Object combination determination method, device, and electronic device

Abstract

The present application discloses a method for determining an object combination, which belongs to the field of computer technology and helps to improve the universality of object combinations. The method includes: determining the first vector representation of each object in the target object combination through a combination relationship learning network in a preset object combination quality recognition model, wherein the network parameters of the combination relationship learning network are initialized according to the corresponding network parameters of the pre-trained combination model, and the combination model is trained based on training samples constructed based on historical object combinations; determining the second vector representation of each object matching in the target object combination; determining the combination quality score of the target object combination based on the first vector representation and the second vector representation of each object in the target object combination through a combination quality classification network in the object combination quality recognition model. The present application helps to improve the universality of high-quality object combinations identified by the model by learning the matching information in the historical combination data and combining the universal feature training model.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of computer technology, and in particular, to a method, device, electronic device, and computer-readable storage medium for determining an object combination. Background Art

[0002] In the prior art, e-commerce platforms or applications usually recommend objects to users in the form of object combinations to improve the efficiency of users in selecting objects. For example, in a dish ordering client, dish combinations are usually displayed to users to improve the efficiency of users in ordering dishes. The method of determining dish combinations and the degree to which the dish combinations displayed to users improve the efficiency of users in ordering dishes are directly affected. One method in the prior art calculates the dish combinations based on the calorie information of the dishes and the calories that the user needs to obtain from the dish combinations, and then selects the dish combinations that match the user's taste based on the taste of the dish combinations, and displays the selected dish combinations to the user for the user to choose, so as to improve the efficiency of users in selecting objects. This dish combination determination method in the prior art is not universal and is not suitable for merchants to determine dish combinations that are generally suitable for users based on the dishes in the menu.

[0003] It can be seen that a more universal method for determining object combinations needs to be provided. Summary of the invention

[0004] The embodiment of the present application provides a method for determining an object combination, which helps to improve the universality of the object combination.

[0005] In order to solve the above problems, in a first aspect, an embodiment of the present application provides a method for determining an object combination, including:

[0006] The first vector representation of each object in the target object combination is determined by a combination relationship learning network in a preset object combination quality recognition model, wherein the network parameters of the combination relationship learning network are initialized according to the corresponding network parameters of the pre-trained combination model, and the combination model is trained by the following method: constructing a plurality of first positive samples according to the historical combination of objects; for each of the first positive samples, performing coding prediction on any object in the first positive sample to obtain a prediction result; determining the similarity between the object corresponding to the prediction result and the coded object; calculating the learning error of the combination model according to the combination probability corresponding to the prediction result and the similarity, and optimizing the network parameters of the combination model with the goal of minimizing the learning error, and training the combination model;

[0007] Determine a second vector representation of a match for each of the objects in the target object combination;

[0008] The combination quality score of the target object combination is determined based on the first vector representation and the second vector representation of each of the objects in the target object combination through the combination quality classification network in the object combination quality identification model.

[0009] In a second aspect, an embodiment of the present application provides an object combination determination device, including:

[0010] A first vector representation determination module is used to determine the first vector representation of each object in the target object combination through a combination relationship learning network in a preset object combination quality recognition model, wherein the network parameters of the combination relationship learning network are initialized according to the corresponding network parameters of the pre-trained combination model, and the combination model is trained by the following method: constructing a number of first positive samples according to the historical combination of objects; for each of the first positive samples, performing coding prediction on any object in the first positive sample to obtain a prediction result; determining the similarity between the object corresponding to the prediction result and the coded object; calculating the learning error of the combination model according to the combination probability corresponding to the prediction result and the similarity, and optimizing the network parameters of the combination model with the goal of minimizing the learning error, and training the combination model;

[0011] A second vector representation determination module, used to determine a second vector representation matching each of the objects in the target object combination;

[0012] A combination quality score determination module is used to determine the combination quality score of the target object combination based on the first vector representation and the second vector representation of each object in the target object combination through the combination quality classification network in the object combination quality identification model.

[0013] In a third aspect, an embodiment of the present application further discloses an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the object combination determination method described in the embodiment of the present application when executing the computer program.

[0014] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the object combination determination method disclosed in the embodiment of the present application.

[0015] The object combination determination method disclosed in the embodiment of the present application determines the first vector representation of each object in the target object combination through a combination relationship learning network in a preset object combination quality recognition model, wherein the network parameters of the combination relationship learning network are initialized according to the corresponding network parameters of the pre-trained combination model, and the combination model is trained by the following method: constructing a number of first positive samples according to the historical combination of objects; for each of the first positive samples, performing coding prediction on any object in the first positive sample to obtain a prediction result; discriminating the similarity between the object corresponding to the prediction result and the coded object; calculating the learning error of the combination model according to the combination probability and the similarity corresponding to the prediction result, and optimizing the network parameters of the combination model with the goal of minimizing the learning error to train the combination model; determining the second vector representation matching each of the objects in the target object combination; determining the combination quality score of the target object combination based on the first vector representation and the second vector representation of each of the objects in the target object combination through the combination quality classification network in the object combination quality recognition model, which helps to improve the universality of object combinations.

[0016] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0018] Figure 1 This is a flow chart of the object combination determination method of the first embodiment of the present application;

[0019] Figure 2 It is a schematic diagram of the structure of the object combination quality identification model in the first embodiment of the present application;

[0020] Figure 3 It is a schematic diagram of the combined model structure in Example 1 of the present application;

[0021] Figure 4 This is a schematic diagram of the application flow of the object combination determination method of the second embodiment of the present application;

[0022] Figure 5This is one of the schematic diagrams of the structure of the object combination determination device of the third embodiment of the present application;

[0023] Figure 6 This is the second structural diagram of the object combination determination device of the third embodiment of the present application;

[0024] Figure 7 A block diagram schematically shows an electronic device for executing the method according to the present application; and

[0025] Figure 8 A storage unit for holding or carrying a program code for implementing the method according to the present application is schematically shown. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0027] Embodiment 1

[0028] The present application discloses a method for determining an object combination, such as Figure 1 As shown, the method includes: steps 110 to 130.

