CN114912015A

CN114912015A - Object recommendation method, model training method, device, equipment and medium

Info

Publication number: CN114912015A
Application number: CN202210334475.9A
Authority: CN
Inventors: 马晓云; 沈磊; 王兵
Original assignee: Alibaba China Co Ltd
Current assignee: Hangzhou Alibaba Overseas Internet Industry Co ltd
Priority date: 2022-03-30
Filing date: 2022-03-30
Publication date: 2022-08-16

Abstract

The application provides an object recommendation method, a model training device, an object recommendation device and a medium, wherein the object recommendation method comprises the following steps: in the object recommendation model, determining a feature vector corresponding to a candidate object according to attribute information of a user and attribute information of the candidate object, wherein the attribute information of the user comprises a target region to which the user belongs, and the attribute information of the candidate object comprises statistical data corresponding to the candidate object in a plurality of regions; in the object recommendation model, based on the characteristic vectors, the candidate objects are scored through a scoring network in the multitask network and network branches connected with the scoring network to obtain target scores of the candidate objects, and different network branches in the multitask network correspond to different regions. Therefore, in consideration of differences of user consumption levels, preferences and the like in different regions, by paying attention to the regions of the users and the statistical data corresponding to the objects in the plurality of regions, the accuracy of object recommendation is improved through the multitask network.

Description

Object recommendation method, model training method, device, equipment and medium

Technical Field

The present application relates to the field of data processing technologies, and in particular, to an object recommendation method, a model training method, an apparatus, a device, and a medium.

Background

With the development of internet technology, internet merchants provide massive commodities and services for users, and user requirements are greatly met. How to accurately recommend interested objects for users in massive commodities and services is one of the challenges facing today.

Generally, the access frequency, the deal frequency, and the like of the object are counted, and based on the access frequency and the deal frequency of the object, the object that the user may be interested in is selected from a large number of objects, for example, the object with the highest access frequency is recommended to the user, and the object with the highest deal frequency is recommended to the user. And selecting objects which are possibly interested by the user from the massive objects according to the historical record data of the user, such as the commodities or services visited by the user, for example, determining the commodities with higher similarity to the commodities visited by the user from the massive objects and recommending the commodities to the user.

However, the object recommendation method described above has low recommendation accuracy.

Disclosure of Invention

The application provides an object recommendation method, a model training device, equipment and a medium. The method is used for solving the problem of low accuracy of object recommendation.

In a first aspect, an embodiment of the present application provides an object recommendation method, including: in the object recommendation model, determining a feature vector corresponding to a candidate object according to attribute information of a user and attribute information of the candidate object, wherein the attribute information of the user comprises a target region to which the user belongs, and the attribute information of the candidate object comprises statistical data corresponding to the candidate object under a plurality of regions respectively; in the object recommendation model, based on the characteristic vectors, scoring is carried out on candidate objects through a scoring network in a multitask network and network branches connected with the scoring network to obtain target scores of the candidate objects, wherein in the multitask network, the number of the network branches is multiple, and different network branches correspond to different regions; and recommending the target object in the candidate objects to the user according to the target scores of the candidate objects.

In a second aspect, an embodiment of the present application provides a model training method, including: determining a training data set, wherein the training data set comprises a plurality of training samples and labels of the training samples, the training samples comprise attribute information of users and attribute information of objects, the attribute information of the users comprises target regions to which the users belong, the attribute information of the objects comprises statistical data corresponding to the objects in the plurality of regions, and the labels of the training samples are used for marking whether the users are interested in the objects in the training samples; and carrying out supervised training on an object recommendation model according to the training data set, wherein the object recommendation model comprises a multitask network for scoring the objects, the multitask network comprises a scoring network and a plurality of network branches connected with the scoring network, and different network branches correspond to different regions.

In a third aspect, an embodiment of the present application provides an object recommendation apparatus, including: the characteristic determining unit is used for determining a characteristic vector corresponding to the candidate object according to attribute information of a user and attribute information of the candidate object in the object recommendation model, wherein the attribute information of the user comprises a target region to which the user belongs, and the attribute information of the candidate object comprises statistical data corresponding to the candidate object under a plurality of regions respectively; the object scoring unit is used for scoring the candidate objects through a scoring network in the multitask network and network branches connected with the scoring network based on the characteristic vectors in the object recommendation model to obtain target scores of the candidate objects, wherein the number of the network branches is multiple in the multitask network, and different network branches correspond to different regions; and the object recommending unit is used for recommending a target object in the candidate objects to the user according to the target scores of the candidate objects.

In a fourth aspect, an embodiment of the present application provides a model training apparatus, including: the data determining unit is used for determining a training data set, the training data set comprises a plurality of training samples and labels of the training samples, the training samples comprise attribute information of users and attribute information of objects, the attribute information of the users comprises target regions to which the users belong, the attribute information of the objects comprises statistical data corresponding to the objects in the plurality of regions, and the labels of the training samples are used for marking whether the users are interested in the objects in the training samples; the training unit is used for carrying out supervised training on an object recommendation model according to a training data set, wherein the object recommendation model comprises a multitask network for scoring the object, the multitask network comprises a scoring network and network branches connected with the scoring network, the network branches are multiple, and different network branches correspond to different regions.

In a fifth aspect, an embodiment of the present application provides a cloud server, including: at least one processor; and a memory communicatively coupled to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the cloud server to execute the object recommendation method provided by the first aspect of the embodiment of the present application and/or to enable the cloud server to execute the model training method provided by the second aspect of the embodiment of the present application.

In a sixth aspect, embodiments of the present application provide a computer-readable storage medium, on which a computer program is stored, where the computer program, when executed by a processor, implements the object recommendation method provided in the first aspect of the embodiments of the present application, and/or where the computer program, when executed by the processor, implements the model training method provided in the second aspect of the embodiments of the present application.

In the embodiment of the application, the object recommendation model comprises a multitask network, the multitask network comprises a scoring network and network branches connected with the scoring network, the number of the network branches in the multitask network is multiple, and different network branches correspond to different regions. The process of object recommendation by using the object recommendation model comprises the following steps: in an object recommendation model, determining a feature vector corresponding to a candidate object according to attribute information of a user and attribute information of the candidate object, wherein the attribute information of the user comprises a target region to which the user belongs, and the attribute information of the candidate object comprises statistical data corresponding to the candidate object under a plurality of regions respectively; then, based on the characteristic vectors corresponding to the candidate objects, scoring the candidate objects through a scoring network and network branches connected with the scoring network in a multitask network included in the object recommendation model to obtain target scores of the candidate objects; and finally, recommending the target object in the candidate objects to the user according to the target scores of the candidate objects.

Therefore, in the process of characteristic design, the region to which the user belongs and the statistical data corresponding to the distribution of the object in a plurality of regions are concerned; in the model structure of the object recommendation model, a multitask model is utilized to provide different network branches for different regions. Therefore, the embodiment of the application starts from feature design and a model structure, the condition that the users in different regions have differences is fully considered, the object recommendation model is adopted, the object recommendation accuracy is effectively improved, and the user experience is improved.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the description below are some embodiments of the present application, and those skilled in the art can obtain other drawings based on the drawings without inventive labor.

