Model training methods, information delivery methods, devices, media and equipment

Abstract

The present disclosure relates to a model training method, an information pushing method, an apparatus, a medium and an equipment. The method comprises: obtaining sample data for training, the sample data comprising sample merchant information labeled with a feature label, sample user operation information of the sample merchant information, and preset sequence information labeled with a revenue label; inputting the sample data into a ranking model, and performing the following operations by the ranking model: generating a plurality of candidate lists according to the sample data, and for each candidate list, determining predicted user operation information of merchant information in the candidate list, predicted feature information of the merchant information in the candidate list, and predicted revenue information corresponding to the merchant information in the candidate list; calculating a target loss value according to the sample user operation information, the predicted user operation information, the feature label, the predicted feature information, the revenue label and the predicted revenue information; and adjusting parameters of the ranking model according to the target loss value.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and more specifically, to a model training method, an information push method, an apparatus, a medium, and a device. Background Technology

[0002] With the development of internet technology, more and more users are obtaining content of interest from information pushed by various business platforms. For example, on food delivery platforms, users can select content of interest from advertised and non-advertising merchants and perform corresponding actions. Meanwhile, if a user clicks on an advertised merchant, the platform earns corresponding advertising revenue; if a user successfully places an order after clicking on an advertised or non-advertising merchant, the platform earns corresponding commission revenue.

[0003] Related technologies, within a reinforcement learning framework, can predict the platform's total revenue under different push orders for advertisers and non-advertisers in different states, and select the push order that maximizes revenue for both advertisers and non-advertisers. However, the input to the models trained under the reinforcement learning framework in these technologies is complex, typically consisting of high-dimensional state-action sequences, which is detrimental to agent learning. Summary of the Invention

[0004] The purpose of this disclosure is to provide a model training method, an information push method, an apparatus, a medium, and a device to solve the problems in related technologies.

[0005] To achieve the above objectives, the first part of this disclosure provides a model training method, the method comprising:

[0006] Obtain sample data for training. The sample data includes sample merchant information labeled with feature tags, sample user operation information of the sample merchant information, and preset order information labeled with revenue tags. The sample merchant information includes advertising merchant information and non-advertising merchant information. The preset order information is used to characterize the combined push order of advertising merchant information and non-advertising merchant information.

[0007] The sample data is input into the sorting model, and the sorting model performs the following operations: Based on the sample data, multiple candidate lists are generated to characterize the push order of each combination of merchant information in the sample merchant information, and for each candidate list, the predicted user operation information, the predicted feature information, and the predicted revenue information corresponding to the merchant information in the candidate list are determined.

[0008] The target loss value is calculated based on the sample user operation information, the predicted user operation information, the feature labels, the predicted feature information, the revenue labels, and the predicted revenue information.

[0009] The parameters of the ranking model are adjusted based on the target loss value.

[0010] Optionally, the sample user operation information of the sample merchant information includes the sample click pass rate of the user on the advertising merchant information in the sample merchant information and the sample pull-down probability of the user on the combined push order of each merchant information in the sample merchant information. Determining the predicted user operation information for the merchant information in the candidate list includes:

[0011] Determine the predicted click-through rate for the advertiser information in the candidate list and the predicted pull-down probability for the combined push order of each advertiser information in the candidate list;

[0012] The step of calculating the target loss value based on the sample user operation information, the predicted user operation information, the feature labels, the predicted feature information, the revenue labels, and the predicted revenue information includes:

[0013] The target loss value is calculated based on the sample click pass rate, the sample drop-down probability, the predicted click pass rate, the predicted drop-down probability, the feature label, the predicted feature information, the revenue label, and the predicted revenue information.

