CN110782308B

CN110782308B - Push method and device for recommended package, electronic equipment and readable storage medium

Info

Publication number: CN110782308B
Application number: CN201910894674.3A
Authority: CN
Inventors: 胡懋地; 石晓文; 王文斌; 罗震
Original assignee: Beijing Sankuai Online Technology Co Ltd
Current assignee: Beijing Sankuai Online Technology Co Ltd
Priority date: 2019-09-20
Filing date: 2019-09-20
Publication date: 2024-07-16
Anticipated expiration: 2039-09-20
Also published as: CN110782308A

Abstract

The embodiment of the application provides a pushing method, a pushing device, electronic equipment and a readable storage medium for recommending packages, and aims to generate and output recommended packages so as to reduce the difficulty of matching dishes by consumers. The method comprises the following steps: determining a target merchant from a plurality of merchants according to input information of a user; obtaining a plurality of candidate dishes for the target merchant; inputting the dish information corresponding to each of the plurality of candidate dishes into a package generation model, extracting characteristics of each dish information through an encoding module of the package generation model to obtain dish characteristics corresponding to the dish information, encoding the plurality of dish characteristics through the encoding module to obtain total dish characteristics corresponding to the plurality of candidate dishes, and decoding the total dish characteristics through a decoding module of the package generation model to obtain a recommended package, wherein the recommended package comprises at least one candidate dish; pushing the recommended packages to the user.

Description

Push method and device for recommended package, electronic equipment and readable storage medium

Technical Field

The embodiment of the application relates to the technical field of data processing, in particular to a pushing method and device for recommending packages, electronic equipment and a readable storage medium.

Background

With the development of internet technology and the popularization of intelligent terminal equipment, more and more off-line merchants enter an e-commerce platform so as to open an information interaction channel between the off-line merchants and consumers through the e-commerce platform. In the field of food take-out, a consumer browses an electronic commerce platform through application software running on terminal equipment of the consumer, selects proper dishes to be combined into packages and places orders according to an electronic menu provided by a resident merchant in the electronic commerce platform, and accordingly online ordering is completed.

In general, in order to improve the competitiveness of the resident merchant in the e-commerce platform, the resident merchant enriches the variety of dishes as much as possible, so that consumers implement more various selection collocation schemes based on the electronic menu. However, with the richness of the dishes, the consumer selection and the matching diversity are enhanced, while the consumer consumption is stimulated, the selection difficulty of the consumer is increased, and the consumer is difficult to quickly select when facing a plurality of potential dishes matching schemes.

To this end, the related art provides a dish classification method of classifying a plurality of dishes of each of a plurality of resident merchants into a plurality of categories, for example: meat dishes, vegetables, fried dishes, stews, cold dishes, marinated dishes, soup, staple food and the like, and aims to enable consumers to quickly lock target dish categories and quickly select proper dishes under the target dish categories. However, based on the method, consumers still need to perform double selection on the dishes and specific dishes under each dish category, and the method cannot fundamentally solve the problem of difficult dish collocation.

Disclosure of Invention

The embodiment of the application provides a pushing method, a pushing device, electronic equipment and a readable storage medium for recommending packages, and aims to improve the intelligent level of a food takeout electronic commerce platform, so as to help consumers reduce the difficulty of matching dishes.

The first aspect of the embodiment of the application provides a pushing method for recommending packages, which comprises the following steps:

Determining a target merchant from a plurality of merchants according to input information of a user;

obtaining a plurality of candidate dishes for the target merchant;

Inputting the dish information corresponding to each of the plurality of candidate dishes into a package generation model, extracting characteristics of each dish information through an encoding module of the package generation model to obtain dish characteristics corresponding to the dish information, encoding the plurality of dish characteristics through the encoding module to obtain total dish characteristics corresponding to the plurality of candidate dishes, and decoding the total dish characteristics through a decoding module of the package generation model to obtain a recommended package, wherein the recommended package comprises at least one candidate dish;

Pushing the recommended packages to the user.

A second aspect of an embodiment of the present application provides a pushing device for recommending packages, where the device includes:

The target merchant determining module is used for determining a target merchant from a plurality of merchants according to the input information of the user;

a candidate dish obtaining module for obtaining a plurality of candidate dishes of the target merchant;

The recommendation package obtaining module is used for inputting the dish information corresponding to each of the plurality of candidate dishes into a package generating model, extracting characteristics of each dish information through an encoding module of the package generating model to obtain dish characteristics corresponding to the dish information, encoding the plurality of dish characteristics through the encoding module to obtain total dish characteristics corresponding to the plurality of candidate dishes, and decoding the total dish characteristics through a decoding module of the package generating model to obtain a recommendation package, wherein the recommendation package comprises at least one candidate dish;

And the recommended package pushing module is used for pushing the recommended packages to the user.

A third aspect of the embodiments of the present application provides a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the method according to the first aspect of the present application.

A fourth aspect of the embodiments of the present application provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, which when executed implements the steps of the method according to the first aspect of the application.

By adopting the pushing method of the recommended package provided by the embodiment of the application, a target merchant is determined according to the input information of the user, and the dish information corresponding to each of a plurality of candidate dishes of the target merchant is input into a package generation model; extracting features of the dish information through an encoding module of the package generation model, encoding the extracted dish features to obtain dish total features corresponding to a plurality of candidate dishes, and decoding the dish total features through a decoding module of the package generation model to obtain recommended packages; and pushing the obtained recommended package to the user. Thus, the intelligent level of the food takeout electronic commerce platform is improved. And on the user side, after the user inputs information, the user can automatically obtain recommended package information, so that the difficulty of matching dishes is reduced, and the ordering efficiency is improved.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings that are needed in the description of the embodiments of the present application will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present application, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flow chart of a push method of recommending packages in an embodiment of the application;

FIG. 2 is a schematic diagram of a package generation model according to an embodiment of the present application;

FIG. 3 is a schematic diagram of outputting recommended packages in an embodiment of the application;

FIG. 4 is a schematic diagram illustrating interaction between application software and a server according to an embodiment of the present application;

FIG. 5 is a training flow diagram of a training package generation model in one embodiment of the application;

FIG. 6 is a schematic diagram of outputting a target package in an embodiment of the application;

Fig. 7 is a schematic diagram of a pushing device for recommending packages according to an embodiment of the present application.

