CN108228782B - A Deep Learning-Based Approach for Implicit Relationship Discovery - Google Patents

Abstract

The invention discloses an implicit relationship discovery method based on deep learning, which belongs to the field of information technology, and specifically includes generating a paper co-authoring network G' from a scholar publishing network G; Co-authored the matrices X _S , X _D , X _T ; proposed the RGRU model; designed and constructed a tARMM model based on RGRU to predict the "mentor-student" relationship. The prediction accuracy of the tARMM model proposed by the invention is higher than other methods on the data set, which can reach about 95%, and has certain reference significance and reference value for other time-dependent social relationship mining.

Description

A Deep Learning-Based Approach for Implicit Relationship Discovery

technical field

The invention belongs to the field of information technology, and in particular relates to a deep learning-based implicit relationship discovery method.

Background technique

With the popularization and promotion of social media such as Facebook, Twitter and WeChat, social media has become an important platform for people to communicate and interact. Different types of social relationships have different effects on people, and people’s life, study and work are changing under the influence of these relationships. For example, in social networks, people’s preferences will be influenced by friends, and students’ research directions will be affected. Influenced by the teacher. At the same time, there is also a lot of additional information implied in these relationships. For example, by studying the "mentor-student" relationship, one can explore academic groups, establish a scientific research community network, further understand the development process of related research topics, and find the next development direction. .

There are many explicit relationships in the network, such as friend relationship, follow relationship, comment relationship, reply relationship, etc. However, there are also many relationships that are implicit in the network, such as: the "tutor-student" relationship is implicit in the co-authorship of the paper. in the network. The paper co-authorship network is a cooperative relationship network gradually formed by researchers in the process of co-publishing documents, such as DBLP. Currently, there are several projects whose goal is to maintain relationships, such as LinkedIn and AI genealogy. The former requires users to label each special object, such as colleagues, mentors, students, etc. The latter also uses manual labeling methods to label the mentor information in the research field. Obviously, these methods rely heavily on manual annotation, which is not only inefficient but also inaccurate, which greatly limits their generalization ability. An ideal solution to this phenomenon is to devise a method that automatically mines or predicts implicit relationships from the network.

In a paper co-authorship network, it can be difficult to tell who is a supervisor just from the publication list. Using heuristic rules can sometimes distinguish relationship types in certain social networks, based on intuitive assumptions. However, the study found that using typical heuristic rules can only achieve an accuracy of 70-80%, even using multiple rules trained on multiple different features combined with a supervised learning model, the accuracy is still only 80% on average, and, in practice training It is often difficult to collect supervisory information.

The "mentor-student" relationship in the paper co-authorship network has the following characteristics:

1. Implicit. The "mentor-student" relationship is hidden in the paper co-authoring network. In the paper co-authoring network, only the co-authors of the paper, the title of the paper, the publication time of the paper, the publication/conference where the paper was published, etc., cannot be explicitly published. Know the "mentor-student" relationship between collaborators.

2. Time dependence. The mentor-student relationship is highly time-dependent. For any author, among its many collaborators, early collaborators are more likely to be their mentors than later ones. Furthermore, a person can transition from a student role to a mentor role without any visible signs of that role transition.

3. Hard to speculate. Since the paper co-authoring network only has relevant information about co-published papers, it is very simple compared to other social media, and because the "tutor-student" relationship is hidden in the paper co-authoring network, this leads to the It is difficult to infer the "tutor-student" relationship manually.

In recent years, the study of social relations has attracted extensive attention in the academic circles. The current research work on social relationships can be divided into three aspects: social relationship prediction, social relationship type identification and relationship interaction prediction.

Social relationship prediction, also known as link prediction, refers to predicting the possibility of an edge between two nodes based on the characteristics of nodes in the network or existing edges. For a specific social network, Liben-Nowell et al. calculated the similarity between nodes using a graph-based similarity measurement method, and then used the similarity to predict the link probability between nodes. Lee et al. proposed a computationally inexpensive model based on social vector clock features to solve the link prediction problem. Cuchao Tu et al. proposed the CANE model, which achieves the goal of link prediction by embedding the user-related text data information in the network. A random walk algorithm based on supervised learning was proposed by Backstrom et al. Zhao et al. proposed a prediction method based on "reliable paths", which is one of the few prediction methods suitable for weighted networks.