Fig. 1 is a diagram illustrating an application scenario provided in an embodiment of the present application;

fig. 2 is a first schematic structural diagram of an object recommendation model provided in an embodiment of the present application;

fig. 3 is a first flowchart of an object recommendation method according to an embodiment of the present application;

fig. 4 is a flowchart illustrating a second object recommendation method according to an embodiment of the present application;

fig. 5 is a schematic structural diagram of an object recommendation model provided in the embodiment of the present application;

fig. 6 is a schematic structural diagram of an object recommendation model provided in the embodiment of the present application;

fig. 7 is a schematic structural diagram of an object recommendation model provided in the embodiment of the present application;

fig. 8 is a third schematic flowchart of an object recommendation method according to an embodiment of the present application;

FIG. 9 is a schematic flowchart of a model training method according to an embodiment of the present disclosure;

fig. 10 is a block diagram illustrating an object recommendation apparatus according to an embodiment of the present application;

fig. 11 is a block diagram of a model training apparatus according to an embodiment of the present disclosure;

fig. 12 is a schematic hardware structure diagram of a cloud server according to an embodiment of the present application.

With the above figures, there are shown specific embodiments of the present application, which will be described in more detail below. These drawings and written description are not intended to limit the scope of the inventive concepts in any manner, but rather to illustrate the inventive concepts to those skilled in the art by reference to specific embodiments.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are some embodiments of the present application, but not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.

The terms "first," "second," "third," and the like in the description and in the claims of the present application and in the above-described drawings are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the application described herein are capable of operation in sequences other than those illustrated or described herein.

Under the influence of factors such as population cardinality, policy, culture, economic level, logistics and the like, the consumption levels and commodity preferences of people in different areas are different, the user traffic distribution in different areas is uneven (for example, the number of visiting users in a certain area per day is 1 ten thousand, the number of visiting users in another area per day is 4 ten thousand), and the concerned business targets and the index values of the business targets in different areas are different. The existence of these differences presents a significant challenge to the object recommendation algorithm.

Taking an object as a commodity as an example, aiming at commodity recommendation, the following scheme can be adopted:

one scheme is as follows: to model a set of recommended models covering users globally (i.e., in all regions), the differences between different regions are not modeled efficiently. In the scheme, areas with more independent visitors (UV) (namely areas with larger user quantity) dominate the statistic characteristics of the commodity, and the statistic characteristics of the commodity are important dependence characteristics of the recommendation model, so that the areas with more UV dominate the recommendation result, and the user requirements of the areas with less UV are difficult to meet and express. In the past, the commodity recommendation can be seriously affected by Martian effect, and the commodity recommendation accuracy is low.

The other scheme is as follows: and (4) carrying out independent model training and model deployment aiming at different regions. Specifically, a recommendation model is independently designed for different regions, the model is trained by using the statistic characteristics of the commodities in the regions, and finally the model is deployed. However, in the actual training, the recommended models corresponding to the regions of the partial regions face the problems of insufficient training samples and sparse features, and the offline training and online validation of the models in the method are complicated.

The other scheme is as follows: the same recommendation model is adopted in different regions, but the recommendation model is optimized (namely model training) independently aiming at different regions, namely the recommendation model is optimized in a different region. However, in this scheme, the UV-rich region model is optimized sufficiently, and the UV-poor region model is optimized insufficiently.

In order to solve the above problems, embodiments of the present application provide an object recommendation method, a model training method, an apparatus, a device, and a medium. According to the object recommendation method, the object recommendation model comprises a multitask network, the multitask network comprises a scoring network and network branches connected with the scoring network, the number of the network branches in the multitask network is multiple, and different network branches correspond to different regions. In the object recommendation model: determining a feature vector corresponding to a candidate object according to attribute information of a user and attribute information of the candidate object, wherein the attribute information of the user comprises a target region to which the user belongs, and the attribute information of the candidate object comprises statistical data corresponding to the candidate object under a plurality of regions respectively; and based on the characteristic vectors, scoring the candidate objects through a scoring network in the multitask network and a network branch connected with the scoring network to obtain target scores of the candidate objects. And finally, recommending the target object in the candidate objects to the user according to the target scores of the candidate objects.

Therefore, on the basis that a set of recommendation algorithm (namely a set of object recommendation model) covers the global user, the multi-task network comprising the scoring network and the network branches corresponding to the regions is designed in the object recommendation model, so that the scoring network is not influenced by insufficient samples and sparse features, the network branches can learn the user features and the object features of different regions in a targeted manner based on the feature vectors, and the accuracy of the object recommendation performed by the object recommendation model is effectively improved.

Exemplarily, an object is taken as an example of a commodity, and fig. 1 is an exemplary diagram of an application scenario provided according to an embodiment of the present application. As shown in fig. 1, in this application scenario, when a user terminal browses a product page, an object recommendation device (fig. 1 takes the object recommendation device as an example of a server) scores products by using a deployed object recommendation model, and recommends recommended products to the user terminal based on the scores of the products.

Referring to fig. 2, fig. 2 is a first structural diagram of an object recommendation model according to an embodiment of the present application. As shown in fig. 2, in the structure of the object recommendation model, the object recommendation model includes an input layer and a multitask network connected to the input layer, the multitask network includes a scoring network and a plurality of network branches, the scoring network is connected to the network branches, and different network branches correspond to different regions, such as the network branch corresponding to region a and the network branch corresponding to region B in fig. 2. The input layer is used for inputting attribute information of a user and attribute information of an object and outputting a feature vector corresponding to the object, and the multitask network is used for scoring the object based on the feature vector corresponding to the object to obtain a score of the object.

Optionally, in the object recommendation model, the input layer may be a feature embedding layer, so as to perform feature embedding on the model input data.

Based on the object recommendation model shown in fig. 2, in the process of scoring the commodities by using the object recommendation model in fig. 1, the commodities can be scored through a scoring network in the multitask network and a network branch connected with the scoring network based on the attribute information of the user and the attribute information of the commodities in the object recommendation model, so that the scores of the commodities are obtained.

The object recommending device can acquire attribute information of the user from the user terminal, also can acquire a user ID from the user terminal, and then acquire the attribute information of the user based on the user ID, wherein the attribute information of the user comprises a region to which the user belongs; the attribute information of the product may be stored in the object recommending apparatus in advance, or may be acquired from another device.

In the following, the technical solution of the present application is described in detail by specific embodiments in conjunction with the application scenario shown in fig. 1. It should be noted that the following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

It should be noted that the execution subject of the embodiment of the present application may be an electronic device, and the electronic device may be a terminal or a server. The terminal may be a Personal Digital Assistant (PDA) device, a handheld device with a wireless communication function (e.g., a smart phone or a tablet), a computing device (e.g., a Personal Computer (PC)), an in-vehicle device, a wearable device (e.g., a smart watch or a smart band), a smart home device (e.g., a smart display device), and the like. The server may be a single server, a server cluster, a distributed server, a centralized server, or a cloud server.

The above is only an exemplary application scenario. The method and the device for recommending the objects in the Internet can be applied to any object recommending scene in the Internet, such as a merchant recommending scene, a logistics service recommending scene and the like.