[0029] Step 110, determining a first vector representation of each object in the target object combination by using a combination relationship learning network in a preset object combination quality recognition model.

[0030] In some embodiments of the present application, the first vector represents a vector expression used to express the object name and / or category information.

[0031] In some embodiments of the present application, Figure 2 As shown, the object combination quality recognition model 200 includes: a combination relationship learning network 210, a combination description vector input network 220 and a combination quality classification network 230. The structure and function of each network are introduced below.

[0032] In some embodiments of the present application, the combination relationship learning network 210 is used to extract the first vector representation of each object in the target object combination input to the object combination quality recognition model. Taking the object as a dish as an example, the combination relationship learning network 210 is used to extract the feature vector of the dish name, primary category information, secondary category information and other category information of each dish in the target dish combination input to the object combination quality recognition model. In some embodiments of the present application, the combination relationship learning network 210 includes a construction based on a convolutional neural network and a coding matrix, wherein the convolutional neural network is used to learn the vector representation that best reflects the object name (such as the dish name); the coding matrix is ​​used to extract the vector representation of the primary category information, secondary category information and other category information of the object. The first vector representation of any object in the object combination is spliced ​​by the vector representation of the object name extracted by the convolutional neural network and the vector representation of the category information extracted by the coding matrix.

[0033] In some embodiments of the present application, the combination description vector input network 220 is used to input the second vector representation of each object in the target object combination to the combination quality classification network 230, wherein the second vector representation is a vector expressing the overall combination description information of the corresponding object in the object combination. Still taking the object as a dish as an example, the second vector representation is used to express the number of dishes in the dish combination, the number of servings of each dish, the price and other continuity information. In some embodiments of the present application, the combination description vector input network 220 can use the encoding method in the prior art to encode the above continuity information to obtain the second vector representation.

[0034] In some embodiments of the present application, the combined quality classification network 230 can adopt a convolutional application network with an attention mechanism to perform weighted feature fusion and feature mapping on the first vector representation output by the combined relationship learning network 210 and the second vector representation output by the combined description vector input network 220, and classify and map the hidden layer vector output by the fully linked layer through an activation function, and finally output the quality recognition result of the object combination.

[0035] Before executing the recognition task corresponding to the object combination quality recognition model, it is first necessary to train the object combination quality recognition model.

[0036] In some embodiments of the present application, in order to improve the training efficiency of the model and fully learn the textual meaning of the object name and the combination information (such as collocation information) implied by the object name in the object combination, the present application obtains the initial data of the network parameters of the combination relationship learning network 210 by means of pre-training the combination model. That is, before training the object combination quality identification model, the combination model is first trained. The combination model includes the combination relationship learning network in the object combination quality identification model. By training the combination model, the network parameters of the combination relationship learning network in the object combination quality identification model can be obtained. The network parameters of the combination relationship learning network are initialized according to the corresponding network parameters of the pre-trained combination model.

[0037] like Figure 3 As shown, the combination model 300 includes: an encoding network 310, a generator network 320 and a discriminator network 330, wherein the encoding network 310 further includes: a first encoding sub-network 3101 and a second encoding sub-network 3102. The structure and connection mode of the encoding network 310 and the generator network 320 of the combination relationship learning network 210 are the same as those of the combination relationship learning network 210 in the object combination quality recognition model 200, therefore, the network parameters of the encoding network 310 and the generator network 320 of the combination model 300 obtained by training can be used to initialize the corresponding network parameters of the combination relationship learning network 210.

[0038] The training process of the combined model 300 is described below.

[0039] In some embodiments of the present application, the combination model is trained by the following method: constructing a number of first positive samples according to historical combinations of objects; for each of the first positive samples, performing coding prediction on any object in the first positive sample to obtain a prediction result; judging the similarity between the object corresponding to the prediction result and the coded object; calculating the learning error of the combination model according to the combination probability corresponding to the prediction result and the similarity, and optimizing the network parameters of the combination model with the goal of minimizing the learning error to train the combination model.

[0040] In order to improve the accuracy of learning object combination information of the combination model, in the embodiment of the present application, object combination samples are determined according to historical object order data, and then the object combination information is learned by training a neural network. Taking the object as a dish as an example, the object is a dish, and the construction of a plurality of first positive samples according to the historical combination of the object includes: constructing a plurality of first positive samples according to the dishes included in the historical dish orders.

[0041] The dish order data of all merchants in a specified historical period can be analyzed, and object combination samples can be generated based on some or all of the dishes in each dish order data. The inventors found through analysis of historical dish order data that the number of dish types in 97% of the dish orders does not exceed 5. Therefore, in order to improve the efficiency of model training and reduce the amount of data processing, in some embodiments of the present application, five dishes in each dish order are taken in descending order of price to form a dish combination sample. If there are less than five dishes in a dish order, all dishes in the dish order are taken to form a dish combination sample.

[0042] In some embodiments of the present application, in order to improve the efficiency of model training and the reliability of training samples, further, after determining the object combination sample (i.e., the dish combination sample), it is also necessary to clean the dish order data. The method of constructing several first positive samples according to the dishes included in the historical dish orders includes: determining several dish combinations according to the dishes included in the historical dish orders; cleaning the several dish combinations based on the click volume and / or order quantity of the dish combinations, and determining the dish combinations that meet the preset conditions as the first positive samples. For example, for each dish combination, the average click rate and average visit purchase rate of the dish combination within the historical 30 days are further determined, and the dish combination whose click rate and visit purchase rate meet the preset condition threshold within the historical 30 days is selected as the training sample for training the combination model. In some other embodiments of the present application, the dish combination of a single merchant can also be cleaned according to the number of orders for the merchant's dish combination. For example, when the number of orders for a certain dish combination of a merchant in the past 30 days is less than the preset order quantity threshold, the dish combination of the merchant is filtered out.

[0043] Each dish combination after cleaning can be used as a sample data of a first positive sample. According to the dish order data, several first positive samples can be constructed according to the above method.