[0014] Optionally, calculating the target loss value based on the sample click-through rate, the sample dropdown probability, the predicted click-through rate, the predicted dropdown probability, the feature label, the predicted feature information, the revenue label, and the predicted revenue information includes:

[0015] Calculate the click pass rate loss value based on the sample click pass rate and the predicted click pass rate;

[0016] Calculate the pull-down loss value based on the sample pull-down probability and the predicted pull-down probability;

[0017] Calculate the key information loss value based on the feature labels and the predicted feature information;

[0018] The target predicted revenue value is determined based on the click pass rate loss value, the drop-down loss value, the key information loss value, and the predicted revenue information;

[0019] Calculate the target loss value based on the target predicted return value and the return label.

[0020] Optionally, determining the predicted revenue information corresponding to the merchant information in the candidate list includes:

[0021] Determine the pre-design fee vector corresponding to the advertiser information in the candidate list, and multiply the predicted click-through rate corresponding to the advertiser information with the pre-design fee vector to obtain the predicted revenue information.

[0022] Optionally, the ranking model includes a deep sub-network and a shallow sub-network. The step of generating multiple candidate lists based on the sample data to characterize the push order of combinations of merchant information in the sample merchant information includes:

[0023] The sample user operation information and the sample merchant information are input into the deep sub-network to obtain the merchant information vector output by the deep sub-network;

[0024] The preset order information is converted into an order information vector, and the order information vector and the merchant information vector are input into the shallow sub-network to generate multiple candidate lists that characterize the push order of each combination of merchant information in the sample merchant information.

[0025] Optionally, the deep subnetwork is a remotely running network model, and the shallow subnetwork is a locally running network model.

[0026] Optionally, the step of generating multiple candidate lists based on the sample data to characterize the push order of combinations of merchant information in the sample merchant information includes:

[0027] Based on the sample user operation information of the non-advertising merchant information in the sample merchant information and the non-advertising merchant information, a non-advertising merchant information vector is generated, and based on the sample user operation information of the advertising merchant information in the sample merchant information and the advertising merchant information, an advertising merchant information vector is generated.

[0028] Based on the non-advertising merchant information vector, the advertising merchant information vector, and the preset order information, a candidate list is generated to characterize the push order of each combination of merchant information in the sample merchant information.

[0029] Optionally, generating a candidate list to characterize the push order of each combination of merchant information in the sample merchant information, based on the non-advertising merchant information vector, the advertising merchant information vector, and the preset order information, includes:

[0030] The preset order information is inverted to obtain the target reverse order information;

[0031] The preset sequence information is converted into a first action matrix, and the target reverse sequence information is converted into a second action matrix;

[0032] The first action matrix is ​​multiplied by the advertising merchant information vector to obtain the first state action matrix, and the second action matrix is ​​multiplied by the non-advertising merchant information vector to obtain the second state action matrix.

[0033] Summing the first state action matrix and the second state action matrix yields the target state action matrix;

[0034] Based on the target state action matrix, a candidate list is generated to characterize the push order of each combination of merchant information in the sample merchant information.

[0035] Optionally, the feature tags of the sample merchant information include at least one of the following tags: merchant delivery fee tag, merchant delivery distance tag, merchant sales volume tag, merchant rating tag, and brand tag used to characterize whether the merchant associated with the sample merchant information is a brand merchant.

[0036] The second part of this disclosure provides an information push method, the method comprising:

[0037] Obtain information on target advertising merchants and target non-advertising merchants to be pushed to;

[0038] The target advertising merchant information and the target non-advertising merchant information are input into the ranking model to obtain the combined push order of the target merchant information and the target non-advertising merchant information, wherein the ranking model is trained by the model training method described in Part 1;

[0039] Based on the combined push order, the target advertising merchant information and the target non-advertising merchant information are pushed.

[0040] A third part of the embodiments of this disclosure provides a model training apparatus, the apparatus comprising:

[0041] The first acquisition module is used to acquire sample data for training. The sample data includes sample merchant information labeled with feature tags, sample user operation information of the sample merchant information, and preset order information labeled with revenue tags. The sample merchant information includes advertising merchant information and non-advertising merchant information. The preset order information is used to characterize the combined push order of advertising merchant information and non-advertising merchant information.