Referring to fig. 3, fig. 3 is a first schematic flowchart of an object recommendation method according to an embodiment of the present application. As shown in fig. 3, the object recommendation method provided in the embodiment of the present application specifically includes the following steps:

s301, in the object recommendation model, determining a feature vector corresponding to the candidate object according to attribute information of a user and attribute information of the candidate object, wherein the attribute information of the user comprises a target region to which the user belongs, and the attribute information of the candidate object comprises statistical data corresponding to the candidate object in a plurality of regions.

The object recommendation model is a pre-trained deep learning network used for scoring the object. The structure of the object recommendation model may refer to fig. 2, and specifically, the object recommendation model multitask network includes a scoring network and a plurality of network branches, the plurality of network branches may be a shallow network, such as a Multi Layer Perceptron (Multi Layer Perceptron), and different network branches correspond to different regions. In the object recommendation model, a scoring network is used for scoring the object, and a plurality of network branches are used for performing regional optimization based on the scoring of the scoring network so as to improve the scoring accuracy of the object.

The multitasking network may be called a multi-task learning (multi-task learning) network, and may utilize the commonality of multiple related learning tasks to accomplish the multiple related learning tasks. In an embodiment of the application, the multiple learning tasks to be performed by the multitask network include providing object scores for users in different regions in a targeted manner, wherein the scoring network is used for learning the communicated characteristics of the objects in the multiple regions, and the multiple network branches are used for learning the uniqueness of the objects in the various regions.

Wherein, the region can be a city, a country, a continent, etc. The target area to which the user belongs can be directly obtained from the user terminal; or, the ID of the user may be acquired from the user terminal, and the user ID may be queried from the user database; alternatively, the device identifier of the user terminal may be obtained, the target area to which the user belongs may be determined based on the device identifier, for example, the device identifier is an IP address, and the target area to which the user belongs may be determined based on the IP address of the user terminal.

Optionally, in addition to the target region to which the user belongs, the attribute information of the user may further include interaction data of the user and an object in a historical time period, where the interaction data may include at least one of: the method comprises the following steps of browsing objects by a user, placing orders by the user, collecting objects by the user, commenting objects by the user and buybacking objects by the user. In addition, the attribute information of the user may further include basic information of the user, such as an age interval to which the user belongs, a gender of the user, and the like. The historical time period is, for example, the past week, the past ten days, or the past month, etc.

The candidate objects can be all or part of the objects in the massive object set; or, to improve the accuracy and efficiency of object recommendation, the candidate object may be an object interacted by the user in the historical time period, for example, an object browsed by the user in the historical time period, an object placed an order, and the like; alternatively, to improve the accuracy and efficiency of object recommendation, the candidate object may be an object interacted by a plurality of users in the target region in a historical time period.

The statistical data corresponding to the candidate object in the plurality of regions may include statistical data of the regions of the candidate object in the feature dimension, and the feature dimension may include at least one of: exposure dimension, click dimension, deal dimension, collection dimension, add shopping list dimension, profit dimension. By way of example, the statistical data of the regions of the candidate object in the exposure dimension includes: the number of times that the candidate is exposed to the users in the area a, the number of times that the candidate is exposed to the users in the area B, etc.; the statistical data of the regions of the candidate objects under the click dimension comprise: the number of times that the candidate object is clicked by the user in the area A and the number of times that the candidate object is clicked by the user in the area B; the remaining dimensions are not given here by way of example.

Optionally, the attribute information of the candidate object further includes identification information of the candidate object to distinguish different candidate objects.

Optionally, the attribute information of the candidate object further includes total statistics of the candidate object in all regions, for example, the sum of the exposure times of the candidate object in all regions, the sum of the click times of the candidate object in all regions, and the like. Therefore, the accuracy of object recommendation is improved by providing the statistical data of the candidate object in all regions and the statistical data of the regions.

In this embodiment, the attribute information of the user may be determined in response to an object recommendation request from the user terminal (for example, when the user browses a product recommendation page, the user terminal sends the object recommendation request to the object recommendation device). And responding to the object recommendation request, and acquiring the attribute information of the candidate object from a pre-collected object database. Thereafter, the attribute information of the user and the attribute information of the candidate object may be input into the object recommendation model. In the object recommendation model, the attribute information of the user and the attribute information of the candidate object are encoded to obtain a feature vector corresponding to the candidate object. If the candidate objects are multiple, for each candidate object, the attribute information of the user and the attribute information of the candidate object can be input into the object recommendation model, and the attribute information of the user and the attribute information of the candidate object are encoded in the object recommendation model to obtain the feature vector corresponding to the candidate object. In this way, a feature vector corresponding to each candidate object is obtained.

Optionally, when the object recommendation model includes the feature embedding layer, the attribute information of the user and the attribute information of the candidate object may be input to the feature embedding layer, and feature vectors associated with the candidate object are obtained by performing feature embedding processing on the attribute information of the user and the attribute information of the candidate object in the feature embedding layer. Therefore, the feature learning effect of the subsequent multi-task network is improved by reducing the dimension of the attribute information of the user and the attribute information of the candidate object into a feature vector with fixed dimension through feature embedding.

S302, in the object recommendation model, based on the feature vectors corresponding to the candidate objects, scoring the candidate objects through a scoring network in the multitask network and network branches connected with the scoring network to obtain target scores of the candidate objects.

The network branches connected with the scoring object comprise network branches corresponding to the target region.

In this embodiment, the feature vector associated with the candidate object is input into the multitask network, and feature processing is performed on the feature vector through the scoring network and the network branch connected to the scoring network, so as to obtain a target score of the candidate object.

In the process of carrying out feature processing on the feature vector through a scoring network and a network branch connected with the scoring network, the feature vector can be subjected to feature processing through the scoring network and the network branch corresponding to the target area, and then, the score obtained through the network branch corresponding to the target area is determined as the target score of the candidate object; alternatively, the feature vector may be subjected to feature processing by a scoring network and a plurality of network branches, and then the scores obtained by the plurality of network branches may be subjected to weighting processing, such as weighted summation, to obtain a target score of the candidate object, and in the weighting processing, a relatively high weight may be corresponding to the network branch corresponding to the target region, so that the target score is mainly influenced by the learning effect of the network branch.

S303, recommending a target object in the candidate objects to the user according to the target scores of the candidate objects.

In this embodiment, after the target scores of the candidate objects are obtained, the target objects in the candidate objects may be recommended to the user in a manner that the target scores are from high to low. Specifically, in one mode, the target object can be determined to be a candidate object with a target score larger than a score threshold value, and the target object is recommended to the user; in yet another way, the candidate objects may be sorted in order of the target scores from high to low, and a preset number of target objects may be selected from the sorted candidate objects. After the target object is determined, the related information (such as pictures, prices and the like) of the target object can be sent to the terminal of the user according to the mode that the target score is from high to low, and the related information of the target object is displayed on a display screen by the terminal of the user, so that object recommendation is realized.