Relationship type identification refers to automatically identifying and mining the relationship types contained in one or more social networks. Coppola et al. proposed a semantic-based automatic relation mining framework. Leskovec et al. used a logarithmic regression model to identify positive or negative relationships in social networks, that is, friend relationships or non-friend relationships. Diehl et al. used a learned ranking function to identify "manager-subordinate" relationships. Pentland et al. proposed several obile data mining models for inferring friend relationships. The problem of "mentor-student" relationship mining in the paper co-authorship network belongs to the problem of relationship type identification. On this issue, Tang Jie et al. proposed a TPFG model to mine the "mentor-mentee" relationship from the paper co-authorship network. In addition, they propose a unified framework based on factor graphs for heterogeneous networks (such as email networks, scientific research collaboration networks, etc.), aiming to solve the identification problem of social relationship types. Li Yongjun et al. used the maximum entropy model to infer the "tutor-student" relationship in the paper co-authorship network.

Relationship interaction prediction mainly studies how a one-way social relationship develops into a two-way social relationship, and the reasons for its change. The most common one-way relationship is that between celebrities and their fans, and the two-way relationship is friendship. Hopcroft et al. explored the problem of relationship interaction prediction, and Lou et al. studied how social relationships developed into ternary closures. Together, they propose a learning framework that abstracts the relational interaction prediction problem as a graph.

SUMMARY OF THE INVENTION

In view of the above technical problems existing in the prior art, the present invention proposes a method for discovering implicit relationships based on deep learning, which has a reasonable design, overcomes the deficiencies of the prior art, and has good effects.

In order to achieve the above object, the present invention adopts the following technical solutions:

An implicit relationship discovery method based on deep learning, which formally defines the implicit relationship mining problem:

Definition 1 Scholars publish network G

表示学者发表网络中所有作者的集合；

是所有论文的集合；E＝{e_ik|1＜＝i＜＝n_a,1＜＝k＜＝n_p,a_i是p_k的作者}，表示学者发表网络中的作者与论文的著作关系；Formally represent the time-dependent scholarly publication network as a bipartite graph, let G=(A,P,E), where

Represents the collection of all authors in the scholar publication network;

is the set of all papers; E={e _ik |1＜=i＜=n _a , 1＜=k＜=n _p , a _i is the author of p _k }, which means the authors and papers published by scholars in the network relation;

Definition 2 Paper Co-authored Network G’

其中，

是作者集合，a₀是一虚拟作者，对于作者a_i，假设其导师为

如果

那么认为

E’＝{e_ij|1<＝i<＝n_a,1<＝j<＝n_a,a_i和a_j具有合作关系且a_i≠a_j}；其中，pn_ij是与e_ij相关的一个向量，pn_ij∈R^1×40表示a_i和a_j在某一个时间域内合著的论文数量；对于单个作者来说，使用pn_i可以表示作者a_i论文发表情况；generated from G

in,

is the set of authors, a ₀ is a virtual author, and for author a _i , suppose that its tutor is

if

then think

E'={e _ij |1<=i<=n _a , 1<=j<=n _a , a _i and a _j have a cooperative relationship and a _i ≠a _j }; where pn _ij is related to e _ij A vector of , pn _ij ∈ R ^1×40 represents the number of papers co-authored by a _i and a _j in a certain time domain; for a single author, pn _i can be used to represent the publication situation of author a _i ’s papers;

Definition 3 Paper Co-authorship Matrix C

则对于作者x，有合著矩阵：For any author x in A, assuming that it has a co-authorship relationship with m authors, the set of collaborators is represented by A _x , A _x ={b ₀ ,b ₁ ,b ₂ ,...,b _m }, where b ₀ =a ₀ ; if in a certain year t, the number of papers co-authored by x and b _j is

Then for author x, there is a co-authorship matrix:

Among them, T is the overall time domain of the author's cooperation. This article takes one year as a time span. If the author's co-authorship time is [1970, 2010], a total of 40 years, then in the above matrix T=39, the co-authorship matrix C∈ R ^(m+1)×40 ;

Definition 4 Tutor-Student Relationship R

Let R={y _ij |0<=i<=n _a , 0<=j<=n _a }, indicating whether there is a "tutor-student" relationship between authors, and its specific values are as follows:

The described deep learning-based implicit relationship discovery method specifically includes the following steps:

Input: Scholars publish network G;

Output: The predicted result of the "Tutor-Student" relationship;

Step 1: Analyze the links in the scholar's publication network G, and generate a paper co-authorship network G' from the scholar's publication network G;

