WO2021077989A1

WO2021077989A1 - Method and device for making recommendation, computer device, and storage medium

Info

Publication number: WO2021077989A1
Application number: PCT/CN2020/118107
Authority: WO
Inventors: 丁子扬; 马文晔
Original assignee: 腾讯科技（深圳）有限公司
Priority date: 2019-10-25
Filing date: 2020-09-27
Publication date: 2021-04-29
Also published as: CN110795625A; CN110795625B

Abstract

A method and device for making a recommendation, a computer device, and a storage medium, related to the technical field of computers. The method comprises: acquiring first characteristic information of a first object; mapping the first characteristic information to a target space on the basis of a mapping model to produce a first mapping vector corresponding to the first object in the target space; acquiring, on the basis of the distance between any two mapping vectors in the target space, a second object corresponding to a second mapping vector from which the distance to the first mapping vector is less than a preset distance, and making a recommendation on the basis of the first object and of the second object. The method does not involve other objects besides the first object and the second object during a recommendation process, that is, the method obviates the need to acquire the other objects besides the first object, and is not limited by the other objects during application, thus expanding the range of applications.

Description

Recommended method, device, computer equipment and storage medium

This application claims the priority of the Chinese patent application filed on October 25, 2019, the application number is 201911026124.6, and the invention title is "Recommended method, device, computer equipment and storage medium", the entire content of which is incorporated into this application by reference .

Technical field

The embodiments of the present application relate to the field of computer technology, and in particular, to a recommendation method, device, computer equipment, and storage medium.

Background technique

With the development of computer technology, more and more users use electronic devices to buy goods, read articles or watch videos, etc. As the scale of data gradually expands, how to recommend products, articles or videos for users has become an urgent solution The problem.

When recommending data for the user, select data similar to the data that the user has processed before, and recommend the similar data to the user. For example, recommend to the user products that are similar to the products that the user has previously purchased.

However, the use of the above-mentioned scheme must obtain the data that the user has processed before in order to make a recommendation, which has strong limitations and small application scope.

Summary of the invention

The embodiments of the present application provide a recommendation method, device, computer equipment, and storage medium, which expand the scope of application. The technical solution is as follows:

On the one hand, a recommendation method is provided, which is applied to a server, and the method includes:

Acquiring first characteristic information of a first object, where the first object belongs to a user identification or candidate data;

Based on the mapping model, the first feature information is mapped to a target space to obtain a first mapping vector corresponding to the first object in the target space, and the target space includes a user mapping vector corresponding to a user identification and The data mapping vector corresponding to the candidate data;

According to the distance between any two mapping vectors in the target space, a recommendation is made based on the first object and the second object, where the second mapping vector is a vector corresponding to the second object in the target space , The distance between the second mapping vector and the first mapping vector is less than a preset distance, and the second mapping vector and the first mapping vector belong to different categories.

In another aspect, a recommendation device is provided, the device including:

A first information acquisition module, configured to acquire first characteristic information of a first object, the first object belonging to a user identification or candidate data;

The first mapping module is configured to map the first feature information to a target space based on a mapping model to obtain a first mapping vector corresponding to the first object in the target space, and the target space includes a user Identify the corresponding user mapping vector and the data mapping vector corresponding to the candidate data;

The recommendation module is configured to make a recommendation based on the first object and the second object according to the distance between any two mapping vectors in the target space, where the second mapping vector is that the second object is in the target space. For a corresponding vector in the space, the distance between the second mapping vector and the first mapping vector is less than a preset distance, and the second mapping vector and the first mapping vector belong to different categories.

In another aspect, a computer device is provided, the computer device includes a processor and a memory, and at least one piece of program code is stored in the memory, and the at least one piece of program code is loaded and executed by the processor to realize the following: The operations performed in the recommended method.

In another aspect, a computer-readable storage medium is provided, and at least one piece of program code is stored in the computer-readable storage medium, and the at least one piece of program code is loaded and executed by a processor, so as to implement The action performed.

In another aspect, a computer program is provided, and at least one program code is stored in the computer program, and the at least one program code is loaded and executed by a processor, so as to implement the operations performed in the recommended method.

The method, device, computer equipment, and storage medium provided by the embodiments of the present application only need to acquire the first object, map the first object to the target space, and then, according to the distance between the respective mapping vectors included in the target space, Obtain the second object for recommendation, and then recommend based on the first object and the second object. No objects other than the first object and the second object are involved in the recommendation process, that is, there is no need to obtain other objects. It is not restricted by other objects and expands the scope of application.

Description of the drawings

In order to more clearly describe the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed in the description of the embodiments. Obviously, the drawings in the following description are only some implementations of the embodiments of the present application. For example, for those of ordinary skill in the art, without creative work, other drawings can be obtained from these drawings.

Fig. 1 is a schematic diagram of an implementation environment provided by an embodiment of the present application.

Fig. 2 is a flowchart of a recommendation method provided by an embodiment of the present application.

Fig. 3 is a schematic diagram of a mapping vector distance provided by an embodiment of the present application.

Fig. 4 is a schematic diagram of another mapping vector distance provided by an embodiment of the present application.

Fig. 5 is a flowchart of another recommendation method provided by an embodiment of the present application.

Fig. 6 is a schematic diagram of a recommendation interface provided by an embodiment of the present application.

Fig. 7 is a schematic diagram of a self-encoder provided by an embodiment of the present application.

Fig. 8 is a schematic diagram of a vector mapping provided by an embodiment of the present application.

Fig. 9 is a schematic diagram of another vector mapping provided by an embodiment of the present application.

Fig. 10 is a schematic diagram of another vector mapping provided by an embodiment of the present application.

FIG. 11 is a schematic diagram of a target space vector distribution provided by an embodiment of the present application.

FIG. 12 is a schematic diagram of a manifold structure provided by an embodiment of the present application.

FIG. 13 is a schematic diagram of a decoding process of a mapping vector provided by an embodiment of the present application.

Fig. 14 is a schematic diagram of a dual autoencoder provided by an embodiment of the present application.

FIG. 15 is a schematic diagram of a reconstruction process of a manifold structure provided by an embodiment of the present application.

FIG. 16 is a schematic diagram of a collaborative metric learning effect provided by an embodiment of the present application.

FIG. 17 is a schematic diagram of an in-depth model provided by an embodiment of the present application.

FIG. 18 is a schematic structural diagram of a recommendation device provided by an embodiment of the present application.

FIG. 19 is a schematic structural diagram of another recommending device provided by an embodiment of the present application.

FIG. 20 is a schematic structural diagram of another recommending device provided by an embodiment of the present application.

FIG. 21 is a schematic structural diagram of a terminal provided by an embodiment of the present application.

FIG. 22 is a schematic structural diagram of a server provided by an embodiment of the present application.

Detailed ways

In order to make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following further describes the embodiments of the present application in detail with reference to the accompanying drawings.

The terms "first", "second", etc. used in this application can be used herein to describe various concepts, but unless otherwise specified, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the present application, the first object may be referred to as the second object, and the second object may be referred to as the first object.

The term "at least one" used in the present application includes one or more than one, and the number of at least one is an integer, for example, the at least one may be 1, 2, 3, etc. The term "plurality" used in the present application includes two or more than two, and the number of the multiple is an integer, for example, the multiple may be 2, 3, 4, etc.

FIG. 1 is a schematic diagram of an implementation environment provided by an embodiment of the present application. The implementation environment includes: at least one terminal 101 and a server 102. At least one terminal 101 is connected to the server 102 and logs in to the server 102 based on a user identifier.

The server 102 stores multiple data, including video data, audio data, text data, or picture data. During the operation of any terminal 101, the server 102 recommends any data to the terminal 101 for display by the terminal 101.

Among them, the terminal 101 is various types of devices such as mobile phones and tablet computers. The server 102 is a server, or a server cluster composed of several servers, or a cloud computing service center.

Fig. 2 is a flowchart of a recommendation method provided by an embodiment of the present application. The execution subject of the embodiment of the present application is a server. Referring to Fig. 2, the method includes:

201. Acquire first characteristic information of a first object and second characteristic information of a second object.

Two types of objects are provided in the embodiments of this application: user identification and candidate data. The characteristic information of the user ID is used to describe the user corresponding to the user ID, and the characteristic information of the user ID includes information such as the age and gender of the user.

In a possible implementation manner, the characteristic information of the user identification also includes the user's interest tag. Optionally, the user's interest tag is obtained according to the candidate data processed before the user identification, for example, the user's interest tag is obtained through the user's product purchase record. Favorite commodity type, obtain the user's favorite article type through the article reading record, obtain the user's favorite video type through the video viewing record, etc.

In addition, the feature information of the candidate data is used to describe the candidate data.

In a possible implementation, if the candidate data is a commodity, the characteristic information is information such as the price and type of the commodity; if the candidate data is an article, the characteristic information is information such as the type and number of words of the article; If the candidate data is a video, the characteristic information is information such as the type and duration of the video.

In the embodiment of this application, the first object belongs to user identification or candidate data, and the second object also belongs to user identification or candidate data. The first object and the second object belong to different categories, that is, if the first object is a user identification, The second object is the candidate data; if the first object is the candidate data, the second object is the user identification.

202. Based on the mapping model, map the first feature information and the second feature information to the target space respectively to obtain the first mapping vector corresponding to the first feature information in the target space, and the second feature information in the target space The corresponding second mapping vector.

In the embodiment of the present application, the mapping model is used to map the feature information, and the feature information is mapped to the target space through the mapping model to obtain the corresponding mapping vector. Among them, compared with the original space of the feature information, the target space has a different dimension from the original space, and the target space is a low-dimensional space or a high-dimensional space.

In addition, the feature information of different types of objects can be mapped to the target space based on the mapping model to obtain the mapping vector corresponding to the feature information. Among them, the feature information of the user identification is mapped to the target space to obtain the corresponding user mapping vector, and the feature information of the candidate data is mapped to the target space to obtain the corresponding data mapping vector.

The first feature information is input to the mapping model to obtain the first mapping vector corresponding to the target space, and the second feature information is input to the mapping model to obtain the second mapping vector corresponding to the target space. Since the first feature information and the second feature information belong to different types of objects, the second mapping vector and the first mapping vector belong to different types of mapping vectors, that is, the first mapping vector is the user mapping vector, and the second mapping vector is the data mapping Vector, or the first mapping vector is a data mapping vector, and the second mapping vector is a user mapping vector.

In one possible implementation, the mapping model is a single mapping model. There is a one-to-one correspondence between the feature information and the mapping vector obtained based on the mapping model. Each feature information has a unique corresponding mapping vector, and each mapping vector There is a unique corresponding feature information. That is, based on the single mapping model, if any feature information is mapped, the mapping vector corresponding to the feature information can be obtained, and there is no other feature information corresponding to the mapping vector.

In a possible implementation manner, the mapping model includes multiple mapping sub-models, and the multiple mapping sub-models are used to map different types of feature information. For example, the mapping model includes a user mapping sub-model and a data mapping sub-model. Among them, the user mapping sub-model is used to map the feature information of the user identification to obtain the user mapping vector, and the data mapping sub-model is used to map the feature information of the candidate data to obtain the data mapping vector. Optionally, the feature information mapping process of the user identification and the feature information mapping process of the candidate data are performed simultaneously or sequentially.

Optionally, based on the mapping model, the first feature information of the first object is mapped to the target space to obtain the first mapping vector corresponding to the first object in the target space, including the following two cases: In the case of the user mapping sub-model, the feature information of the user identification is mapped to the target space, and the user mapping vector corresponding to the user identification in the target space is obtained; in the case that the first object is candidate data, based on the data mapping The sub-model maps the feature information of the candidate data to the target space, and obtains the data mapping vector corresponding to the candidate data in the target space.

Optionally, based on the mapping model, the first feature information of the second object is mapped to the target space to obtain the first mapping vector corresponding to the second object in the target space, including the following two cases: In the case of the user mapping sub-model, the feature information of the user identification is mapped to the target space, and the user mapping vector corresponding to the user identification in the target space is obtained; in the case that the second object is candidate data, based on the data mapping The sub-model maps the feature information of the candidate data to the target space, and obtains the data mapping vector corresponding to the candidate data in the target space.

In one possible implementation, the server stores the feature information and the mapping vector obtained based on the mapping model mapping, the feature information of the user identification and the corresponding user mapping vector are stored correspondingly, and the feature information of the candidate data and the corresponding data mapping vector Corresponding storage, so that two mapping vectors can be determined later, and when a recommendation is made based on the objects corresponding to the two mapping vectors, the feature information corresponding to the mapping vector can be easily obtained.

It should be noted that, in the embodiment of the present application, the first feature information and the second feature information are mapped to the target space based on the mapping model. Optionally, the first feature information and the second feature information are simultaneously mapped based on the mapping model. To the target space, or based on the mapping model, first map the second feature information to the target space, and then map the first feature information to the target space. You only need to ensure that the distance between the first mapping vector and the second mapping vector is determined. It suffices that the second mapping vector exists in the target space.

203. Obtain the distance between the first mapping vector and the second mapping vector.

In the embodiment of the present application, the distance between the first mapping vector and the second mapping vector needs to be measured to determine the distance between the two. Therefore, it is necessary to define a metric to measure the distance between any two mapping vectors in the target space. The metric needs to meet at least the following conditions:

(1) Satisfy the various requirements of measurement in mathematics, namely the identity of non-negativity, indistinguishability, symmetry and triangle inequality.

(2) The metric can be calculated by Euclidean metric.

(3) On the premise of ensuring the distance between the mapping vectors, it is possible to embed as many mapping vectors as possible.

A consistent metric is defined in the embodiments of the present application, and the consistent metric is used to measure the distance between any two mapping vectors. Defined consistent metric

for:

among them,

Is any two mapping vectors in the target space;

Is the Chebyshev distance, that is, L _∞ distance; a is the preset distance, and a>0; sup{·} is the supremum function, i takes different values, and |x _i -y _i | can get multiple different Numerical value, sup{·} represents the minimum upper bound corresponding to the multiple different values. If there is a maximum value among multiple values, the maximum value is the minimum upper bound. If the largest multiple values are infinitely close to a certain value , This value is the minimum upper bound; min{·} represents the minimum value among multiple values.

In the consistency metric _{, the maximum value between any two mapping vectors is calculated by L ∞} distance, the maximum value is compared with a, and the minimum value is selected.

For example, to

Indicates a certain mapping vector. There are multiple mapping vectors in the target space. The multiple mapping vectors are regarded as multiple points. The line in Figure 3 refers to the

The distance between is the line formed by connecting the points of a, the top view of Fig. 3 is shown in Fig. 4, the circle and the mapping vector in Fig. 4

The distance between is a, it is considered that any mapping vector in the part indicated by diagonal lines is the same as the mapping vector

The distance between the two is relatively close, and any mapping vector in the area outside the circle is considered to be the same as the mapping vector

The distance between them is far.

Prove that the defined consistent metric satisfies the above metric conditions

The proof process for the consistent measurement to prove that the above condition (1) is satisfied is as follows:

Regarding non-negative proof: due to

And a>0,

Therefore, when the minimum value is taken, non-negativity is established.

Proof of the identity of the indistinguishable: if

Then

And get

in case

Then

And get

Therefore, the identity of the indistinguishable is established.

Regarding the proof of symmetry: due to

Is a well-defined metric, so it has symmetry, that is

Then

And get

Therefore, symmetry holds.

About the proof of triangle inequality: if there are three vectors

Then the sum of the distance between any one of the vectors and the other two vectors should be greater than the distance between the other two vectors.

The sum of the distance between any vector and the other two vectors is:

Define a new metric d(x, y) = min{d(x, y), a}, the non-negativity of the metric, the identity and symmetry of the indistinguishable are obvious, and its triangular inequality properties:

Therefore, d(x,y)=min{d(x,y),a} is a well-defined metric. The above inequality is written as:

Through the above proof, it can be determined that the consistent metric is a well-defined metric that meets the requirements of the metric definition.

Since this consistency metric is a metric based on Euclidean space, it satisfies the above condition (2).

This consistent metric is different from other Euclidean space metrics. The metric topological space induced by this metric is considered to be

The distance between the vector is not less than a and

The distance between is a, so even in a low-dimensional space, many vectors with equal distances can be mapped. If you need to adjust the mapping ability of the space, you only need to adjust a. The smaller a, the stronger the mapping ability; the larger the a, the weaker the mapping ability. A is equal to 0. The space is completely equivalent to the one induced _{by l ∞.} Measure the topological space. Therefore, the above condition (3) is satisfied.

The distance between the first mapping vector and the second mapping vector is measured by the defined consistency metric, and the distance between the two mapping vectors is obtained.

204. If the distance is less than the preset distance, make a recommendation based on the first object and the second object.

In the embodiment of the present application, if the distance between the first mapping vector and the second mapping vector is less than the preset distance, the recommendation is made based on the first object and the second object.

If the distance between the first mapping vector and the second mapping vector in the target space is not less than the preset distance, no recommendation is made based on the first object and the second object.

In one possible implementation, the first object is the user identification and the second object is the candidate data, or the first object is the candidate data and the second object is the user identification, and recommendations are made based on the first object and the second object , Including: recommending candidate data to the user ID.

In a possible implementation manner, recommending candidate data to the user ID includes: sending the candidate data by the server to the terminal logged in with the user ID, and displaying the candidate data by the terminal, which can be viewed by the user.

For example, the recommendation interface displayed by the terminal is shown in FIG. 6, and the recommendation interface includes a user avatar, following options, and recommendation options. Click the user’s avatar to view user information such as the user ID, click the follow option to view the articles published by other user IDs that the user ID follows, and click the recommendation option. The recommendation interface displays articles that may be of interest to the user recommended by the user, as well as some popular articles. Interested articles are recommended based on the characteristic information identified by the user.

Among them, the preset distance is the minimum distance used to indicate that the user corresponding to the user identifier is interested in the candidate data. The preset distance is randomly determined by the server or set according to needs. If the recommended accuracy rate is higher, it is recommended If the candidate data is more in line with the user's interest, a smaller preset distance is set; if it is necessary to obtain as much recommended candidate data as possible, a larger preset distance is set.

For example, if the user and the product are known, it is determined whether to recommend the product to the user. First, obtain the user characteristic information corresponding to the user and the product characteristic information corresponding to the product, and input the user characteristic information and the product characteristic information into the mapping model respectively. Obtain the user mapping vector corresponding to the user and the product mapping vector corresponding to the product. Based on the consistent measurement, obtain the distance between the user mapping vector and the product mapping vector. If the distance is less than the preset distance, it means that the user is interested in the product. Recommend the product to the user. If the distance is not less than the preset distance, it means that the user is not interested in the product and there is no need to recommend the product to the user.

It should be noted that, in another embodiment, the feature information of the user identification can be input into the mapping model, and similar users can be obtained through a method similar to the embodiment of this application, and then the recommendation can be made; the feature information of the candidate data can be input To the mapping model, similar data is obtained by a method similar to the embodiment of the present application, and then the recommendation is made.

For example, for two users in an application client that can add friends, input the characteristic information of the two users into the mapping model to obtain two corresponding user mapping vectors, and the difference between the two user mapping vectors is If the distance is less than the preset distance, the two users are considered to be similar users, and one user can be recommended to the other user.

In the method provided by the embodiments of the present application, only the first object is acquired, the first object is mapped to the target space, and the recommended second object can be acquired according to the distance between the mapping vectors included in the target space. Therefore, recommendations are made based on the first object and the second object. The recommendation process does not involve other objects except the first object and the second object, that is, there is no need to obtain other objects, and it is not restricted by other objects during application. The scope of application.

For example, for a user and a product, in related technologies, it is necessary to obtain the product that the user has previously purchased, and determine whether to recommend the product to the user according to whether the product is similar to the purchased product. In the embodiment of the present application, it is only necessary to obtain the distance between the user mapping vector and the product mapping vector in the target space according to the user feature information of the user and the product feature information of the product, and then make recommendations based on the distance. If the distance is less than the preset distance , The product is recommended to the user, and there is no need to indirectly determine whether to recommend through other products.

Fig. 5 is a flowchart of another recommendation method provided by an embodiment of the present application. The execution subject of the embodiment of the present application is a server. Referring to FIG. 5, the method includes:

501. Acquire first characteristic information of a first object.

The specific implementation is similar to the implementation of step 201 in the foregoing embodiment, and will not be repeated here.

502. Based on the mapping model, map the first feature information to the target space, and obtain a first mapping vector corresponding to the first feature information in the target space.

The target space includes a user mapping vector corresponding to the user identification and a data mapping vector corresponding to the candidate data.

The implementation manner in which the first feature information is mapped to obtain the corresponding first mapping vector in the embodiment of the present application is similar to the implementation manner of step 202 in the foregoing embodiment, and will not be repeated here.

It should be noted that in this embodiment of the application, only the first feature information needs to be mapped to the target space to obtain the corresponding first mapping vector, and the mapping method for other mapping vectors in the target space except the user mapping vector is not used. Limited, optionally, the mapping model in the embodiment of the present application is used for mapping, or other methods are used for mapping.

503. Determine at least one third mapping vector that belongs to a different category from the first mapping vector in the target space.

504. Obtain a distance between the first mapping vector and each third mapping vector.

Among them, the third mapping vector and the first mapping vector belong to different categories. If the first mapping vector is the mapping vector identified by the user, the third mapping vector is the mapping vector of the candidate data; if the first mapping vector is the mapping vector of the candidate data Mapping vector, the third mapping vector is the mapping vector of the user identification.

The target space includes at least one third mapping vector, the position of each third mapping vector in the space is determined, and the distance between the first mapping vector and each third mapping vector is obtained according to the consistent metric defined in the target space .

The distance acquisition method in the embodiment of the present application is similar to the implementation method in step 203 described above, and will not be repeated here.

505. From at least one third mapping vector, select a second mapping vector whose distance to the first mapping vector is less than a preset distance.

After obtaining the distance between the first mapping vector and each third mapping vector according to step 504, the second mapping vector is selected from the third mapping vector whose distance from the first mapping vector is less than the preset distance.

In a possible implementation manner, when selecting a second mapping vector whose distance from the first mapping vector is less than a preset distance, optionally, there are one or more second mapping vectors selected. Among them, the number of selected second mapping vectors is set as required.

506. Determine a second object corresponding to the second mapping vector, and make a recommendation based on the first object and the second object.

According to the selected second mapping vector, the second object corresponding to the second mapping vector is determined, and recommendations are made based on the first object and the second object.

In a possible implementation manner, the server stores the corresponding relationship between each mapping vector and the corresponding object, and by querying the corresponding relationship, the object corresponding to each mapping vector is determined.

In another possible implementation manner, based on the inverse mapping model, the second mapping vector is inversely mapped to obtain the second object corresponding to the second mapping vector. Among them, the de-mapping model is used to de-map the mapping vector, and the mapping vector is de-mapped to the original space through the de-mapping model to obtain corresponding feature information.

In one possible implementation, the de-mapping model is a SLR mapping model, that is, each mapping vector has a one-to-one correspondence with the feature information obtained based on the de-mapping model de-mapping, and each mapping vector has a unique corresponding feature Information, each feature information also has a unique corresponding mapping vector.

In another possible implementation, the anti-mapping model is a non-SLR mapping model, that is, each mapping vector has unique corresponding feature information, but one feature information may correspond to multiple mapping vectors.

In addition, if the second object is a user identifier, the de-mapping model is a user de-mapping model, and if the second object is candidate data, the de-mapping model is a data de-mapping model.

One thing to note is that if the first object is a user ID and the second object is candidate data, the characteristic information of the user ID is obtained, and based on the mapping model, the characteristic information is mapped to the target space, and the user mapping corresponding to the characteristic information is obtained. Vector, and then determine the data mapping vector of at least one candidate data in the target space, obtain the distance between the user mapping vector and each data mapping vector, and select the distance to the user mapping vector from at least one data mapping vector For the data mapping vector less than the preset distance, the candidate data corresponding to the selected data mapping vector is determined, and the selected candidate data is recommended to the user identification.

If the first object is the candidate data and the second object is the user identification, obtain the feature information of the candidate data, map the feature information to the target space based on the mapping model, obtain the data mapping vector corresponding to the feature information, and then determine the target The user mapping vector of at least one user identification in the space, the distance between the data mapping vector and each user mapping vector is obtained, and the user whose distance to the data mapping vector is less than the preset distance is selected from at least one user mapping vector The mapping vector determines the user identification corresponding to the selected user mapping vector, and then recommends the candidate data to the selected user identification.

Another point that needs to be explained is that, in another embodiment, for objects of the same category, similar user identifications or similar candidate data can be obtained through a method similar to the embodiment of the present application.

For example, if the first object and the second object are both user identifications, then the feature information of the first user identification is obtained, based on the mapping model, the feature information is mapped to the target space, and the first user mapping vector corresponding to the feature information is obtained, and then Determine the second user mapping vector of at least one second user identifier in the target space, obtain the distance between the first user mapping vector and each second user mapping vector, and select the second user mapping vector from the at least one second user mapping vector. A second user mapping vector whose distance between the user mapping vectors is less than the preset distance is determined, and the second user identification corresponding to the selected second user mapping vector is determined, and the user represented by the first user identification is considered to be the same as the second user identification. The interests of the users are similar, and the selected second user ID is recommended to the first user ID.

For example, if the candidate data is a commodity, and the first object and the second object are both commodities, the characteristic information of the first commodity is obtained, and the characteristic information is mapped to the target space based on the mapping model to obtain the first data corresponding to the characteristic information Mapping vector, and then determine the second data mapping vector of at least one second product in the target space, obtain the distance between the first data mapping vector and each second data mapping vector, from the at least one second data mapping vector, Select the second data mapping vector whose distance from the first data mapping vector is less than the preset distance, and determine the second product corresponding to the selected second data mapping vector. The first product and the second product are considered to be similar. The second product is recommended to users who have purchased the first product.

Moreover, in related technologies, when recommending candidate data for a user ID, it is necessary to obtain data that the user has previously processed. For a new user ID or new candidate data, it is impossible to obtain a new user ID and new candidate data and data. The relationship between other user identifications or candidate data, it is impossible to recommend candidate data to the new user identification, or to recommend new candidate data to the user identification. However, when recommending candidate data to a user ID in this application, it does not involve other user IDs or candidate data before the user ID and the candidate data. Therefore, it can also be used for new user IDs or new candidate data. Recommendations are made to expand the scope of application.

Moreover, if the first object is a user ID and the second object is candidate data, only the characteristic information of the user ID needs to be mapped to the target space, and the distance from the user mapping vector of the user ID is less than the preset value. The data mapping vector of the distance is used to determine the candidate data that the user is interested in corresponding to the user ID, without the need to indirectly obtain the candidate data that the user is interested in corresponding to the user ID based on other user IDs or candidate data. The scope of application. If the first object is the candidate data and the second object is the user identification, you only need to map the feature information of the candidate data to the target space, and the distance from the data mapping vector of the candidate data is less than the preset The user mapping vector of the distance is used to determine the user identification interested in the candidate data, without the need to obtain the user identification interested in the candidate data indirectly based on other candidate data or user identification, which expands the scope of application.

In addition, this method can also absoluteize the interest points of the user identification, make the user interest more clear, and realize the inference of the characteristics of the candidate data that the user likes when there is no candidate data.

In the above embodiments, the mapping model and the de-mapping model are involved. In order to facilitate the training of these two models, an autoencoder can be used. The autoencoder includes an encoding model and a decoding model. The encoding model is used as the mapping model, and the decoding model is used as the de-mapping model. The following describes the training process of the autoencoder.

(1) Obtain sample data.

Obtain sample information, the sample information includes the characteristic information of the sample user identification, the characteristic information of the sample data and the sample label, and the sample label is used to indicate whether to identify the recommended sample data to the sample user.

Optionally, the sample label is 1 or -1. 1 indicates that the sample user ID and sample data have a positive relationship, which means that the sample data is recommended to the user ID; -1 indicates that the sample user ID and the sample data have a negative relationship, which means that there is no Recommend sample data to user identification.

The feature information of the sample user identification is similar to the feature information of the above-mentioned user identification, and the feature information of the sample data is similar to the feature information of the above-mentioned candidate data, and will not be repeated here.

(2) According to the sample data, train the autoencoder.

The feature information of the sample user identification and the feature information of the sample data are input to the autoencoder, and the feature information of the predicted user identification or the feature information of the predicted sample data is output based on the self-encoder, and the predicted feature information and the corresponding input feature are based on The loss value generated between the information, the parameters of the autoencoder are adjusted, so that the adjusted loss value between the predicted feature information output by the autoencoder and the corresponding input feature information is reduced, so as to achieve the training of the The purpose of the self-encoder.

Among them, the structure of the autoencoder is shown in Figure 7, including the encoding model and the decoding model, and the feature vector

Input to the coding model and get a corresponding mapping vector

Use the decoding model to map the vector

Decode the corresponding prediction feature vector

Optionally, the encoding model and the decoding model further include multiple hidden layers.

In a possible implementation, when the sample labels are input to the autoencoder, in the training process, after the mapping vector is obtained based on the coding model, it is predicted between the mapping vector corresponding to the user identification and the mapping vector corresponding to the sample data The distance between the sample user ID and the sample data is negative or positive, and the obtained relationship is compared with the relationship represented by the input sample label, and the parameters of the autoencoder are adjusted to make the autoencoder after adjustment. The prediction relationship of the encoder is the same as the relationship represented by the sample label, which achieves the purpose of training the autoencoder.

In one possible implementation, the loss function for training the autoencoder includes the following:

The first:

The embodiment of the present application provides two loss functions, and the first loss function is:

Among them, L _neck1 is the loss value of the mapping model,

Is the sample label, its value is 1 or -1,

Is the mapping vector corresponding to the characteristic information of the sample user identification,

Is the mapping vector corresponding to the feature information of the sample data.

According to the consistent metric defined in the above embodiment, the distance between the mapping vector corresponding to the user identifier and the mapping vector corresponding to the candidate data under the consistent metric is obtained, and the distance is multiplied by the corresponding label data as the first loss function .

If the first loss function is used for training, considering a special situation, that is, the sample user ID and sample data in the training sample show a positive relationship. Based on the coding model in the untrained autoencoder, the sample user ID and The sample data is mapped to the metric space, and the mapping vector corresponding to the sample user ID is obtained

And the mapping vector corresponding to the sample data

get

with

Refer to Figure 8 for the consistent measurement distance of the two vectors. The distance is greater than the preset distance. At this time, the gradient is 0. On the basis of this distance, the method of decreasing gradient cannot be used to continue training.

The second loss function is:

Among them, L _neck2 is the first loss value of the mapping model, and λ _margin is the preset parameter,

Is the sample label, its value is 1 or -1,

When the second type of loss function, namely hinge loss (a loss function) is used for training, for the case where the sample user ID and the sample data show a negative relationship, see Figure 9, where the loss value of the loss function in the diagonal area is higher. Small, the loss value of the loss function in the blank area is large, and the arrow direction indicates the hope vector

The direction of movement so that

versus

The distance between them is as far as possible. The picture on the left shows the training using the first loss function, and the dashed circle is and

The distance between a

The figure on the right shows the second loss function for training. The dashed circle is the target safety limit of the negative sample. The target safety limit is the distance obtained by adding a certain value to the distance a. The use of the target safety limit makes the relationship between the user identification obtained by training and the sample data more accurate. In this case, the results of training with the two loss functions are the same.

For the case where the sample user ID and the sample data show a positive relationship, see Figure 10, where the loss value of the loss function in the oblique area is small, the loss value of the loss function in the blank area is larger, and the arrow direction indicates the desired vector

The direction of movement so that

versus

The distance between them is as close as possible. The picture on the left shows the use of the first loss function for training, and the picture on the right shows the use of the second loss function for training. In this case, the left picture shows the situation where training cannot be performed as shown in FIG. 8, and the second loss function in the right picture can avoid the situation where training cannot be performed.

In summary, from the mathematical point of view, directly training the autoencoder based on the distance between two vectors defined by a consistent metric has the same effect as training the autoencoder using hinge loss. In practical applications, the second loss function, namely hinge loss, is used for training, and the preset distance a in the consistent metric is trained to obtain a better training effect.

The second type:

Considering that when the feature information is mapped to a low-dimensional space, the distribution of the mapping vector in the space needs to be distributed as evenly as possible, and the following loss function is used for training:

Among them, L _cov is the second loss value of the mapping model, N is the number of sample data, and E is

with

The formed matrix, Cov(E) is the covariance matrix of matrix E, ||·|| _f is the transposition function, and diag(·) is the diagonal element extraction function of the matrix.

In the encoding process, when the feature information corresponding to the sample user ID or the feature information corresponding to the sample data is mapped to a low-dimensional space, the distribution of the corresponding mapping vector in the low-dimensional space is shown in Figure 11, where the triangle represents The mapping vector of one category, and the circle represents the mapping vector of another category. The picture on the left shows the high collinearity distribution obtained after mapping. This situation wastes space and will cause the autoencoder to be trained. In order to map more mapping vectors, the dimension of the space can only be increased, which will lead to The current training samples can be perfectly predicted, but the prediction effect for new samples is poor. Using this loss function to train the autoencoder can achieve the effect of the right figure, so that the mapping vector obtained by the mapping is evenly distributed in the space.

The third type:

It is necessary that the input feature information of the autoencoder is as same as the output feature information, and the following loss function is used for training:

Among them, L _reconstruct is the loss value of the self-encoder,

Is the characteristic information identified by the sample user or the characteristic information of the sample data,

for

Based on the feature information output after processing by the autoencoder.

Or, combine the above three loss functions to get an overall loss function:

L=αL _neck2 +βL _reconstruct +γL _cov ;

In the above training process, if the mapped space is infinite, the distance between all the mapping vectors in the space cannot be measured. Therefore, the low-dimensional space that needs to be mapped is a limited space. Therefore, the activation function that enters the embedding layer in the last layer of the coding model needs to be a bounded activation function, such as a bounded function such as a Sigmoid (a bounded activation function) function, a tanh (hyperbolic tangent) function, etc.

Moreover, since the output feature information includes numeric features and binary features, it needs to be standardized during processing, that is, the value obtained is between 0 and 1. Therefore, in the decoding model, the value range of the activation function of the last layer entering the output layer needs to be between 0 and 1. Optionally, the activation function is a Sigmoid function or other functions.

In a possible implementation manner, the autoencoder trained by the above method is used to reconstruct the user identification or the interest manifold structure of the candidate data in a low-dimensional space, as shown in Figure 12, where the triangle represents a category of The mapping vector, the circle represents the mapping vector of another category. What forms the interest manifold structure is the mapping vector corresponding to all user identities or candidate data. Among them, for any two mapping vectors, regardless of whether the two mapping vectors are of the same category, the distance between the two mapping vectors can directly indicate the "favoring" relationship of the two mapping vectors. The closer the distance, the stronger the favoring relationship. The farther the distance, the weaker the relationship. If the distance is less than the preset distance, the two mapping vectors are in a positive relationship, and if the distance is not less than the preset distance, the two mapping vectors are in a negative relationship. Due to the nature of the triangular inequality of the consistent measurement, each mapping vector will have a clustering effect, that is, similar user identifications will be aggregated, similar candidate data will be aggregated, and the user identification and the user identification The recommended candidate data will also be aggregated.

Moreover, after the interest manifold structure is obtained, the mapping vector included in the interest manifold structure can be decoded through the decoding model to obtain the feature information of the corresponding user identification or the feature information of the candidate data. Refer to FIG. 13 for the decoding process, where the circular area represents a part of the interest manifold structure after mapping, and the mapping vector in this area obtains the decoded manifold structure through the decoding model. The manifold structure obtained after decoding has continuity, so the interest manifold structure will not lose the mapping vector due to the decoding process. And if the decoding model is not a single decoding model, that is, after multiple mapping vectors of the same category are decoded by the decoding model, the same feature information may be obtained. Therefore, the structure obtained after decoding may have a cross region.

The first point that needs to be explained is that the embodiment of the present application only takes the training process of an autoencoder as an example for description. When a self-encoder is used in the above embodiment, a double-self-encoder can be used. The structure of the double-self-encoder is shown in Figure 14. One of the self-encoders is used to encode and decode the characteristic information of the user identification, and the other is used to Encode and decode the feature information of the candidate data. In another embodiment, if multiple categories of feature information are included, a corresponding number of multiple autoencoders can be used to respectively encode and decode the feature information of each category.

The second point that needs to be explained is that in another embodiment, different types of auto-encoders can be used according to different data types, such as VAE (Variational Auto-Encoder) and Contrastive auto-encoders. Wait. The use of different types of encoders needs to follow corresponding principles, and the characteristic information needs to be mapped into a bounded space with a clear metric definition.

For example, when using VAE (Variational Auto-Encoder), the target space is a high-dimensional Gaussian probability distribution space, in which KLD (Kullback–Leibler divergence, KL divergence) pairs are defined in The distance between the distributions is defined.

The third point that needs to be explained is that for different data sparsity, the embedding layer can be added to the original model, or other models such as wide&deep (a deep learning model) can be used to replace the basic MLP (Multilayer Perceptron, artificial neural network). ) To better obtain the information in the coefficient data.

The fourth point that needs to be explained is that if the input data is time-series data, a neural network that changes over time can be used. For example, from the aspect of deep learning, RNN (Recurrent Neural Network) or LSTM (Long Short-Term Memory) is used; from the aspect of statistical learning, Bayesian prior The posterior timing update, or the use of Kalman filter for time series data learning.

In the related technology, CML (Collaborative Metric Learning) is used for recommendation. This method is based on the known relationship between the user identifier and the candidate data in the original space and the target space. The corresponding vector is moved to obtain a distance relationship similar to the original space. This method is used for fixed user identification and candidate data. This method has a small range of use, and needs to obtain the relationship between the user identification and candidate data in the original space. When recommending new user identification or candidate data, This method is not applicable.

For example, using collaborative metric learning to recommend products for users, see Figure 16. The circle in the figure represents the user, the triangle represents the product the user likes, the rectangle represents the product the user dislikes, and the arrow is used to indicate the direction of the product. The left picture is the original position of the product and the user. Based on the collaborative measurement learning, the result in the right picture is obtained, so that the product that the user likes is close to the user, and the product that the user does not like is far away from the user. When this method is adopted, the users and commodities in the space are fixed, and only fixed commodities can be recommended for the fixed users. If there are no commodities in the space, it is impossible to infer the commodities that the user may like.

In the above CML method, for multiple user IDs and multiple candidate data, the relationship between some user IDs and some candidate data must be known, otherwise the vector cannot be moved in the target space according to the known relationship , And for a new user ID or new candidate data, because the relationship between the new user ID and new candidate data and other user IDs and candidate data is unknown, it is impossible to determine the new user ID and new candidate data. The position of the vector corresponding to the candidate data in the target space cannot be recommended. However, the embodiment of the present application does not need to obtain the relationship between the user identification and the candidate data in advance, and is applicable to any user identification or candidate data, which expands the scope of application.

In related technologies, t-SNE (student-t Stochastic Neighborhood Embedding, an algorithm) is used to reconstruct the data manifold structure. The principle of the algorithm is: the distance relationship between any two feature vectors in the high-dimensional space should be similar to the distance relationship between any two mapping vectors in the low-dimensional space. Assuming that there are two feature vectors in the high-dimensional space, the The two eigenvectors are far apart in the high-dimensional space, so in the low-dimensional space, the two mapping vectors corresponding to the two eigenvectors should also be far away, and vice versa. If there are multiple feature vectors in the original high-dimensional space, such as n feature vectors, then the low-dimensional space will have n corresponding mapping vectors. The effect achieved by this method is shown in Figure 15. The first figure on the left is the manifold structure composed of multiple feature vectors in the original high-dimensional space, and the second figure is the mapping of multiple feature vectors in the original high-dimensional space to the low-dimensional The manifold structure formed by the mapping vector obtained after the space, and then the manifold structure of the third graph and the fourth graph are obtained in turn, and finally the manifold structure of the fifth graph is obtained, which realizes the reconstruction of the high-dimensional space in the low-dimensional space. Manifold structure of n eigenvectors in a dimensional space.

In related technologies, VaeCF (Variational Autoencoder Collaborative Filtering, an in-depth model) is also used for data recommendation. Refer to Figure 17 for the in-depth model. This method can accurately obtain the relationship between user identification and candidate data. However, When using this method, if only the characteristic information of the user ID is used, and the characteristic information of the candidate data is not given, the candidate data recommended for the user ID cannot be obtained, that is, the user's interest cannot be obtained based on the user ID.

In the method provided in the embodiments of the present application, according to the characteristic information of any user identification, the candidate data recommended for the user identification can be obtained based on the self-encoder, or, according to the characteristic information of any candidate data, it can be based on the self-encoder. The encoder obtains the user identification that is interested in the candidate data, and then recommends the candidate data. This method has a wide range of use and can make recommendations based on the user identification or the feature information of one of the candidate data. In related technologies, the characteristic information of the user identification and the characteristic information of the candidate data must be obtained, otherwise the problem of recommendation cannot be made.

FIG. 18 is a schematic structural diagram of a recommendation device provided by an embodiment of the present application. Referring to Figure 18, the device includes:

The first information obtaining module 1801 is configured to obtain first characteristic information of a first object, and the first object belongs to user identification or candidate data;

The first mapping module 1802 is configured to map the first feature information to the target space based on the mapping model to obtain the first mapping vector corresponding to the first object in the target space. The target space includes the user mapping vector corresponding to the user identification and The data mapping vector corresponding to the candidate data;

The recommendation module 1803 is configured to make recommendations based on the first object and the second object according to the distance between any two mapping vectors in the target space, where the second mapping vector is a vector corresponding to the second object in the target space, and the first The distance between the second mapping vector and the first mapping vector is less than the preset distance, and the second mapping vector and the first mapping vector belong to different categories.

Optionally, referring to FIG. 19, the device further includes:

The second information obtaining module 1804 is configured to obtain second characteristic information of the second object;

The second mapping module 1805 is configured to map the second feature information to the target space based on the mapping model to obtain a second mapping vector corresponding to the second object in the target space;

The recommended module 1803 also includes:

The first distance obtaining unit 18031 is configured to obtain the distance between the first mapping vector and the second mapping vector;

The first recommendation unit 18032 is configured to make a recommendation based on the first object and the second object if the distance is less than the preset distance.

Optionally, referring to FIG. 19, the recommendation module 1803 includes:

The vector determining unit 18033 is configured to determine at least one third mapping vector in the target space, where the third mapping vector and the first mapping vector belong to different categories;

The second distance obtaining unit 18034 is configured to obtain the distance between the first mapping vector and each third mapping vector;

The vector selecting unit 18035 is configured to select, from at least one third mapping vector, a second mapping vector whose distance from the first mapping vector is less than a preset distance;

The second recommendation unit 18036 is configured to determine the second object corresponding to the second mapping vector, and make recommendations based on the first object and the second object.

Optionally, the second recommendation unit 18036 is further configured to perform inverse mapping on the second mapping vector based on the inverse mapping model to obtain second feature information corresponding to the second mapping vector, and determine the second object to which the second feature information belongs.

Optionally, referring to FIG. 19, the device further includes:

The first sample acquisition module 1806 is used to acquire sample information, the sample information includes the characteristic information of the sample user identification, the characteristic information of the sample data, and the sample label, and the sample label is used to indicate whether to identify the recommended sample data to the sample user;

The first training module 1807 is used to train the mapping model according to the sample information.

Optionally, the device further includes a loss function for training the mapping model, including at least one of the following:

Among them, L _neck is the first loss value of the mapping model, and λ _margin is the preset parameter,

Is the sample label,

Is the mapping vector corresponding to the sample user ID,

Is the mapping vector corresponding to the sample data;

Among them, L _cov is the second loss value of the mapping model, N is the number of sample information, and E is

with

Optionally, referring to Figure 20, the mapping model is an encoding model in the autoencoder; the device further includes:

The second sample acquisition module 1808 is used to acquire sample information, the sample information includes the characteristic information of the sample user identification, the characteristic information of the sample data, and the sample label, and the sample label is used to indicate whether to identify the recommended sample data to the sample user;

The second training module 1809 is used to train the autoencoder according to the sample information.

Optionally, the device further includes a loss function for training the autoencoder, including at least:

Among them, L _reconstruct is the loss value of the self-encoder,

Characteristic information identified by the sample user or characteristic information of the sample data,

for

Based on the feature information output after processing by the autoencoder.

Optionally, the first object is a user identification and the second object is candidate data, or the first object is candidate data and the second object is a user identification;

The recommendation module 1803 is also used to recommend candidate data to the user identification.

Optionally, the mapping model includes a user mapping sub-model and a data mapping sub-model;

The user mapping sub-model is used to map the characteristic information of the user identification to obtain the user mapping vector;

The data mapping sub-model is used to map the feature information of the candidate data to obtain the data mapping vector.

It should be noted that the recommendation device provided in the above embodiment only uses the division of the above functional modules for illustration. In practical applications, the above functions can be allocated by different functional modules to complete all or all of the above descriptions according to needs. Part of the function. In addition, the recommending device provided in the foregoing embodiment and the recommending method embodiment belong to the same concept, and the implementation process is detailed in the method embodiment, which will not be repeated here.

FIG. 21 is a schematic structural diagram of a terminal 2100 provided by an embodiment of the present application.

Generally, the terminal 2100 includes a processor 2101 and a memory 2102.

The processor 2101 includes one or more processing cores, such as a 4-core processor, an 8-core processor, and so on. The processor 2101 adopts at least one hardware form among DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array, Programmable Logic Array). . The processor 2101 also includes a main processor and a coprocessor. The main processor is a processor used to process data in the awake state, also called a CPU; the coprocessor is used to process data in a standby state. Low-power processor for processing. In some embodiments, the processor 2101 is integrated with a GPU (Graphics Processing Unit, image processing interactor), and the GPU is used to render and draw content that needs to be displayed on the display screen. In some embodiments, the processor 2101 further includes an AI (Artificial Intelligence) processor, and the AI processor is used to process computing operations related to machine learning.

The memory 2102 includes one or more computer-readable storage media, which are non-transitory. The memory 2102 may also include high-speed random access memory and non-volatile memory, such as one or more magnetic disk storage devices and flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 2102 is used to store at least one instruction, and the at least one instruction is used by the processor 2101 to implement the recommended method provided in the method embodiment of the present application. .

In some embodiments, the terminal 2100 may optionally further include: a peripheral device interface 2103 and at least one peripheral device. The processor 2101, the memory 2102, and the peripheral device interface 2103 are connected by a bus or signal line. Each peripheral device is connected to the peripheral device interface 2103 through a bus, a signal line or a circuit board. Optionally, the peripheral device includes: at least one of a radio frequency circuit 2104, a display screen 2105, a camera component 2106, an audio circuit 2107, a positioning component 2108, and a power supply 2109.

The peripheral device interface 2103 may be used to connect at least one peripheral device related to I/O (Input/Output) to the processor 2101 and the memory 2102. In some embodiments, the processor 2101, the memory 2102, and the peripheral device interface 2103 are integrated on the same chip or circuit board; in some other embodiments, any one of the processor 2101, the memory 2102, and the peripheral device interface 2103 or The two are implemented on separate chips or circuit boards, which are not limited in this embodiment.

The radio frequency circuit 2104 is used to receive and transmit RF (Radio Frequency, radio frequency) signals, also called electromagnetic signals. The radio frequency circuit 2104 communicates with a communication network and other communication devices through electromagnetic signals. The radio frequency circuit 2104 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals into electrical signals. Optionally, the radio frequency circuit 2104 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, and so on. The radio frequency circuit 2104 communicates with other terminals through at least one wireless communication protocol. The wireless communication protocol includes, but is not limited to: metropolitan area networks, various generations of mobile communication networks (2G, 3G, 4G, and 8G), wireless local area networks, and/or WiFi (Wireless Fidelity, wireless fidelity) networks. In some embodiments, the radio frequency circuit 2104 also includes a circuit related to NFC (Near Field Communication), which is not limited in this application.

The display screen 2105 is used to display UI (User Interface, user interface). The UI includes graphics, text, icons, videos, and any combination of them. When the display screen 2105 is a touch display screen, the display screen 2105 also has the ability to collect touch signals on or above the surface of the display screen 2105. The touch signal is input to the processor 2101 as a control signal for processing. At this time, the display screen 2105 is also used to provide virtual buttons and/or virtual keyboards, also called soft buttons and/or soft keyboards. In some embodiments, one display screen 2105 is provided with the front panel of the terminal 2100; in other embodiments, there are at least two display screens 2105, which are respectively provided on different surfaces of the terminal 2100 or in a folding design; In the embodiment, the display screen 2105 is a flexible display screen, which is arranged on the curved surface or the folding surface of the terminal 2100. Furthermore, the display screen 2105 can also be set as a non-rectangular irregular pattern, that is, a special-shaped screen. The display screen 2105 is made of materials such as LCD (Liquid Crystal Display) and OLED (Organic Light-Emitting Diode).

The camera assembly 2106 is used to capture images or videos. Optionally, the camera assembly 2106 includes a front camera and a rear camera. Generally, the front camera is set on the front panel of the terminal 2100, and the rear camera is set on the back of the terminal 2100. In some embodiments, there are at least two rear cameras, each of which is a main camera, a depth-of-field camera, a wide-angle camera, and a telephoto camera, so as to realize the fusion of the main camera and the depth-of-field camera to realize the background blur function, the main camera Integrate with the wide-angle camera to realize panoramic shooting and VR (Virtual Reality) shooting functions or other fusion shooting functions. In some embodiments, the camera assembly 2106 also includes a flash. Optionally, the flash is a single-color temperature flash or a dual-color temperature flash. Dual color temperature flash refers to a combination of warm light flash and cold light flash used for light compensation under different color temperatures.

The audio circuit 2107 includes a microphone and a speaker. The microphone is used to collect sound waves of the user and the environment, and convert the sound waves into electrical signals and input them to the processor 2101 for processing, or input to the radio frequency circuit 2104 to implement voice communication. For the purpose of stereo collection or noise reduction, there are multiple microphones, which are respectively set in different parts of the terminal 2100. The microphone is an array microphone or an omnidirectional acquisition microphone. The speaker is used to convert the electrical signal from the processor 2101 or the radio frequency circuit 2104 into sound waves. Optionally, the speaker is a traditional thin-film speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it not only converts the electrical signal into human audible sound waves, but also converts the electrical signal into human inaudible sound waves for purposes such as distance measurement. In some embodiments, the audio circuit 2107 also includes a headphone jack.

The positioning component 2108 is used to locate the current geographic location of the terminal 2100 to implement navigation or LBS (Location Based Service, location-based service). The positioning component 2108 is a positioning component based on the GPS (Global Positioning System, Global Positioning System) of the United States, the Beidou system of China, the Granus system of Russia, or the Galileo system of the European Union.

The power supply 2109 is used to supply power to various components in the terminal 2100. The power source 2109 is alternating current, direct current, disposable batteries or rechargeable batteries. When the power source 2109 includes a rechargeable battery, the rechargeable battery supports wired charging or wireless charging. The rechargeable battery is also used to support fast charging technology.

In some embodiments, the terminal 2100 further includes one or more sensors 2110. The one or more sensors 2110 include, but are not limited to: an acceleration sensor 2111, a gyroscope sensor 2112, a pressure sensor 2113, a fingerprint sensor 2114, an optical sensor 2115, and a proximity sensor 2116.

The acceleration sensor 2111 detects the magnitude of acceleration on the three coordinate axes of the coordinate system established by the terminal 2100. For example, the acceleration sensor 2111 is used to detect the components of gravitational acceleration on three coordinate axes. The processor 2101 controls the display screen 2105 to display the user interface in a horizontal view or a vertical view according to the gravitational acceleration signal collected by the acceleration sensor 2111. The acceleration sensor 2111 is also used for the collection of game or user motion data.

The gyroscope sensor 2112 detects the body direction and rotation angle of the terminal 2100, and the gyroscope sensor 2112 and the acceleration sensor 2111 cooperate to collect the user's 3D actions on the terminal 2100. The processor 2101 implements the following functions based on the data collected by the gyroscope sensor 2112: motion sensing (such as changing the UI according to the user's tilt operation), image stabilization during shooting, game control, and inertial navigation.

The pressure sensor 2113 is arranged on the side frame of the terminal 2100 and/or the lower layer of the display screen 2105. When the pressure sensor 2113 is arranged on the side frame of the terminal 2100, the user's holding signal of the terminal 2100 is detected, and the processor 2101 performs left and right hand recognition or quick operation according to the holding signal collected by the pressure sensor 2113. When the pressure sensor 2113 is arranged on the lower layer of the display screen 2105, the processor 2101 controls the operability controls on the UI interface according to the user's pressure operation on the display screen 2105. The operability control includes at least one of a button control, a scroll bar control, an icon control, and a menu control.

The fingerprint sensor 2114 is used to collect the user's fingerprint. The processor 2101 identifies the user's identity according to the fingerprint collected by the fingerprint sensor 1414, or the fingerprint sensor 2114 identifies the user's identity according to the collected fingerprint. When it is recognized that the user's identity is a trusted identity, the processor 2101 authorizes the user to have related sensitive operations, including unlocking the screen, viewing encrypted information, downloading software, paying, and changing settings. The fingerprint sensor 2114 is provided on the front, back or side of the terminal 2100. When a physical button or manufacturer logo is provided on the terminal 2100, the fingerprint sensor 2114 is integrated with the physical button or manufacturer logo.

The optical sensor 2115 is used to collect the ambient light intensity. In an embodiment, the processor 2101 controls the display brightness of the display screen 2105 according to the ambient light intensity collected by the optical sensor 2115. Optionally, when the ambient light intensity is high, the display brightness of the display screen 2105 is increased; when the ambient light intensity is low, the display brightness of the display screen 2105 is decreased. In another embodiment, the processor 2101 also dynamically adjusts the shooting parameters of the camera assembly 2106 according to the ambient light intensity collected by the optical sensor 2115.

The proximity sensor 2116, also called a distance sensor, is usually arranged on the front panel of the terminal 2100. The proximity sensor 2116 is used to collect the distance between the user and the front of the terminal 2100. In one embodiment, when the proximity sensor 2116 detects that the distance between the user and the front of the terminal 2100 gradually decreases, the processor 2101 controls the display screen 2105 to switch from the on-screen state to the off-screen state; when the proximity sensor 2116 detects When the distance between the user and the front of the terminal 2100 gradually increases, the processor 2101 controls the display screen 2105 to switch from the screen-on state to the screen-on state.

Those skilled in the art understand that the structure shown in FIG. 21 does not constitute a limitation to the terminal 2100, and can also include more or less components than those shown in the figure, or combine some components, or adopt different component arrangements.

FIG. 22 is a schematic structural diagram of a server provided by an embodiment of the present application. The server 2200 may have relatively large differences due to different configurations or performance, including one or more processors (Central Processing Units, CPU) 2201 and one or There are more than one memory 2202, where at least one instruction is stored in the memory 2202, and at least one instruction is loaded and executed by the processor 2201 to implement the methods provided by the foregoing method embodiments. Of course, the server also has components such as a wired or wireless network interface, a keyboard, and an input and output interface for input and output. The server also includes other components for implementing device functions, which will not be repeated here.

The server 2200 is configured to execute the steps executed by the server in the above-mentioned recommendation method.

The embodiment of the present application also provides a computer device, the computer device includes a processor and a memory, at least one piece of program code is stored in the memory, and the at least one piece of program code is loaded and executed by the processor to implement the following steps:

Acquire first characteristic information of the first object, where the first object belongs to user identification or candidate data;

Based on the mapping model, the first feature information is mapped to the target space, and the first mapping vector corresponding to the first object in the target space is obtained. The target space includes the user mapping vector corresponding to the user identification and the data mapping corresponding to the candidate data vector;

According to the distance between any two mapping vectors in the target space, recommend based on the first object and the second object, where the second mapping vector is the vector corresponding to the second object in the target space, and the second mapping vector is the same as the first object. The distance between the mapping vectors is less than the preset distance, and the second mapping vector and the first mapping vector belong to different categories.

Optionally, the at least one piece of program code is loaded and executed by the processor to implement the following steps:

Acquiring second characteristic information of the second object;

Based on the mapping model, map the second feature information to the target space to obtain a second mapping vector corresponding to the second object in the target space;

Obtaining the distance between the first mapping vector and the second mapping vector;

If the distance is less than the preset distance, a recommendation is made based on the first object and the second object.

Determine at least one third mapping vector in the target space, where the third mapping vector and the first mapping vector belong to different categories;

Obtaining the distance between the first mapping vector and each third mapping vector;

From at least one third mapping vector, selecting a second mapping vector whose distance from the first mapping vector is less than a preset distance;

The second object corresponding to the second mapping vector is determined, and recommendations are made based on the first object and the second object.

Based on the inverse mapping model, the second mapping vector is inversely mapped to obtain the second feature information corresponding to the second mapping vector, and the second object to which the second feature information belongs is determined.

Obtain sample information. The sample information includes the characteristic information of the sample user identification, the characteristic information of the sample data, and the sample label. The sample label is used to indicate whether to identify and recommend the sample data to the sample user;

According to the sample information, the mapping model is trained.

Optionally, the loss function used to train the mapping model includes at least one of the following:

Is the sample label,

Is the mapping vector corresponding to the sample user ID,

Is the mapping vector corresponding to the sample data;

with

Optionally, the mapping model is an encoding model in the autoencoder; when the at least one piece of program code is loaded and executed by the processor, the following steps are implemented:

According to the sample information, the autoencoder is trained.

Optionally, the loss function used to train the autoencoder includes at least:

Among them, L _reconstruct is the loss value of the self-encoder,

for

Based on the feature information output after processing by the autoencoder.

Optionally, the first object is a user identification and the second object is candidate data, or the first object is candidate data and the second object is a user identification; the at least one piece of program code is loaded and executed by the processor to achieve The following steps:

Recommend alternative data to the user ID.

Optionally, the mapping model includes a user mapping sub-model and a data mapping sub-model; the user mapping sub-model is used to map the feature information of the user identification to obtain the user mapping vector; the data mapping sub-model is used to map the feature information of the candidate data Perform the mapping to obtain the data mapping vector.

The embodiment of the present application also provides a computer-readable storage medium, in which at least one piece of program code is stored, and the at least one piece of program code is loaded and executed by a processor to implement the following steps:

According to the distance between any two mapping vectors in the target space, recommendation is made based on the first object and the second object. The second mapping vector is the vector corresponding to the second object in the target space. The distance between the mapping vectors is less than the preset distance, and the second mapping vector and the first mapping vector belong to different categories.

Acquiring second characteristic information of the second object;

According to the sample information, the mapping model is trained.

Is the sample label,

Is the mapping vector corresponding to the sample user ID,

Is the mapping vector corresponding to the sample data;

with

According to the sample information, the autoencoder is trained.

Optionally, the loss function used to train the autoencoder includes at least:

Among them, L _reconstruct is the loss value of the self-encoder,

for

Based on the feature information output after processing by the autoencoder.

Recommend alternative data to the user ID.

The embodiment of the present application also provides a computer program in which at least one piece of program code is stored, and the at least one piece of program code is loaded and executed by a processor to implement the following steps:

Acquiring second characteristic information of the second object;

According to the sample information, the mapping model is trained.

Is the sample label,

Is the mapping vector corresponding to the sample user ID,

Is the mapping vector corresponding to the sample data;

with

According to the sample information, the autoencoder is trained.

Optionally, the loss function used to train the autoencoder includes at least:

Among them, L _reconstruct is the loss value of the self-encoder,

for

Based on the feature information output after processing by the autoencoder.

Recommend alternative data to the user ID.

Those of ordinary skill in the art can understand that all or part of the steps in the foregoing embodiments can be implemented by hardware, or by a program to instruct relevant hardware to be completed. The program can be stored in a computer-readable storage medium. The storage medium can be read-only memory, magnetic disk or optical disk, etc.

The foregoing descriptions are only optional embodiments of the embodiments of the present application, and are not intended to limit the embodiments of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the embodiments of the present application shall be It is included in the scope of protection of this application.

Claims

A recommendation method, characterized in that it is applied to a server, and the method includes:

Acquiring first characteristic information of a first object, where the first object belongs to a user identification or candidate data;

Based on the mapping model, the first feature information is mapped to a target space to obtain a first mapping vector corresponding to the first object in the target space, and the target space includes a user mapping vector corresponding to a user identification and The data mapping vector corresponding to the candidate data;

According to the distance between any two mapping vectors in the target space, a recommendation is made based on the first object and the second object, where the second mapping vector is a vector corresponding to the second object in the target space , The distance between the second mapping vector and the first mapping vector is less than a preset distance, and the second mapping vector and the first mapping vector belong to different categories.
The method according to claim 1, wherein the method further comprises:

Acquiring second characteristic information of the second object;

Mapping the second feature information to the target space based on the mapping model to obtain the second mapping vector;

The performing recommendation based on the first object and the second object according to the distance between any two mapping vectors in the target space includes:

Acquiring the distance between the first mapping vector and the second mapping vector;

If the distance is less than the preset distance, a recommendation is made based on the first object and the second object.
The method according to claim 1, wherein the recommending based on the first object and the second object according to the distance between any two mapping vectors in the target space comprises:

Determining at least one third mapping vector in the target space, where the third mapping vector and the first mapping vector belong to different categories;

Acquiring the distance between the first mapping vector and each third mapping vector;

Selecting, from the at least one third mapping vector, a second mapping vector whose distance from the first mapping vector is less than the preset distance;

The second object corresponding to the second mapping vector is determined, and a recommendation is made based on the first object and the second object.
The method according to claim 3, wherein the determining the second object corresponding to the second mapping vector comprises:

Based on the inverse mapping model, perform inverse mapping on the second mapping vector to obtain second feature information corresponding to the second mapping vector, and determine the second object to which the second feature information belongs.
The method according to claim 1, characterized in that, based on the mapping model, the first feature information is mapped to a target space to obtain that the first object is before the corresponding first mapping vector in the target space , The method further includes:

Acquiring sample information, where the sample information includes the characteristic information of the sample user identification, the characteristic information of the sample data, and the sample label, where the sample label is used to indicate whether to recommend the sample data to the sample user identification;

According to the sample information, the mapping model is trained.
The method according to claim 5, wherein the loss function used for training the mapping model comprises at least one of the following:

Wherein, L neck is the first loss value of the mapping model, and λ margin is a preset parameter,
Is the sample label,
Is the mapping vector corresponding to the sample user ID,
Is the mapping vector corresponding to the sample data;

Wherein, L cov is the second loss value of the mapping model, N is the number of sample information, and E is the
And said
The formed matrix, Cov(E) is the covariance matrix of matrix E, ||·|| f is the transposition function, and diag(·) is the diagonal element extraction function of the matrix.
The method according to claim 1, wherein the mapping model is an encoding model in an autoencoder;

Before the mapping the first feature information to the target space based on the mapping model to obtain the first mapping vector corresponding to the first object in the target space, the method further includes:

Acquiring sample information, where the sample information includes the characteristic information of the sample user identification, the characteristic information of the sample data, and the sample label, where the sample label is used to indicate whether to recommend the sample data to the sample user identification;

According to the sample information, the autoencoder is trained.
The method according to claim 7, wherein the loss function used for training the autoencoder at least comprises:

Among them, L reconstruct is the loss value of the self-encoder,
Is the characteristic information identified by the sample user or the characteristic information of the sample data,
As said
Based on the feature information output after processing by the self-encoder.
The method according to any one of claims 1-8, wherein the first object is a user ID, the second object is candidate data, or the first object is candidate data, so The second object is a user ID;

The recommending based on the first object and the second object includes: recommending the candidate data to the user identifier.
The method according to claim 1, wherein the mapping model includes a user mapping sub-model and a data mapping sub-model;

The user mapping sub-model is used to map the characteristic information of the user identification to obtain a user mapping vector;

The data mapping sub-model is used to map the feature information of the candidate data to obtain a data mapping vector.
A recommendation device, characterized in that the device comprises:

A first information acquisition module, configured to acquire first characteristic information of a first object, the first object belonging to a user identification or candidate data;

The first mapping module is configured to map the first feature information to a target space based on a mapping model to obtain a first mapping vector corresponding to the first object in the target space, and the target space includes a user Identify the corresponding user mapping vector and the data mapping vector corresponding to the candidate data;

The recommendation module is configured to make a recommendation based on the first object and the second object according to the distance between any two mapping vectors in the target space, where the second mapping vector is that the second object is in the target space. For a corresponding vector in the space, the distance between the second mapping vector and the first mapping vector is less than a preset distance, and the second mapping vector and the first mapping vector belong to different categories.
The device according to claim 11, wherein the device further comprises:

A second information acquiring module, configured to acquire second characteristic information of the second object;

A second mapping module, configured to map the second feature information to the target space based on the mapping model to obtain the second mapping vector;

Recommended modules also include:

A first distance acquiring unit, configured to acquire the distance between the first mapping vector and the second mapping vector;

The first recommendation unit is configured to make a recommendation based on the first object and the second object if the distance is less than the preset distance.
The device according to claim 11, wherein the recommendation module comprises:

A vector determining unit, configured to determine at least one third mapping vector in the target space, where the third mapping vector and the first mapping vector belong to different categories;

A second distance acquiring unit, configured to acquire the distance between the first mapping vector and each third mapping vector;

A vector selecting unit, configured to select, from the at least one third mapping vector, a second mapping vector whose distance from the first mapping vector is less than the preset distance;

The second recommendation unit is configured to determine the second object corresponding to the second mapping vector, and perform recommendation based on the first object and the second object.
The device according to claim 13, wherein the second recommendation unit is further configured to perform inverse mapping on the second mapping vector based on an inverse mapping model to obtain a second feature corresponding to the second mapping vector Information, determining the second object to which the second characteristic information belongs.
The device according to claim 11, wherein the device further comprises:

The first sample acquisition module is used to acquire sample information. The sample information includes the characteristic information of the sample user identification, the characteristic information of the sample data, and the sample label. Describe the sample data;

The first training module is used to train the mapping model according to the sample information.
The device according to claim 15, wherein the device further comprises a loss function for training the mapping model, comprising at least one of the following:

Wherein, L neck is the first loss value of the mapping model, and λ margin is a preset parameter,
Is the sample label,
Is the mapping vector corresponding to the sample user ID,
Is the mapping vector corresponding to the sample data;

Wherein, L cov is the second loss value of the mapping model, N is the number of sample information, and E is the
And said
The formed matrix, Cov(E) is the covariance matrix of matrix E, ||·|| f is the transposition function, and diag(·) is the diagonal element extraction function of the matrix.
The device according to claim 11, wherein the mapping model is an encoding model in a self-encoder, and the device further comprises:

The second sample acquisition module is used to acquire sample information, the sample information includes the characteristic information of the sample user identification, the characteristic information of the sample data and the sample label, and the sample label is used to indicate whether to recommend the sample user identification to the sample user sample;

The second training module is used to train the autoencoder according to the sample information.
The device according to claim 17, wherein the device further comprises a loss function for training the autoencoder, at least comprising:

Among them, L reconstruct is the loss value of the self-encoder,
Is the characteristic information identified by the sample user or the characteristic information of the sample data,
As said
Based on the feature information output after processing by the self-encoder.
The device according to any one of claims 11-18, wherein the first object is a user identifier, and the second object is candidate data, or the first object is candidate data, and the second object Is the user ID;

The recommendation module is further configured to recommend the candidate data to the user identifier.
A computer device, wherein the computer device includes a processor and a memory, and at least one piece of program code is stored in the memory, and the at least one piece of program code is loaded and executed by the processor to implement The operation performed in the recommendation method according to any one of claims 1 to 10.
A computer-readable storage medium, wherein at least one piece of program code is stored in the computer-readable storage medium, and the at least one piece of program code is loaded and executed by a processor to implement any one of claims 1 to 10 The operation performed in the recommendation method described in the claim.