In the embodiment of the application, a target region to which a user belongs is concerned, data statistics of the candidate object is carried out in a regional mode, in an object recommendation model, a multi-task network comprising a scoring network and a plurality of network branches is designed, and the commonality and the difference of feature expressions of the candidate object in the plurality of regions are learned through the scoring network and the plurality of network branches in the multi-task network. Therefore, the object recommendation model can be applied to all regions, and users in all regions can provide scores of objects in a targeted manner, namely, the objects are recommended in a targeted manner, and the problem of poor model effect caused by less training samples in the regions with less UV is solved. Therefore, the object recommendation accuracy of each region is effectively improved.

Referring to fig. 4, fig. 4 is a schematic flowchart diagram of a second object recommendation method provided in the embodiment of the present application. As shown in fig. 4, the object recommendation method provided in the embodiment of the present application specifically includes the following steps:

s401, in the object recommendation model, determining a feature vector corresponding to a candidate object according to attribute information of a user and attribute information of the candidate object, wherein the attribute information of the user comprises a target region to which the user belongs, and the attribute information of the candidate object comprises statistical data corresponding to the candidate object in a plurality of regions.

The implementation principle and the technical effect of S401 may refer to the foregoing embodiments, and are not described again.

S402, inputting the feature vector corresponding to the candidate object into a scoring network in the multitask network, and scoring the candidate object in the scoring network based on the feature vector to obtain a score output by the scoring network.

The scoring network may also be referred to as an expert network, and the scoring network may include a plurality of network layers, and further, a network layer in the scoring network may be a fully connected network layer. The input data of the scoring network is a feature vector corresponding to the candidate object, and the output data of the scoring network is a score corresponding to the feature vector.

In this embodiment, the feature vectors corresponding to the candidate objects are input to a scoring network, and in the scoring network, the feature vectors are subjected to feature processing of a plurality of network layers, for example, convolution processing. Finally, a score output by the scoring network is obtained, and the score can be regarded as an initial score of the candidate object.

And S403, inputting the score output by the scoring network into a network branch, and optimizing the score output by the scoring network in the network branch to obtain a target score.

Wherein, the scoring network can be connected with one or more network branches and provides scores for the one or more network branches. In a multitasking network, different network branches correspond to different regions, and the network branch corresponding to one region is obtained through training based on the difference between the preference condition of an object by a user in the region and the score of the object by an object recommendation model.

In this embodiment, the score output by the scoring network is input to the network branch connected to the scoring network, and the score output by the scoring network is optimized through a plurality of network layers in the network branch. For example, the score after the optimization processing of the network branch corresponding to the region a is the score of the candidate object in the region a, and the score after the optimization processing of the network branch corresponding to the region B is the score of the candidate object in the region B. And finally, obtaining the target score of the candidate object based on the score after the network branch optimization processing.

In a possible model structure, referring to fig. 5, fig. 5 is a schematic structural diagram of an object recommendation model provided in an embodiment of the present application, where a scoring network is multiple, and different network branches share all scoring networks, in other words, each network branch is connected to all scoring networks. In the object recommendation process, the scoring network is equivalent to a panel maker, the scoring accuracy of a plurality of panel makers is obviously higher than that of one scoring, therefore, a plurality of scoring networks connected with all network branches are arranged in the object recommendation model, the panel maker is equivalent to providing a plurality of panel makers for all regions, and different network branches share all the scoring networks, which shows that the scoring networks can be fully trained based on samples of all the regions, and the problems of insufficient samples and sparse features do not exist, so that the model structure can effectively improve the recommendation accuracy of the object recommendation model.

Based on the object recommendation model shown in fig. 5, in one possible implementation, as shown in fig. 4, S402 includes: s4021, respectively inputting the eigenvectors corresponding to the candidate objects into each scoring network in the multitask network, and scoring the candidate objects in each scoring network based on the eigenvectors to obtain the scores output by each scoring network. S403 comprises: s4031, for each network branch, all scores output by the scoring network are input to the network branch, and the scores of all scoring networks are fused in the network branch to obtain the scores output by the network branch; and S4032, obtaining a target score according to the score output by the network branch. Therefore, the characteristic expressions of the candidate objects in the whole area are learned by using the plurality of scoring networks, and the characteristic expressions of the candidate objects in each area are learned through the network branches corresponding to the areas, so that the object recommendation accuracy is improved.

The scores of all scoring networks are merged in the network branch, for example, all scoring networks are weighted and adjusted in the network branch according to the network parameters in the network branch.

Optionally, based on any model result of the object recommendation model, for example, based on the object recommendation model shown in fig. 5, as shown in fig. 5, the network branch includes a weighting network and an optimization network, the weighting network is used to fuse scores from multiple scoring networks, and the optimization network is used to further optimize the scores fused by the weighting network, so as to improve the scoring accuracy of the object recommendation model by way of weighting and further optimization.

In still another possible model structure, referring to fig. 6, fig. 6 is a schematic structural diagram of an object recommendation model provided in the embodiment of the present application, and the multitasking network further includes a mask network, where the mask network is configured to filter output data of multiple network branches, so as to filter, from the output data of the multiple network branches, output data of a network branch corresponding to a target region, that is, to mask output data of other network branches. The network parameters of the mask network are determined by the target region to which the user belongs.

Based on the object recommendation model shown in fig. 6, in one possible implementation, S4032 includes: initializing a mask network according to the target area; in the multitask network, inputting the scores output by all network branches into an initialized mask network, and screening the scores output by the target network branches corresponding to a target area from the initialized mask network; and determining a target score according to the score output by the target network branch.

For the sake of brevity, the network branch corresponding to the target area is referred to as a target network branch.

In this implementation, the network parameters of the mask network include network parameters corresponding to each network branch. In the process of initializing the mask network, in the network parameters of the mask network, the network parameters corresponding to the target network branch are initialized to 1, the network parameters corresponding to the remaining network branches are initialized to 0, and then the network parameters in the mask network and the scores output by the corresponding network branches are weighted to screen the scores output by the target network branches. The target score of the candidate object may then be determined as the score of the target network branch output. Therefore, in the implementation manner, under the condition that the users belong to the target region and the interests of the users in different regions for the object are different, the scores of the candidate objects in the target region are obtained through the network structure, so that the scoring accuracy is improved, and the object recommendation accuracy is further improved.

S404, recommending a target object in the candidate objects to the user according to the target scores of the candidate objects.

The implementation principle and the technical effect of S404 may refer to the foregoing embodiments, and are not described again.

In the embodiment of the application, a target region to which a user belongs is concerned, data statistics of the candidate object is carried out in a regional mode, in an object recommendation model, a multitask network comprising a scoring network and a plurality of network branches is designed, the candidate object is scored through the scoring network in the multitask network, and scores output by the scoring network are optimized in a regional mode through the network branches. Therefore, a set of recommendation algorithm is suitable for all regions, and the object scores are provided for users in all regions in a targeted manner, namely the compatibility of the object recommendation model to the region difference is realized, the scoring network can be fully trained under a global sample, fewer samples are needed for network branch training of a shallow network, and the problem that the model effect is poor due to fewer training samples in the regions with less UV is solved. Therefore, the object recommendation accuracy of each region is effectively improved.

