)
wherein, yiRepresents the output of the i-th fully-connected layer, WiAnd biThe distribution represents the projection matrix and offset vector in the ith layer,
Figure FDA00026969786100000212
structural one-hot representation of the representation input, dsFor the dimension of the one-hot representation of the structure, σ is the sigmoid function,
Figure FDA00026969786100000213
denotes dsA complex space of dimensions,;
the decoding module inverts the process of the coding module to obtain output
Figure FDA00026969786100000214
The parameters are trained along with the learning method through reconstruction, and the loss function of the structural multi-layer self-encoder is
Figure FDA00026969786100000215
Figure FDA00026969786100000216
Wherein n issAs to the number of the nodes,
Figure FDA0002696978610000031
the structure input for the ith node is expressed uniquely,
Figure FDA0002696978610000032
for the structural representation resulting from the decoding by the encoder,
Figure FDA0002696978610000033
is a penalty parameter;
paper node p and attribute neighbor CApStructural independent thermal representation input structure multilayer self-encoder AE ofsObtaining a preliminary structural representation of the nodes of the thesis at the hidden layer
Figure FDA0002696978610000034
And structural representation of attribute neighbors
Figure FDA0002696978610000035
Wherein
Figure FDA0002696978610000036
Obtaining a preliminary structure representation for the ith attribute neighbor from the hidden layer;
step 4-2, establishing an attribute multilayer self-encoder AEaPaper node p and structure neighbor CspAttribute-independent representation of input attributes multilayer self-encoderAEaObtaining a preliminary attribute representation of a paper node p at a hidden layer
Figure FDA0002696978610000037
And attribute representation of structure neighbors
Figure FDA0002696978610000038
Wherein
Figure FDA0002696978610000039
And l is 1-T for the initial attribute representation obtained from the hidden layer for the ith structure neighbor.
6. The method of claim 5, wherein step 5 comprises:
step 5-1, calculating a structural similarity coefficient through point multiplication by a point multiplication attention mechanism method, and representing according to the primary structure of the thesis node p
Figure FDA00026969786100000310
And structural representation of attribute neighbors
Figure FDA00026969786100000311
Calculating point-by-point similarity:
Figure FDA00026969786100000312
wherein ep，jNormalizing the point-by-point similarity between the thesis node p and the attribute neighbor K to obtain a similarity coefficient alphap，j：
Figure FDA00026969786100000313
Wherein exp represents a natural index; normalization by the inverse mechanism yields a new inverse attention coefficient
Figure FDA00026969786100000314
Figure FDA00026969786100000315
Wherein the reverse attention coefficient indicates the attention degree of the paper node to each neighbor;
neighbor information representation weighted by attention coefficient calculation
Figure FDA00026969786100000316
Figure FDA00026969786100000317
Step 5-2, obtaining the fusion structure representation of the thesis node
Figure FDA00026969786100000318
Step 5-3, calculating to obtain weighted neighbor information representation
Figure FDA0002696978610000041
Step 5-4, obtaining the fusion attribute representation of the paper node
Figure FDA0002696978610000042
7. The method of claim 6, wherein step 5-2 comprises: the importance of the neighbors is controlled using a weighted sum mechanism:
Figure FDA0002696978610000043
obtaining a fusion of paper nodesStructural representation
Figure FDA0002696978610000044
Wherein the parameter lambdasAnd the value is 0-1, and the method is used for controlling the importance degree represented by the neighbor information.
8. The method of claim 7, wherein step 5-3 comprises: the weighted neighbor information representation is calculated using the following formula
Figure FDA0002696978610000045
Figure FDA0002696978610000046
9. The method of claim 8, wherein steps 5-4 comprise: the importance of the neighbors is controlled using a weighted sum mechanism:
Figure FDA0002696978610000047
obtaining a fused attribute representation of a paper node
Figure FDA0002696978610000048
Wherein the parameter lambdaaAnd the value is 0-1, and the method is used for controlling the importance degree represented by the neighbor information.
10. The method of claim 9, wherein step 6 comprises:
step 6-1, build structure-to-attribute decoder, let the fused structure representation in step 5-2
Figure FDA0002696978610000049
Deriving a reconstructed property representation by a structure-to-property decoder
Figure FDA00026969786100000410
Interacting through reconstruction operations to obtain the loss of structure-to-attribute decoder
Figure FDA00026969786100000411
Figure FDA00026969786100000412
Wherein n isaIn order to be the number of attributes,
Figure FDA00026969786100000413
for the property-one-hot representation of node p,
Figure FDA00026969786100000414
is a penalty coefficient;
step 6-2, build attribute to structure decoder, let step 5-4 fuse attribute representation
Figure FDA00026969786100000415
Deriving a reconstructed structure representation by an attribute-to-structure decoder
Figure FDA00026969786100000416
Interacting through reconstruction operations to obtain attribute-to-structure decoder losses
Figure FDA0002696978610000051
Figure FDA0002696978610000052
Wherein n issIn order to the number of nodes of the paper,
Figure FDA0002696978610000053
the structural one-hot representation of the representation node p,
Figure FDA0002696978610000054
is a penalty factor.
CN202011009122.9A 2020-09-23 2020-09-23 Paper classification and recommendation method Active CN112148876B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011009122.9A CN112148876B (en) 2020-09-23 2020-09-23 Paper classification and recommendation method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011009122.9A CN112148876B (en) 2020-09-23 2020-09-23 Paper classification and recommendation method