Detailed Description

The following description of the embodiments of the present application will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are some, but not all embodiments of the application. All other embodiments, which can be made by those skilled in the art based on the embodiments of the present application without making any inventive effort, shall fall within the scope of the present application.

In the field of food take-out, more and more off-line merchants enter a food take-out electronic commerce platform, and after consumers (hereinafter, collectively referred to as users) enter the electronic commerce platform through application software running on terminal equipment of the consumers, proper dishes are selected to be combined into packages and ordered according to an electronic menu provided by the on-line merchants in the electronic commerce platform, so that on-line ordering operation is completed. In general, in order to improve the competitiveness of a resident merchant in an e-commerce platform, the resident merchant enriches the variety of dishes as much as possible, so that users can implement more various selection collocation schemes based on electronic menus of the resident merchant. However, with the richness of the dishes, the user selection and the diversity of the collocation are enhanced, while the consumption of the user is stimulated, the difficulty of the selection of the user is intangibly increased, and the user is difficult to quickly make the selection when facing a plurality of potential dishes collocation schemes.

To this end, the related art provides a dish classification method of classifying a plurality of dishes of a plurality of resident merchants into a plurality of categories, for example: meat dishes, vegetables, fried dishes, stews, cold dishes, marinated dishes, soup, staple food and the like, and aims to enable a user to quickly lock a target dish category and quickly select proper dishes under the target dish category. However, based on the method, the user still needs to perform double selection on the dishes and the specific dishes under each dish category, and the method cannot fundamentally solve the problem of difficult dish collocation.

Therefore, the current food takeout electronic commerce platform has low intelligent level, can not effectively help users to reduce difficulty in matching dishes, and is difficult to effectively improve ordering efficiency of the users. In view of this, the embodiments of the present application provide: determining a target merchant according to input information of a user, and inputting dish information corresponding to each of a plurality of candidate dishes of the target merchant into a package generation model; extracting features of the dish information through an encoding module of the package generation model, encoding the extracted dish features to obtain dish total features corresponding to a plurality of candidate dishes, and decoding the dish total features through a decoding module of the package generation model to obtain recommended packages; and pushing the obtained recommended package to the user.

Referring to fig. 1, fig. 1 is a flowchart of a method for pushing recommended packages according to an embodiment of the present application. As shown in fig. 4, the pushing method includes the following steps:

Step S110: the target merchant is determined from a plurality of merchants based on the user input.

In this embodiment, the input information of the user generally carries information of the target merchant. For example, when a user clicks a certain merchant on the homepage of the food takeout electronic commerce platform to enter the page of the merchant, information of the merchant, such as a name, a number, a character string and the like, is usually carried in the information input along with the clicking operation of the user. In this way, the server can determine the merchant that the user is attempting to enter from a plurality of registered merchants according to the input information of the user, and the merchant is the target merchant.

For another example, in the case that the user has entered a certain merchant page, there is a package recommendation key in the page, and when the user clicks the package recommendation key to request the server to push packages to the user, the information input along with the clicking operation of the user may also carry the information of the merchant. Therefore, the server side can determine the merchant where the user is currently located from a plurality of registered merchants according to the input information of the user, and the merchant is the target merchant.

Step S120: a plurality of candidate dishes for the target merchant are obtained.

In this embodiment, the electronic menu of the target merchant may be obtained in real time, and dishes included in the electronic menu may be determined as a plurality of candidate dishes. Or may read from a database a plurality of candidate dish information pre-recorded for the target merchant. It should be appreciated that the present application is not limited as to how the plurality of candidate dishes are obtained.

Wherein each candidate dish carries dish information. Typically, the dish information for one candidate dish includes at least two of: dish name, dish category, dish price, dish sales. The names and categories of dishes are usually text, so the names and categories of dishes are usually text information, such as dish names of "tomato fried eggs", "corn stewed pork ribs", "Kung Pao Chicken" and the like, and dish categories such as "meat dish", "vegetable dish", "stewed dish", "eggs", "fish" and the like. While the price and sales of dishes are typically specific numbers, the price and sales of dishes are thus typically digital information.

Step S130: inputting the dish information corresponding to each of the plurality of candidate dishes into a package generation model, extracting features of each dish information through an encoding module of the package generation model to obtain dish features corresponding to the dish information, encoding the plurality of dish features through the encoding module to obtain total dish features corresponding to the plurality of candidate dishes, and decoding the total dish features through a decoding module of the package generation model to obtain a recommended package, wherein the recommended package comprises at least one candidate dish.

In this embodiment, the package generation model used in step S130 may be a package generation model obtained after training the first preset model through the following training process including steps S510 to S550. The package generation model utilized in step S130 may be a package generation model obtained through other training methods or through other channels.

Referring to fig. 2, fig. 2 is a schematic structural diagram of a package generation model according to an embodiment of the present application. As shown in fig. 2, the package generation model includes an encoding module and a decoding module. The coding module mainly comprises a memory sub-network and a feature extraction sub-network. The Memory sub-network may be a recurrent neural network RNN (Recurrent Neural Network), a Long Short-Term Memory network LSTM (Long Short-Term Memory), or a gated recurrent unit network GRU (Gate Recurrent Unit). The feature extraction sub-network may be a convolutional neural network CNN (Convolutional Neural Networks). The decoding module can also select a cyclic neural network RNN, a long and short-term memory network LSTM or a gate-controlled cyclic unit network GRU, and the like. It should be noted that, for the two feature extraction sub-networks shown in fig. 2, the nature thereof is the same feature extraction sub-network, which processes a plurality of dishes in sequence.

In the step S130, regarding one candidate dish of the plurality of candidate dishes, considering that the dish information of the candidate dish may include text information such as a dish name and a dish category, the dish name and the dish category may be first classified from the dish information thereof, and then input to the word vector conversion module embedding shown in fig. 2, thereby converting the dish name and the dish category into word vectors, respectively. In addition, considering the influence of factors such as regions, living habits and dialects on the names of the dishes, different merchants slightly differ in the names of the dishes of the same sample, so that when the encoding module of the package generation model performs feature extraction on the dish information of the candidate dishes in step S130, the following extraction mode can be specifically adopted:

Substep S130-1: and extracting the characteristics of the text information in the dish information through the characteristic extraction sub-network to obtain text characteristics corresponding to the text information.