In still another possible model structure, referring to fig. 7, fig. 7 is a schematic structural diagram of an object recommendation model provided in this embodiment of the present application, where there are multiple multitasking networks, and different multitasking networks correspond to different business objectives, for example, business objectives related to access, deal and profit, respectively. Each of the multitasking networks includes a scoring network and a plurality of network branches, and in each of the multitasking networks, different network branches correspond to different regions. Through the multitask network, the target score of the object under the service target corresponding to the multitask network can be obtained.

Wherein the structure of the different multitasking networks may be the same.

Based on the object recommendation model shown in fig. 7, referring to fig. 8, fig. 8 is a third schematic flowchart of the object recommendation method provided in the embodiment of the present application. The object recommendation method provided by the embodiment of the application specifically comprises the following steps:

s801, in the object recommendation model, determining a feature vector corresponding to a candidate object according to attribute information of a user and attribute information of the candidate object, wherein the attribute information of the user comprises a target region to which the user belongs, and the attribute information of the candidate object comprises statistical data corresponding to the candidate object in a plurality of regions.

The implementation principle and technical effect of S801 may refer to the foregoing embodiments, and are not described again.

S802, in the object recommendation model, the feature vectors are input into each multitask network, and based on the feature vectors, the candidate objects are scored through a scoring network in the multitask network and a network branch connected with the scoring network in the multitask network, so that target scores of the candidate objects under the service targets corresponding to the multitask network are obtained.

In this embodiment, the feature vectors corresponding to the candidate objects are respectively input into each multitask network of the object recommendation model, and in each multitask network, the feature vectors are processed through a scoring network and a plurality of network branches, so that the score of the candidate objects is achieved, and the scores respectively output by each multitask network are obtained, that is, the target scores of the candidate objects under the service targets corresponding to each multitask network are obtained. The process of scoring the candidate object through the scoring network and the plurality of network branches in the multitasking network may refer to the foregoing embodiments, and details are not described here.

And S803, determining the final target score of the candidate object according to the target scores of the candidate object under the service targets corresponding to the multi-task networks.

In this embodiment, the target scores of the candidate objects may be determined to include target scores of the candidate objects under the service targets corresponding to the respective multitask networks, or the target scores of the candidate objects under the service targets corresponding to the respective multitask networks may be fused, for example, operation may be performed through some preset operation formulas to obtain final target scores of the candidate objects.

Optionally, the service targets include Click-Through-Rate (CTR) and Conversion Rate (Conversion Rate), the statistical data of the candidate objects in the multiple regions includes the number of Click users of the candidate objects in the multiple regions (for example, the number of Click users of the candidate objects in region a and the number of Click users of the candidate objects in region B) and the number of transaction users of the candidate objects in the multiple regions, and as shown in fig. 7, the multitask network includes a first multitask network corresponding to the Click-Through Rate and a second multitask network corresponding to the Conversion Rate. At this time, the candidate objects can be respectively scored at the click rate and the conversion rate by using the first multitask network and the second multitask network based on the number of click users and the number of transaction users of the candidate objects in a plurality of regions.

The click rate and the conversion rate are two basic service targets, and other service targets can be obtained through conversion based on the click rate and the conversion rate, so that a first multitask network corresponding to the click rate and a second multitask network corresponding to the conversion rate are designed in the object recommendation model, target scores of the candidate objects under the click rate and the conversion rate can be obtained, and the target scores of the candidate objects under the other service targets can be further obtained through calculation based on the target scores of the candidate objects under the click rate and the conversion rate. Furthermore, the object recommendation model can be suitable for different service scenes and different regions, and the application range of the object recommendation model and the accuracy of object recommendation based on the object recommendation model are improved.

Based on this alternative, in one possible implementation, S802 includes: inputting the characteristic vector into a first multi-task network, and scoring the candidate object through a scoring network in the first multi-task network and a network branch connected with the scoring network in the first multi-task network to obtain a target score of the candidate object under the click rate; and inputting the characteristic vector into a second multi-task network, and scoring the candidate object through a scoring network in the second multi-task network and a network branch connected with the scoring network in the second multi-task network to obtain a target score of the candidate object under the conversion rate. At this time, S803 includes: and obtaining the final target score of the candidate object based on the target score of the candidate object under the click rate and the target score of the candidate object under the conversion rate.

In one example, as shown in fig. 7, the target score of the candidate object at the click rate is multiplied by the target score of the candidate object at the conversion rate to obtain the final target score of the candidate object.

S804, recommending a target object in the candidate objects to the user according to the target scores of the candidate objects.

The implementation principle and the technical effect of S804 may refer to the foregoing embodiments, and are not described again.

In the embodiment of the application, a target region to which a user belongs is concerned, data statistics of the candidate object is carried out in a regional mode, a plurality of multi-task networks are designed in an object recommendation model, each multi-task network comprises a scoring network and a plurality of network branches, the candidate object is scored under a plurality of service targets in the multi-task networks, regional differences and diversity of the service targets are considered, the object recommendation method is applicable to various recommendation scenes, and object recommendation accuracy of each region is improved.

Based on any of the foregoing embodiments, optionally, the Model architecture of the object recommendation Model may adopt a multitask Learning Task relationship Model (MMOE) based on Multi-expert mixing, and the multitask network may adopt an Entire Space Multitask Model (ESMM). In other words, the scoring tasks under multiple business objectives can be fitted separately with regionally tailored MMOE structures within the framework of ESMM. Further, in the case that the business objectives include CTR (click through rate) and CVR (conversion rate), within the framework of ESMM, the scoring tasks under CTR and CVR are fitted with regionally customized MMOE structures, respectively. Therefore, the method combines the advantages that the MMOE can be compatible with a plurality of related and different multitask networks and the ESMM has better performance on solving the problems of data sparseness and sample selection deviation, promotes the recommendation relevance and accuracy, and enlarges the area of the user-used group.

Referring to fig. 9, fig. 9 is a schematic flowchart of a model training method according to an embodiment of the present application. The model training method provided by the embodiment of the application specifically comprises the following steps:

s901, determining a training data set, wherein the training data set comprises a plurality of training samples and labels of the training samples, the training samples comprise attribute information of users and attribute information of objects, the attribute information of the users comprises target regions to which the users belong, the attribute information of the objects comprises statistical data corresponding to the objects in the plurality of regions, and the labels of the training samples are used for marking whether the users are interested in the objects in the training samples.

The attribute information of the user, the attribute information of the object, and the statistical data corresponding to the object in the multiple regions may refer to the description of the attribute information of the user, the attribute information of the candidate object, and the statistical data corresponding to the candidate object in the multiple regions in the foregoing embodiment, which is not described herein again. Whether the user is interested in the object in the training sample can be represented as different user behaviors in different recommendation scenes, for example, whether the user is interested in the object is represented as whether the user clicks the object (namely whether the user accesses the object) in the recommendation scene for improving the click rate of the object, and whether the user is interested in the object is represented as whether the user purchases the object or not in the recommendation scene for improving the deal rate of the object.

In this embodiment, a pre-collected training data set may be obtained from a database.

S902, performing supervised training on an object recommendation model according to a training data set, wherein the recommendation model comprises a multitask network for scoring the object, the multitask network comprises a scoring network and network branches connected with the scoring network, the network branches are multiple, and different network branches correspond to different regions.

The object recommendation model may be applied to the object recommendation method provided in any of the foregoing embodiments, and the model structure thereof may refer to the description of any of the foregoing embodiments, which is not described again.

In this embodiment, a training sample in a training data set is used as input data, and an object in the training sample is scored through an object recommendation model to obtain a target score of the object in the training sample. Wherein the target score of the object reflects a probability that the user is interested in the object. Therefore, based on the target scores of the objects and the labels of the training samples, the model parameters of the object recommendation model can be adjusted, and supervised training of the object recommendation model is achieved. The object recommendation model can be subjected to multiple supervised training to improve the scoring accuracy of the object recommendation model, namely the accuracy of the object recommendation model for object recommendation.

In one possible implementation, S902 includes: in the object recommendation model, determining a feature vector corresponding to an object in a training sample according to attribute information of a user and attribute information of the object in the training sample; in the object recommendation model, based on the feature vectors corresponding to the objects in the training samples, scoring the objects in the training samples through a scoring network in the multitask network and network branches connected with the scoring network to obtain target scores of the objects in the training samples; determining a loss value of the object recommendation model according to the label of the training sample and the target score of the object in the training sample; and adjusting the model parameters of the object recommendation model according to the loss value of the object recommendation model.

The object recommendation model is used to score the object, and the process of scoring the candidate object by using the object recommendation model in the foregoing embodiment may be referred to, which is not described herein again.

In this implementation, based on the attribute information of the user and the attribute information of the object in the training sample, the object in the training sample is scored through the object recommendation model, and after the target score of the object in the training sample is obtained, the loss value of the object recommendation model can be determined based on the difference between the label of the training sample and the target score of the object in the training sample. And then, based on the loss value, adjusting the model parameters of the object recommendation model by using an optimization algorithm.

The optimization algorithm is, for example, a gradient descent method, and is not limited herein.

In yet another possible implementation manner, in the case that the object recommendation model includes a plurality of multitask networks, S902 includes: in the object recommendation model, determining a feature vector corresponding to an object in a training sample according to attribute information of a user and attribute information of the object in the training sample; in the object recommendation model, based on the feature vectors corresponding to the objects in the training sample, scoring the objects in the training sample through scoring networks in the multi-task networks and network branches connected with the scoring networks to obtain target scores of the objects in the training sample under the service targets corresponding to the multi-task networks; determining a plurality of loss values of an object recommendation model according to the labels of the training samples and the target scores of the objects in the training samples under the service targets corresponding to the multi-task networks; model parameters of the object recommendation model are adjusted according to the plurality of loss values.

The process of scoring the candidate object by using the multiple multitask networks in the object recommendation model in the foregoing embodiment may be referred to, and details are not repeated herein. The labels of the training samples comprise actual index values of the objects in the training samples under the business targets respectively corresponding to the multiple multitask networks.

In this implementation manner, for each multitask network, network parameters of the multitask network may be adjusted based on a difference between an actual index value of an object in a training sample under a service target corresponding to the multitask network and a target score of the object in the training sample under the service target, so as to improve accuracy of parameter adjustment of the multitask network.

Further, in the case that the business objective includes click rate and conversion rate, the label of the training sample may include actual click rate and actual conversion rate of the object in the training sample, the objective score of the object at click rate is equivalent to the predicted click rate of the object, and the objective score of the object at conversion rate is equivalent to the predicted conversion rate of the object. Therefore, the network parameters of the first multitask network corresponding to the click rate can be adjusted based on the difference between the target score of the object under the click rate and the actual click rate of the object, and the network parameters of the second multitask network can be adjusted based on the difference between the target score of the object under the conversion rate and the actual conversion rate of the object.

Based on any one of the foregoing parameter adjustment schemes, optionally, since different network branches in the multitasking network correspond to different regions, in the process of adjusting the model parameter of the object recommendation model, the scoring network and the target network branch corresponding to the target region in the multitasking network may be adjusted according to the loss value.

In this optional manner, after the object is scored by using the object recommendation model based on the attribute information of the user located in the target area and the attribute information of the object, and the target score of the object is obtained, the target score of the object reflects the target score of the object in the target area, that is, reflects the interest level of the user in the target area in the object. Considering the difference of users in different regions, a loss value can be determined based on the difference between a target score of an object in a training sample and a label corresponding to the training sample, and a scoring network in the multitask network and a target network branch corresponding to the target region are adjusted based on the loss value, so that the target network branch learns characteristics related to the preference of the user in the target region for the object.

The estimation optimization process of the object C in different areas (area a and area B) is taken as an example and explained as follows:

based on the attribute information of the object C (such as the total statistics of all the areas of the object C and the statistics of the object C in the area a) and the attribute information of the users in the area a, the target scores (under two business targets, namely the click rate and the conversion rate) of the object C in the area a can be respectively obtained through an object recommendation sub-network of the area a (for example, in the case that the business target includes the click rate and the conversion rate, the object recommendation sub-network of the area a includes a multi-task network corresponding to the click rate, a multi-task network corresponding to the conversion rate and a feature embedding layer); the difference between the actual index (actual click rate and actual transaction rate) of the object C in the area A and the target score can correspondingly calculate a loss value; based on the loss value, parameter adjustment is performed on the scoring network and the network branch corresponding to the area a in the multitask network corresponding to the click rate and the multitask network corresponding to the conversion rate (which may be referred to as a click rate prediction network and a conversion rate prediction network, respectively).

Similarly, the target score of the object C in the area B may also be input into the object recommendation sub-network of the area B from attribute information of the object C (for example, including total statistical data of the object C in all areas and statistical data of the object C in the area B) and attribute information of the user in the area B, so as to obtain a difference between the target score obtained by the object recommendation sub-network of the object C in the area B and an actual index of the object C in the area B, and may correspondingly calculate a loss value of the object C in the area B; based on the loss value, the parameters of the scoring network and the network branch corresponding to the area B are adjusted in the multitask network.

The scoring network is shared by a plurality of network branches, and the network branches are independent of the region, so that the scoring network shared by all the network branches can be updated when the parameters are updated, and the network branches independent of the region can be updated only by the loss of the region. That is, the loss of the area a cannot update the network branch of the area B, but only the network branch of the area a; the loss of zone B cannot update the network branch of zone a, but only of zone B.

Based on any one of the foregoing parameter adjustment schemes, optionally, when the object recommendation model includes the feature embedding layer, the network parameter of the feature embedding layer may also be adjusted based on the loss value, so as to improve accuracy of embedding and representing data input into the object recommendation model, and further improve accuracy of object recommendation.

The object recommendation device provided in the embodiment of the present application is used for executing the technical solution in any method embodiment of the object recommendation method, and the implementation principle and the technical effect are similar, which are not described herein again.

Referring to fig. 10, fig. 10 is a block diagram of an object recommendation device 100 according to an embodiment of the present application. As shown in fig. 10, an object recommendation apparatus 100 according to an embodiment of the present application includes: a feature determination unit 101, an object scoring unit 102, and an object recommendation unit 103, wherein:

a feature determining unit 101, configured to determine, in an object recommendation model, a feature vector corresponding to a candidate object according to attribute information of a user and attribute information of the candidate object, where the attribute information of the user includes a target region to which the user belongs, and the attribute information of the candidate object includes statistical data corresponding to the candidate object in multiple regions, respectively;

the object scoring unit 102 is configured to score the candidate object through a scoring network in the multitask network and a network branch connected with the scoring network based on the feature vector in the object recommendation model to obtain a target score of the candidate object, where in the multitask network, the number of network branches is multiple, and different network branches correspond to different regions;

and the object recommending unit 103 is used for recommending a target object in the candidate objects to the user according to the target scores of the candidate objects.

In an embodiment of the present application, the object scoring unit 102 is specifically configured to: inputting the feature vectors into a scoring network, and scoring the candidate objects in the scoring network based on the feature vectors to obtain scores output by the scoring network; and inputting the score output by the scoring network into the network branch, and optimizing the score output by the scoring network in the network branch to obtain the target score.

In an embodiment of the present application, the scoring network is multiple, different network branches share all scoring networks, and when a score output by a scoring network is input to a network branch, the object scoring unit 102 is specifically configured to: aiming at each network branch, all scores output by the scoring network are input into the network branch, and the scores of all scoring networks are fused in the network branch to obtain the score output by the network branch; and obtaining a target score according to the score output by the network branch.

In an embodiment of the application, the multitasking network further includes a mask network, and the object scoring unit 102 is specifically configured to: initializing a mask network according to the target area; in the multitask network, inputting the scores output by all network branches into an initialized mask network, and screening the scores output by the target network branches corresponding to a target area from the initialized mask network; and determining a target score according to the score output by the target network branch.

In an embodiment of the present application, there are multiple multitask networks, different multitask networks correspond to different service targets, and the object scoring unit 102 is specifically configured to: in the object recommendation model, inputting the feature vectors into each multitask network, and scoring the candidate objects through a scoring network in the multitask network and a network branch connected with the scoring network in the multitask network based on the feature vectors to obtain target scores of the candidate objects under the service targets corresponding to the multitask network; and determining the final target score of the candidate object according to the target scores of the candidate object under the service targets corresponding to the multi-task networks.

In an embodiment of the present application, the service target includes a click rate and a conversion rate, the statistical data includes a number of users clicking and a number of users making a transaction, the multitask network includes a first multitask network corresponding to the click rate and a second multitask network corresponding to the conversion rate, and the object scoring unit 102 is specifically configured to: inputting the characteristic vector into a first multi-task network, and scoring the candidate object through a scoring network in the first multi-task network and a network branch connected with the scoring network in the first multi-task network to obtain a target score of the candidate object under the click rate; and inputting the characteristic vector into a second multi-task network, and scoring the candidate object through a scoring network in the second multi-task network and a network branch connected with the scoring network in the second multi-task network to obtain a target score of the candidate object under the conversion rate.

Referring to fig. 11, fig. 11 is a block diagram of a model training apparatus 110 according to an embodiment of the present disclosure. As shown in fig. 11, an object recommendation apparatus 110 provided in an embodiment of the present application includes: a determination unit 111 and a training unit 112, wherein:

the data determining unit 111 is configured to determine a training data set, where the training data set includes a plurality of training samples and tags of the training samples, the training samples include attribute information of a user and attribute information of an object, the attribute information of the user includes a target region to which the user belongs, the attribute information of the object includes statistical data corresponding to the object in a plurality of regions, and the tags of the training samples are used to mark whether the user is interested in the object in the training samples;

the training unit 112 is configured to perform supervised training on an object recommendation model according to a training data set, where the object recommendation model includes a multitask network for scoring the object, the multitask network includes a scoring network and network branches connected to the scoring network, the network branches are multiple, and different network branches correspond to different regions.

In an embodiment of the present application, the training unit 112 is specifically configured to: in the object recommendation model, determining a feature vector corresponding to an object in a training sample according to attribute information of a user and attribute information of the object in the training sample; in the object recommendation model, based on the characteristic vector, scoring the objects in the training sample through a scoring network in the multitask network and a network branch connected with the scoring network to obtain a target score of the objects in the training sample; determining a loss value of the object recommendation model according to the label and the target score of the training sample; and adjusting the model parameters of the object recommendation model according to the loss value.

In an embodiment of the present application, the training unit 112 is specifically configured to: and according to the loss value, carrying out parameter adjustment on a scoring network in the multitask network and a target network branch corresponding to a target region in the multitask network.

The model training device provided in the embodiment of the present application is used for executing the technical solution in any method embodiment of the above model training method, and the implementation principle and the technical effect are similar, which are not described herein again.

The technical scheme provided by the embodiment of the application can be realized on a cloud server.

Fig. 12 is a schematic structural diagram of a cloud server according to an exemplary embodiment of the present application. The cloud server is used for operating an object recommendation method, scoring the candidate objects through an object recommendation model based on the attribute information of the user and the attribute information of the candidate objects to obtain target scores of the candidate objects, and recommending the target objects in the candidate objects to the user based on the target scores of the candidate objects; and/or the cloud server is used for operating a model training method and is used for training an object recommendation model. As shown in fig. 12, the cloud server includes: a memory 123 and a processor 124.

The memory 123 is used for storing computer programs and may be configured to store other various data to support operations on the cloud server. The Storage 123 may be an Object Storage Service (OSS).

The memory 123 may be implemented by any type or combination of volatile or non-volatile memory devices, such as Static Random Access Memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic or optical disks.

A processor 124, coupled to the memory 123, for executing the computer program in the memory 123, so as to execute the object recommendation method and/or the model training method provided by any of the foregoing embodiments

Further, as shown in fig. 12, the cloud server further includes: firewall 121, load balancer 122, communication component 125, power component 126, and other components. Only some of the components are schematically shown in fig. 12, and the cloud server is not meant to include only the components shown in fig. 12.

Accordingly, the present application also provides a computer readable storage medium storing a computer program, which when executed by a processor causes the processor to implement the steps in the above method embodiments.

Accordingly, the present application also provides a computer program product, which includes a computer program/instruction, when the computer program/instruction is executed by a processor, the processor is caused to implement the steps in the above method embodiments.

The communications component 125 of fig. 12 described above is configured to facilitate communications between the device in which the communications component resides and other devices in a wired or wireless manner. The device in which the communication component 125 is located may access a wireless network based on a communication standard, such as WiFi, a mobile communication network such as 2G, 3G, 4G/LTE, 5G, or a combination thereof. In an exemplary embodiment, the communication component 125 receives a broadcast signal or broadcast related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 125 further includes a Near Field Communication (NFC) module to facilitate short-range communications. For example, the NFC module may be implemented based on Radio Frequency Identification (RFID) technology, infrared data association (IrDA) technology, Ultra Wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

The power supply module 126 of fig. 12 provides power to the various components of the device in which the power supply module 126 is located. The power components 126 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the device in which the power components 126 are located.

As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present invention has been described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

In a typical configuration, a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.