[0042] The first input module is used to input the sample data into the sorting model and perform the following operations through the sorting model: generate multiple candidate lists based on the sample data to characterize the push order of each combination of merchant information in the sample merchant information, and for each candidate list, determine the predicted user operation information, the predicted feature information of the merchant information in the candidate list, and the predicted revenue information corresponding to the merchant information in the candidate list;

[0043] The calculation module is used to calculate the target loss value based on the sample user operation information, the predicted user operation information, the feature label, the predicted feature information, the revenue label, and the predicted revenue information;

[0044] An adjustment module is used to adjust the parameters of the ranking model based on the target loss value.

[0045] A fourth part of the embodiments of this disclosure provides an information push device, the device comprising:

[0046] The second acquisition module is used to acquire information on the target advertising merchants and the target non-advertising merchants to be pushed;

[0047] The second input module is used to input the target advertising merchant information and the target non-advertising merchant information into the ranking model to obtain the combined push order of the target merchant information and the target non-advertising merchant information, wherein the ranking model is trained by the model training method described in the first part;

[0048] The push module is used to push the target advertising merchant information and the target non-advertising merchant information based on the combined push order.

[0049] The fifth part of this disclosure provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the methods described in the first or second part above.

[0050] A sixth part of this disclosure provides an electronic device, including:

[0051] A memory on which computer programs are stored;

[0052] A processor for executing the computer program in the memory to implement the steps of the method described in the first or second part above.

[0053] The above technical solution involves inputting sample merchant information labeled with feature tags, sample user operation information of the sample merchants, and preset order information labeled with revenue tags into the ranking model. This allows for the full utilization of prior business knowledge, including sample merchant information and sample user operation information, to train the ranking model and obtain the predicted revenue information output by the model. Based on this, a target loss value can be calculated using this prior business knowledge, the predicted revenue information, and the revenue tags. The parameters of the ranking model can then be adjusted according to the target loss value, resulting in a ranking model with higher accuracy and better application performance. In other words, this method fully utilizes prior business knowledge to train the model, guides feature dimensionality reduction through supervised information in the prior business knowledge, improves model training efficiency, and thus enhances the model's accuracy and application performance.

[0054] Other features and advantages of this disclosure will be described in detail in the following detailed description section. Attached Figure Description

[0055] The accompanying drawings are provided to further illustrate the present disclosure and form part of the specification. They are used together with the following detailed description to explain the present disclosure, but do not constitute a limitation thereof. In the drawings:

[0056] Figure 1 This is a flowchart illustrating a model training method according to an exemplary embodiment of the present disclosure.

[0057] Figure 2 This is a schematic diagram illustrating a sorting model according to an exemplary embodiment of the present disclosure.

[0058] Figure 3 This is a flowchart illustrating an information push method according to an exemplary embodiment of the present disclosure.

[0059] Figure 4 This is a block diagram illustrating a model training apparatus according to an exemplary embodiment of the present disclosure.

[0060] Figure 5 This is a block diagram illustrating an information push device according to an exemplary embodiment of the present disclosure.

[0061] Figure 6 This is a block diagram illustrating an electronic device according to an exemplary embodiment of the present disclosure. Detailed Implementation

[0062] The specific embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this disclosure.

[0063] It should be noted that all actions involving the acquisition of signals, information, or data in this disclosure are performed in accordance with the relevant data protection laws and regulations of the country where the location is situated, and with authorization from the owner of the corresponding device. That is to say, before applying the technical solutions disclosed in the embodiments of this disclosure, the user should be informed of the types of information involved in this disclosure, the scope of use, and the usage scenarios in accordance with data protection laws and regulations. Furthermore, the acquisition of corresponding signals, information, or data is only permitted with the user's authorization.

[0064] For example, authorization information can be displayed directly on the user's browsing page, or in a pop-up window, or similar methods, to send a prompt to the user and provide options to confirm or revoke authorization, allowing the user to agree to or refuse authorization. Alternatively, the user can be informed in other ways that comply with data protection regulations and policies, and the technical solutions provided in this disclosure can be implemented with the user's authorization; this disclosure does not specifically limit this approach.