Step 2: According to the paper co-authorship network G', calculate the paper publication matrix C, D, S, and then calculate the paper's co-author matrix X _S , X _D , X _T ;

Step 3: establish a tARMM (time-aware Advisor-advise Relationship Mining Model, time-aware advisor-student relationship mining model) model;

Step 4: Process the co-authorship matrix through the tARMM model;

Step 4.1: Calculate probability P _T using RGRU;

Step 4.2: Calculate probability P _F using DNN;

Step 4.3: Calculate the final mentor probability P;

Step 5: The candidate tutor with the highest probability in P is the predicted tutor of x, so as to obtain the prediction result of the "tutor-student" relationship.

Preferably, in step 2, the co-authorship of the paper is analyzed from the following two aspects:

The first aspect is to analyze the details of the co-authorship. For the author x, the co-authorship matrix C is used to represent the publication of the co-authored papers between x and his candidate supervisor;

The publication status of candidate supervisors’ papers is denoted by D:

The publication situation of author x's paper pn _x is represented by S:

S = (S ₀ ... S _T-1 ) (2.3);

The co-authorship matrix C is normalized by the publications of the authors and candidate supervisors respectively:

X _S =C · S (2.5);

X _D = D · S (2.6);

Among them, X _S is the co-authored sub-matrix based on students, X _Sij ∈ X _S , indicating the ratio of the number of papers co-authored by author x and his candidate supervisor bi to the total number of papers of author x in year _j in the jth year; X _D is the co-authored submatrix based on the supervisor, X _Dij ∈ X _D , which represents the proportion of the number of papers co-authored by the author x and his candidate supervisor _bi in the _jth year to the total number of papers in the jth year by the candidate supervisor bi;

In the second aspect, from the perspective of co-authoring time, the time structure of co-authoring situations is represented in the form of a matrix according to the co-authoring matrix C, which is specifically defined as follows:

X _T is a co-authorship sub-matrix based on time structure, which means that the time structure of co-authored papers between author x and his candidate supervisor b _i is represented in the form of a matrix.

Preferably, in step 4.1, in the tARMM model, the RNN (Recursive Neural Network, Recurrent Neural Network) is transformed to generate an update gate recurrent unit RGRU (Refresh Gate Recurrent Unit, update gate recurrent unit), by updating the gate recurrent unit RGRU, process X _T to obtain the mentor probability P _T ;

For time t, there are:

r _t =σ(w _r h _t+1 +w _h x _t + _br ) (2.9);

h _t =(1-r _t )h _t+1 +r _t x _t (2.10);

Among them, σ refers to the sigmoid activation function, r _t is the state of the update gate at time t, w _r w _r and w _h are the weight matrices of the update gate, b _r is the offset of the update gate, h _t+1 is the state of the update gate unit at time t+1, x _t is the input matrix at time t, and h _t is the state of the update gate unit at time t;

RGRU-based mentor probability P _T :

P _T =h _T (2.11);

Among them, h _T is the state of the update gate unit at time T; its formula is the same as h _t ;

Specific steps are as follows:

Input: paper co-author matrix X _T ;

Output: RGRU-based mentor probability P _T ;

Step 4.1.1: Initialize P _T to zero matrix;

Step 4.1.2: Calculate the state r _t of the update gate in year t by formula (2.9);

Step 4.1.3: Calculate the state h _t of the update gate unit in year t by formula (2.10);

Step 4.1.4: Calculate the mentor probability P _T of x by formula (2.11).

Preferably, in step 4.2, through the tARMM model, a deep neural network is used to process X _S and X _D to obtain a mentor probability P _F based on the class diagram matrix;

Combine X _S and X _D to form a two-color channel bitmap, which is called the class diagram matrix X; the goal is to find the row number of the specific graphic in the class diagram matrix X; since this is a pixel-level target Positioning problem, so build a DNN for identification, according to the calculation formula of the perceptron, for each node in the DNN, the output is:

Among them, w _i , b are the weight and offset parameters of the model, and p _i is the probability value predicted by each node;

Then the tutor probability P _F based on the class graph matrix finally generated by DNN is the output of the last layer of DNN:

P _F =Relu(f(X _S ,X _D )) (2.13);

Specific steps are as follows:

Input: paper co-author matrix X _S and X _D ;

Output: tutor probability P _F based on class diagram matrix;

Step 4.2.1: Initialize _PF to zero matrix;

Step 4.2.2: Calculate the output of each node in the DNN by formula (2.12);

Step 4.2.3: Calculate probability PF by formula ( _2.13 ).