[0044] Next, a specific implementation plan of the combined model for training the above-mentioned first positive samples is described. In some embodiments of the present application, the combined model is implemented based on the principle of generator and discriminator. The generator is an object in the combination of random mask objects, and then the model predicts the object in the masked position. The discriminator is used to discriminate the category similarity between the predicted object and the real object, and the combined value of the two constitutes the training target of the model. Therefore, the training process of the combined model is a process of repeatedly executing coding prediction, result discrimination, and parameter optimization.

[0045] First, the network parameters of the combined model are randomly initialized and the learning task is performed based on the current network parameters.

[0046] In some embodiments of the present application, coding prediction is performed on any object in the first positive sample to obtain a prediction result, including: determining the third vector representation of each object in the first positive sample through the encoding network in the combined model; splicing the third vector representation of each object in the first positive sample, and coding the third vector representation of any object in the first positive sample in the spliced ​​vector to obtain a fourth vector representation corresponding to the first positive sample; predicting the coded object in the first positive sample based on the fourth vector representation through the generator network in the combined model to obtain a prediction result. Taking the first positive sample as the dish combination Item1 as an example, the technical solution for coding prediction of any dish in the dish combination Item1 is described in detail below.

[0047] Assume that there are five dishes (i.e., objects), namely: rice, soup, Kung Pao chicken, stir-fried pork with chili peppers, and blanched lettuce.

[0048] First, by executing the learning task corresponding to the first encoding subnetwork 3101, the names of the five dishes in the dish combination Item1 are feature extracted and mapped to obtain the dish name vector representation F of each of the five dishes. name At the same time, by executing the learning task corresponding to the second encoding subnetwork 3102, the first-level category labels and second-level category labels of the five dishes in the dish combination Item1 are respectively extracted and mapped to obtain the vector representation F of the category information of each of the five dishes. cate , the vector representation of each dish is F name and F cate After splicing, the third vector representation of each dish is obtained, which is denoted as F in the following text. sep_i .

[0049] Then, the third vector representation of the five dishes in the dish combination Item1 (such as F sep_1 、F sep_2 、F sep_3 、F sep_4 and F sep_5 ) to obtain a concatenated vector, for example, represented by F Item1 =(F sep_1 , F sep_2 , F sep_3 , F sep_4 , F sep_5 ). Then, for any third vector representation in the concatenated vector (such as the third vector representation F of the second dish), sep_2 ) is coded to obtain the fourth vector representation corresponding to the dish combination Item1, for example, represented as F Item1 '=(F sep_1 , H, Fsep_3 , F sep_4 , F sep_5 ). Wherein, H is a preset vector for coding, for example, a vector whose values ​​of each dimension are all 0.

[0050] Then, the fourth vector representation is predicted by the generator network 320 in the combined model 300. In some embodiments of the present application, the generator network in the combined model predicts the coded object in the first positive sample based on the fourth vector representation to obtain a prediction result, including: performing feature weighted fusion on the fourth vector representation by the generator network in the combined model to obtain a vector representation of the predicted object corresponding to the coded position vector; and determining the prediction result according to the vector distance between the vector representation of the predicted object and the candidate predicted object. The candidate predicted objects include the coded object and a predicted number of objects randomly selected from the prediction object list except for the objects in the first positive sample.

[0051] For example, the fourth vector representation (such as F Item1 ') perform feature weighted fusion, and represent the third vector corresponding to the coded generator network 320, that is, F sep_2 The output at the position is used as a vector representation of the predicted dish, for example, represented as h mask . Further, the candidate predicted dishes of the coded dishes in the dish combination Item1 are determined according to the preset dish list. For example, any three dishes in the preset dish list other than the dishes in the dish combination Item1 are determined as candidate predicted dishes, and at the same time, the coded dishes in the dish combination Item1 are also used as candidate predicted dishes, and the vector representation h of the predicted dishes output by the calculation generator network 320 is calculated. mask The vector distances of the vector representations of the candidate predicted dishes are respectively determined, and the candidate predicted dish with the smallest vector distance (i.e., the highest prediction combination probability) is determined as the prediction result. In the specific implementation of this application, the prediction result may be a coded dish or other dish. The prediction result can be represented by a dish identifier. mask The smaller the vector distance, the higher the quality of the dish combination composed of the dish and the dishes in the dish combination Item1 except the coded dish. For example, when the third vector of the dish "Kung Pao Chicken" in the dish combination composed of the dishes "Rice", "Soup", "Kung Pao Chicken", "Stir-fried Pork with Spicy Peppers", and "Boiled Lettuce" is coded, h mask The corresponding predicted candidate dishes include: "Kung Pao Chicken", "Fish-flavored Shredded Pork", "Coke" and "Vinegar-fried Shredded Potatoes". After calculation, it is found that the vector representation of "Fish-flavored Shredded Pork" is consistent with h maskThe vector distance of is the smallest, indicating that "Fish-flavored Shredded Pork" and "Rice", "Soup", "Stir-fried Pork with Spicy Peppers" and "Boiled Lettuce" may also form a high-quality dish combination.

[0052] Next, the similarity of the prediction results is determined by the discriminator network of the combined network 300. In some embodiments of the present application, the similarity between the object corresponding to the prediction result and the coded object is determined by: using the discriminator network in the combined model, according to the similarity between the category information of the prediction result and the category information of the coded object, the similarity between the object corresponding to the prediction result and the coded object is determined. Taking the object as a dish as an example, the secondary category of the dish in the prediction result (such as the aforementioned h mask The corresponding dish "Fish-flavored Shredded Pork" has a secondary category of "meat dish") and the secondary category of the coded dish (such as the dish "Kung Pao Chicken" has a secondary category of "meat dish"), the greater the similarity, the greater the similarity between the predicted result and the coded dish, which means the higher the accuracy of the predicted result.