[0065] Related technologies, within a reinforcement learning framework, can predict the platform's total revenue under different push orders for ads and merchants in different states, and select the push order that maximizes revenue. Here, "state" can refer to information such as the number and quality of advertising merchants, user information, advertising merchant information, and non-advertising merchant information. The push order refers to the actions corresponding to the ads and merchants under the reinforcement learning framework. Since ads and merchants in different states can have various actions, the state actions input when training the reinforcement learning model are usually quite complex. Furthermore, in food delivery platforms, the demand for advertising merchant pushes is significant, with up to 50 merchants requiring push order decisions; in this case, the state actions input to the model are even more complex. Faced with high-dimensional state-action sequences, the agent struggles to learn.

[0066] To address the problems in related technologies, this disclosure provides a model training method, information push method, apparatus, medium, and device to fully utilize prior business knowledge to train the model, guide feature dimensionality reduction through supervised information in the prior business knowledge, improve model training efficiency, and thereby enhance the model's accuracy and application effect.

[0067] Figure 1 This is a flowchart illustrating a model training method according to an exemplary embodiment of this disclosure. (Refer to...) Figure 1 The model training method includes:

[0068] Step S101: Obtain sample data for training.

[0069] The sample data includes sample merchant information labeled with feature tags, sample user operation information of sample merchant information, and preset order information labeled with revenue tags. The sample merchant information includes advertising merchant information and non-advertising merchant information. The preset order information is used to characterize the combined push order of advertising merchant information and non-advertising merchant information.

[0070] It should be noted that the model training method provided in this embodiment can be model training performed within a reinforcement learning framework. Within the reinforcement learning framework, sample merchant information labeled with feature tags, sample user operation information of the sample merchant information, and preset order information labeled with revenue tags can be used as sample data.

[0071] The sample merchant information can include both advertising and non-advertising merchant information. The corresponding feature tags for the sample merchant information can refer to multi-dimensional descriptive information about the sample merchant, such as merchant rating tags, sales volume tags, etc. The sample user operation information for the sample merchant information refers to the user's actions performed on the sample merchant information pushed by the business platform.

[0072] The preset order information can be used to characterize the combined push order of advertising merchant information and non-advertising merchant information. The revenue tag corresponding to the preset order information can refer to the actual revenue obtained by the business platform when pushing sample merchant information according to the combined push order characterized by the preset order information. There can be multiple preset order information, and different preset order information can characterize different combined push orders. Each combined push order can refer to the combined push order of sample merchant information pushed by the business platform within a screen range. It can be understood that a screen range can refer to the range that the screen of an electronic device can display. For example, for the screen range of a mobile terminal, the corresponding number of sample merchant information can be displayed according to the setting of the display range of sample merchant information. Usually, the sample merchant information set to be displayed within the screen range of a mobile terminal is 5. The sample merchant information set to be displayed within the screen range of each terminal can be determined according to the actual situation, and this disclosure does not make specific limitations on this.

[0073] It should also be noted that the sample data used for training can be obtained from the historical data of the business platform. Furthermore, this historical data was recorded with user authorization.

[0074] Step S102: Input the sample data into the ranking model, and perform the following operations through the ranking model:

[0075] Based on the sample data, multiple candidate lists are generated to characterize the push order of various combinations of merchant information in the sample merchant information. For each candidate list, the predicted user operation information, the predicted feature information, and the predicted revenue information corresponding to the merchant information in the candidate list are determined.

[0076] It should be noted that after obtaining the sample data for training, the sample data can be input into the ranking model. This allows the generation of multiple candidate lists representing the push order of various merchant information combinations within the sample data, based on the sample merchant information, sample user operation information of the sample merchants, and preset order information. The preset order information indicates the positions of advertising and non-advertising merchant information within the combined push order. Correspondingly, merchant information of the corresponding type can be determined from the sample merchant information included in the sample data and filled into the corresponding positions in the preset order information. This results in multiple candidate lists. Each candidate list can include advertising and / or non-advertising merchant information selected from the sample merchant information and displayed within a single screen, as well as the corresponding combined push order of the advertising and / or non-advertising merchant information.

[0077] Based on this, for each candidate list, we can determine the predicted user operation information, the predicted feature information, and the predicted revenue information corresponding to the merchant information in the candidate list.

[0078] Step S103: Calculate the target loss value based on the sample user operation information, predicted user operation information, feature labels, predicted feature information, profit labels, and predicted profit information.

[0079] Step S104: Adjust the parameters of the ranking model according to the target loss value.

[0080] It is easy to understand that a target loss value can be calculated based on sample user operation information, predicted user operation information, feature labels, predicted feature information, revenue labels, and predicted revenue information. The target loss value can be a revenue loss value calculated based on the predicted revenue information. Based on this, the parameters of the ranking model can be adjusted according to the calculated target loss value, thereby obtaining a ranking model with higher accuracy in revenue prediction (i.e., the revenue that the business platform can obtain by pushing information from various merchants). In other words, the technical solution provided in this disclosure improves the accuracy of the ranking model in predicting the revenue that the business platform can obtain when pushing advertising merchant information and merchant information in different states according to different combined push orders.

[0081] Building upon this foundation, for the target advertising merchant information and target non-advertising merchant information to be pushed, a ranking model can be used to predict the revenue information of the target advertising merchant information and target non-advertising merchant information under different combined push orders, and determine the target combination push order that yields the highest revenue. In this way, the business platform can obtain the maximum revenue when pushing target advertising merchant information and target non-advertising merchant information according to the target combination push order.

[0082] The technical solution provided in this disclosure uses sample merchant information labeled with feature tags, sample user operation information of the sample merchants, and preset order information labeled with revenue tags as sample data. This sample data is then input into a ranking model for training. This process fully utilizes prior business knowledge, including sample merchant information and sample user operation information, and guides feature dimensionality reduction through supervised information within this prior knowledge. This avoids the difficulty ranking models face in learning high-dimensional state-action sequences, improving model training efficiency, accuracy, and application effectiveness, thereby increasing platform revenue.

[0083] Optionally, the sample user operation information of the sample merchant information includes the sample click-through rate of users on the advertising merchant information in the sample merchant information and the sample pull-down probability of users on the combined push order of various merchant information in the sample merchant information. Based on this, in step S102 above, determining the predicted user operation information for merchant information in the candidate list may include:

[0084] Determine the predicted click-through rate for advertiser information in the candidate list and the predicted drop-down probability for the combined push order of information from each advertiser in the candidate list.

[0085] The sample click-through rate (CTR) of users viewing advertised merchant information within the sample merchant information can be defined as the CTR, which is the number of actual clicks a user makes on advertised merchant information divided by the number of ads pushed to the screen. The sample pull-down probability refers to the probability that a user will pull down a screen and move to the next screen when presented with merchant information within a given push order. In possible implementations, this probability can be determined based on historical data (e.g., merchant information across all screens from the start to the end of a user's browsing journey). In other words, the sample pull-down probability can be the probability of a user pulling down the screen based on the merchant information within a given screen and the push order of those merchant information.

[0086] It should be noted that the sample click-through rate can also include the click-through rate of non-advertising merchant information within the sample merchant information. By analyzing the user click-through rate of non-advertising merchant information, the probability of a user placing an order can be preliminarily predicted, thus revealing the commission revenue that the business platform can earn by pushing non-advertising merchant information.

[0087] Furthermore, since the sample click-through rate is descriptive information for each merchant, it can be used to introduce more supervisory information. This more supervisory information helps to represent high-dimensional state-action sequences in a low-dimensional way. This improves the training effect of the ranking model, thereby increasing its accuracy and application performance, and ultimately boosting platform revenue.

[0088] Based on this, step S103 above may include:

[0089] The target loss value is calculated based on the sample click pass rate, sample drop-down probability, predicted click pass rate, predicted drop-down probability, feature labels, predicted feature information, revenue labels, and predicted revenue information.

[0090] Optionally, the target loss value can be calculated based on the sample click pass rate, sample dropdown probability, predicted click pass rate, predicted dropdown probability, feature labels, predicted feature information, revenue labels, and predicted revenue information. This may include:

[0091] Calculate the click pass rate loss value based on the sample click pass rate and the predicted click pass rate;

[0092] Calculate the drop-down loss value based on the sample drop-down probability and the predicted drop-down probability;

[0093] Calculate the key information loss value based on the feature labels and predicted feature information;

[0094] The target predicted revenue value is determined based on the click pass rate loss value, drop-down loss value, key information loss value, and predicted revenue information.

[0095] Calculate the target loss value based on the target predicted return value and the return label.

[0096] It should be noted that the click pass rate loss value corresponding to the sample click pass rate and the predicted click pass rate, the drop-down loss value corresponding to the sample drop-down probability and the predicted drop-down probability, and the key information loss value corresponding to the feature label and the predicted feature information can be calculated using the cross-entropy loss function.

[0097] For example, the click pass rate loss value can be calculated using the following formula:

[0098]

[0099] Where L1 represents the click pass rate loss value, and y1 represents the sample click pass rate. This indicates the predicted click-through rate.

[0100] It is understandable that the calculation process for dropdown loss value and key information loss value is similar to that for click pass rate loss value, so it will not be repeated here.

[0101] It should also be noted that after calculating the click-through rate loss value, dropdown loss value, key information loss value, and predicted revenue information, the target predicted revenue value can be determined based on these factors. On this basis, the target loss value can be calculated using the target predicted revenue value and revenue tags. The target loss value can be the revenue loss value calculated based on the predicted revenue information.

[0102] Optionally, in step S102 above, determining the predicted revenue information corresponding to the merchant information in the candidate list may include:

[0103] Determine the pre-design fee vector corresponding to the advertiser information in the candidate list, and multiply the predicted click-through rate corresponding to the advertiser information with the pre-design fee vector to obtain the predicted revenue information.

[0104] The pre-design cost vector can refer to CPC (Cost Per Click), which is the price per click for an advertiser's information. Predicted revenue can be obtained by multiplying the predicted click-through rate by the pre-design cost vector.

[0105] Optionally, the feature tags of the sample merchant information include at least one of the following tags: merchant delivery fee tag, merchant delivery distance tag, merchant sales volume tag, merchant rating tag, and brand tag used to characterize whether the merchant associated with the sample merchant information is a brand merchant.

[0106] The merchant delivery fee tag indicates whether the merchant's delivery fee exceeds a delivery fee threshold. For example, if the merchant's delivery fee exceeds the threshold, the merchant delivery fee tag can be set to 1; otherwise, it can be set to 0. The merchant delivery distance tag indicates whether the merchant's delivery distance exceeds a delivery distance threshold. The merchant sales volume tag indicates whether the merchant's sales volume exceeds a sales volume threshold. The merchant rating tag indicates whether the merchant's rating exceeds a rating threshold. The brand tag indicates whether the brand corresponding to the merchant information is a preset brand. All of the above thresholds can be determined according to actual circumstances, and this disclosure does not impose specific limitations on them.

[0107] Accordingly, the predicted feature information may include: predicted delivery fee feature information corresponding to whether the merchant's delivery fee exceeds a delivery fee threshold; predicted delivery distance feature information corresponding to whether the merchant's delivery distance exceeds a delivery distance threshold; predicted merchant sales volume feature information corresponding to whether the merchant's sales volume exceeds a sales volume threshold; predicted merchant rating feature information corresponding to whether the merchant's rating exceeds a rating threshold; and predicted brand feature information corresponding to whether the merchant's brand is a preset brand. Based on this, the calculated key information loss values ​​may include delivery fee loss value, delivery distance loss value, sales volume loss value, rating loss value, and brand loss value. It can be understood that the delivery fee loss value is calculated based on the merchant's delivery fee tag and the predicted delivery fee feature information; similarly, the delivery distance loss value, sales volume loss value, rating loss value, and brand loss value can be obtained.