Publications (2)

Publication Number Publication Date
CN112148876A true CN112148876A (en) 2020-12-29
CN112148876B CN112148876B (en) 2023-10-13

Family

ID=73897877

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011009122.9A Active CN112148876B (en) 2020-09-23 2020-09-23 Paper classification and recommendation method

Country Status (1)

Country Link
CN (1) CN112148876B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112749757A (en) * 2021-01-21 2021-05-04 厦门大学 Paper classification model construction method and system based on gated graph attention network
CN112836050A (en) * 2021-02-04 2021-05-25 山东大学 Citation network node classification method and system aiming at relation uncertainty

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108228728A (en) * 2017-12-11 2018-06-29 北京航空航天大学 A kind of paper network node of parametrization represents learning method
CN109376857A (en) * 2018-09-03 2019-02-22 上海交通大学 A kind of multi-modal depth internet startup disk method of fusion structure and attribute information
US20200082531A1 (en) * 2018-09-07 2020-03-12 3Mensio Medical Imaging B.V. Method, Device and System for Dynamic Analysis from Sequences of Volumetric Images

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108228728A (en) * 2017-12-11 2018-06-29 北京航空航天大学 A kind of paper network node of parametrization represents learning method
CN109376857A (en) * 2018-09-03 2019-02-22 上海交通大学 A kind of multi-modal depth internet startup disk method of fusion structure and attribute information
US20200082531A1 (en) * 2018-09-07 2020-03-12 3Mensio Medical Imaging B.V. Method, Device and System for Dynamic Analysis from Sequences of Volumetric Images

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
CHUNYANG TAN: "MSGE: A Multi-step Gated Model for Knowledge Graph Completion", 《SPRINGER》 *
余传明;林奥琛;钟韵辞;安璐;: "基于网络表示学习的科研合作推荐研究", 情报学报, no. 05 *
谭春阳: "复杂网络结构的表示学习研究", 《硕士电子期刊》 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112749757A (en) * 2021-01-21 2021-05-04 厦门大学 Paper classification model construction method and system based on gated graph attention network
CN112749757B (en) * 2021-01-21 2023-09-12 厦门大学 Thesis classification model construction method and system based on gating graph annotation force network
CN112836050A (en) * 2021-02-04 2021-05-25 山东大学 Citation network node classification method and system aiming at relation uncertainty
CN112836050B (en) * 2021-02-04 2022-05-17 山东大学 Citation network node classification method and system aiming at relation uncertainty

Also Published As

Publication number Publication date
CN112148876B (en) 2023-10-13

Similar Documents

Publication Publication Date Title
CN110321494B (en) Socialized recommendation method based on matrix decomposition and network embedding combined model
CN112380435B (en) Document recommendation method and system based on heterogeneous graph neural network
CN109753589A (en) A kind of figure method for visualizing based on figure convolutional network
CN111709518A (en) Method for enhancing network representation learning based on community perception and relationship attention
Qu et al. Curriculum learning for heterogeneous star network embedding via deep reinforcement learning
CN111079409A (en) Emotion classification method by using context and aspect memory information
CN112100485B (en) Comment-based scoring prediction article recommendation method and system
CN109447261B (en) Network representation learning method based on multi-order proximity similarity
CN112597296A (en) Abstract generation method based on plan mechanism and knowledge graph guidance
CN112148876A (en) Paper classification and recommendation method
CN112559764A (en) Content recommendation method based on domain knowledge graph
Song et al. Session-based recommendation with hierarchical memory networks
CN110781271A (en) Semi-supervised network representation learning model based on hierarchical attention mechanism
CN107491782A (en) Utilize the image classification method for a small amount of training data of semantic space information
Luo et al. ResumeNet: A learning-based framework for automatic resume quality assessment
Jin et al. Deepwalk-aware graph convolutional networks
CN112784118A (en) Community discovery method and device in graph sensitive to triangle structure
Cheng et al. Dynamic embedding on textual networks via a gaussian process
CN109815335A (en) A kind of paper domain classification method suitable for document network
Li et al. Neural architecture search via proxy validation
CN117150041A (en) Small sample knowledge graph completion method based on reinforcement learning
Bao et al. HTRM: a hybrid neural network algorithm based on tag-aware
CN116821371A (en) Method for generating scientific abstracts of multiple documents by combining and enhancing topic knowledge graphs
CN116610874A (en) Cross-domain recommendation method based on knowledge graph and graph neural network
CN117009674A (en) Cloud native API recommendation method integrating data enhancement and contrast learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant