CN109522954A - Heterogeneous Information network linking prediction meanss - Google Patents

Abstract

A heterogeneous information network link prediction device, the device includes: a setting unit, suitable for setting a meta-path between node pairs in a heterogeneous network to be predicted, the maximum length of the meta-path and the setting corresponding to each meta-path type type label; construction unit, suitable for extracting heterogeneous topological features between node pairs based on meta-paths, constructing sample vectors, and forming a sample set; the sample set includes a training set and a test set; a classification learning unit, suitable for The training set and test set in the sample set are used for multi-label classification learning, and the corresponding multi-label classifier is obtained; the prediction unit is suitable for using the multi-label classifier obtained by training to predict the unknown relationship between the nodes in the heterogeneous network to be predicted. . The above solution can improve the accuracy of link prediction of heterogeneous information networks.

Description

Heterogeneous Information Network Link Prediction Device

technical field

The invention belongs to the technical field of data analysis, and in particular relates to a heterogeneous information network link prediction device.

Background technique

Many complex systems in the real world can be formalized as networks, with nodes representing objects and links representing interactions between objects. Most of these networks are heterogeneous networks that contain various types of objects and relationships, often consisting of multiple sub-networks. For example, the online social network Twitter contains information about types such as user basic information, user location, and user tweeting actions, with types of post/reply/retweet, follow/follow, check-in, and the like.

As a key problem in link mining, link prediction aims to predict the formation of future links based on current or historical networks. It has wider applications in bibliographic networks, biological networks, social networks, etc. Most existing link prediction methods are designed for homogeneous information networks whose nodes and links are of the same type. Recently, there has been great interest in advancing link prediction in heterogeneous networks because of its broader application prospects.

However, the link prediction in the heterogeneous network in the prior art has the problem of low prediction accuracy.

SUMMARY OF THE INVENTION

The technical problem solved by the present invention is how to improve the accuracy of link prediction of heterogeneous information networks.

In order to achieve the above object, an embodiment of the present invention also provides a heterogeneous information network link prediction device, the device includes:

a setting unit, adapted to set the meta-path between the node pairs in the heterogeneous network to be predicted, the maximum length of the meta-path and the type label corresponding to each meta-path type setting;

a construction unit, adapted to extract heterogeneous topological features between pairs of nodes based on the meta-path, construct a sample vector, and form a sample set; the sample set includes a training set and a test set;

a classification learning unit, adapted to perform multi-label classification learning based on the training set and the test set in the sample set to obtain a corresponding multi-label classifier;

The prediction unit is suitable for using the multi-label classifier obtained by training to predict the unknown relationship between the nodes in the heterogeneous network to be predicted.

Optionally, the classification learning unit is adapted to select a training subset corresponding to the tag pair formed by every two type tags in the set type tags from the training set, and respectively select a training subset for the selected training subset. Carry out two-class learning, and obtain multiple two-classifiers corresponding to each label pair one-to-one; input the test set into the multiple two-classifiers obtained by training, and calculate the instances corresponding to the samples in the test set in each type The first vote obtained on the label; the corresponding virtual label is added to each sample in the corresponding training subset to obtain the training subset corresponding to the label pair formed by the corresponding type label and the virtual label, and the obtained training The subsets are trained to obtain a plurality of auxiliary binary classifiers corresponding to each type label one-to-one; the virtual label is used to mark the segmentation point of the type label that is related and unrelated to the samples in the corresponding training subset; The test set is respectively input into a plurality of auxiliary binary classifiers obtained by training, and the second vote obtained on each type label and the third vote obtained on the virtual label for the instances corresponding to the samples in the test set are calculated respectively; The first vote and the second vote obtained on each type label and the third vote obtained on the virtual label for the instance corresponding to the test sample determine the final multi-label classifier.

Optionally, the heterogeneous topology features between the node pairs include path number features and random walk features.

Optionally, the classification learning unit is adapted to use the following formula to calculate and obtain the first vote obtained on each type label for the instance corresponding to the sample in the test set:

where ζ( _xi , l _j ) represents the votes obtained by the instance _xi on the label l _j , Clf _jk represents the binary classifier corresponding to the label pair (l _j , l _k ), indicates that the sample is correctly predicted as a negative example in the training subset, when Indicates that samples are correctly predicted as positives in the training subset.

Optionally, the classification learning unit is adapted to use the following formula to calculate the second vote obtained on each type label for the instance corresponding to the sample in the test set:

where ζ( _xi , l _j ) represents the unupdated votes obtained by the instance _xi on the label l _j , Clf _jk represents the binary classifier corresponding to the label pair (l _j , l _k ), Indicates that samples are correctly predicted as positives in the training subset.

Optionally, the classification learning unit is adapted to use the following formula to calculate the second vote obtained on each type label for the instance corresponding to the sample in the test set:

where ζ ^* ( _xi , _ls ) represents the votes obtained by instance _xi on the virtual label _ls , Indicates that samples are correctly predicted as negatives in the training subset.

Optionally, the classification learning unit is adapted to adopt the following formula based on the first vote and the second vote obtained on each type label and the third vote obtained on the virtual label based on the instance corresponding to the test sample: Vote to determine the final multi-label classifier:

h(x)={l _j |ζ ^* (x, l _j )>ζ ^* (x, l _s )};

where h(x) denotes the multi-label classifier, l _j denotes the j-th type label, ζ ^* (x, l _j ) denotes the final vote obtained by instance x _i on label l _j , ζ ^* (x, l _s ) represents the votes obtained by instance _xi on the virtual label l _s .

Optionally, the device may also include:

Calculate the output unit, calculate the dependency score between the label pairs and output.

Compared with the prior art, the beneficial effects of the present invention are:

In the above scheme, a corresponding multi-label classifier is obtained by performing multi-label classification learning based on the training set and the test set in the sample set, and the multi-label classifier obtained by training is used to predict unknown unknowns between nodes in a heterogeneous network. Relationship prediction can not only predict the direct links between nodes, but also predict other relationships between nodes, that is, meta-paths, so that instances have multiple rather than unique type labels, so it can improve the Accuracy of Link Prediction in Heterogeneous Networks.

Further, by calculating and outputting the dependency scores between the tag pairs, suggestions on how to form new links can be provided, helping to explore the rules of link and relationship formation in complex networks, which is convenient and practical.

Description of drawings

In order to illustrate the technical solutions in the embodiments of the present application more clearly, the following briefly introduces the drawings that are used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative labor.

1 is a schematic flowchart of a method for predicting a heterogeneous information network link according to an embodiment of the present invention;

2 is a schematic flowchart of a training method for a multi-label classifier according to an embodiment of the present invention;

FIG. 3 shows a schematic structural diagram of an apparatus for predicting a link of a heterogeneous information network in an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application. The relevant directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship between the various components under a certain posture (as shown in the accompanying drawings). Movement conditions, etc., if the specific posture changes, the directional indication also changes accordingly.

As mentioned in the background art, the link prediction of heterogeneous networks has the following problems:

(1) Representation of topological features: Due to the heterogeneity of objects and links, topological features used in homogeneous networks are not directly available. The neighbors of two nodes can be of different types, so the number of common neighbors cannot represent this heterogeneity.

(2) Interdependence between different relationships: In the present invention, a relationship not only refers to an explicit link, but also refers to an indirect link. Modeling correlations between different types of relationships is important because they may affect each other. For example, in a co-authorship network, two scholars can establish co-authorship (one type of relationship) by attending the same conference (another type of relationship).

The simple approach of studying isomorphic projections of networks alone avoids the first problem, but loses information because it ignores patterns of dependencies between types. Although heterogeneity brings many difficulties to link prediction, it also provides rich and valuable information to understand the underlying mechanisms of link formation. Research on link prediction follows the common sense that several links can be generated randomly, but most links are generated in latent patterns. The information of the target link depends on the potential relationship between the nodes of the target link.

If not only direct links between prediction nodes are considered, but also other relationships between prediction nodes, namely meta-paths, the traditional two-class supervision method will no longer be applicable, and multi-label learning needs to be used.

Objects in the real world often do not have unique semantics, but may be ambiguous. For example, a picture may convey various information such as "blue sky", "little river", "cow" and "smoke". Ambiguity objects no longer have unique semantics, which makes it difficult for traditional supervised learning frameworks with single semantics to achieve good results.

The technical scheme of the present invention obtains the corresponding multi-label classifier by performing multi-label classification learning based on the training set and the test set in the sample set, and uses the multi-label classifier obtained by training to predict the relationship between nodes in the heterogeneous network. Predicting unknown relationships can not only predict the direct links between nodes, but also predict other relationships between nodes, that is, meta-paths, so that instances have multiple rather than unique type labels, so it can be Improve the accuracy of link prediction in heterogeneous networks.

In order to make the above objects, features and beneficial effects of the present invention more clearly understood, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

FIG. 1 is a schematic flowchart of a method for predicting a link of a heterogeneous information network according to an embodiment of the present invention. Referring to Fig. 1, a method for predicting a heterogeneous information network link may specifically include the following steps:

Step S101: Set the meta-path between the node pairs in the heterogeneous network to be predicted, the maximum length of the meta-path, and the type label corresponding to each meta-path type setting.

In a specific implementation, the meta-path is a path from one node to another node through a series of type connections in a heterogeneous network. For example, a meta-path _Pk =(V ₁ E ₁ V ₂ E ₂ ... E _n-1 V _n ) means that node V ₁ is linked through a series of types E _i (i=1,2,...,n- 1) The path to reach the node _Vn can be recorded as

For simplicity, in this embodiment of the present invention, the first capital letter of the English name of the corresponding node is used for representation. For example, in the DBLP co-author network, P represents the paper, A represents the author, T represents the topic, V represents the conference, and U represents the user. The meta-path used to describe the citation relationship between authors can be abbreviated as "APPA".

In the specific implementation, the meta-path between the node pairs in the heterogeneous network to be predicted, the maximum length of the meta-path, and the type label corresponding to each meta-path type are set, and those skilled in the art can perform according to actual needs. Setting, there is no restriction here.

Step S102: Extract heterogeneous topological features between node pairs based on the meta-path, construct a sample vector, and form a sample set; the sample set includes a training set and a test set.

In an embodiment of the present invention, the heterogeneous topology features between the node pairs include path number features and random walk features. Among them, the number of paths represents the total number of meta-paths of the corresponding type between node pairs, and the random walk features between nodes are in one-to-one correspondence with the meta-path types, that is, each meta-path type has a corresponding random walk feature. , which can be calculated by the following formula:

in, represents the number of k-th meta-paths whose origin is v _i and the end is node v _j , Represents the total number of _k -th meta-paths whose origin is vi.

It can be seen from the above description that when the type of meta-path in the heterogeneous network to be predicted is K, and each meta-path between node pairs can be quantified as two heterogeneous topological features, the number of paths and random walks, then each Each node pair can be described by 2K heterogeneous topological features. In other words, each node pair is described by a corresponding 2K-dimensional space vector, and the 2K-dimensional space vector is called an instance corresponding to the node pair.

Assuming that X represents a 2K-dimensional instance space, and L ₌ {l ₁ , l ₂ , . In the sample set composed of the constructed sample vectors, D={(x _i , Li )|1≤i≤m _} represents the multi-label training set, m represents the number of samples in the multi-label training set, and ( _{xi ,} Li ) is a multi-label sample, where To describe the 2K-dimensional feature vector of the node pair (vi, _{vj ), each dimension in x i can} be a heterogeneous feature or a homogeneous feature, is the label set of _{xi; T={(x i ,} L _i )|m+1≤i≤n} is the multi-label test set, and n represents the total number of samples in the sample set.

Step S103: Perform multi-label classification learning based on the training set and the test set in the sample set to obtain a corresponding multi-label classifier.

In a specific implementation, when the corresponding training set and test set are constructed, the obtained multi-label training set D can be used to carry out multi-label classification learning and training of the multi-label classifier, and the multi-label classifier obtained by training can be used for The multi-label test set T is used for optimization. For details, please refer to the detailed introduction in FIG. 2 , which will not be repeated here.

Step S104: Use the multi-label classifier obtained by training to predict the unknown relationship between the nodes in the heterogeneous network to be predicted.

In a specific implementation, when the corresponding multi-label classifier is trained, the obtained multi-label classifier can be used to predict the unknown part of the network to be predicted, that is, the relationship between unconnected target node pairs in the network to be predicted. The unknown connection relationship is predicted, and the predicted new network graph is obtained.

In an embodiment of the present invention, the method may further include:

Step S105: Calculate and output the dependency scores between the tag pairs.

In an embodiment of the present invention, by using a chi-square test tool (Chi-Square Test tool), the dependency score between the tag pairs can be calculated and output, and the user can be provided with suggestions on how to form new links between node pairs , which helps to explore the laws of formation of links and relationships in complex networks, which is convenient and practical.

FIG. 2 is a schematic flowchart of a training method for a multi-label classifier according to an embodiment of the present invention. Referring to Figure 2, a training method for a multi-label classifier may specifically include the following steps:

Step S201: respectively select a training subset corresponding to the label pair formed by every two type labels in the set type labels from the training set, and perform two-class learning on the selected training subset respectively, and obtain a Multiple binary classifiers with one-to-one correspondence of labels.

In a specific implementation, for the label space L ₌ {l ₁ , l ₂ , . can be defined as:

D _jk = {(x _i , ψ(L _i , l _j , l _k ))|φ(L _i , l _j )≠φ(L _i , l _k )} (2)

Among them, D _jk represents the training subset corresponding to the label pair (l _j , l _k ).

In a specific implementation, when the training subset D _jk corresponding to the label pair (l _j , l _k ) is selected from the multi-label training set D, the classification learning algorithm can be used to learn the two-class classification, and each label pair (l _j , _lk ) one-to-one correspondence of multiple binary classifiers.

Step S202: Input the test set into a plurality of binary classifiers obtained by training respectively, and calculate the first vote obtained on each type label of the instance corresponding to the sample in the test set.

In an embodiment of the present invention, the first vote obtained on each type label for the instance corresponding to the sample in the test set is calculated by using the following formula:

where ζ( _xi , l _j ) represents the votes obtained by the instance _xi on the label l _j , Clf _jk represents the binary classifier corresponding to the label pair (l _j , l _k ), indicates that the sample is correctly predicted as a negative example in the training subset, when Indicates that samples are correctly predicted as positives in the training subset.

Step S203: Add the corresponding virtual label to each sample in the corresponding training subset respectively, obtain the training subset corresponding to the label pair formed by the corresponding type label and the virtual label, and use the obtained training subset to train respectively to obtain Multiple auxiliary binary classifiers corresponding to each type label one-to-one.

In a specific implementation, a virtual label _ls is added to each sample to mark the segmentation points of labels related to x _i and unrelated labels, respectively obtained for each new label pair (l _j , _ls ) The corresponding training set D _is , and the corresponding two-class learning algorithm is used to learn the corresponding training set D _js obtained from each new label pair (l _j , l _s ) respectively, to obtain the corresponding K auxiliary two-class loader Clf _js .

Step S204: Input the test set into a plurality of auxiliary binary classifiers obtained by training respectively, and calculate the second vote obtained on each type label and the first vote obtained on the virtual label for the instances corresponding to the samples in the test set. Three votes.

In an embodiment of the present invention, the following formula is used to calculate the second vote obtained by the instance corresponding to the sample in the test set on each type label:

where ζ( _xi , l _j ) represents the unupdated votes obtained by the instance _xi on the label l _j , Clf _jk represents the binary classifier corresponding to the label pair (l _j , l _k ), Indicates that samples are correctly predicted as positives in the training subset.

In an embodiment of the present invention, the following formula is used to calculate the third vote obtained on each type label for the instance corresponding to the sample in the test set:

where ζ ^* ( _xi , _ls ) represents the votes obtained by instance _xi on the virtual label _ls , Indicates that samples are correctly predicted as negatives in the training subset.

Step S205: Determine a final multi-label classifier based on the first vote and the second vote obtained on the each type label and the third vote obtained on the virtual label for the instance corresponding to the test sample.

In an embodiment of the present invention, the determined final multi-label classifier can be expressed as:

h(x)={l _j |ζ ^* (x, l _j )>ζ ^* (x, l _s )} (8)

where h(x) denotes the multi-label classifier, l _j denotes the j-th type label, ζ ^* (x, l _j ) denotes the final vote obtained by instance x _i on label l _j , ζ ^* (x, l _s ) represents the votes obtained by instance _xi on the virtual label l _s .

The method for predicting a link of a heterogeneous information network in the embodiment of the present invention has been described in detail above, and an apparatus corresponding to the above method will be introduced below.

FIG. 3 shows a schematic structural diagram of an apparatus for predicting a link of a heterogeneous information network in an embodiment of the present invention. 3, a heterogeneous information network link prediction apparatus 30 may include a setting unit 301, a construction unit 302, a classification learning unit 303 and a prediction unit 304, wherein:

The setting unit 301 is adapted to set the meta-path between node pairs in the heterogeneous network to be predicted, the maximum length of the meta-path, and the type label corresponding to each meta-path type setting.

The construction unit 302 is adapted to extract heterogeneous topological features between node pairs based on the meta-path, construct a sample vector, and form a sample set; the sample set includes a training set and a test set.

The classification learning unit 303 is adapted to perform multi-label classification learning based on the training set and the test set in the sample set to obtain a corresponding multi-label classifier.

The predicting unit 304 is adapted to use the multi-label classifier obtained by training to predict the unknown relationship between nodes in the heterogeneous network to be predicted.

In a specific implementation, the classification learning unit 303 is adapted to select a training subset corresponding to a label pair formed by every two type labels in the set type labels from the training set, and perform a Two-class learning is carried out on the set respectively, and multiple two-classifiers corresponding to each label pair are obtained; the test set is respectively input into the multiple two-classifiers obtained by training, and the instances corresponding to the samples in the test set are calculated in The first vote obtained on each type label; the corresponding virtual label is added to each sample in the corresponding training subset, and the corresponding training subset of the label pair composed of the corresponding type label and the virtual label is obtained, and the obtained The training subsets are trained to obtain a plurality of auxiliary binary classifiers corresponding to each type label one-to-one; the virtual label is used to mark the segmentation point of the type label related and unrelated to the sample in the corresponding training subset; the The test set is respectively input to a plurality of auxiliary binary classifiers obtained by training, and the second vote obtained on each type label and the third vote obtained on the virtual label for the instances corresponding to the samples in the test set are calculated respectively; based on The first vote and the second vote obtained on each type label and the third vote obtained on the virtual label for the instance corresponding to the test sample determine the final multi-label classifier.

In a specific implementation, the heterogeneous topological features between the node pairs include path number features and random walk features.

In an embodiment of the present invention, the classification learning unit 303 is adapted to use the following formula to calculate and obtain the first vote obtained on each type label for the instance corresponding to the sample in the test set:

where ζ( _xi , l _j ) represents the votes obtained by the instance _xi on the label l _j , Clf _jk represents the binary classifier corresponding to the label pair (l _j , l _k ), indicates that the sample is correctly predicted as a negative example in the training subset, when Indicates that samples are correctly predicted as positives in the training subset.

In an embodiment of the present invention, the classification learning unit 303 is adapted to use the following formula to calculate the second vote obtained on each type label for the instance corresponding to the sample in the test set:

where ζ( _xi , l _j ) represents the unupdated votes obtained by the instance _xi on the label l _j , Clf _jk represents the binary classifier corresponding to the label pair (l _j , l _k ), Indicates that samples are correctly predicted as positives in the training subset.

In an embodiment of the present invention, the classification learning unit 303 is adapted to use the following formula to calculate the second vote obtained on each type label for the instance corresponding to the sample in the test set:

where ζ ^* ( _xi , _ls ) represents the votes obtained by instance _xi on the virtual label _ls , Indicates that samples are correctly predicted as negatives in the training subset.

In an embodiment of the present invention, the classification learning unit 303 is adapted to adopt the following formula based on the first vote and the second vote obtained on each type label based on the instance corresponding to the test sample and the virtual label The third vote obtained on the final multi-label classifier:

h(x)={l _j |ζ ^* (x, l _j )>ζ ^* (x, l _s )}; where h(x) represents the multi-label classifier, and l _j represents the j-th type label , ζ ^* (x, l _j ) denotes the final vote obtained by instance _xi on label l _j , and ζ ^* (x, l _s ) denotes the vote obtained by instance _xi on virtual label l _s .

In an embodiment of the present invention, the apparatus 30 may further include 305, wherein:

The calculation and output unit 305 is adapted to calculate and output the dependency scores between the tag pairs.

Embodiments of the present invention further provide a computer-readable storage medium, which stores computer instructions, and when the computer instructions are run, executes the steps of the method for predicting a link between heterogeneous information networks. Wherein, for the method for predicting the link of heterogeneous information network, please refer to the introduction in the above-mentioned relevant part, and will not be repeated here.

An embodiment of the present invention further provides a terminal, including a memory and a processor, where the memory stores computer instructions that can be run on the processor, and the processor executes the exception when running the computer instructions. Steps to construct an information network link prediction method. Wherein, for the method for predicting the link of heterogeneous information network, please refer to the introduction in the above-mentioned relevant part, and will not be repeated here.

Using the above solution in the embodiment of the present invention, a corresponding multi-label classifier is obtained by performing multi-label classification learning based on the training set and the test set in the sample set, and the multi-label classifier obtained by training is used to predict the heterogeneous network Predicting the unknown relationship between nodes in the network can not only predict the direct links between nodes, but also predict other relationships between nodes, that is, new types of meta-paths, so it can improve the link prediction of heterogeneous networks. accuracy.