[0108] Furthermore, it should be noted that the sample merchant information may also include user profile information (such as the user's gender, age, and other identity information), contextual information (such as the time, location, and entry point for the user's click or order), and other relevant information. This information helps to represent high-dimensional state-action sequences in a low-dimensional way. This can improve the training effect of the ranking model, thereby increasing the accuracy and application effectiveness of the trained ranking model, and ultimately increasing platform revenue.

[0109] Figure 2 This is a schematic diagram illustrating a sorting model according to an exemplary embodiment of the present disclosure. (Refer to...) Figure 2 The ranking model 200 includes a deep sub-network 201 and a shallow sub-network 202. The deep sub-network 201 may include a parameter-sharing network to extract feature information of each merchant from sample merchant information and sample user operation information, thereby obtaining a merchant information vector. The shallow sub-network 202 may apply an attention mechanism to obtain multiple candidate lists representing the push order of combinations of merchant information in the sample merchant information based on the merchant information vector and preset order information.

[0110] The sample merchant information S1 may include user profile information, contextual information, etc. The sample merchant information S2 may include the merchant's delivery fee, delivery distance, sales volume, rating, and brand, as well as the corresponding feature tags.

[0111] For example, sample user operation information, advertising merchants, non-advertising merchants, sample merchant information S1, and sample merchant information S2 are input into the deep sub-network. A parameter-sharing network is used to obtain the vectors of each merchant information. Then, Concat (vector concatenation) is used to concatenate these vectors with the sample merchant information S2 to obtain the final merchant information vector. Based on this, the merchant information vectors and the corresponding sequence information vectors for a preset order are input into the shallow sub-network. In the shallow sub-network, a State and Action Cross Unit (2021) is used to obtain the state and action matrix corresponding to the merchant information vectors and the preset order information. Based on this state and action matrix, a candidate list is generated to represent the push order of the various combinations of merchant information in the sample merchant information. For each candidate list, the click-through rate loss (CTR_Loss), dropdown loss (P_Loss), key information loss (R(f)_Loss, where f can be used to represent predicted feature information, such as predicted delivery fee feature information, predicted delivery distance feature information, etc.), and predicted revenue information for each merchant information in the candidate list can be calculated. The predicted revenue information can be calculated by multiplying the predicted click-through rate (CTR) and the pre-designed cost vector (CPC). Based on this, the target predicted revenue value (Q(s,a)) can be determined using the CTR loss value, dropdown loss value, key information loss value, and predicted revenue information. Then, the target loss value (RL_Loss) can be calculated based on the target predicted revenue value and revenue labels. In this way, the parameters of the ranking model can be adjusted using the target loss value to obtain a ranking model with higher accuracy and better application performance, thereby improving the revenue of the business platform.

[0112] Optionally, the ranking model includes deep sub-networks and shallow sub-networks. In step S102 above, generating multiple candidate lists based on the sample data to characterize the push order of various combinations of merchant information in the sample merchant information may include:

[0113] The sample user operation information and the sample merchant information are input into the deep sub-network to obtain the merchant information vector output by the deep sub-network;

[0114] The preset order information is converted into an order information vector, and the order information vector and the merchant information vector are input into the shallow sub-network to generate multiple candidate lists that characterize the push order of each combination of merchant information in the sample merchant information.

[0115] There can be multiple preset order information vectors, and correspondingly, there can also be multiple order information vectors. Since a candidate list can be generated from a single order information vector and a single merchant information vector, parallel computation can be performed on multiple order information vectors and merchant information vectors to generate multiple candidate lists.

[0116] Furthermore, in a possible implementation, the merchant information vector output by the deep subnetwork can be obtained by concatenating the vector corresponding to the sample merchant information and the vector corresponding to the sample user operation information. In specific implementations, the merchant information vectors corresponding to the sample merchant information and sample user operation information can also be obtained in other ways. This disclosure does not impose specific limitations on this, as long as the obtained merchant information vector is a low-dimensional representation vector.