Substep S130-2: and splicing the text features with the digital information in the dish information to obtain dish features corresponding to the dish information.

As shown in fig. 2, in the substep S130-1, the word vector of the dish name is specifically input into the feature extraction sub-network, and the feature extraction sub-network extracts the feature that can reflect the nature of the dish for the word vector of the dish name that is input. Feature extraction is carried out on word vectors of dish names through a feature extraction sub-network, so that influence of naming differentiation on package pushing accuracy caused by factors such as regions, living habits and dialects is eliminated, and rationality and applicability of recommended packages are improved. In addition, as shown in FIG. 2, the coding module of the package generation model may also include a pooling layer pooling. The features extracted by the feature extraction sub-network aiming at the word vector of the dish name are input into the pooling layer pooling, and after pooling processing of the pooling layer pooling, the features of the dish name are obtained. Note that, for the two pooling layers pooling shown in fig. 2, the same pooling layer pooling is essential, and the pooling layer pooling sequentially processes a plurality of dishes.

As shown in fig. 2, in the substep S130-2, the feature of the dish name of the candidate dish, the word vector of the dish category, the dish price, the dish sales, and the like may be spliced into a larger total feature, i.e., the dish feature is obtained.

In the step S130, after the dish features are obtained, when the encoding module encodes the plurality of dish features, the following encoding method may be specifically adopted:

Substep S130-3: and respectively inputting the plurality of dish features into the memory sub-network in different rounds to obtain the total dish features output by the memory sub-network.

For example, as shown in fig. 2, after obtaining a dish feature of a candidate dish, the dish feature of the candidate dish may be input to the memory sub-network as a single input to the memory sub-network. And then based on the same way, obtaining the dish characteristics of the next candidate dish, and inputting the dish characteristics into the memory sub-network as another input for the memory sub-network. After the dish characteristics of each of the plurality of candidate dishes are sequentially input into the memory sub-network, the memory sub-network finishes encoding each candidate dish and outputs one dish total characteristic corresponding to all the candidate dishes.

In the step S130, the decoding module of the package generation model decodes the total features of the dishes, and when the recommended packages are obtained, the following decoding method may be specifically adopted:

Substep S130-4: inputting the total characteristics of the dishes into the decoding module, and obtaining the respective selected probabilities of the candidate dishes output by the decoding module and the selected probabilities of a first stop bit, wherein the first stop bit indicates that the recommended package is generated.

Substep S130-5: according to the respective selected probabilities of the plurality of candidate dishes and the selected probability of the first termination bit, selecting N selected probabilities with larger probability; wherein N is an integer greater than or equal to 1.

Substep S130-6: and determining that the dish corresponding to each selected probability is a target candidate dish according to each selected probability in the N selected probabilities, and inputting the characteristics of the target candidate dish into the decoding module.

Substep S130-7: obtaining the probability of selection of the first termination bit and the probability of selection of each of the plurality of candidate dishes under the condition that the target candidate dish is selected and outputted by the decoding module, and returning to the step: according to the respective selected probabilities of the plurality of candidate dishes and the selected probability of the first termination bit, selecting N selected probabilities with larger probability; and determining the dish corresponding to each selected probability as a target candidate dish according to each selected probability in the N selected probabilities.

Substep S130-8: and when the N selected probabilities selected each time contain the selected probabilities corresponding to the first termination bit, combining dishes corresponding to the selected probabilities before the first termination bit is selected to obtain a recommended package.

Referring to fig. 3, fig. 3 is a schematic diagram of an output recommended package according to an embodiment of the present application. As shown in fig. 3, after receiving the total characteristics of the dishes (i.e. codes) input by the coding module, the decoding module outputs a first set of selected probabilities, including: the probability of each of the plurality of candidate dishes being selected (each candidate dish being represented by a circle in fig. 3), and the probability of the first termination bit being selected (the first termination bit being represented by a box in fig. 3).

As shown in fig. 3, taking N equal to 2 as an example, 2 selected probabilities with a larger probability are selected from among the plurality of candidate dishes according to their respective selected probabilities and the selected probability of the first termination bit.

As shown in fig. 3, after the first round of selection, the candidate dish 1 and the candidate dish 4 are input to the decoding module as the selected target candidate dish, and their respective features (for simplicity of the drawing, a process in which the respective features of the candidate dish 1 and the candidate dish 4 are input to the decoding module is not shown in fig. 3). The decoding module outputs two second sets of selected probabilities based on the input features, each second set of selected probabilities including: the probability of each of the plurality of candidate dishes being selected (each candidate dish being represented by a circle in fig. 3), and the probability of the first termination bit being selected (the first termination bit being represented by a box in fig. 3).

As shown in fig. 3, the selection is made for the two second set of selected probabilities based on the same manner. After the second round of selection, the candidate dish 2 is selected in the second set of selection probabilities on the left, and the candidate dish 2 is also selected in the second set of selection probabilities on the right. Two candidate dishes 2 are input to the decoding module as the selected target candidate dishes, with their respective characteristics. The decoding module outputs left and right third sets of selected probabilities based on the input features, each third set of selected probabilities including: the probability of each of the plurality of candidate dishes being selected (each candidate dish being represented by a circle in fig. 3), and the probability of the first termination bit being selected (the first termination bit being represented by a box in fig. 3).

As shown in fig. 3, the selection is made for the two third sets of selected probabilities based on the same manner. After the third round of selection, the first termination bit in the third set of selection probabilities on the left is selected and the candidate dish 3 in the third set of selection probabilities on the right is selected. For the first termination bit selected, a recommended package is formed using candidate dish 1 and candidate dish 2 that have been selected prior to selection. For the selected candidate dish 3, the selected candidate dish is used as the selected target candidate dish, and the characteristics of the selected candidate dish are input to the decoding module. The decoding module outputs a fourth set of selected probabilities based on the input features, the fourth set of selected probabilities comprising: the probability of each of the plurality of candidate dishes being selected (each candidate dish being represented by a circle in fig. 3), and the probability of the first termination bit being selected (the first termination bit being represented by a box in fig. 3).

As shown in fig. 3, the selection is made for this fourth set of selected probabilities in the same manner. After the fourth round of selection, the first termination bit in the fourth set of selection probabilities is selected. For the first termination bit selected, another recommended package is assembled using candidate dish 4, candidate dish 2, and candidate dish 3 that have been selected prior to selection thereof.

In this way, a total of two recommended packages are obtained, which are output by the package generation model. It should be understood that the above decoding manner including the sub-steps S130-4 to S130-8 is merely an example of various embodiments of the present application, and is not intended to limit the present application. For example, to obtain a recommended package, the approach shown in FIG. 6 may also be used. For the description of fig. 6, please refer to the following.

Step S140: pushing the recommended packages to the user.

For example, the server may return information of the recommended packages to software running on the user terminal, through which the recommended packages are presented to the user. Means of presentation include, but are not limited to: text, images, and/or animation.

By executing the pushing method of the recommended packages comprising the steps S110 to S140, the intelligent level of the food takeout electronic commerce platform is improved. And on the user side, after the user inputs information, the user can automatically obtain recommended package information, so that the difficulty of matching dishes is reduced, and the ordering efficiency is improved.

In addition, in order to meet the personalized requirements of the users, the embodiment of the application can further comprise the following steps of:

Presenting a plurality of recommendation types to a user; information of a target recommendation type input by a user is obtained, wherein the target recommendation type is one of the recommendation types. Then, in S130, when the dish information corresponding to each of the plurality of candidate dishes is input into the package generation model, specifically, for the information of the target recommendation type input by the user, the dish information corresponding to each of the plurality of candidate dishes is input into the package generation model, so as to generate the package corresponding to the target recommendation type through the package generation model.

Referring to fig. 4, fig. 4 is an interaction diagram of application software and a server according to an embodiment of the present application. As shown in FIG. 4, the application software user interface presents to the user a plurality of recommendation types for merchant A, such as "light taste", "meat and vegetables collocation", "full reduction preference", "hot-box" and the like. Correspondingly, the server side at least comprises four package generation models, and the four package generation models are respectively used for generating the four recommended packages.

For example, after the user clicks the "meat and vegetable matching" recommendation type on the application software user interface, the application software obtains the information of the target recommendation type input by the user, that is, obtains the information of the recommendation type of the "meat and vegetable matching". The application software submits the recommendation type information of meat and vegetable collocation to the server. And the server inputs the dish information corresponding to each of the plurality of candidate dishes of the merchant A into the package generation model 2 according to the information. Wherein, the package generation model 2 is used for generating a recommended package of a meat and vegetable collocation type.

Or, for example, after the user clicks the recommendation type of "light taste" on the application user interface, the application obtains information of the target recommendation type input by the user, that is, obtains information of the recommendation type of "light taste". The application software submits the recommendation type information of the 'light taste' to the server. And the server inputs the dish information corresponding to each of the plurality of candidate dishes of the merchant A into the package generation model1 according to the information. Wherein package generation model1 is used to generate recommended packages of the "light taste" type.

In order to more intelligently implement the pushing method provided by the embodiment of the application, before implementing the method, a first preset model can be pre-built, training samples are collected for the first preset model, then the first preset model is trained based on the collected training samples, and the trained first preset model is used as a package generation model. The package generation model can be used as an optional means for executing part of the steps in the pushing method provided by the embodiment of the application.

Referring to fig. 5, fig. 5 is a training flowchart of a training set generation model according to an embodiment of the present application. As shown in fig. 5, the training process includes the following steps:

Step S510: dish characteristics of a plurality of sample dishes are input into a first preset model.

Illustratively, the structure of the first preset model may be selected from the model structure shown in fig. 2. That is, the first preset model also includes an encoding module and a decoding module. The coding module mainly comprises a memory sub-network and a feature extraction sub-network. The Memory sub-network may be a recurrent neural network RNN (Recurrent Neural Network), a Long Short-Term Memory network LSTM (Long Short-Term Memory), or a gated recurrent unit network GRU (Gate Recurrent Unit). The feature extraction sub-network may be a convolutional neural network CNN (Convolutional Neural Networks). The decoding module can also select a cyclic neural network RNN, a long and short-term memory network LSTM or a gate-controlled cyclic unit network GRU, and the like. It should be noted that, for the two feature extraction sub-networks shown in fig. 2, the nature thereof is the same feature extraction sub-network, which processes a plurality of dishes in sequence.

In this embodiment, the plurality of sample dishes may be from an electronic menu of a merchant, each sample dish having attribute information, and the attribute information of one sample dish includes at least two of the following: sample dish name, sample dish category, sample dish price, sample dish sales. The names and categories of the sample dishes are usually text, so the names and categories of the sample dishes are usually text information, for example, the sample dishes are named as "tomato fried eggs", "corn stewed spareribs", "Kung Pao Chicken" and the like, and the sample dishes are named as "meat dishes", "vegetables", "stews", "eggs", "fish" and the like. The price and sales of the sample dishes are typically specific numbers, so the price and sales of the sample dishes are typically digital information.

In this embodiment, for each of a plurality of sample dishes, attribute information of the sample dish may be classified. Illustratively, the classification refers to: and classifying attribute information such as sample dish names, sample dish categories, sample dish prices, sample dish sales and the like from the original information of the sample dishes. The categorized attribute information of the sample dish may then be used as a characteristic of the sample dish. In this way, in the above step S510, the dish characteristics of the sample dishes are input into the first preset model, that is: and inputting the classified attribute information into a first preset model.

As shown in fig. 2, for the text class information such as the sample dish name and the sample dish category of one sample dish, the word vector conversion module embedding may be used to convert the sample dish name and the sample dish category into word vectors respectively before inputting the first preset model. Considering the influence of factors such as regions, living habits and dialects on the names of the dishes, different merchants slightly differ in the names of the same sample dishes, so that the feature extraction sub-network shown in fig. 2 can be utilized to extract the features of the word vectors of the names of the sample dishes. The feature extraction sub-network extracts features which can reflect the essence of the dishes aiming at word vectors of the input sample dish names. For example, for the word vector of the two sample dish names of "tomato-fried egg soup" and "large tomato-egg soup", the two features extracted by the feature extraction sub-network are two features that are similar to each other.