[0053] Afterwards, the learning error of the combination model is calculated based on the combination probability corresponding to the prediction result and the similarity, and the network parameters of the combination model are optimized with the goal of minimizing the learning error to train the combination model. The combination probability corresponding to the prediction result is used to indicate the probability of obtaining a high-quality object combination by combining the prediction result with objects other than the coded object in the object combination; the similarity is used to indicate the category similarity between the prediction result and the coded object. In some embodiments of the present application, the combined value of the combination probability corresponding to the prediction result and the similarity is used as the training target of the combination model. For example, the loss function of the combination model can be expressed as:

[0054] Wherein, i is the label of the first positive sample, N is the number of the first positive samples, LG(i) represents the loss value of the generator network corresponding to the i-th first positive sample, that is, the error between the prediction result and the coded object (for example, it can be represented by the combined probability difference), and LD(i) represents the loss value of the discriminator network corresponding to the i-th first positive sample, that is, the error between the prediction result and the category information of the coded object (for example, it can be represented by (1-similarity)).

[0055] In some embodiments of the present application, the combined probability corresponding to the i-th first positive sample can be calculated by the following formula:

[0056]

[0057] Among them, F input represents the input feature of the generator network corresponding to the current first positive sample, h maskrepresents the output vector of the generator network 320 corresponding to the coded third vector representation position, θ G is the generator network parameter, g target The third vector representation represents the coded object, h is the prediction result, and H is the vector set used for coding.

[0058] The loss function of the model is the learning error of the combined model. Next, with the goal of minimizing the learning error, the random gradient descent method or other methods in the prior art are used to optimize the network parameters of the combined model so that the calculated learning error of the combined model reaches a preset threshold, thereby completing the training of the combined model.

[0059] During the online application process, the network parameters of the combination relationship learning network in the object combination quality recognition model are initialized through the network parameters of the encoding network and the generator network of the combination model, and the other network parameters of the object combination quality recognition model are trained based on the combination relationship learning network with the network parameters initialized.

[0060] It can be seen from the above description that, through the combination relationship learning network in the object combination quality recognition model, the first vector representation of each object in the target object combination determined includes the name feature and category feature of each object.

[0061] After the training process of the combination model 300 is completed, the network parameters of the encoding network 310 and the generator network 320 of the combination model 300 are used to initialize the corresponding network parameters of the combination relationship learning network 210. Based on the initialized combination relationship learning network 210, other networks in the object combination quality recognition model are trained, such as: the combination quality classification network 230 and the combination description vector input network 220.

[0062] In some embodiments of the present application, the object combination quality recognition model is a classification model. The object combination quality recognition model is trained by the following method: determining a sample set consisting of the plurality of negative samples and the plurality of positive samples, wherein the positive sample includes a first positive sample; for each training sample in the sample set, determining the first vector representation of each object in the training sample through a combination relationship learning network in a preset object combination quality recognition model, and obtaining the second vector representation of each object in the training sample through the combination description vector input network, wherein the network parameters of the combination relationship learning network are initialized according to the corresponding network parameters of the combination model; for each training sample in the sample set, obtaining the combination quality estimation value of the training sample based on the first vector representation and the second vector representation of each object in the training sample through the combination quality classification network; calculating the model error of the object combination quality recognition model according to the error between the combination quality estimation value of each training sample in the sample set and the combination quality true value indicated by the sample label of the training sample, and optimizing the network parameters of the object combination quality recognition model with the goal of minimizing the model error, and training the object combination quality recognition model.

[0063] First, training samples need to be constructed.

[0064] Training a classification model requires positive samples and negative samples. The positive samples described in the embodiment of the present application at least include the first positive sample of the training combination model.

[0065] In some embodiments of the present application, the positive sample also includes: a second positive sample, and the determination of the sample set consisting of the several negative samples and the several positive samples includes: constructing several second positive samples according to the object combination configured by the merchant and whose order quantity meets the preset conditions. For example, a dish combination configured by the merchant is determined as a candidate positive sample, and then the candidate positive samples are screened according to the historical order quantity of the dish combination, and the candidate positive samples whose historical order quantity is greater than a preset quantity threshold (such as the historical order quantity is greater than 0) are selected as the second positive sample.

[0066] The determining of the sample set consisting of the plurality of negative samples and the plurality of positive samples further includes: constructing a plurality of negative samples according to the object history combination blacklist. For example, a dish combination with a history order quantity of 0 may be added to the dish combination blacklist, and / or a dish combination blacklist may be generated according to user review data, and then a negative sample may be generated according to the dish combination in the dish combination blacklist.

[0067] In some embodiments of the present application, positive samples and negative samples can be identified by sample labels, wherein the sample label of the positive sample is used to indicate that the object combination in the positive sample is a high-quality object combination, and the sample label of the negative sample is used to indicate that the object combination in the negative sample is a low-quality object combination.

[0068] After determining a sample set consisting of a first positive sample, a second positive sample and a negative sample, the network parameters of the combination relationship learning network 210 in the object combination quality recognition model 200 are initialized according to the corresponding network parameters of the combination model 300, and the network parameters of the combination description vector input network 220 and the combination quality classification network 230 in the object combination quality recognition model 200 are randomly initialized. After that, the object combination quality recognition model 200 is trained based on the determined sample set.

[0069] During the training process of the object combination quality recognition model 200 , the network parameters of the combination relationship learning network 210 remain unchanged. The training process of the object combination quality recognition model 200 is actually a combined training process of the combination description vector input network 220 and the combination quality classification network 230 .

[0070] Taking dishes as an example, during the training process of the object combination quality recognition model 200, for each training sample in the sample set, the first vector representation of each dish in the training sample is determined through the combination relationship learning network 210, and the second vector representation of each dish in the training sample is obtained through the combination description vector input network 220; then, the first vector representation and the second vector representation of each dish are concatenated and input into the combination quality classification network 230; then, the combination quality classification network 230 performs weighted fusion based on the concatenated input first vector representation and the second vector representation, performs feature mapping through the fully linked layer, and finally, obtains the combination quality estimate of the input training sample through classification function mapping.

[0071] During the training process of the object combination quality recognition model 200, the model error of the object combination quality recognition model is calculated according to the error between the combination quality estimation value of each training sample in the sample set and the combination quality true value of the training sample. For example, by formula:

[0072] The model error is calculated, where N represents the total number of samples, x represents the input features of the object combination (such as dish name, category information, number of objects in the combination, combination price, etc.), y represents the true value of the combination quality of the i-th training sample, and p(x) represents the estimated value of the combination quality of the i-th training sample. Afterwards, the network parameters other than the combination relationship learning network 210 in the object combination quality recognition model can be optimized by using the prior art methods such as the gradient descent method or the error back conduction method to reduce the model error until the model error converges to a preset range, completing the model training process.

[0073] Among them, the combination quality estimation value p(x) of the training sample is determined according to the first vector representation output by the combination relationship learning network 210 and the description information of each object in the training sample. The network parameter value learned through the training process of the object combination quality identification model 200 reflects the influence weight of the description information, name and category information of each object in the object combination on the combination.

[0074] Step 120: Determine a second vector representation of a match between each of the objects in the target object combination.

[0075] In some embodiments of the present application, the second vector representation is used to express the descriptive information of the object in the combination, and the descriptive information includes: the number of objects in the combination, the number of portions of the current object, and one or more of the combination price. The descriptive information of the objects in the combination described in the embodiments of the present application can be encoded by the encoding method in the prior art to obtain the second vector representation. In some embodiments of the present application, the descriptive information of each object in the target object combination is encoded by the combination description vector input network 220 in the object combination quality identification model. Taking the object as a dish as an example, the descriptive information of the object may include the number of portions of the dish in the dish combination (such as 2 portions of rice), the sum of the prices corresponding to the respective category labels under the secondary label of the dish (such as the total price of rice), and the number of portions corresponding to the respective category labels under the secondary label of the dish. The present application does not limit the network structure of the combination description vector input network 220.

[0076] The second vector representation of the object matching described in the embodiment of the present application reflects the overall characteristics of the object in the object combination.

[0077] Step 130: Determine a combination quality score of the target object combination based on the first vector representation and the second vector representation of each of the objects in the target object combination through a combination quality classification network in the object combination quality identification model.

[0078] After obtaining the first vector representation of the target combination object through the combination relationship learning network 210 of the object combination quality recognition model and obtaining the second vector representation of the target combination object through the combination description vector input network 220, the first vector representation and the second vector representation of each object in the target object combination are concatenated into a vector, and then input into the combination quality classification network in the object combination quality recognition model. The combination quality classification network performs weighted fusion and feature mapping on the input vector based on the trained network parameters, and finally outputs a combination quality estimation value.

[0079] In some embodiments of the present application, the combination quality estimate is a two-dimensional vector, which is used to represent the probability that the target object combination is a high-quality combination and a low-quality combination, respectively.

[0080] The object combination determination method disclosed in the embodiment of the present application determines the first vector representation of each object in the target object combination through a combination relationship learning network in a preset object combination quality recognition model, wherein the network parameters of the combination relationship learning network are initialized according to the corresponding network parameters of the pre-trained combination model, and the combination model is trained by the following method: constructing a number of first positive samples according to the historical combination of objects; for each of the first positive samples, performing coding prediction on any object in the first positive sample to obtain a prediction result; discriminating the similarity between the object corresponding to the prediction result and the coded object; calculating the learning error of the combination model according to the combination probability and the similarity corresponding to the prediction result, and optimizing the network parameters of the combination model with the goal of minimizing the learning error to train the combination model; determining the second vector representation matching each of the objects in the target object combination; determining the combination quality score of the target object combination based on the first vector representation and the second vector representation of each of the objects in the target object combination through the combination quality classification network in the object combination quality recognition model, which helps to improve the universality of object combinations.

[0081] The object combination determination method disclosed in the embodiment of the present application learns historical combination data to learn the collocation information of the combination, and trains a neural network model in combination with the common features of the historical combinations. The target object combination is then identified based on the trained neural network model to determine the probability that the target object combination is a high-quality object combination, so that the high-quality object combination determined by this method is suitable for most users, rather than just individual users.

[0082] Furthermore, by training an end-to-end network, we learn the collocation knowledge between combined objects and transfer the learned knowledge to estimate the quality of object combinations. The enriched knowledge can improve the recognition accuracy and robustness of the object combination quality recognition model.

[0083] Embodiment 2

[0084] An embodiment of the present application discloses a method for determining an object combination in a specific application scenario. In a specific application scenario, for a pre-established object set, the object set includes multiple objects. By executing the object combination determination method described in the embodiment, a high-quality object combination composed of the objects in the object set is determined; or, given one or more target objects, a high-quality object combination including the target object can be determined.

[0085] like Figure 4 As shown, the method includes: steps 410 to 470.

[0086] Step 410: taking the target object as the currently determined combination object.

[0087] In a specific application process, a target object is first selected from the object set, and then other objects that form a high-quality object combination with the object are determined, and the target object is the currently determined combination object.

[0088] Taking the scenario of determining a dish combination as an example, the preset object set may be a menu or a dish list. The merchant may specify to generate a dish combination including a specified dish, and specify the number of dishes included in the dish combination. Further, the specified dish may be determined as the currently determined combination object.

[0089] Step 420: Determine objects in a preset object set other than the currently determined combined object as a candidate object set.

[0090] For example, other dishes in the merchant's menu except the designated dishes are taken as a candidate dish set.

[0091] Step 430: determine each object in the candidate object set and combine it with the target object composed of the currently determined combined objects.

[0092] Next, each dish in the candidate dish set is combined with one or more currently selected dishes to form a target combination object. For example, if the candidate object set includes 10 dishes, 10 target combination objects can be formed.

[0093] Step 440: For each target object combination, determine the combination quality score of the target object combination by using a preset object combination quality recognition model.

[0094] Next, for each target combination object, a combination quality score of the target object combination is determined by using a preset object combination quality recognition model.

[0095] Among them, for each target object combination, the combination quality score of the target object combination is determined by a preset object combination quality recognition model, including: determining a first vector representation of each object in the target object combination through a combination relationship learning network in the preset object combination quality recognition model; determining a second vector representation of a match of each object in the target object combination; and determining a combination quality score of the target object combination based on the first vector representation and the second vector representation of each object in the target object combination through a combination quality classification network in the object combination quality recognition model.

[0096] For example, for the 10 dish combinations determined in the above steps, the input features of each dish combination are determined respectively, and the input features of each dish combination are respectively input into the object combination quality recognition model, and the object combination quality recognition model will output the combination quality score corresponding to each dish combination. Among them, the input features of each dish combination include: the name of each dish in the dish combination, category information, number of dishes, total price of the dish combination, and other information.