The memory may include forms of volatile memory in a computer readable medium, Random Access Memory (RAM) and/or non-volatile memory, such as Read Only Memory (ROM) or flash memory (flash RAM). Memory is an example of a computer-readable medium.

Computer-readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), Digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. As defined herein, a computer readable medium does not include a transitory computer readable medium such as a modulated data signal and a carrier wave.

It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

The above description is only an example of the present application and is not intended to limit the present application. Various modifications and changes may occur to those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. An object recommendation method, comprising:

in an object recommendation model, determining a feature vector corresponding to a candidate object according to attribute information of a user and attribute information of the candidate object, wherein the attribute information of the user comprises a target region to which the user belongs, and the attribute information of the candidate object comprises statistical data corresponding to the candidate object under a plurality of regions respectively;

in the object recommendation model, based on the feature vector, scoring the candidate object through a scoring network in a multitask network and a network branch connected with the scoring network to obtain a target score of the candidate object, wherein in the multitask network, the number of the network branches is multiple, and different network branches correspond to different regions;

and recommending the target object in the candidate objects to the user according to the target scores of the candidate objects.

2. The object recommendation method according to claim 1, wherein in the object recommendation model, scoring the candidate objects through a scoring network in a multitasking network and a network branch connected to the scoring network based on the feature vector to obtain target scores of the candidate objects comprises:

inputting the feature vector into the scoring network, and scoring the candidate object in the scoring network based on the feature vector to obtain a score output by the scoring network;

and inputting the score output by the scoring network into the network branch, and optimizing the score output by the scoring network in the network branch to obtain the target score.

3. The object recommendation method according to claim 2, wherein the scoring network is a plurality of scoring networks, different network branches share all scoring networks, the scoring output by the scoring network is input to the network branches, and the scoring output by the scoring network is optimized in the network branches to obtain the target scoring, and the method comprises:

for each network branch, inputting the scores output by all the scoring networks into the network branch, and fusing the scores of all the scoring networks in the network branch to obtain the scores output by the network branch;

and obtaining the target score according to the score output by the network branch.

4. The object recommendation method of claim 3, wherein the multitasking network further comprises a mask network, and the obtaining the target score according to the score output by the network branch comprises:

initializing the mask network according to the target region;

in the multitask network, inputting the scores output by all network branches into an initialized mask network, and screening out the scores output by the target network branches corresponding to the target region from the initialized mask network;

and determining the target score according to the score output by the target network branch.

5. The object recommendation method according to any one of claims 1-4, wherein the multitask network is multiple, different multitask networks correspond to different business targets, and the scoring the candidate object based on the feature vector in the object recommendation model through a scoring network in the multitask network and a network branch connected to the scoring network to obtain a target score of the candidate object comprises:

inputting the characteristic vectors into each multitask network in an object recommendation model, and scoring the candidate objects through a scoring network in the multitask network and a network branch connected with the scoring network in the multitask network based on the characteristic vectors to obtain target scores of the candidate objects under service targets corresponding to the multitask network;

and determining the final target score of the candidate object according to the target scores of the candidate object under the service targets corresponding to the multi-task networks.

6. The object recommendation method according to claim 5, wherein the service targets include click through rates and conversion rates, the statistical data include the number of click through users and the number of transaction users, the multitask network includes a first multitask network corresponding to the click through rates and a second multitask network corresponding to the conversion rates, the object recommendation model inputs the feature vectors into each multitask network, and based on the feature vectors, the candidate objects are scored through a scoring network in the multitask network and a network branch connected to the scoring network in the multitask network, so as to obtain target scores of the candidate objects under the service targets corresponding to the multitask network, and the method includes:

inputting the feature vector into the first multitask network, and scoring the candidate object through a scoring network in the first multitask network and a network branch connected with the scoring network in the first multitask network to obtain a target score of the candidate object under the click rate;

inputting the feature vector into the second multitask network, and scoring the candidate object through a scoring network in the second multitask network and a network branch connected with the scoring network in the second multitask network to obtain a target score of the candidate object under the conversion rate.

7. A method of model training, comprising:

determining a training data set, wherein the training data set comprises a plurality of training samples and labels of the training samples, the training samples comprise attribute information of users and attribute information of objects, the attribute information of the users comprises target regions to which the users belong, the attribute information of the objects comprises statistical data corresponding to the objects in the plurality of regions, and the labels of the training samples are used for marking whether the users are interested in the objects in the training samples;

and carrying out supervised training on an object recommendation model according to the training data set, wherein the object recommendation model comprises a multitask network for scoring the object, the multitask network comprises a scoring network and network branches connected with the scoring network, the network branches are multiple, and different network branches correspond to different regions.

8. The model training method of claim 7, wherein the supervised training of the object recommendation model based on the training dataset comprises:

in the object recommendation model, determining a feature vector corresponding to an object in the training sample according to attribute information of the user and attribute information of the object in the training sample;

in the object recommendation model, based on the feature vectors, scoring the objects in the training sample through a scoring network in the multitask network and network branches connected with the scoring network to obtain target scores of the objects in the training sample;

determining a loss value of the object recommendation model according to the label of the training sample and the target score;

and adjusting the model parameters of the object recommendation model according to the loss value.

9. The model training method of claim 8, wherein the adjusting the model parameters of the object recommendation model according to the loss value comprises:

and according to the loss value, carrying out parameter adjustment on a scoring network in the multitask network and a target network branch corresponding to the target area in the multitask network.

10. An object recommendation apparatus, comprising:

the characteristic determining unit is used for determining a characteristic vector corresponding to a candidate object according to attribute information of a user and attribute information of the candidate object in an object recommendation model, wherein the attribute information of the user comprises a target region to which the user belongs, and the attribute information of the candidate object comprises statistical data corresponding to the candidate object under a plurality of regions respectively;

the object scoring unit is used for scoring the candidate objects through a scoring network in a multitask network and network branches connected with the scoring network based on the feature vectors in the object recommendation model to obtain target scores of the candidate objects, wherein the number of the network branches in the multitask network is multiple, and different network branches correspond to different regions;

and the object recommending unit is used for recommending a target object in the candidate objects to the user according to the target scores of the candidate objects.

11. A model training apparatus, comprising:

the data determining unit is used for determining a training data set, the training data set comprises a plurality of training samples and labels of the training samples, the training samples comprise attribute information of users and attribute information of objects, the attribute information of the users comprises target regions to which the users belong, the attribute information of the objects comprises statistical data corresponding to the objects in the plurality of regions, and the labels of the training samples are used for marking whether the users are interested in the objects in the training samples;

the training unit is used for carrying out supervised training on an object recommendation model according to the training data set, wherein the object recommendation model comprises a multitask network used for scoring objects, the multitask network comprises a scoring network and network branches connected with the scoring network, the network branches are multiple, and different network branches correspond to different regions.

12. A cloud server, comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein,

the memory stores instructions executable by the at least one processor to enable the cloud server to perform the object recommendation method of any one of claims 1 to 6 and/or to enable the cloud server to perform the model training method of any one of claims 7 to 9.

13. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the object recommendation method of any one of claims 1 to 6 and/or which, when being executed by a processor, carries out the model training method of any one of claims 7 to 9.