Furthermore, as shown in fig. 2, the encoding module may also include a pooling layer pooling. The features extracted by the feature extraction sub-network aiming at the word vector of the sample dish name are input into the pooling layer pooling, and the features of the sample dish name are obtained after pooling treatment of the pooling layer pooling. Thus, as shown in fig. 2, the feature of the sample dish name of one sample dish and the word vector of the sample dish category are obtained, and then the feature of the sample dish name of the sample dish, the word vector of the sample dish category, the price of the sample dish, the sales of the sample dish and the like can be spliced into a larger total feature, and then the total feature is input into the memory sub-network as one input for the memory sub-network. Note that, for the two pooling layers pooling shown in fig. 2, the same pooling layer pooling is essential, and the pooling layer pooling sequentially processes a plurality of dishes.

In the same way, as shown in fig. 2, its total characteristics are obtained for the next sample dish, which is then input to the memory subnetwork as another input to the memory subnetwork. After the total characteristics of each of the plurality of sample dishes are sequentially input into the memory sub-network, the memory sub-network finishes encoding each sample dish.

Step S520: and obtaining a target package output by the first preset model, wherein the target package comprises at least one sample dish in the plurality of sample dishes.

In this embodiment, the first preset model determines at least one sample dish from the plurality of sample dishes, and composes and outputs a target package.

In order to obtain the target package, the embodiment of the application proposes a specific way for obtaining the target package, which comprises the following substeps:

Step S520-1: and obtaining the respective selected probabilities of the plurality of sample dishes output by the first preset model and the selected probabilities of a second termination bit, wherein the second termination bit represents that the target package is generated.

Step S520-2: and selecting a target sample dish from the plurality of sample dishes as the dish contained in the target package or selecting the second termination bit as the target package according to the respective selection probability of the plurality of sample dishes and the selection probability of the second termination bit.

Step S520-3: under the condition that the target sample dish is selected, inputting the characteristics of the target sample dish into the first preset model, obtaining the selected probability of the second termination bit and the respective selected probabilities of the plurality of sample dishes under the condition that the target sample dish is selected and outputting by the first preset model, and returning to the steps: and selecting a target sample dish from the plurality of sample dishes as the dish contained in the target package or selecting the second termination bit as the target package according to the respective selection probability of the plurality of sample dishes and the selection probability of the second termination bit.

Step S520-4: and determining the combination of the selected sample dishes as the target package and outputting the target package when the second ending bit is selected each time.

Referring to fig. 6, fig. 6 is a schematic diagram of an output target package according to an embodiment of the present application. As shown in fig. 6, after receiving the code input by the coding module, the decoding module outputs a first set of selected probabilities, including: the probability of each of the plurality of sample dishes being selected (each sample dish being represented by a circle in fig. 6), and the probability of each of the second termination bits being selected (the second termination bits being represented by boxes in fig. 6).

As shown in fig. 6, when the plurality of sample dishes and the second termination bit are selected according to the respective selection probabilities of the plurality of sample dishes and the selection probability of the second termination bit, one of which the probabilities are the largest may be selected. Or each candidate is selected with a corresponding probability of being selected, for example, in the first set of probabilities of being selected output by the decoding module in fig. 6, the probability of being selected for sample dish 1 is 0.62, the probability of being selected for the second termination bit is 0.05, so that the probability of sample dish 1 being selected in the round of selection is 0.62, and the probability of being selected for the second termination bit in the round of selection is 0.05.

As shown in fig. 6, after the first round of selection, sample dish 1 is input to the decoding module as the selected target sample dish, and features thereof. The decoding module outputs a second set of selected probabilities based on the characteristics of the input sample dish 1, including: the probability of each of the plurality of sample dishes being selected (each sample dish being represented by a circle in fig. 6), and the probability of each of the second termination bits being selected (the second termination bits being represented by boxes in fig. 6).

As shown in fig. 6, the selection is made for the second set of selected probabilities based on the same manner. After a second round of selection, sample dish 3 is input to the decoding module as the selected target sample dish, which is characterized by the sample dish. The decoding module outputs a third set of selected probabilities based on the characteristics of the input sample dish 3, including: the probability of each of the plurality of sample dishes being selected (each sample dish being represented by a circle in fig. 6), and the probability of each of the second termination bits being selected (the second termination bits being represented by boxes in fig. 6).

As shown in fig. 6, the selection is made for the third set of selected probabilities based on the same manner. After a third round of selection, the second termination bit is selected. In this way, the sample dish 1 and the sample dish 3, which have been selected before the second termination bit is selected, are combined into a target package, and output.

It should be understood that the above-described specific manner of obtaining a target package including sub-steps S120-1 through S120-4 is merely an example of one of the various embodiments of the present application and is not intended to limit the present application. For example, in order to obtain a target package, the obtaining manner shown in fig. 3 may also be adopted. For the description of fig. 3, please refer to the above.

Step S530: and determining a feedback value of the target package, wherein the feedback value of the target package represents the matching degree of the target package and the expected package.

In this embodiment, one possible case is: the feedback value is positively correlated with the degree of matching. In other words, the higher the matching of a target package to a desired package, the greater the feedback value of the target package. Another possible case is: the feedback value is inversely related to the degree of matching. In other words, the higher the matching of a target package to a desired package, the smaller the feedback value of the target package. The following will take the positive correlation between the feedback value and the matching degree as an example.

In this embodiment, a specific way to determine the feedback value of the target package may be: inputting the target package into the feedback function, and determining the output of the feedback function as the feedback value of the target package.

The model training personnel can select feedback functions corresponding to different training targets according to different training targets. For example, if the training objective for the first preset model is: a package generation model for recommending meat and vegetable matching packages is trained. Based on the training objective, the corresponding feedback function should have the following characteristics: when a target package input with meat vegetables and vegetables is provided, outputting a higher feedback value; when a target package input with meat only, or vegetables only, a lower feedback value is output.