Preferably, in step 4.3, the final mentor probability matrix is generated by PT and _PF through the fully connected layer, and the highest probability value _P is selected from it, and the corresponding candidate mentor is the predicted mentor of x;

P=σ(P _F · P _T ) (2.14).

Beneficial technical effects brought by the present invention:

The invention draws on the theory of the variant cyclic neural network (RNN) model such as the long short-term memory model (LSTM) and the logic gate recurrent unit (GRU), transforms the RNN, and proposes an update gate recurrent unit (RGRU) for processing co-authored Temporal structure in the matrix; since RGRU has only one gate unit, it is structurally simpler than LSTM and GRU, but has higher accuracy on the problem of mining the "mentor-student" relationship;

The invention adopts the idea of deep learning to deal with the problem of "mentor-student" relationship mining in the paper co-authoring network, and proposes a time-dependent "mentor-student" relationship mining neural network (tARMM), which can predict the accuracy of the model on the data set. Compared with other methods, it can reach about 95%, which has certain reference significance and reference value for other time-dependent social relationship mining.

Description of drawings

Figure 1 is a schematic diagram of the "mentor-student" relationship mining.

Figure 2 is a schematic diagram of tARMM.

FIG. 3 is a schematic diagram of the RGRU.

Figure 4(a) is a schematic diagram of the class diagram matrix of _XS .

Figure 4(b) is a schematic diagram of the class diagram matrix of _XD .

Figure 5 is a schematic diagram of the DNN.

Figure 6 is a schematic diagram of a fully connected layer.

Detailed ways

The present invention is described in further detail below in conjunction with the accompanying drawings and specific embodiments:

1. Formal definition of the problem

In this part, some basic symbols and definitions in this paper are given.

Table 1.1 Main symbols and their meanings

Definition 1 Scholars publish network G

表示学者发表网络中所有作者的集合；

是所有论文的集合；E＝{e_ik|1＜＝i＜＝n_a,1＜＝k＜＝n_p,a_i是p_k的作者}，表示学者发表网络中的作者与论文的著作关系。Formally represent the time-dependent scholarly publication network as a bipartite graph, let G=(A,P,E), where

Represents the collection of all authors in the scholar publication network;

is the set of all papers; E={e _ik |1＜=i＜=n _a , 1＜=k＜=n _p , a _i is the author of p _k }, which means the authors and papers published by scholars in the network relation.

Definition 2 Paper Co-authored Network G’

其中，

是作者集合，a₀是一虚拟作者，对于作者a_i，假设其导师为

如果

那么认为

E’＝{e_ij|1<＝i<＝n_a,1<＝j<＝n_a,a_i和a_j具有合作关系且a_i≠a_j}。pn_ij是与e_ij相关的一个向量，pn_ij∈R^{1 ×40}表示a_i和a_j在某一个时间域内合著的论文数量。对于单个作者来说，使用pn_i可以表示作者a_i论文发表情况。generated from G

in,

is the set of authors, a ₀ is a virtual author, and for author a _i , suppose that its tutor is

if

then think

E'={e _ij |1<=i<=n _a , 1<=j<=n _a , a _i and a _j have a cooperative relationship and a _i ≠a _j }. pn _ij is a vector related to e _ij , and pn _ij ∈ R ^{1 ×40} represents the number of papers co-authored by a _i and a _j in a certain time domain. For a single author, the use of pn _i can represent the publication status of author a _i .

Definition 3 Paper Co-authorship Matrix C

则对于作者x，有合著矩阵：For any author x in A, assuming that it has a co-authorship relationship with m authors, the set of collaborators is represented by A _x , A _x ={b ₀ ,b ₁ ,b ₂ ,...,b _m }, where b ₀ = a ₀ . If in a certain year t, the number of papers co-authored by x and b _j is

Then for author x, there is a co-authorship matrix:

Among them, T is the overall time domain of the author's cooperation. This article takes one year as a time span. If the author's co-authorship time is [1970, 2010], a total of 40 years, then in the above matrix T=39, the co-authorship matrix C∈ R ^(m+1)×40 .

Definition 4 Tutor-Student Relationship R

Let R={y _ij |0<=i<=n _a , 0<=j<=n _a }, indicating whether there is a "tutor-student" relationship between authors, and its specific values are as follows:

The research goal of this paper is to predict the mentor y _x of x from C, which needs to solve the two problems of who is the mentor of x and what probability is the mentor of x.

2 Model building

2.1 Construction of co-authorship matrix

In order to mine the "mentor-student" relationship, the paper co-authorship network G' is firstly generated from the original scholar publication network G, and then the co-authorship matrix (as shown in Figure 1) is extracted from it.

For the co-authorship of the paper, it can be analyzed from the following two aspects:

The first aspect analyzes the details of the co-authorship. For an author x, there is a co-authorship matrix C that represents the publication of co-authored papers between x and his candidate supervisors.

The publication status of candidate supervisors’ papers is denoted by D:

The publication situation of author x's paper pn _x is represented by S:

S = (S ₀ ... S _T-1 ) (2.3);

Therefore, the co-authorship matrix C is normalized by using the papers published by the authors and candidate supervisors respectively:

X _S =C · S (2.5);

X _D = D · S (2.6);

X _S and X _D are the student-based co-authorship sub-matrix and the tutor-based co-author sub-matrix, respectively. For X _Sij ∈ X _S , it represents the ratio of the number of papers co-authored by author x and his candidate supervisor bi in the _jth year to the total number of papers of the author x in the jth year. X _Dij ∈ X _D represents the ratio of the number of papers co-authored by author x and his candidate supervisor _bi in the _jth year to the total number of papers in the jth year by the candidate supervisor bi.

In the second aspect, from the perspective of co-authoring time, the time structure of co-authoring situations is represented in the form of a matrix according to the co-authoring matrix C, which is specifically defined as follows:

X _T becomes a co-authorship sub-matrix based on time structure, which means that the time structure of co-authored papers between author x and his candidate supervisor is represented in the form of a matrix.

2.2 Time-dependent relationship mining model construction

In this section, we propose a time-dependent relation mining neural network model tARMM (as shown in Figure 2), which processes X _T and X _S and X _D respectively to obtain a mentor based on time structure and class diagram matrix. probability matrix, and then generate the final tutor probability matrix through the fully connected layer. When processing X _T , a reverse-time update gate cyclic unit is designed, and when processing X _S and X _D , a deep neural network is used.

2.2.1 Probability calculation method based on RGRU

In this paper, an update gate unit is added on the basis of the standard RNN to form a recurrent neural network with only one update gate, which is called the update gate recurrent unit RGRU (as shown in Figure 3). For the co-authorship matrix X _T based on publication time, formula (2.7) shows that in the matrix, the higher the column where the non-zero element is located, the higher the probability of the candidate mentor represented by the row. Therefore, the matrix X _T is reversely processed by RGRU in units of columns, and the RGRU-based tutor probability matrix is obtained.

For time t, there are:

r _t =σ(w _r h _t+1 +w _h x _t + _br ) (2.9);

h _t =(1-r _t )h _t+1 +r _t x _t (2.10);

Among them, r _t is the state of the update gate at time t, w _r w _h is the weight matrix of the update gate, b _r is the offset of the update gate, and h _t+1 is the state of the update gate unit at time t+1 , x _t is the input matrix at time t, h _t is the state of the update gate unit at time t;

RGRU-based mentor probability P _T :

P _T =h _T (2.11);

Among them, h _T is the state of the update gate unit at time T; the formula of h _T is the same as that of h _t .

To sum up, the mentor probability calculation based on the modified update gate recurrent unit (RGRU) is shown in Algorithm 1.

2.2.2 Probability calculation method based on class diagram matrix

X _S and X _D characterize the co-authorship situation from the perspective of students and candidate supervisors, respectively. Taking _XS as an example, the student-based co-authorship matrix is regarded as a 66×40 grayscale image and displayed in bitmap mode. It can be found that when the candidate tutor represented by a row is the actual tutor, the There will be a continuous segment of pixel values that constitute a special image similar to "one", but have different characteristics in different bitmaps, so this paper processes the bitmap through a deep neural network, extracts its feature matrix, and mines special images. the location. Figure 4(a) is the _XS class diagram matrix; Figure 4(b) is the _XD class diagram matrix.