[0097] For a dish combination, the specific implementation method of determining the combination quality score of the target object combination by using a preset object combination quality recognition model is shown in Example 1, which will not be described in detail in this embodiment.

[0098] Step 450: determine all objects in the target object combination with the highest combination quality score as currently determined combination objects.

[0099] After determining the combined quality scores of the above-mentioned multiple target object combinations (such as the above-mentioned 10 dish combinations), the target object combination with the highest combined quality score is selected as the current optimal object combination, and all objects in the current optimal object combination are used as the currently determined combination objects.

[0100] Step 460 , determining whether the number of combined objects reaches a preset object number threshold, if so, executing step 470 , otherwise, jumping to executing step 420 .

[0101] Usually, an object combination includes multiple objects. Taking the number of dishes L in the dish combination specified by the merchant as an example, after determining the current optimal object combination, it is necessary to further determine whether the number of dishes included in the current optimal object combination is greater than or equal to the number of combined dishes L specified by the merchant. If the number of dishes included in the previous optimal object combination is less than the number of combined dishes L specified by the merchant, jump to execution step 420 and continue to select dishes to add to the combination. If the number of dishes included in the previous optimal object combination is greater than or equal to the number of combined dishes L specified by the merchant, the determination process of the current dish combination ends.

[0102] Step 470: In response to the number of the currently determined combination objects being equal to a preset number value, outputting the currently determined combination objects.

[0103] In some embodiments of the present application, after determining the current optimal dish combination including the number of dishes that meet the merchant's set conditions, all currently determined combination objects included in the current optimal dish combination can be output in a preset manner, for example, displayed on the merchant's client.

[0104] The object combination determination method disclosed in the embodiment of the present application, by taking the target object as the currently determined combination object, determining the objects in the preset object set other than the currently determined combination object as the candidate object set; determining each object in the candidate object set, respectively, and the target object combination composed of the currently determined combination object; then, for each target object combination, determining the combination quality score of the target object combination through the preset object combination quality recognition model, and determining all the objects in the target object combination with the highest combination quality score as the currently determined combination object; in response to the number of the currently determined combination objects being less than the preset number value, jumping to the step of executing the determination of the objects in the preset object set other than the currently determined combination objects as the candidate object set; in response to the number of the currently determined combination objects being equal to the preset number value, outputting the currently determined combination objects, it is possible to output high-quality object combinations according to a given object set and a specified number of combination objects. For object providers through a network platform, high-quality object combinations can be output according to the object sets matched by each object provider, so that the object providers can be displayed to users to improve the efficiency of user object selection.

[0105] Embodiment 3

[0106] The present application discloses a device for determining a combination of objects, such as Figure 5 As shown, the device comprises:

[0107] The first vector representation determination module 510 is used to determine the first vector representation of each object in the target object combination through the combination relationship learning network in the preset object combination quality recognition model, wherein the network parameters of the combination relationship learning network are initialized according to the corresponding network parameters of the pre-trained combination model, and the combination model is trained by the following method: constructing a plurality of first positive samples according to the historical combination of objects; for each of the first positive samples, performing coding prediction on any object in the first positive sample to obtain a prediction result; determining the similarity between the object corresponding to the prediction result and the coded object; calculating the learning error of the combination model according to the combination probability corresponding to the prediction result and the similarity, and optimizing the network parameters of the combination model with the goal of minimizing the learning error, and training the combination model;

[0108] A second vector representation determination module 520, configured to determine a second vector representation matching each of the objects in the target object combination;

[0109] The combination quality score determination module 530 is used to determine the combination quality score of the target object combination based on the first vector representation and the second vector representation of each object in the target object combination through the combination quality classification network in the object combination quality identification model.

[0110] In some embodiments of the present application, performing coding prediction on any object in the first positive sample to obtain a prediction result includes:

[0111] Determine a third vector representation of each object in the first positive sample through the encoding network in the combined model;

[0112] splicing the third vector representations of each object in the first positive sample, and performing coding processing on the third vector representation of any object in the first positive sample in the spliced ​​vector to obtain a fourth vector representation corresponding to the first positive sample;

[0113] The coded object in the first positive sample is predicted based on the fourth vector representation by the generator network in the combined model to obtain a prediction result.

[0114] In some embodiments of the present application, the determining the similarity between the object corresponding to the prediction result and the coded object includes:

[0115] The discriminator network in the combined model is used to discriminate the similarity between the object corresponding to the prediction result and the coded object according to the similarity between the category information of the prediction result and the category information of the coded object.

[0116] In some embodiments of the present application, the object is a dish, and the step of constructing a plurality of first positive samples according to the object history combination includes:

[0117] According to the dishes included in the historical dish orders, a number of first positive samples are constructed.

[0118] In some embodiments of the present application, constructing a plurality of first positive samples according to dishes included in historical dish orders includes:

[0119] Determine a number of dish combinations based on dishes included in historical dish orders;

[0120] Based on the number of clicks and / or the number of orders for the dish combinations, the several dish combinations are cleaned, and a dish combination that meets a preset condition is determined as a first positive sample.

[0121] In some embodiments of the present application, the object combination quality recognition model is trained by the following method:

[0122] Determine a sample set consisting of the plurality of negative samples and a plurality of positive samples, wherein the positive samples include a first positive sample;

[0123] For each training sample in the sample set, determine a first vector representation of each object in the training sample through a combination relationship learning network in a preset object combination quality recognition model, and obtain a second vector representation of each object in the training sample through the combination description vector input network, wherein network parameters of the combination relationship learning network are initialized according to corresponding network parameters of the combination model;

[0124] For each training sample in the sample set, obtaining, by the combined quality classification network, a combined quality estimate of the training sample based on the first vector representation and the second vector representation of each object in the training sample;

[0125] The model error of the object combination quality recognition model is calculated based on the error between the combination quality estimate of each training sample in the sample set and the true value of the combination quality indicated by the sample label of the training sample, and the network parameters of the object combination quality recognition model are optimized with the goal of minimizing the model error to train the object combination quality recognition model.