For another example, if the training objective for the first preset model is: a package generation model for recommending packages with higher conversion rates is trained. Based on the training objective, the corresponding feedback function should have the following characteristics: outputting a higher feedback value when a target package with higher conversion rate is input; when a target package input with a lower conversion rate is inputted, a lower feedback value is outputted. Where conversion rate refers to a historical rate of packages, in other words, conversion rate may reflect the degree of heat in a package, or may reflect the likelihood that a user is placing a package again.

In order to provide more choices to the user, a plurality of first preset models may be preset, with different types of feedback functions being employed when training for different first preset models. Thus, after training is finished, a plurality of package generation models can be obtained, and the package types generated by the package generation models are different from each other. For example, four first preset models are pre-established, and different types of feedback functions are adopted when training is performed for different first preset models. Thus, after the training is completed, four package generation models shown in fig. 4 are obtained.

Because the feedback function is not bound with a specific merchant, a feedback value is determined based on the feedback function, and a model updating step which is described later is executed according to the feedback value, so that deviation of package quality evaluation in the dimension of the merchant can be eliminated, and generalization performance of a finally obtained package generation model is improved.

In addition, the embodiment of the application provides a specific way for obtaining the feedback function: obtaining a plurality of sample packages, wherein each sample package carries a pre-marked feedback value, and the feedback value carried by the sample package represents the matching degree of the sample package and the expected package; and training the second preset model by taking a plurality of sample packages as input, and determining the second preset model at the end of training as a feedback function.

Along with the above example, if the training objective for the first preset model is: a package generation model for recommending packages with higher conversion rates is trained. A plurality of sample packages may be determined based on the travel log and the next log of a large number of users. Then for each sample package, determining the conversion rate of the sample package; and the conversion of the sample package is marked as a feedback value carried by the sample package. Where conversion may be represented by the frequency of orders for a sample package by the user. For example, a merchant's electronic menu is browsed 20000 times in the last month, wherein the sample package consisting of dish 5, dish 8 and dish 16 is ordered 124 times, the order frequency of the sample package is 124/20000, and thus the conversion rate of the sample package is equal to 124/20000. In addition, the conversion rate may also be directly represented by the number of times the user places a sample package, for example, in an electronic menu of a certain merchant, the sample package consisting of dish 5, dish 8 and dish 16 is placed 124 times in total in the last month, and then the conversion rate of the sample package is equal to 124.

After inputting the sample package carrying the conversion into the second preset model, the second preset model outputs the predicted conversion for the sample package. Thus, the loss value can be determined according to the conversion rate carried by the sample package and the predicted conversion rate, and the second preset model is updated based on the loss value. After updating the second preset model a number of times, the final second preset model may be used as a feedback function. After the feedback function is obtained, the feedback value of the target package may be determined using the feedback function, i.e., its conversion rate may be predicted for the target package.

Step S540: and updating the first preset model according to the feedback value of the target package.

In this embodiment, in the case where the feedback value and the matching degree are positively correlated, if the feedback value is larger, it is indicated that the matching degree between the target package and the desired package output by the first preset model is larger, and accordingly, the first preset model is closer to the package generation model desired by the training person. Therefore, the first preset model can be updated according to the feedback value of the target package based on the training mode of the strategy gradient.

Step S550: taking the updated model as the first preset model, and returning to the step: and inputting the characteristics of each of the plurality of sample dishes into a first preset model, and ending training until the feedback value of the target package output by the first preset model meets a preset condition so as to obtain a package generation model.

In other words, after the step S540, the updated model is used as the first preset model again, and the steps S510 to S540 are circularly performed. In step S550, when the feedback value of the target package output by the first preset model meets the preset condition, it is indicated that the first preset model is close to or meets the training target of the training person, so that the training can be ended, and the trained first preset model is used as the package generating model.

The preset condition may be, for example: the N feedback values output by the first preset model in the continuous N rounds of training are all greater than a first preset threshold, where N is a preset value, for example 100. Or the first preset model has M (M is less than or equal to N) feedback values which are greater than a first preset threshold value in N feedback values output in continuous N rounds of training, wherein M/N is greater than a second preset threshold value.

Up to this point, by executing the training flow including steps S510 to S550 described above, a package generation model is obtained. As described above, the training process is performed for a first preset model to obtain a package generation model. The training process is executed for a plurality of first preset models, and different types of feedback functions are adopted when training is performed for different first preset models. Thus, a plurality of package generation models can be obtained, and the package types generated by the plurality of package generation models are different from each other. These package generation models may be used as an optional means for executing some of the steps in the pushing method provided by the embodiment of the present application.

It should be appreciated that the basic structure of the package generation model obtained at the end of training is the same as the basic structure of the first preset model. The package generation model differs from the first preset model in that the model parameters in the structure are different.

Based on the same inventive concept, an embodiment of the present application provides a push device for recommending packages. Referring to fig. 7, fig. 7 is a schematic diagram of a pushing device for recommending packages according to an embodiment of the present application. As shown in fig. 7, the apparatus includes:

The target merchant determining module 71 is configured to determine a target merchant from a plurality of merchants according to input information of a user.

A candidate dish obtaining module 72 is configured to obtain a plurality of candidate dishes for the target merchant.

The recommended package obtaining module 73 is configured to input the package generating model with the package information corresponding to each of the plurality of candidate dishes, perform feature extraction on each of the dish information through the encoding module of the package generating model to obtain a dish feature corresponding to the dish information, encode the plurality of dish features through the encoding module to obtain a total dish feature corresponding to the plurality of candidate dishes, and decode the total dish feature through the decoding module of the package generating model to obtain a recommended package, where the recommended package includes at least one candidate dish.

A recommended packages pushing module 74 for pushing the recommended packages to the user.

Optionally, the apparatus further comprises:

and the recommendation type display module is used for displaying a plurality of recommendation types to the user.

And the target recommendation type obtaining module is used for obtaining information of a target recommendation type input by a user, wherein the target recommendation type is one of the recommendation types.

The recommended package obtaining module includes:

And the package generation sub-module is used for inputting the dish information corresponding to each of the plurality of candidate dishes into the package generation model aiming at the information of the target recommendation type input by the user so as to generate packages corresponding to the target recommendation type through the package generation model.