Further, by calculating and outputting the dependency scores between the tag pairs, suggestions on how to form new links can be provided, helping to explore the rules of link and relationship formation in complex networks, which is convenient and practical.

The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above-mentioned embodiments, and the descriptions in the above-mentioned embodiments and the description are only to illustrate the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have Various changes and improvements, the claimed scope of the present invention is defined by the appended claims, description and their equivalents.

Claims (8)

1. a heterogeneous information network link prediction device, is characterized in that, comprises: a setting unit, adapted to set the meta-path between the node pairs in the heterogeneous network to be predicted, the maximum length of the meta-path and the type label corresponding to each meta-path type setting; a construction unit, adapted to extract heterogeneous topological features between pairs of nodes based on the meta-path, construct a sample vector, and form a sample set; the sample set includes a training set and a test set; a classification learning unit, adapted to perform multi-label classification learning based on the training set and the test set in the sample set to obtain a corresponding multi-label classifier; The prediction unit is suitable for using the multi-label classifier obtained by training to predict the unknown relationship between the nodes in the heterogeneous network to be predicted. 2 . The heterogeneous information network link prediction device according to claim 1 , wherein the classification learning unit is adapted to select from the training set and each two type labels in the set type labels respectively. 3 . The training subsets corresponding to the label pairs, and the selected training subsets are respectively subjected to binary classification learning to obtain multiple binary classifiers corresponding to each label pair one-to-one; The second classifier calculates the first vote obtained on each type label for the instance corresponding to the sample in the test set; adds the corresponding virtual label to each sample in the corresponding training subset, and obtains the corresponding type label and virtual label. The training subset corresponding to the label pair formed by the label is used, and the obtained training subset is used to train to obtain a plurality of auxiliary binary classifiers corresponding to each type label one-to-one; the virtual label is used to mark the corresponding training sub-classifier. The split points of the relevant and irrelevant type labels of the samples in the set; the test set is respectively input into a plurality of auxiliary binary classifiers obtained by training, and the instances corresponding to the samples in the test set are calculated respectively obtained on each type label. The second vote and the third vote obtained on the virtual label; the first vote and the second vote obtained on the each type label based on the instance corresponding to the test sample and the third vote obtained on the virtual label, Determine the final multi-label classifier. 3 . The link prediction device for heterogeneous information networks according to claim 2 , wherein the heterogeneous topology features between the node pairs include path number features and random walk features. 4 . 4 . The heterogeneous information network link prediction device according to claim 3 , wherein the classification learning unit is adapted to obtain the instance corresponding to the sample in the test set by calculating the following formula on each type label. 5 . First vote received: where ζ( _xi , l _j ) represents the votes obtained by the instance _xi on the label l _j , Clf _jk represents the binary classifier corresponding to the label pair (l _j , l _k ), indicates that the sample is correctly predicted as a negative example in the training subset, when Indicates that samples are correctly predicted as positives in the training subset. 5. The heterogeneous information network link prediction device according to claim 4, wherein the classification learning unit is adapted to use the following formula to calculate the instance corresponding to the sample in the test set and obtain on each type label The second vote: where ζ( _xi , l _j ) represents the unupdated votes obtained by the instance _xi on the label l _j , Clf _jk represents the binary classifier corresponding to the label pair (l _j , l _k ), Indicates that samples are correctly predicted as positives in the training subset. 6 . The heterogeneous information network link prediction device according to claim 5 , wherein the classification learning unit is adapted to calculate the instance corresponding to the sample in the test set using the following formula and obtain on each type label. 7 . The second vote: where ζ ^* ( _xi , _ls ) represents the votes obtained by instance _xi on the virtual label _ls , Indicates that samples are correctly predicted as negatives in the training subset. 7. The heterogeneous information network link prediction device according to any one of claims 3 to 6, wherein the classification learning unit is adapted to use the following formula in the The first and second votes obtained on each type label and the third vote obtained on the virtual label determine the final multi-label classifier: h(x)={l _j |ζ ^* (x, l _j )>ζ ^* (x, l _s )} where h(x) denotes the multi-label classifier, l _j denotes the j-th type label, ζ ^* (x, l _j ) denotes the final vote obtained by instance x _i on label l _j , ζ ^* (x, l _s ) represents the votes obtained by instance _xi on the virtual label l _s . 8. The heterogeneous information network link prediction device according to any one of claims 3-6, further comprising: A calculation output unit, adapted to calculate and output the dependency scores between the tag pairs.