[0126] In some embodiments of the present application, the positive sample further includes: a second positive sample, and the determining of a sample set consisting of the plurality of negative samples and the plurality of positive samples includes:

[0127] Constructing a number of negative samples according to the object history combination blacklist; and constructing a number of second positive samples according to the object combination configured by the merchant and whose order quantity meets the preset conditions.

[0128] In some embodiments of the present application, Figure 6 As shown, the device also includes:

[0129] A first combination object determination module 540, configured to use the target object as the currently determined combination object;

[0130] A candidate object set determination module 550 is used to determine objects in a preset object set other than the currently determined combined object as a candidate object set;

[0131] The candidate target object combination determination module 560 is used to determine the target object combination composed of each object in the candidate object set and the currently determined combination object; and

[0132] A second combination object determination module 570 is configured to, after determining the combination quality score of the target object combination, determine all objects in the target object combination with the highest combination quality score as currently determined combination objects;

[0133] The judgment output module 580 is used to jump to the step of determining objects in the preset object set other than the currently determined combined objects as a candidate object set in response to the number of the currently determined combined objects being less than a preset number value; and output the currently determined combined objects in response to the number of the currently determined combined objects being equal to the preset number value.

[0134] The object combination determination device disclosed in the embodiment of the present application is used to implement the object combination determination method described in the first embodiment of the present application. The specific implementation methods of each module of the device will not be repeated here, and reference can be made to the specific implementation methods of the corresponding steps in the method embodiment.

[0135] The object combination determination device disclosed in the embodiment of the present application determines the first vector representation of each object in the target object combination through a combination relationship learning network in a preset object combination quality recognition model, wherein the network parameters of the combination relationship learning network are initialized according to the corresponding network parameters of the pre-trained combination model, and the combination model is trained by the following method: constructing a plurality of first positive samples according to the object history combination; for each of the first positive samples, performing coding prediction on any object in the first positive sample to obtain a prediction result; discriminating the similarity between the object corresponding to the prediction result and the coded object; calculating the learning error of the combination model according to the combination probability and the similarity corresponding to the prediction result, and optimizing the network parameters of the combination model with the goal of minimizing the learning error to train the combination model; determining the second vector representation matching each of the objects in the target object combination; determining the combination quality score of the target object combination based on the first vector representation and the second vector representation of each of the objects in the target object combination through the combination quality classification network in the object combination quality recognition model, which helps to improve the universality of object combination.

[0136] The object combination determination device disclosed in the embodiment of the present application learns historical combination data, learns the matching information of the combination, and trains a neural network model in combination with the common features of the historical combinations. Then, based on the trained neural network model, the target object combination is identified to determine the probability that the target object combination is a high-quality object combination, so that the high-quality object combination determined by this method is suitable for most users, rather than just for individual users.

[0137] Furthermore, by training an end-to-end network, we learn the collocation knowledge between combined objects and transfer the learned knowledge to estimate the quality of object combinations. The enriched knowledge can improve the recognition accuracy and robustness of the object combination quality recognition model.

[0138] On the other hand, through the object combination determination device disclosed in the embodiment of the present application, it is possible to output a high-quality object combination according to a given object set and a specified number of combined objects. For object providers through the network platform, a high-quality object combination can be output according to the object set matched by each object provider, so that the object provider can show it to the user to improve the efficiency of the user's object selection.

[0139] Each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0140] The above is a detailed introduction to a method and device for determining an object combination provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and a core idea. At the same time, for a person skilled in the art, according to the idea of ​​the present application, there will be changes in the specific implementation method and scope of application. In summary, the content of this specification should not be understood as a limitation on the present application.

[0141] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0142] The various component embodiments of the present application can be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It should be understood by those skilled in the art that a microprocessor or digital signal processor (DSP) can be used in practice to implement some or all functions of some or all components in the electronic device according to the embodiment of the present application. The present application can also be implemented as a device or apparatus program (e.g., computer program and computer program product) for executing part or all of the methods described herein. Such a program implementing the present application can be stored on a computer-readable medium, or can have the form of one or more signals. Such a signal can be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0143] For example, Figure 7An electronic device that can implement the method according to the present application is shown. The electronic device can be a PC, a mobile terminal, a personal digital assistant, a tablet computer, etc. The electronic device traditionally includes a processor 710 and a memory 720 and a program code 730 stored on the memory 720 and can be run on the processor 710, and the processor 710 implements the method described in the above embodiment when executing the program code 730. The memory 720 can be a computer program product or a computer-readable medium. The memory 720 can be an electronic memory such as a flash memory, an EEPROM (electrically erasable programmable read-only memory), an EPROM, a hard disk or a ROM. The memory 720 has a storage space 7201 for the program code 730 of the computer program for executing any method step in the above method. For example, the storage space 7201 for the program code 730 can include various computer programs for implementing the various steps in the above method respectively. The program code 730 is a computer-readable code. These computer programs can be read from one or more computer program products or written into the one or more computer program products. These computer program products include program code carriers such as hard disks, compact disks (CDs), memory cards or floppy disks. The computer program includes a computer readable code, and when the computer readable code is executed on an electronic device, the electronic device is caused to execute the method according to the above embodiment.

[0144] The embodiment of the present application further discloses a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the object combination determination method described in the first embodiment of the present application are implemented.

[0145] Such a computer program product may be a computer-readable storage medium, which may have a computer-readable storage medium having a Figure 7 The memory 720 in the electronic device shown in the figure is similar to the storage segment, storage space, etc. The program code can be compressed and stored in the computer-readable storage medium in an appropriate form. The computer-readable storage medium is generally as shown in the reference Figure 8 The portable or fixed storage unit. Generally, the storage unit includes computer readable code 730', which is a code read by a processor. When the code is executed by the processor, each step in the method described above is implemented.

[0146] The term "one embodiment", "embodiment" or "one or more embodiments" herein means that a particular feature, structure or characteristic described in conjunction with the embodiment is included in at least one embodiment of the present application. In addition, please note that the examples of the term "in one embodiment" here do not necessarily all refer to the same embodiment.