Optionally, the encoding module in the package generation model includes a feature extraction subnetwork; the recommended package obtaining module includes:

and the feature extraction sub-module is used for carrying out feature extraction on the text information in the dish information through the feature extraction sub-network to obtain text features corresponding to the text information.

And the splicing sub-module is used for splicing the text characteristics with the digital information in the dish information to obtain dish characteristics corresponding to the dish information.

Optionally, the coding module in the package generation model includes a memory subnetwork; the recommended package obtaining module includes:

And the dish feature input sub-module is used for inputting the plurality of dish features into the memory sub-network respectively in different rounds to obtain the total dish features output by the memory sub-network.

Optionally, the recommended package obtaining module includes:

And the dish total feature input sub-module is used for inputting the dish total feature into the decoding module, and obtaining the respective selected probabilities of the plurality of candidate dishes output by the decoding module and the selected probability of a first termination bit, wherein the first termination bit indicates that the recommended package is generated.

The probability selection sub-module is used for selecting N selected probabilities with larger probability according to the selected probabilities of the candidate dishes and the selected probability of the first termination bit; wherein N is an integer greater than or equal to 1.

The target candidate dish feature input sub-module is used for determining the dish corresponding to each selected probability of the N selected probabilities as a target candidate dish and inputting the feature of the target candidate dish into the decoding module.

The first circulation sub-module is used for obtaining the selected probability of the first termination bit and the respective selected probabilities of the plurality of candidate dishes under the condition that the target candidate dishes are selected and output by the decoding module, and returning to the steps: according to the respective selected probabilities of the plurality of candidate dishes and the selected probability of the first termination bit, selecting N selected probabilities with larger probability; and determining the dish corresponding to each selected probability as a target candidate dish according to each selected probability in the N selected probabilities.

And the dish combination sub-module is used for combining dishes corresponding to each selected probability before the first termination bit is selected when the N selected probabilities selected each time contain the selected probability corresponding to the first termination bit, so as to obtain a recommended package.

Optionally, the apparatus further comprises:

and the training module is used for training the package generation model.

The training module comprises:

and the sample dish feature input sub-module is used for inputting dish features of a plurality of sample dishes into the first preset model.

And the target package obtaining sub-module is used for obtaining a target package output by the first preset model, wherein the target package comprises at least one sample dish in the plurality of sample dishes.

And the feedback value determining submodule is used for determining the feedback value of the target package, and the feedback value of the target package represents the matching degree of the target package and the expected package.

And the model updating sub-module is used for updating the first preset model according to the feedback value of the target package.

The second circulation sub-module is used for taking the updated model as the first preset model, and returning to the steps: and inputting the characteristics of each of the plurality of sample dishes into a first preset model, and ending training until the feedback value of the target package output by the first preset model meets a preset condition so as to obtain a package generation model.

Optionally, the apparatus further comprises:

The sample package obtaining module is used for obtaining a plurality of sample packages, each sample package carries a feedback value marked in advance, and the feedback value carried by the sample package represents the matching degree of the sample package and the expected package.

And the feedback function training module is used for taking a plurality of sample packages as input, training the second preset model and determining the second preset model at the end of training as a feedback function.

The feedback value determination submodule includes:

And the target package input subunit is used for inputting the target package into the feedback function and determining the output of the feedback function as the feedback value of the target package.

Optionally, the apparatus further comprises:

And the conversion rate determining module is used for determining the conversion rate of each sample package.

And the conversion rate marking module is used for marking the conversion rate of the sample package as a feedback value carried by the sample package.

Optionally, the target package obtaining submodule includes:

The probability obtaining subunit is configured to obtain the respective selected probabilities of the plurality of sample dishes output by the first preset model and the selected probability of a second termination bit, where the second termination bit characterizes that the target package has been generated.

And the selecting subunit is used for selecting a target sample dish from the plurality of sample dishes as the dish contained in the target package or selecting the second termination bit as the target package according to the respective selection probability of the plurality of sample dishes and the selection probability of the second termination bit.

The circulation subunit is configured to input, when the target sample dish is selected, a feature of the target sample dish into the first preset model, obtain, when the target sample dish is selected, a probability of selecting the second termination bit and a probability of selecting each of the plurality of sample dishes, which are output by the first preset model, and return to the steps: and selecting a target sample dish from the plurality of sample dishes as the dish contained in the target package or selecting the second termination bit as the target package according to the respective selection probability of the plurality of sample dishes and the selection probability of the second termination bit.

And the sample dish combination subunit is used for determining the combination of the selected sample dishes as the target package and outputting the combination when the second termination bit is selected each time.

Optionally, the sample dish feature input submodule includes:

An attribute information classification extraction subunit, configured to classify attribute information of each sample dish in the plurality of sample dishes, where attribute information of one sample dish includes at least two of: sample dish name, sample dish category, sample dish price, sample dish sales.

And the attribute information input subunit is used for inputting the classified attribute information into the first preset model.

Based on the same inventive concept, another embodiment of the present application provides a computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the method according to any of the above embodiments of the present application.

Based on the same inventive concept, another embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the steps in the method according to any one of the foregoing embodiments of the present application.

For the device embodiments, since they are substantially similar to the method embodiments, the description is relatively simple, and reference is made to the description of the method embodiments for relevant points.

In this specification, each embodiment is described in a progressive manner, and each embodiment is mainly described by differences from other embodiments, and identical and similar parts between the embodiments are all enough to be referred to each other.

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

Embodiments of the present application are described with reference to flowchart illustrations and/or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations 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 terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, 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.

While preferred embodiments of the present application have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. It is therefore intended that the following claims be interpreted as including the preferred embodiment and all such alterations and modifications as fall within the scope of the embodiments of the application.

Finally, it is further noted that relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or terminal 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 terminal. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or terminal device that comprises the element.

The push method, the device, the electronic equipment and the readable storage medium for recommending packages provided by the application are described in detail, and specific examples are applied to illustrate the principle and the implementation of the application, and the description of the above examples is only used for helping to understand the method and the core idea of the application; meanwhile, as those skilled in the art will have variations in the specific embodiments and application scope in accordance with the ideas of the present application, the present description should not be construed as limiting the present application in view of the above.

