CN114036330A - Object feature matrix determination method, apparatus, device and storage medium - Google Patents

Abstract

The present disclosure relates to a method, apparatus, device and storage medium for determining an object feature matrix. The aforementioned method includes: acquiring heterogeneous feature association matrices and corresponding association relationship indication matrices of various types of objects in the object data set, and multi-layer attribute heterogeneous network matrices of various types of objects; based on heterogeneous feature association matrices and corresponding associations The relationship indicator matrix constructs the heterogeneous feature correlation function; based on the multi-layer attribute heterogeneous network matrix, the attribute heterogeneous network function is constructed; based on the heterogeneous feature correlation function and the attribute heterogeneous network function, the minimized objective function is constructed; based on the minimized objective function calculation Low-rank feature matrices of various types of objects; low-rank feature matrices are used to reconstruct object feature matrices of various types of objects in the object dataset. By adopting the solutions provided by the embodiments of the present disclosure, the information loss caused by the homogeneous transformation of the attribute heterogeneous network can be reduced.

Description

Object feature matrix determination method, apparatus, device and storage medium

technical field

The present disclosure relates to the technical field of data processing, and in particular, to a method, apparatus, device and storage medium for determining an object feature matrix.

Background technique

Effective classification of objects such as images, biomolecules, users of social networking sites, etc., relies on building effective image classifiers. In the method of object classification using a discriminant model, the premise of constructing an effective object classifier is to extract an object feature matrix that effectively characterizes the characteristics of the sample object.

In the related art, a solution is proposed in which a low-rank feature matrix of an object is determined based on matrix decomposition, and an object feature matrix of a sample object is constructed based on the low-rank feature matrix. The scheme of constructing object feature matrix based on matrix decomposition can guarantee the internal structure of heterogeneous data sources. However, due to the incompleteness of object data and the limitations of model assumptions and experimental design, existing methods still suffer from information loss caused by homogeneous data conversion and incomplete known heterogeneous associated data.

SUMMARY OF THE INVENTION

In order to solve the above technical problems or at least partially solve the above technical problems, embodiments of the present disclosure provide a method, apparatus, device, and storage medium for determining an object characteristic matrix.

In a first aspect, an embodiment of the present disclosure provides a method for determining an object feature matrix, including:

Obtain the heterogeneous feature correlation matrix and the corresponding correlation indication matrix of various types of objects in the object dataset, as well as the multi-layer attribute heterogeneous network matrix of various types of objects;

constructing a heterogeneous feature correlation function based on the heterogeneous feature correlation matrix and the corresponding correlation indication matrix;

constructing an attribute heterogeneous network function based on the multi-layer attribute heterogeneous network matrix;

based on the heterogeneous feature correlation function and the attribute heterogeneous network function, constructing a minimization objective function;

Calculate low-rank feature matrices of various types of objects based on the minimized objective function;

The low-rank feature matrix is used to reconstruct object feature matrices of various types of objects in the object dataset.

Optionally, the method further includes:

constructing a first constraint function based on the first weight matrix in the heterogeneous feature correlation function;

constructing a second constraint function based on the second weight matrix in the attribute heterogeneous network function;

Based on the heterogeneous feature correlation function and the attribute heterogeneous network function, a minimization objective function is constructed, including:

The minimization objective function is constructed based on the heterogeneous feature correlation function, the attribute heterogeneous network function, the first constraint function and the second constraint function.

Optionally, the method further includes:

Also includes: constructing an unknown noise correlation constraint function;

Based on the heterogeneous feature correlation function, the attribute heterogeneous network function, the first constraint function and the second constraint function, constructing the minimization objective function includes:

The minimization objective function is constructed based on the heterogeneous feature correlation function, the attribute heterogeneous network function, the first constraint function, the second constraint function and the unknown noise correlation constraint function.

Optionally, determining low-rank feature matrices of various types of objects based on the minimization objective function, including:

The minimization objective function is solved by the alternating direction multiplier method, and the low-rank feature matrix of various types of objects is determined.

The method of using alternating direction multipliers to solve the minimization objective function includes:

The alternating direction multiplier method is used to alternately solve the first weight matrix, the second weight matrix, the low-rank feature matrix, the heterogeneous correlation network matrix, and the multi-layer attribute heterogeneous network matrix in the minimization objective function. basis matrix until a preset number of iterations or convergence is reached.

Optionally, the alternating direction multiplier method is used to alternately solve the first weight matrix, the second weight matrix, the low-rank feature matrix, the heterogeneous association network matrix, and the multi-layer attribute heterogeneity in the minimization objective function. The basis matrix of the network matrix, including:

Set the first weight matrix, the second weight matrix, the low-rank feature matrix, and the heterogeneous association network matrix as fixed values, and obtain a first bias for the basis matrix of the multi-layer attribute heterogeneous network matrix. derivative, when the first partial derivative value is zero, determine the basis matrix of the multi-layer attribute heterogeneous network matrix;

Setting the basis matrix of the first weight matrix, the second weight matrix, the low-rank feature matrix and the multi-layer attribute heterogeneous network matrix as a fixed value, and obtaining a second partial derivative for the heterogeneous association network matrix, When the second partial derivative value is zero, determining the heterogeneous correlation network matrix;

Setting the base matrices of the first weight matrix, the second weight matrix, the heterogeneous association network matrix and the multi-layer attribute heterogeneous network matrix as a fixed value, and obtaining a third partial derivative for the low-rank feature matrix, When the third partial derivative value is zero, determining the low-rank feature matrix;

Set the basis matrix of the first weight matrix, the low-rank feature matrix, the heterogeneous association network matrix, and the multi-layer attribute heterogeneous network matrix as a fixed value, and obtain a fourth bias for the second weight matrix. derivative, when the fourth partial derivative value is zero, determine the second weight matrix;

Set the base matrix of the second weight matrix, the low-rank feature matrix, the heterogeneous association network matrix, and the multi-layer attribute heterogeneous network matrix as a fixed value, and obtain a fifth bias for the first weight matrix The first weight matrix is determined when the value of the fifth partial derivative is zero.

Optionally, the object dataset is an image dataset, and the objects in the object dataset are images.

In a second aspect, an embodiment of the present disclosure provides an apparatus for determining an object characteristic matrix, including:

The matrix acquisition unit is used to acquire the heterogeneous feature correlation matrix of various types of objects in the object dataset and the corresponding correlation indication matrix, as well as the multi-layer attribute heterogeneous network matrix of various types of objects;

a function construction unit, configured to construct a heterogeneous feature correlation function based on the heterogeneous feature correlation matrix and the corresponding correlation indication matrix; build an attribute heterogeneous network function based on the multi-layer attribute heterogeneous network matrix; and, based on The heterogeneous feature correlation function and the attribute heterogeneous network function are used to construct a minimization objective function;

a solving unit for calculating low-rank feature matrices of various types of objects based on the minimization objective function;

An object feature matrix determining unit, configured to use the low-rank feature matrix to reconstruct object feature matrices of various types of objects in the object dataset.

In a third aspect, an embodiment of the present disclosure provides an electronic device, including:

one or more processors;

storage means for storing one or more programs,

The one or more programs, when executed by the one or more processors, cause the one or more processors to implement a method as described in any one of the first aspects.

In a fourth aspect, an embodiment of the present disclosure provides a computer-readable storage medium on which a computer program is stored, characterized in that, when the program is executed by a processor, the method according to any one of the first aspects is implemented.

Compared with the prior art, the technical solutions provided by the embodiments of the present disclosure have the following advantages:

The present disclosure implements the solution provided by constructing a heterogeneous feature correlation function through a heterogeneous feature correlation matrix and a corresponding correlation relationship indication matrix, so as to realize the use of the correlation relationship indicator matrix to distinguish heterogeneous heterogeneous correlations and potential heterogeneous correlations. The attribute heterogeneous network function is constructed by using the low-rank feature matrix corresponding to various types of objects and the decomposed basis matrix of the attribute heterogeneous network of various types of objects. The objective function is to use the objective function to minimize the first weight matrix, low-rank feature matrix, heterogeneous correlation network matrix, second weight matrix and the attribute difference of various types of objects in the heterogeneous feature correlation function and the attribute heterogeneous network function. The basis matrices decomposed by the qualitative network are co-graded to obtain low-rank feature matrices of various types of objects, and the object feature matrices of various types of objects are determined based on the low-rank feature matrices of various types of objects. By adopting the solution provided by the embodiments of the present disclosure, the topology structure of the network and the attribute information of the nodes can be fused together, so as to make up for the cold start problem caused by the lack of known associations, thereby reducing the information caused by the homogeneous transformation of the heterogeneous network with attributes. loss.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the disclosure and together with the description serve to explain the principles of the disclosure.

In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the accompanying drawings that are required to be used in the description of the embodiments or the prior art will be briefly introduced below. In other words, on the premise of no creative labor, other drawings can also be obtained from these drawings.

1 is a flowchart of a method for determining an object feature matrix provided by an embodiment of the present disclosure;

2 is a flowchart of a method for determining an object feature matrix provided by an embodiment of the present disclosure;

3 is a flowchart of a method for determining an object feature matrix provided by an embodiment of the present disclosure;

FIG. 4 is a schematic structural diagram of an object determination apparatus provided by an embodiment of the present disclosure;

FIG. 5 is a schematic structural diagram of an electronic device provided by an embodiment of the present disclosure.

Detailed ways

In order to more clearly understand the above objects, features and advantages of the present disclosure, the solutions of the present disclosure will be further described below. It should be noted that the embodiments of the present disclosure and the features in the embodiments may be combined with each other under the condition of no conflict.

Many specific details are set forth in the following description to facilitate a full understanding of the present disclosure, but the present disclosure can also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only a part of the embodiments of the present disclosure, and Not all examples.

The embodiments of the present disclosure provide a method for determining an object feature matrix, which is used to solve the information loss caused by the conversion of homogeneous data during the fusion (that is, matrix decomposition) process of the existing method for determining object features based on matrix decomposition, and the known differences It solves the problem of incomplete correlation of qualitative data, and then more accurately characterizes the characteristics of objects in the object dataset. In specific embodiments, the objects in the embodiments of the present disclosure may be objects of types such as images, biomolecules, users of social networking sites, and the like.

Wherein, the object feature matrix determination method of the present disclosure is executed by an electronic device or an application program in the electronic device or the like. The electronic device may be a tablet computer, a mobile phone, a notebook computer, a server and other devices, and the present disclosure does not impose any limitation on the specific type of the electronic device. The present disclosure does not limit the type of the operating system of the electronic device. For example, Android system, Linux system, Windows system, iOS system, etc.

FIG. 1 is a flowchart of a method for determining an object feature matrix provided by an embodiment of the present disclosure. As shown in FIG. 1 , the method for determining an object feature matrix provided by an embodiment of the present disclosure includes steps S101-S106.

Step S101: Acquire the heterogeneous feature correlation matrix of various types of objects in the object data set and the corresponding correlation relationship indication matrix, and the multi-layered attribute heterogeneous network matrix of various types of objects.

In the embodiment of the present disclosure, the electronic device can process the original data of various types of objects in the object data set, and obtain the heterogeneous feature correlation matrix, the corresponding correlation indication matrix and the multi-layer attribute heterogeneous network matrix of various types of objects, The aforementioned matrix can also be obtained by reading pre-stored data in storage.

表示第i种类型对象的n_i个样本之间的异质特征关联矩阵，采用H_ij表示与R_ij对应的并且维度同R_ij相同的关联关系指示矩阵。其中关联关系指示矩阵H_ij用于将一致特征关联关系矩阵中的观察到的关联关系和未观察到的关联关系区分开。具体的，如果R_ij(s,t)>0，那么H_ij＝1，否则H_ij＝0。The embodiment of the present disclosure adopts

Represents the heterogeneous feature correlation matrix between n _i samples of the i-th type of object, using H _ij to represent the correlation indicating matrix corresponding to R _ij and having the same dimension as R _ij . The correlation indicator matrix H _ij is used to distinguish the observed correlation from the unobserved correlation in the consistent feature correlation matrix. Specifically, if _Rij (s,t)>0, then _Hij =1, otherwise _Hij =0.

表示第i种类型对象的多种属性异质网络矩阵，其中t_i表示第i种类型对象共收集到t_i种来源的属性异质网络，d_it表示第t种属性异质网络有d_it种属性。The embodiment of the present disclosure adopts

Represents the heterogeneous network matrix of various attributes of the ith type of object, where t _i represents the attribute heterogeneous network of t _i sources collected by the ith type of object, and d _it indicates that the t-th attribute heterogeneous network has d _it a property.

Step S102 : constructing a heterogeneous feature correlation function based on the heterogeneous feature correlation matrix and the corresponding correlation relationship indication matrix.

In addition to the aforementioned heterogeneous feature correlation matrix and the corresponding correlation management indicator matrix, the heterogeneous feature correlation function also includes a first weight matrix, a low-rank feature matrix corresponding to various types of objects, and the heterogeneity between various types of objects. The qualitative correlation network matrix.

In an embodiment of the present disclosure, the heterogeneous feature correlation function is represented by formula 1.1.

表示第i种或第j种类型对象之间的异构关联网络指导的低秩特征矩阵，两个低秩标识矩阵分别表示了在压缩的ki,k_j维空间以实数化的形式描述第i种类型对象的n_i个对象的属性信息和第j种类型对象的n_j个对象的属性信息；k_i和k_j分别表示第i种或者第j种低秩特征矩阵的维度；

表示了异质关联网络矩阵，其相比于R_ij规模小很多；

表示对

个异质关联网络所赋予的第一权重矩阵，对于

可近似为重构损失并被用于区别已观察到的关联关系和未观察到的关联关系，并使得已观察到的关联在重构后的R_ij中得以保持。Among them: e represents the Hadamard product;

Represents the low-rank feature matrix guided by the heterogeneous association network between the i-th or j-th types of objects, and the two low-rank identification matrices respectively represent the compressed ki, k _j -dimensional space to describe the i-th in the form of real numbers The attribute information of n _i objects of the type of object and the attribute information of the n _j objects of the j-th type of object; k _i and k _j respectively represent the dimension of the i-th or j-th low-rank feature matrix;

represents the heterogeneous correlation network matrix, which is much smaller than R _ij ;

express right

The first weight matrix assigned by a heterogeneous association network, for

can be approximated as a reconstruction loss and is used to distinguish between observed and unobserved associations, and allows the observed associations to be preserved in the reconstructed _Rij .

Step S103 : constructing an attribute heterogeneity network function based on the multi-layer attribute heterogeneity network matrix.

In the embodiment of the present disclosure, the attribute heterogeneity network function includes, in addition to the aforementioned various attribute heterogeneity network matrices, a second weight matrix, a low-rank feature matrix corresponding to various types of objects, and attribute heterogeneity of various types of objects The basis matrix after network decomposition.

In an embodiment of the present disclosure, the attribute heterogeneity network function is represented by formula 1.2.

表示对多种对象的属性异质网络矩阵所赋予的第二权重矩阵，对于X_it，如果t>max_it_i,

G_i表示第i种类型对象的属性异质网络指导的低秩特征矩阵；

表示第i种类型对象的属性异质网络分解后的基矩阵。in,

Represents the second weight matrix assigned to the attribute heterogeneous network matrix of various objects, for X _it , if t>max _i t _i ,

G _i represents the low-rank feature matrix guided by the attribute heterogeneous network of the i-th type of object;

Represents the basis matrix after the decomposition of the attribute heterogeneous network of the ith type of object.

In the embodiment of the present disclosure, the attribute heterogeneous network function can be used to directly decompose the attribute heterogeneous data, so as to avoid information loss caused by homogeneous transformation.

Step S104: Construct a minimization objective function based on the heterogeneous feature correlation function and the attribute heterogeneous network function.

In the embodiment of the company, the minimization objective function is constructed based on the heterogeneous feature correlation function and the attribute heterogeneous network function, which may be the addition of the heterogeneous feature correlation function and the attribute heterogeneous network function.

In an embodiment of the present disclosure, after adding the heterogeneous feature correlation function and the attribute heterogeneous network function, the obtained minimization objective function is expressed by formula 1.3

stω ^r ≥0, ω ^h ≥0, ∑vec(ω ^r )=1, ∑vec(ω ^h )=1 (1.3)

Step S105: Calculate low-rank feature matrices of various types of objects based on the minimization objective function.

In this embodiment of the present disclosure, calculating low-rank feature matrices of various types of objects based on the minimization objective function is the calculation of the first weight matrix in the minimized objective function, the low-rank feature matrices corresponding to various types of objects, and the difference between the various types of objects. The heterogeneous correlation network matrix, the second weight matrix, and the basis matrix after the heterogeneous network decomposition of the attributes of various types of objects are decomposed collaboratively, and the low-rank feature matrix G of various types of objects is obtained in the process of co-decomposition.

In the embodiment of the present disclosure, the minimization objective function is non-convex on G, S, U, ω ^r , ω ^h , so the Alternating Direction Method of Multipliers (Alternating Direction Method of Multipliers, ADMM) to optimize it. Specifically, four parameters in G, S, U, ω ^r , ω ^h can be set as constants, and the other one can be optimized at the same time, and iteratively iterate until all the parameters are solved. Specifically, the calculation of G, S, U, ω ^r , ω ^h by the alternate direction multiplier method is described below.

Step S106: Reconstructing object feature matrices of various types of objects in the object dataset by using the low-rank feature matrix.

After determining the low-rank feature matrices of various types of objects, the low-rank feature matrices can be used to reconstruct the object feature matrices. In some embodiments of the present disclosure, after determining low-rank feature matrices of various types of objects, the low-rank feature matrices may be directly used as object feature matrices of corresponding types of objects.

The present disclosure implements the object feature matrix determination method provided by the heterogeneous feature correlation matrix and the corresponding correlation indication matrix to construct a heterogeneous feature correlation function, so as to realize the heterogeneous correlation and potential heterogeneity using the correlation indication matrix. Association to distinguish. The attribute heterogeneous network function is constructed by using the low-rank feature matrix corresponding to various types of objects and the decomposed basis matrix of the attribute heterogeneous network of various types of objects. The objective function is to use the objective function to minimize the first weight matrix, low-rank feature matrix, heterogeneous correlation network matrix, second weight matrix and the attribute difference of various types of objects in the heterogeneous feature correlation function and the attribute heterogeneous network function. The basis matrices decomposed by the qualitative network are co-graded to obtain low-rank feature matrices of various types of objects, and the object feature matrices of various types of objects are determined based on the low-rank feature matrices of various types of objects.

By using the method provided by the embodiments of the present disclosure, the topology structure of the network and the attribute information of the nodes can be fused together, so as to make up for the cold start problem caused by the lack of known associations, thereby reducing the information caused by the homogeneous transformation of the heterogeneous network with attributes. loss.

FIG. 2 is a flowchart of a method for determining an object feature matrix provided by an embodiment of the present disclosure. As shown in FIG. 2 , the method provided by the embodiment of the present disclosure includes steps S201-S207.

Step S201: Obtain the heterogeneous feature correlation matrix and the corresponding correlation relationship indication matrix of various types of objects in the object data set, and the multi-layer attribute heterogeneous network matrix of various types of objects.

Step S202 : constructing a heterogeneous feature correlation function based on the heterogeneous feature correlation matrix and the corresponding correlation relationship indication matrix.

Step S203 : constructing an attribute heterogeneity network function based on the multi-layer attribute heterogeneity network matrix.

The foregoing steps S201-S203 are the same as the steps S101-S103 in the foregoing embodiment. For details, refer to the foregoing embodiment, which will not be repeated here.

Step S204 : constructing a first constraint function based on the first weight matrix in the heterogeneous feature correlation function; and constructing a second constraint function based on the second weight matrix in the attribute heterogeneous network function.

Step S205 : constructing a minimization objective function based on the heterogeneous feature correlation function, the attribute heterogeneous network function, the first constraint function and the second constraint function.

时，R_ij的权重将为

此时其他异质关联矩阵将全都被忽略。同样的，当X_it具有最少的近似损失

此时倾向于将

分配给X_it。也就是说，所有其他的同质矩阵的贡献都会被忽略。When solving the minimization objective function shown in Equation 1.3, when R _ij has the smallest approximation loss

, the weight of R _ij will be

At this point the other heterogeneous correlation matrices will all be ignored. Likewise, when X _it has the least approximation loss

tend to be

assigned to X _it . That is, the contributions of all other homogeneous matrices are ignored.

At the same time, numerous studies have confirmed that different data sources can provide complementary information to each other. Therefore, using only a single heterogeneous correlation matrix and a single homogeneous correlation matrix may not give reliable predictions.

In order to make up for the aforementioned defective weight assignment, in the embodiment of the present disclosure, a first constraint function is first constructed based on the first weight matrix in the heterogeneous feature correlation function, and a second constraint function is constructed based on the second weight matrix in the attribute heterogeneous network function. .

第二约束函数为

前述vec(ω^r)是将ω^r的行堆叠拼接后的向量，vec(ω^h)是将ω^h的行堆叠拼接后的向量。α>0,β>0被用来控制vec(ω^r)和vec(ω^h)的复杂度。同时，α,β还可以帮助选择性的集成不同的异质关联数据源和属性异质数据源。In the embodiment of the present disclosure, both the first constraint function and the second constraint function are regular terms based on the l2 norm. The first constraint function is

The second constraint function is

The aforementioned vec(ω ^r ) is a vector obtained by stacking and splicing rows of ω ^r , and vec(ω ^h ) is a vector obtained by stacking and splicing rows of ω ^h . α>0, β>0 are used to control the complexity of vec(ω ^r ) and vec(ω ^h ). At the same time, α, β can also help to selectively integrate different heterogeneous association data sources and attribute heterogeneous data sources.

In the case of adding the first constraint function and the second constraint function, the minimized objective function obtained in the embodiment of the present disclosure is expressed by formula 1.4.

Step S206: Calculate low-rank feature matrices of various types of objects based on the minimization objective function.

Step S207: Reconstructing object feature matrices of various types of objects in the object dataset by using the low-rank feature matrix.

The foregoing steps S206-S207 are the same as the steps S105-S106 in the foregoing embodiment. For details, refer to the foregoing embodiment, which will not be repeated here.

In the method for determining an object feature matrix provided by the embodiment of the present disclosure, a first constraint function is constructed based on a first weight matrix in a heterogeneous feature correlation function, and a second constraint function is constructed based on a second weight matrix in the attribute heterogeneous network function, through The first constraint function and the second constraint function construct the minimized objective function with the heterogeneous feature correlation function and the attribute heterogeneous network function, and calculate the characteristics of each object based on the minimized objective function, which makes up for the use of only a single heterogeneous correlation matrix. and a single homogeneous correlation matrix may not give reliable predictions for the flaw.

FIG. 3 is a flowchart of a method for object determination provided by an embodiment of the present disclosure. As shown in FIG. 3 , the method provided by the embodiment of the present disclosure includes steps S301-S308.

Step S301: Obtain the heterogeneous feature correlation matrix and the corresponding correlation relationship indication matrix of various types of objects in the object data set, and the multi-layer attribute heterogeneous network matrix of various types of objects.

Step S302: Construct a heterogeneous feature correlation function based on the heterogeneous feature correlation matrix and the corresponding correlation relationship indication matrix.

Step S303 : constructing an attribute heterogeneity network function based on the multi-layer attribute heterogeneity network matrix.

Step S304 : constructing a first constraint function based on the first weight matrix in the heterogeneous feature correlation function; and constructing a second constraint function based on the second weight matrix in the attribute heterogeneous network function.

The foregoing steps S301-S304 are the same as the steps S201-S204 in the foregoing embodiment. For details, refer to the foregoing embodiment, which will not be repeated here.

Step S305: Construct an unknown noise correlation constraint function.

In a specific embodiment, the association indicator matrix H _ij distinguishes observed associations from unobserved associations, but only constrains observed associations, so that it is likely to be left at random to some extent A large number of unobserved associations are left unconstrained, resulting in a large amount of noise in the target association matrix.

In order to avoid the aforementioned problems, in the embodiment of the present disclosure, an unknown noise correlation constraint function is also constructed, and a minimization objective function is constructed based on the unknown noise correlation constraint function.

其中

是对于未知噪声关联的约束，γ被用来控制该项的复杂度。In the embodiment of the present disclosure, the unknown noise correlation constraint function is

in

is the constraint on the unknown noise association, and γ is used to control the complexity of the term.

It should be noted that the aforementioned steps S304 and S305 may be performed sequentially, or may be performed in parallel.

Step S306: Construct a minimization objective function based on the heterogeneous feature correlation function, the attribute heterogeneous network function, the first constraint function, the second constraint function, and the unknown noise correlation constraint function.

After the unknown noise correlation constraint function is determined, the minimized objective function constructed in the embodiment of the present disclosure is:

Step S307: Calculate low-rank feature matrices of various types of objects based on the minimization objective function.

Step S308: Reconstructing object feature matrices of various types of objects in the object dataset by using the low-rank feature matrix.

The foregoing steps S307-S308 are the same as the steps S206-S207 in the foregoing embodiment. For details, refer to the foregoing embodiment, which will not be repeated here.

As mentioned above, in the implementation process of this solution, the alternating direction multiplier method can be used to solve the minimization objective function, and the low-rank feature matrix of various types of objects can be determined. Specifically, the alternating-direction multiplier method is used to calculate the low-rank feature matrices of various types of objects, and the alternating-direction multiplier method is used to alternately solve the first weight matrix, the second weight matrix, and the low-rank feature matrix in the minimized objective function. , the heterogeneous association network matrix, and the basis matrix of each multi-layer attribute heterogeneous network matrix, until the preset number of iterations or convergence is reached.

In the embodiment of the present disclosure, using the alternating direction multiplier method to solve the minimization objective function may include steps S401-S405. The following takes solving formula 1.5 as an example to analyze the alternating direction multiplier algorithm adopted in the embodiment of the present disclosure.

因此公式1.5可以等价为公式1.6。Before calculating Equation 1.5, first introduce the Lagrangian multiplier with constraint G _i ≥ 0

So Equation 1.5 can be equivalent to Equation 1.6.

Steps S401-S405 may then be performed using an alternating direction multiplier algorithm based on Equation 1.6.

Step S401: Set the first weight matrix, the second weight matrix, the low-rank feature matrix, and the heterogeneous correlation network matrix as fixed values, and obtain the first partial derivative of the basis matrix of the multi-layer attribute heterogeneous network matrix. When the value is zero, determines the basis matrix of the multi-layer attribute heterogeneous network matrix.

Specifically, assuming that G, S, ω ^r , ω ^h are known, U _it can be optimized, so the partial derivative with respect to U _it is calculated from Equation 1.6:

使得

可以得到

即多层属性异质网络矩阵的基矩阵为

for

make

can get

That is, the basis matrix of the multi-layer attribute heterogeneous network matrix is

Step S402: Set the basis matrix of the first weight matrix, the second weight matrix, the low-rank feature matrix and the multi-layer attribute heterogeneous network matrix as a fixed value, and obtain the second partial derivative for the heterogeneous correlation network matrix, and in the second partial derivative When the value is zero, the heterogeneous association network matrix is determined.

Specifically, assuming that G, U, ω ^r , ω ^h are known, S _ij can be optimized, so the partial derivative with respect to S _ij is obtained from equation 1.6:

使得

可以得到：

即异质关联网络矩阵为

for

make

You can get:

That is, the heterogeneous correlation network matrix is

Step S403: Set the basis matrix of the first weight matrix, the second weight matrix, the heterogeneous association network matrix and the multi-layer attribute heterogeneous network matrix as a fixed value, and obtain the third partial derivative for the low-rank feature matrix. When the value is zero, a low-rank feature matrix is determined.

Specifically, assuming that S, U, ω ^r , ω ^h are known, the partial derivative with respect to G _i can be obtained from Equation 1.6:

获得，由Karush-Kuhn-Tucker(KKT)条件可得：The polynomial factor λ _i can be obtained by making

can be obtained from the Karush-Kuhn-Tucker (KKT) condition:

where e represents the Hadamard product, Equation 1.10 is a fixed-point equation and the solution must satisfy the convergence conditions, so that:

for

For t=1,2,K,max _{i ti} :

和

因此，低秩特征矩阵G可更新为：

The positive and negative signs in equations 1.11, 1.12 and 1.13 can be defined as

and

Therefore, the low-rank feature matrix G can be updated as:

Step S404: Set the basis matrix of the first weight matrix, the low-rank feature matrix, the heterogeneous association network matrix and the multi-layer attribute heterogeneous network matrix as a fixed value, and obtain the fourth partial derivative for the second weight matrix. When the value is zero, the second weight matrix is determined.

Specifically, when calculating the partial derivative of ω ^h , the first, third, and fifth parts on the right side of Equation 1.6 have nothing to do with ω ^h , so they can be ignored. And thus get:

代表属性异质网络X_it的重构损失，前述公式可以化简为：make

Representing the reconstruction loss of the attribute heterogeneous network X _it , the aforementioned formula can be simplified as:

The solution of formula 1.15 can be regarded as a quadratic programming problem about vec(ω ^h ), which can be solved by introducing Lagrangian multipliers based on the Selective Non-Matrix Factorization (SNMF) algorithm.

Step S405: Set the basis matrix of the second weight matrix, the low-rank feature matrix, the heterogeneous association network matrix, and the multi-layer attribute heterogeneous network matrix as a fixed value, obtain the fifth partial derivative for the first weight matrix, and obtain the fifth partial derivative at the fifth partial derivative. When the value is zero, the first weight matrix is determined.

Specifically, when the partial derivative of ω ^r is obtained, the parts 2, 4, and 5 on the right side of Equation 1.6 have nothing to do with ω ^r , so they can be ignored. And thus get:

代表异质关联R_ij的重构损失，公式1.16可简化为：make

Representing the reconstruction loss of heterogeneous correlation R _ij , Equation 1.16 can be simplified to:

Equation 1.17 can be regarded as a quadratic programming problem about vec(ω ^r ), and it can also be solved by introducing Lagrange multipliers based on the algorithm of selective matrix decomposition. It should be noted that the foregoing specific embodiments of steps S401-S405 are not limited in order.

On the basis of the foregoing embodiments, the embodiments of the present disclosure provide an object determination apparatus. FIG. 4 is a schematic structural diagram of an object determination apparatus provided by an embodiment of the present disclosure. As shown in FIG. 4 , the object determination apparatus provided by the embodiment of the present disclosure includes a matrix acquisition unit 401 , a function construction unit 402 , a solution unit 403 , and an object feature matrix determination unit 404 .

The matrix acquisition unit 401 is configured to acquire heterogeneous feature correlation matrices of various types of objects in the object data set and corresponding correlation relationship indication matrices, as well as multi-layer attribute heterogeneous network matrices of various types of objects;

The function construction unit 402 is configured to construct a heterogeneous feature correlation function based on the heterogeneous feature correlation matrix and the corresponding correlation indication matrix; construct an attribute heterogeneous network function based on the multi-layer attribute heterogeneous network matrix; and, based on the heterogeneous feature correlation function and Attribute heterogeneity network function to construct the minimization objective function;

The solving unit 403 is configured to calculate low-rank feature matrices of various types of objects based on the minimization objective function;

The object feature matrix determining unit 404 is configured to reconstruct object feature matrices of various types of objects in the object dataset by using the low-rank feature matrix.

In the device for determining the object feature matrix provided by the embodiment of the present disclosure, a heterogeneous feature correlation function is constructed by using the heterogeneous feature correlation matrix and the corresponding correlation relationship indication matrix, so as to realize the heterogeneous correlation and potential correlation between the heterogeneous features using the correlation relationship indicator matrix. Heterogeneous associations are distinguished. The attribute heterogeneous network function is constructed by using the low-rank feature matrix corresponding to various types of objects and the decomposed basis matrix of the attribute heterogeneous network of various types of objects. The objective function is to use the objective function to minimize the first weight matrix, low-rank feature matrix, heterogeneous correlation network matrix, second weight matrix and the attribute difference of various types of objects in the heterogeneous feature correlation function and the attribute heterogeneous network function. The basis matrices decomposed by the qualitative network are co-graded to obtain low-rank feature matrices of various types of objects, and the object feature matrices of various types of objects are determined based on the low-rank feature matrices of various types of objects.

By using the device provided by the embodiment of the present disclosure, the topology structure of the network and the attribute information of the nodes can be fused together, so as to make up for the cold start problem caused by the lack of known associations, thereby reducing the information caused by the homogeneous transformation of the heterogeneous network with attributes. loss.

In some embodiments of the present disclosure, the function constructing unit 402 is further configured to construct a first constraint function based on the first weight matrix in the heterogeneous feature correlation function, and construct a second constraint based on the second weight matrix in the attribute heterogeneous network function function. Correspondingly, the function constructing unit 402 constructs the minimization objective function based on the heterogeneous feature correlation function, the attribute heterogeneous network function, the first constraint function and the second constraint function.

In some embodiments of the present disclosure, the function construction unit 402 is further configured to construct an unknown noise correlation constraint function. Correspondingly, the function construction unit 402 constructs the minimization objective function based on the heterogeneous feature correlation function, the attribute heterogeneous network function, the first constraint function, the second constraint function and the unknown noise correlation constraint function.

In some embodiments of the present disclosure, the solving unit 403 uses the alternating direction multiplier method to solve the minimization objective function, and determines low-rank feature matrices of various types of objects.

Specifically, in some embodiments of the present disclosure, the solving unit 403 adopts the alternate direction multiplier method to alternately solve the first weight matrix, the second weight matrix, the low-rank feature matrix, and the heterogeneous correlation network matrix in the minimization objective function in turn. , and the basis matrix of each multi-layer attribute heterogeneous network matrix until the preset number of iterations or convergence is reached.

In actual implementation, the solving unit 403 obtains the first partial derivative for the basis matrix of the multi-layer attribute heterogeneous network matrix by setting the first weight matrix, the second weight matrix, the low-rank feature matrix and the heterogeneous correlation network matrix as fixed values, When the first partial derivative value is zero, determine the basis matrix of the multi-layer attribute heterogeneous network matrix;

In actual implementation, the solving unit 403 obtains the second partial derivative for the heterogeneous association network matrix by setting the basis matrix of the first weight matrix, the second weight matrix, the low-rank feature matrix and the multi-layer attribute heterogeneous network matrix as a fixed value, When the second partial derivative value is zero, determine the heterogeneous correlation network matrix;

In actual implementation, the solving unit 403 obtains the third partial derivative for the low-rank feature matrix by setting the basis matrix of the first weight matrix, the second weight matrix, the heterogeneous association network matrix and the multi-layer attribute heterogeneous network matrix as a fixed value, When the third partial derivative value is zero, determine the low-rank feature matrix;

In actual implementation, the solving unit 403 obtains the fourth partial derivative for the second weight matrix by setting the basis matrix of the first weight matrix, the low-rank feature matrix, the heterogeneous association network matrix and the multi-layer attribute heterogeneous network matrix as a fixed value, When the fourth partial derivative value is zero, determine the second weight matrix;

In actual implementation, the solving unit 403 obtains the fifth partial derivative for the first weight matrix by setting the basis matrix of the second weight matrix, the low-rank feature matrix, the heterogeneous association network matrix and the multi-layer attribute heterogeneous network matrix as a fixed value, When the fifth partial derivative value is zero, the first weight matrix is determined.

In this embodiment of the present disclosure, the object dataset may be an image dataset, and the objects in the object dataset are images.

The apparatus provided by the embodiment of the present invention can execute the method provided by any embodiment of the present invention, and has corresponding functional modules and beneficial effects for executing the method.

FIG. 5 is a schematic structural diagram of an electronic device provided by an embodiment of the present disclosure. As shown in FIG. 5 , the electronic device includes a processor 501, a memory 502, an input device 503, and an output device 504; the number of processors 501 in the electronic device There may be one or more, and a processor 501 is taken as an example in FIG. 5; the processor 501, the memory 502, the input device 503, and the output device 504 in the electronic device can be connected by a bus or in other ways. Connect as an example.

As a computer-readable storage medium, the memory 502 may be used to store software programs, computer-executable programs, and modules, such as program instructions/modules corresponding to the methods in the embodiments of the present invention. The processor 501 executes various functional applications and data processing of the electronic device by running the software programs, instructions and modules stored in the memory 502, that is, implements the methods provided by the embodiments of the present invention.

The memory 502 may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function; the storage data area may store data created according to the use of the terminal, and the like. Additionally, memory 502 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid state storage device. In some examples, memory 502 may further include memory located remotely from processor 501, which may be connected to the computer device through a network. Examples of such networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

The input device 503 can be used to receive input digital or character information, and generate key signal input related to user settings and function control of the electronic device, which can include a keyboard, a mouse, etc., and the output device 504 can include a display device such as a display screen.

Embodiments of the present disclosure further provide a storage medium containing computer-executable instructions, and the computer-executable instructions, when executed by a computer processor, are used to implement the methods provided by the embodiments of the present disclosure.

Certainly, a storage medium containing computer-executable instructions provided by an embodiment of the present invention, the computer-executable instructions of which are not limited to the above method operations, and can also perform related operations in the methods provided by any embodiment of the present invention.

It should be noted that, in this document, relational terms such as "first" and "second" etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply these There is no such actual relationship or sequence between entities or operations. Moreover, the terms "comprising", "comprising" or any other variation thereof are intended to encompass non-exclusive inclusion such that a process, method, article or device comprising a list of elements includes not only those elements, but also includes not explicitly listed or other elements inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in the process, method, article, or device that includes the element.

The above are only specific embodiments of the present disclosure, so that those skilled in the art can understand or implement the present disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the present disclosure is not to be limited to the embodiments herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (10)

1. An object feature matrix determination method, comprising:

acquiring heterogeneous characteristic incidence matrixes and corresponding incidence relation indication matrixes of various types of objects in an object data set and multilayer attribute heterogeneous network matrixes of various types of objects;

constructing a heterogeneous characteristic correlation function based on the heterogeneous characteristic correlation matrix and the corresponding incidence relation indication matrix;

constructing an attribute heterogeneous network function based on the multilayer attribute heterogeneous network matrix;

constructing a minimized objective function based on the heterogeneous characteristic correlation function and the attribute heterogeneous network function;

calculating low-rank feature matrices of various types of objects based on the minimized objective function;

and reconstructing an object feature matrix of each type of object in the object data set by using the low-rank feature matrix.

2. The method of claim 1, further comprising:

constructing a first constraint function based on a first weight matrix in the heterogeneous characteristic correlation function;

constructing a second constraint function based on a second weight matrix in the attribute heterogeneous network function;

constructing a minimization objective function based on the heterogeneous characteristic correlation function and the attribute heterogeneous network function, wherein the minimization objective function comprises the following steps:

constructing the minimized objective function based on the heterogeneous feature correlation function, the attribute heterogeneous network function, the first constraint function and the second constraint function.

3. The method of claim 2, further comprising: constructing an unknown noise correlation constraint function;

constructing the minimization objective function based on the heterogeneous feature correlation function, the attribute heterogeneous network function, the first constraint function and the second constraint function, including:

and constructing the minimized objective function based on the heterogeneous characteristic correlation function, the attribute heterogeneous network function, the first constraint function, the second constraint function and the unknown noise correlation constraint function.

4. The method according to any of claims 1-3, wherein said determining a low rank feature matrix for each type of object based on said minimization objective function comprises:

and solving the minimized objective function by adopting an alternating direction multiplier method, and determining low-rank characteristic matrixes of various types of objects.

5. The method of claim 4, wherein solving the minimized objective function using an alternating direction multiplier method comprises:

and adopting the alternating direction multiplier method to sequentially and alternately solve the first weight matrix, the second weight matrix, the low-rank characteristic matrix, the heterogeneous correlation network matrix and the base matrix of each multilayer attribute heterogeneous network matrix in the minimized objective function until the preset iteration times or convergence is reached.

6. The method of claim 5, wherein the alternating direction multiplier method is adopted to sequentially and alternately solve the first weight matrix, the second weight matrix, the low rank feature matrix, the heterogeneous correlation network matrix and the base matrix of the multi-layer property heterogeneous network matrix in the minimized objective function, and comprises:

setting the first weight matrix, the second weight matrix, the low-rank feature matrix and the heterogeneous correlation network matrix as fixed values, solving a first partial derivative of a base matrix of the multilayer attribute heterogeneous network matrix, and determining the base matrix of the multilayer attribute heterogeneous network matrix when the first partial derivative value is zero;

setting the first weight matrix, the second weight matrix, the low-rank characteristic matrix and the base matrix of the multi-layer attribute heterogeneous network matrix as fixed values, solving a second partial derivative of the heterogeneous correlation network matrix, and determining the heterogeneous correlation network matrix when the second partial derivative value is zero;

setting the first weight matrix, the second weight matrix, the heterogeneous correlation network matrix and the base matrix of the multilayer attribute heterogeneous network matrix as fixed values, solving a third partial derivative of the low-rank feature matrix, and determining the low-rank feature matrix when the third partial derivative value is zero;

setting the first weight matrix, the low-rank feature matrix, the heterogeneous correlation network matrix and the base matrix of the multilayer attribute heterogeneous network matrix as fixed values, solving a fourth partial derivative of the second weight matrix, and determining the second weight matrix when the fourth partial derivative value is zero;

setting the second weight matrix, the low-rank feature matrix, the heterogeneous correlation network matrix and the base matrix of the multilayer attribute heterogeneous network matrix as fixed values, solving a fifth partial derivative of the first weight matrix, and determining the first weight matrix when the fifth partial derivative value is zero.

7. The method according to any of claims 1-3, wherein the object data set is an image data set and the object in the object data set is an image.

8. An object feature matrix determination apparatus, comprising:

the matrix acquisition unit is used for acquiring a heterogeneous characteristic incidence matrix and a corresponding incidence relation indication matrix of various types of objects in the object data set and a multilayer attribute heterogeneous network matrix of various types of objects;

the function construction unit is used for constructing a heterogeneous characteristic incidence function based on the heterogeneous characteristic incidence matrix and the corresponding incidence relation indication matrix; constructing an attribute heterogeneous network function based on the multilayer attribute heterogeneous network matrix; constructing a minimized objective function based on the heterogeneous characteristic correlation function and the attribute heterogeneous network function;

the solving unit is used for calculating low-rank characteristic matrixes of various types of objects based on the minimized objective function;

and the object feature matrix determining unit is used for reconstructing the object feature matrix of each type of object in the object data set by adopting the low-rank feature matrix.

9. An electronic device, comprising:

one or more processors;

a storage device for storing one or more programs,

when executed by the one or more processors, cause the one or more processors to implement a method as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, on which a computer program is stored, which program, when being executed by a processor, carries out the method of any one of claims 1 to 7.

Images