Combine X _S and X _D to form a two-color channel bitmap, which is called the class map matrix X. So the goal of the next step is to discover the row number of a particular figure in the class diagram matrix X. Since this is a pixel-level object localization problem, a DNN (as shown in Figure 5) is constructed for recognition. According to the calculation formula of the perceptron, for each node in the DNN, its output is:

Let y' be the output of the last layer in the DNN. Then the tutor probability P _F based on the class diagram matrix finally generated by DNN is:

P _F =Relu(f(X _S ,X _D ))=y'(2.13);

In summary, the implementation process of DNN can be described by the following algorithm.

Finally, the final mentor probability matrix is generated by _{PT and PF through} the fully connected layer (as shown in Figure 6), and the highest probability value is selected from it, and the corresponding candidate mentor is the predicted mentor.

P=σ(P _F ·P _T ) (2.14);

2.3 Learning algorithm of the model

This section will introduce the learning algorithm of the model, including the loss function and parameter update method. The model proposed in this paper uses cross entropy as the loss function, as follows:

In terms of parameter update, this paper adopts Adam method to optimize parameters after all parameters are initialized. The Adam method is an adaptive learning rate learning method that calculates its own learning rate for each parameter. Its formula is as follows:

m _t =β ₁ m _t-1 +(1-β ₁ )g _t (2.16);

的估计，

和

是对m_t和v_t的校正，近似为对期望的无偏估计。

是学习率的一个动态约束。Among them, m _t is the first-order moment estimation of the gradient, which can be regarded as the estimation of the expectation E|g _t |, and v _t is the second-order moment estimation of the matrix, which can be regarded as the estimation of the expected E|g t |

's estimate,

and

is the correction for m _t and v _t , which approximates an unbiased estimate of the expectation.

is a dynamic constraint on the learning rate.

2.4 Algorithm Description

The complete algorithm of the tARMM model proposed in this paper is described as follows:

3 Experimental design and analysis

3.1 Experimental setup

data set. The DBLP computer science literature database developed by Michael Ley was used as the experimental data set to infer the "mentor-student" relationship. Taking the section from 1970 to 2010, it contains 654,628 authors and 1,076,946 publications. As label data, the union of MAN, MathGP, and AIGP datasets is used as the validation dataset, where MAN is obtained by crawling on the tutor's personal homepage, MathGP is obtained by crawling from the MathematicsGenealogy project, and AIGP is obtained from Crawled from the AI Genealogy project.

A series of experiments are done to explore the correctness and effectiveness of the model on the problem of "mentor-student" relationship mining. Randomly select some data from the dataset to train the model, and then randomly select a dataset from the dataset for testing.

In order to compare the predicted results intuitively, this paper uses the most commonly used evaluation index for classification algorithms: the accuracy rate ACC, whose calculation formula is as follows:

Among them, TP is the number of true examples, and FP is the number of false positive examples.

The experimental environment is: Intel Core i5-2520M dual-core (2.5GHz), windows10 64-bit, 8G memory, NVIDIA GeForce GT635M graphics card. The programming languages are: Matlab and Python, using the TensorFlow framework.

3.2 Programming techniques

The data preprocessing stage uses Matlab to write the code, and the tARMM model implementation part is written in python, which uses the TensorFlow machine learning framework. The display part of the results is implemented by JavaWeb, and the open source ECharts component of Baidu is mainly used on the web page.

(1) TensorFlow

TensorFlow is Google's open-source second-generation software library for numerical computing. It is a very flexible machine learning framework that can run on single or multiple CPUs and GPUs on servers or personal computers and even mobile devices.

TensorFlow is a processing framework based on data flow graphs. The nodes in the data flow graph represent mathematical operations, and the edges represent the data interaction between operation nodes. In TensorFlow, Tensor represents the data transmitted between nodes, and Flow represents the data flow, that is, Tensor enters each node of the data operation graph in the form of a flow.

When programming, you need to use a graph to represent computing tasks, and then execute the graph in a context called a session. At the same time, use tensor to represent data, maintain state through variables, and use feed or fetch as Arbitrary operations assign values or get data from them.

(2)Echarts

Echarts is a pure javascript icon library, which can be embedded in html web pages and can run smoothly on computers and mobile devices. It is compatible with most current browsers. The underlying implementation relies on the lightweight Canvas class library Zrender, which provides Vivid, intuitive, interactive, and highly personalizable data visualization icons. The text uses Echarts for the presentation of experimental results.