Claims

1. A push method of recommending packages, the method comprising:

obtaining a plurality of candidate dishes for the target merchant;

Pushing the recommended packages to a user;

wherein:

the decoding module for decoding the total features of the dishes through the package generation model to obtain recommended packages comprises:

inputting the total characteristics of the dishes into the decoding module, and obtaining the respective selected probabilities of the plurality of candidate dishes output by the decoding module and the selected probability of a first termination bit, wherein the first termination bit indicates that a recommended package is generated;

according to the respective selected probabilities of the plurality of candidate dishes and the selected probability of the first termination bit, selecting N selected probabilities with larger probability; wherein N is an integer greater than or equal to 1;

Determining that a dish corresponding to each selected probability is a target candidate dish according to each selected probability in the N selected probabilities, and inputting the characteristics of the target candidate dish into the decoding module;

Obtaining the probability of selection of the first termination bit and the probability of selection of each of the plurality of candidate dishes under the condition that the target candidate dish is selected and outputted by the decoding module, and returning to the step: according to the respective selected probabilities of the plurality of candidate dishes and the selected probability of the first termination bit, selecting N selected probabilities with larger probability; aiming at each of the N selected probabilities, determining that the dish corresponding to the selected probability is a target candidate dish;

And when the N selected probabilities selected each time contain the selected probabilities corresponding to the first termination bit, combining dishes corresponding to the selected probabilities before the first termination bit is selected to obtain a recommended package.

2. The method according to claim 1, wherein the method further comprises:

presenting a plurality of recommendation types to a user;

obtaining information of a target recommendation type input by a user, wherein the target recommendation type is one of the recommendation types;

Inputting the dish information corresponding to each of the plurality of candidate dishes into a package generation model, including:

And inputting the dish information corresponding to each of the plurality of candidate dishes into the package generation model aiming at the information of the target recommendation type input by the user so as to generate packages corresponding to the target recommendation type through the package generation model.

3. The method of claim 1, wherein the encoding module in the package generation model comprises a feature extraction subnetwork; the feature extraction is performed on each dish information by the coding module of the package generation model to obtain dish features corresponding to the dish information, including:

extracting features of text information in the dish information through the feature extraction sub-network to obtain text features corresponding to the text information;

And splicing the text features with the digital information in the dish information to obtain dish features corresponding to the dish information.

4. The method of claim 1, wherein the coding module in the package generation model comprises a memory subnetwork; the coding module codes the plurality of dish features to obtain dish total features corresponding to the plurality of candidate dishes, including:

and respectively inputting the plurality of dish features into the memory sub-network in different rounds to obtain the total dish features output by the memory sub-network.

5. The method according to any one of claims 1 to 4, further comprising: training a package generation model;

The training package generation model comprises the following steps:

Inputting dish characteristics of a plurality of sample dishes into a first preset model;

obtaining a target package output by the first preset model, wherein the target package comprises at least one sample dish in the plurality of sample dishes;

Determining a feedback value of the target package, wherein the feedback value of the target package represents the matching degree of the target package and the expected package;

Updating the first preset model according to the feedback value of the target package;

Taking the updated model as the first preset model, and returning to the step: and inputting the characteristics of each of the plurality of sample dishes into a first preset model, and ending training until the feedback value of the target package output by the first preset model meets a preset condition so as to obtain a package generation model.

6. The method of claim 5, wherein the method further comprises:

Obtaining a plurality of sample packages, wherein each sample package carries a pre-marked feedback value, and the feedback value carried by the sample package represents the matching degree of the sample package and the expected package;

Training a second preset model by taking a plurality of sample packages as input, and determining the second preset model after training is finished as a feedback function;

The determining the feedback value of the target package includes:

inputting the target package into the feedback function, and determining the output of the feedback function as the feedback value of the target package.

7. The method of claim 6, wherein the method further comprises:

determining, for each of the sample packages, a conversion rate for the sample package;

the conversion of the sample package is marked as a feedback value carried by the sample package.

8. The method of claim 5, wherein obtaining the target package output by the first predetermined model comprises:

Obtaining the respective selected probabilities of the plurality of sample dishes output by the first preset model and the selected probabilities of a second termination bit, wherein the second termination bit represents that the target package is generated;

Selecting a target sample dish from the plurality of sample dishes as the dish contained in the target package or selecting the second termination bit as the target package according to the respective selection probability of the plurality of sample dishes and the selection probability of the second termination bit;

under the condition that the target sample dish is selected, inputting the characteristics of the target sample dish into the first preset model, obtaining the selected probability of the second termination bit and the respective selected probabilities of the plurality of sample dishes under the condition that the target sample dish is selected and outputting by the first preset model, and returning to the steps: selecting a target sample dish from the plurality of sample dishes as the dish contained in the target package or selecting the second termination bit as the target package according to the respective selection probability of the plurality of sample dishes and the selection probability of the second termination bit;

And determining the combination of the selected sample dishes as the target package and outputting the target package when the second ending bit is selected each time.

9. The method of claim 5, wherein inputting the characteristics of each of the plurality of sample dishes into the first predetermined model comprises:

Classifying attribute information of each sample dish of the plurality of sample dishes, the attribute information of one sample dish including at least two of: sample dish name, sample dish category, sample dish price, sample dish sales;

and inputting the classified attribute information into the first preset model.

10. A push device for recommending packages, the device comprising:

The recommended package obtaining module is configured to input the dish information corresponding to each of the plurality of candidate dishes into a package generating model, perform feature extraction on each of the dish information through an encoding module of the package generating model to obtain dish features corresponding to the dish information, encode the plurality of dish features through the encoding module to obtain a total dish feature corresponding to the plurality of candidate dishes, decode the total dish feature through a decoding module of the package generating model to obtain a recommended package, wherein the recommended package includes at least one candidate dish, and decode the total dish feature through a decoding module of the package generating model to obtain the recommended package, and the recommended package comprises:

When the N selected probabilities selected each time contain the selected probabilities corresponding to the first termination bit, combining dishes corresponding to the selected probabilities selected before the first termination bit is selected to obtain a recommended package;

11. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any one of claims 1 to 9.

12. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor performs the steps of the method according to any one of claims 1 to 9.