[0147] In the description provided herein, a large number of specific details are described. However, it is understood that the embodiments of the present application can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0148] In the claims, any reference signs placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present application may be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for determining an object combination, characterized in that: include: The first vector representation of each object in the target object combination is determined by a combination relationship learning network in a preset object combination quality recognition model, wherein the network parameters of the combination relationship learning network are initialized according to the corresponding network parameters of the pre-trained combination model, and the combination model is trained by the following method: Constructing a number of first positive samples according to the object history combination; For each of the first positive samples, perform coding prediction on any object in the first positive sample to obtain a prediction result; Determining the similarity between the object corresponding to the prediction result and the coded object; Calculating the learning error of the combined model according to the combined probability corresponding to the prediction result and the similarity, optimizing the network parameters of the combined model with the goal of minimizing the learning error, and training the combined model; Determine a second vector representation of a match for each of the objects in the target object combination; Determining a combination quality score of the target object combination based on the first vector representation and the second vector representation of each of the objects in the target object combination through a combination quality classification network in the object combination quality identification model; The object is a dish, and the step of constructing a plurality of first positive samples according to the object history combination includes: Constructing a number of first positive samples according to the dishes included in the historical dish orders; The step of constructing a plurality of first positive samples according to the dishes included in the historical dish orders includes: Determine a number of dish combinations based on dishes included in historical dish orders; Based on the number of clicks and / or the number of orders for the dish combinations, the several dish combinations are cleaned, and a dish combination that meets a preset condition is determined as a first positive sample.

2. The method according to claim 1, characterized in that The step of performing coding prediction on any object in the first positive sample to obtain a prediction result includes: Determine a third vector representation of each object in the first positive sample through the encoding network in the combined model; splicing the third vector representations of each object in the first positive sample, and performing coding processing on the third vector representation of any object in the first positive sample in the spliced ​​vector to obtain a fourth vector representation corresponding to the first positive sample; The coded object in the first positive sample is predicted based on the fourth vector representation by the generator network in the combined model to obtain a prediction result.

3. The method according to claim 1, characterized in that The step of determining the similarity between the object corresponding to the prediction result and the coded object comprises: The discriminator network in the combined model is used to discriminate the similarity between the object corresponding to the prediction result and the coded object according to the similarity between the category information of the prediction result and the category information of the coded object.

4. The method according to claim 1, characterized in that: The object combination quality recognition model is trained by the following method: Determine a sample set consisting of a plurality of negative samples and a plurality of positive samples, wherein the positive samples include a first positive sample; For each training sample in the sample set, determine a first vector representation of each object in the training sample through a combination relationship learning network in a preset object combination quality recognition model, and obtain a second vector representation of each object in the training sample through a combination description vector input network, wherein network parameters of the combination relationship learning network are initialized according to corresponding network parameters of the combination model; For each training sample in the sample set, obtaining, by the combined quality classification network, a combined quality estimate of the training sample based on the first vector representation and the second vector representation of each object in the training sample; The model error of the object combination quality recognition model is calculated based on the error between the combination quality estimate of each training sample in the sample set and the true value of the combination quality indicated by the sample label of the training sample, and the network parameters of the object combination quality recognition model are optimized with the goal of minimizing the model error to train the object combination quality recognition model.

5. The method according to claim 4, characterized in that The positive sample further includes: a second positive sample, and the step of determining a sample set consisting of the plurality of negative samples and the plurality of positive samples includes: Construct several negative samples based on the object history combination blacklist; And, construct a plurality of second positive samples according to the object combinations configured by the merchant and whose order quantities meet the preset conditions.

6. The method according to any one of claims 1 to 5, characterized in that: Before the step of determining the first vector representation of each object in the target object combination by using the combination relationship learning network in the preset object combination quality recognition model, the method further includes: The target object is the currently determined composite object; Determine objects in a preset object set other than the currently determined combined object as a candidate object set; Determine each object in the candidate object set and combine them with the target object composed of the currently determined combination objects; After the step of determining the combined quality score of the target object combination, the method further includes: Determine all objects in the target object combination with the highest combination quality score as currently determined combination objects; In response to the number of the currently determined combined objects being less than a preset number value, jumping to the step of determining objects in the preset object set other than the currently determined combined objects as a candidate object set; In response to the number of the currently determined combination objects being equal to a preset number value, the currently determined combination objects are output.

7. An object combination determination device, characterized in that: include: The first vector representation determination module is used to determine the first vector representation of each object in the target object combination through a combination relationship learning network in a preset object combination quality recognition model, wherein the network parameters of the combination relationship learning network are initialized according to the corresponding network parameters of the pre-trained combination model, and the combination model is trained by the following method: Constructing a number of first positive samples according to the object history combination; For each of the first positive samples, perform coding prediction on any object in the first positive sample to obtain a prediction result; Determining the similarity between the object corresponding to the prediction result and the coded object; Calculating the learning error of the combined model according to the combined probability corresponding to the prediction result and the similarity, optimizing the network parameters of the combined model with the goal of minimizing the learning error, and training the combined model; A second vector representation determination module, used to determine a second vector representation matching each of the objects in the target object combination; a combination quality score determination module, configured to determine a combination quality score of the target object combination based on the first vector representation and the second vector representation of each of the objects in the target object combination through a combination quality classification network in the object combination quality identification model; The object is a dish, and the step of constructing a plurality of first positive samples according to the object history combination includes: Constructing a number of first positive samples according to the dishes included in the historical dish orders; The step of constructing a plurality of first positive samples according to the dishes included in the historical dish orders includes: Determine a number of dish combinations based on dishes included in historical dish orders; Based on the number of clicks and / or the number of orders for the dish combinations, the several dish combinations are cleaned, and a dish combination that meets a preset condition is determined as a first positive sample.

8. An electronic device comprising a memory, a processor, and a program code stored in the memory and executable on the processor, wherein: When the processor executes the program code, the object combination determination method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium having a program code stored thereon, characterized in that: When the program code is executed by a processor, the steps of the object combination determination method described in any one of claims 1 to 6 are implemented.

Images