3.3 Experimental results

For deep learning, different optimization methods will have different effects on the efficiency and effectiveness of training. In general, gradient descent is commonly used as a training method for models. Gradient descent has a variety of classifications. Among them, batch gradient descent method BGD is the most primitive form of gradient descent method. The specific idea is to use all samples to update each parameter. Since batch gradient descent requires all training samples to update each parameter, the training process becomes slower and slower as the number of samples increases. The stochastic gradient descent method SGD is proposed to solve the drawback of the batch gradient descent method. This paper adopts the stochastic gradient descent method to train the model.

3.3.1 Effectiveness of RGRU

To demonstrate the effectiveness of RGRU, experiments were conducted with RNN, LSTM and update gate recurrent unit RGRU alone, and the results were as follows:

Table 3.2 Performance of Different Neural Networks

Through experiments, it can be seen that RGRU has higher accuracy than the simple recurrent neural network and long-short-term memory model in the problem of "mentor-student" relationship mining. Prove that the improvements to the recurrent neural network are correct and effective.

3.3.2 Comparison of tARMM with other algorithms

In this part, the TPFG model and the SVM model for classification proposed for the "mentor-student" relationship are selected for comparison with the tARMM model proposed in this paper. Carry out multiple tests and take the average value. The results are as follows:

Table 3.3 Comparison of Results of Different Algorithms

It can be seen through experiments that the accuracy of the tARMM model is higher than that of SVM and TPFG, which further proves the correctness of the tARMM model.

4 Summary and Outlook

This paper investigates the problem of identifying “mentor-student” relationships in paper co-authorship networks. To solve this problem, firstly, the co-authorship matrix is generated by preprocessing the data, and then a tARMM model is established to process the co-authorship matrix to mine the "mentor-student" relationship. The RNN is transformed in the tARMM model to generate RGRU, a unit that mines time-dependent relationships. Experiments using the data in DBLP demonstrate the correctness and effectiveness of the tARMM model.

In this study, there is a certain error because the labeled dataset cannot cover the entire DBLP database. In this regard, the model will be improved by expanding the labeled dataset in the later stage to improve the accuracy of the model. At the same time, this model has certain expansibility for time-dependent social relations, and the model will be further improved for different social media in the later stage to improve the generality of the model.

Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the essential scope of the present invention should also belong to the present invention. The scope of protection of the invention.

Claims (1)

1. A method for discovering implicit relationships based on deep learning, characterized in that: a formal definition is made to the implicit relationship mining problem: Definition 1 Scholars publish network G Formally represent the time-dependent scholarly publication network as a bipartite graph, let G=(A, P, E), where

Represents the collection of all authors in the scholar publication network;

is the set of all papers; E={e _ik |1＜=i＜=n _a , 1＜=k＜=n _p , a _i is the author of p _k }, which means the authors and papers published by scholars in the network relation; Definition 2 Paper Co-authored Network G’ Generate G'=(A', E', {pn _ij }e _ij∈E' ) from G, where,

is the set of authors, a ₀ is a virtual author, and for author a _i , suppose that its tutor is

if

then think

E'={e _ij |1<=i<=n _a , 1<=j<=n _a , a _i and a _j have a cooperative relationship and a _i ≠a _j }; where pn _ij is related to e _ij A vector of , pn _ij ∈ R ^1×40 represents the number of papers co-authored by a _i and a _j in a certain time domain; for a single author, pn _i can be used to represent the publication situation of author a _i ’s papers; Definition 3 Paper Co-authorship Matrix C For any author x in A, assuming that it has a co-authorship relationship with m authors, the set of collaborators is represented by A _x , A _x ={b ₀ , b ₁ , b ₂ ,..., b _m }, where b ₀ = a ₀ ; if in a certain year t, the number of papers co-authored by x and b _j is

Then for author x, there is a co-authorship matrix:

Among them, T is the overall time domain of the author's cooperation, and this paper takes one year as a time span; Definition 4 Tutor-Student Relationship R Let R={y _ij |0＜=i＜=n _a , 0＜=j＜ ₌ na }, indicating whether there is a “tutor-student” relationship between authors, and its specific values are as follows:

The described deep learning-based implicit relationship discovery method specifically includes the following steps: Input: Scholars publish network G; Output: The predicted result of the "Tutor-Student" relationship; Step 1: Analyze the links in the scholar's publication network G, and generate a paper co-authorship network G' from the scholar's publication network G; Step 2: According to the paper co-authorship network G', calculate the paper publication matrix C, D, S, and then calculate the paper's co-author matrix X _S , X _D , X _T ; The co-authorship of the paper is analyzed from the following two aspects: The first aspect is to analyze the details of the co-authorship. For the author x, the co-authorship matrix C is used to represent the publication of the co-authored papers between x and his candidate supervisor; The publication status of candidate supervisors’ papers is denoted by D:

The publication situation of author x's paper pn _x is represented by S: S = (S ₀ ... S _T-1 ) (2.3);

The co-authorship matrix C is normalized by the publications of the authors and candidate supervisors respectively: X _S =C · S (2.5); X _D = D · S (2.6); Among them, X _S is the student-based co-authorship sub-matrix, X _Sij ∈ X _S , indicating the ratio of the number of papers co-authored by author x and his candidate supervisor bi in the _jth year to the total number of papers of the author x in the jth year; X _D is the co-authored submatrix based on the supervisor, X _Dij ∈ X _D , indicating the ratio of the number of papers co-authored by the author x and his candidate supervisor _bi in the _jth year to the total number of papers in the jth year by the candidate supervisor bi; In the second aspect, from the perspective of co-authoring time, the time structure of co-authoring situations is represented in the form of a matrix according to the co-authoring matrix C, which is specifically defined as follows:

X _T is a co-authored sub-matrix based on time structure, which means that the time structure of co-authored papers between author x and his candidate supervisor b _i is represented in the form of a matrix; Step 3: Establish a tARMM model, a time-aware mentor-student relationship mining model; Step 4: Process the co-authorship matrix through the tARMM model; Step 4.1: Use the update gate recurrent unit RGRU to calculate the mentor probability P _T ; In the tARMM model, the RNN is transformed to generate an update gate cyclic unit RGRU, and by updating the gate cyclic unit RGRU, X _T is processed to obtain a mentor probability P _T ; For time t, there are: r _t =σ(w _r h _t+1 +w _h x _t + _br ) (2.9); h _t =(1-r _t )h _t+1 +r _t x _t (2.10); Among them, σ refers to the sigmoid activation function, r _t is the state of the update gate at time _t , _wr and wh are the weight matrices of the update gate, b _r is the offset of the update gate, and h _t+1 is the update gate The state of the gate unit at time t+1, x _t is the input matrix at time t, h _t is the state of the updated gate unit at time t; RGRU-based mentor probability P _T : P _T =h _T (2.11); Among them, h _T is the state of the update gate unit at time T; its formula is the same as h _t ; Specific steps are as follows: Input: paper co-author matrix X _T ; Output: RGRU-based mentor probability P _T ; Step 4.1.1: Initialize P _T to zero matrix; Step 4.1.2: Calculate the state r _t of the update gate in year t by formula (2.9); Step 4.1.3: Calculate the state h _t of the update gate unit in year t by formula (2.10); Step 4.1.4: Calculate the RGRU-based tutor probability P _T of x by formula (2.11); Step 4.2: Use DNN to calculate the mentor probability _PF based on the class diagram matrix; Through the tARMM model, a deep neural network is used to process X _S and X _D to obtain the mentor probability P _F based on the class diagram matrix; Combine X _S and X _D to form a two-color channel bitmap, which is called the class diagram matrix X; the goal is to find the row number of the specific graphic in the class diagram matrix X; since this is a pixel-level target Positioning problem, so build a DNN for identification, according to the calculation formula of the perceptron, for each node in the DNN, the output is:

Among them, w _i , b are the weight and offset parameters of the model, and p _i is the probability value predicted by each node; Then the tutor probability P _F based on the class graph matrix finally generated by DNN is the output of the last layer of DNN: P _F =Relu(f(X _S , X _D )) (2.13); Specific steps are as follows: Input: paper co-author matrix X _S and X _D ; Output: tutor probability P _F based on class diagram matrix; Step 4.2.1: Initialize _PF to zero matrix; Step 4.2.2: Calculate the output of each node in the DNN by formula (2.12); Step 4.2.3: Calculate the tutor probability P _F based on the class diagram matrix by formula (2.13); Step 4.3: Calculate the final mentor probability P; The RGRU-based mentor probability _{P T} and the class graph matrix-based mentor probability PF are generated through the full connection layer to generate the final mentor probability matrix, and the highest probability value P is selected from it, and its corresponding candidate mentor is the predicted mentor of x; P=σ(P _F ·P _T ) (2.14); Step 5: The candidate tutor with the highest probability in P is the predicted tutor of x, so as to obtain the prediction result of the "tutor-student" relationship.

Images