[0117] Optionally, the deep sub-network is a remotely running network model, and the shallow sub-network is a locally running network model.

[0118] It should be noted that deep subnetworks have a relatively high network layer and can be used to learn low-dimensional vector representations of information from various merchants. Shallow subnetworks have a relatively low network layer and a small size, so they can be deployed locally for easy access and reduced processing time.

[0119] Optionally, in step S102 above, generating multiple candidate lists based on the sample data to characterize the push order of various combinations of merchant information in the sample merchant information may include:

[0120] Based on the sample user operation information and non-advertising merchant information in the sample merchant information, generate a non-advertising merchant information vector, and based on the sample user operation information and advertising merchant information in the sample merchant information, generate an advertising merchant information vector.

[0121] Based on the non-advertising merchant information vector, the advertising merchant information vector, and the preset order information, a candidate list is generated to represent the push order of each combination of merchant information in the sample merchant information.

[0122] One possible implementation is to concatenate the vector corresponding to the non-advertising merchant information in the sample merchant information with the vector corresponding to the sample user operation information in the same sample merchant information. The vector for advertising merchant information can be obtained similarly, and will not be elaborated upon here.

[0123] In specific implementations, non-advertising merchant information vectors and advertising merchant information vectors can also be obtained through other means. This disclosure does not impose specific limitations on this, as long as the obtained vectors are low-dimensional representation vectors.

[0124] Continue Figure 2For example, using the State and Action Cross Unit 2021, after obtaining the non-advertising merchant information vector, the advertising merchant information vector, and the preset order information, a state action matrix corresponding to the merchant information vector and the preset order information can be obtained. Based on this state action matrix, a candidate list is generated to represent the push order of each combination of merchant information in the sample merchant information.

[0125] Optionally, based on the non-advertising merchant information vector, the advertising merchant information vector, and preset order information, a candidate list is generated to characterize the push order of each combination of merchant information in the sample merchant information, which may include:

[0126] Invert the preset sequence information to obtain the target reverse sequence information;

[0127] The preset sequence information is converted into a first action matrix, and the target reverse sequence information is converted into a second action matrix;

[0128] The first action matrix is ​​multiplied by the advertising merchant information vector to obtain the first state action matrix, and the second action matrix is ​​multiplied by the non-advertising merchant information vector to obtain the second state action matrix.

[0129] Summing the first-state action matrix and the second-state action matrix yields the target-state action matrix;

[0130] Based on the target state action matrix, a candidate list is generated to characterize the push order of each combination of merchant information in the sample merchant information.

[0131] The advertiser information vector can represent multiple advertisers to be pushed to the platform, and the non-advertiser information vector can represent multiple non-advertisers to be pushed to the platform. Preset order information can be used to represent the combined push order of advertiser and non-advertiser information.

[0132] For example, the preset order information could be

[10010] , where 1 represents the position of advertising merchant information and 0 represents the position of non-advertising merchant information. Inverting the preset order information yields the target reverse order information

[01101] , where 0 represents the position of advertising merchant information and 1 represents the position of non-advertising merchant information. Based on this, a first-state action matrix can be obtained by multiplying the advertising information vector and the preset order information, and a second-state action matrix can be obtained by multiplying the non-advertising merchant information vector and the second action matrix. Then, the target-state action matrix can be obtained by summing the first and second-state action matrices. Based on this, the target-state matrix can be flattened (i.e., reducing the multi-dimensional target-state matrix to one dimension) and multi-channel processed (i.e., cutting the target-state matrix into multiple channels, where each channel focuses on only one aspect, such as click-through rate). Then, an attention mechanism can be applied to obtain a candidate list representing the push order of various merchant information combinations in the sample merchant information from the target-state action matrix.

[0133] This disclosure inputs sample merchant information labeled with feature tags, sample user operation information of the sample merchants, and preset order information labeled with revenue tags into a ranking model. This allows for the full utilization of prior business knowledge, including sample merchant information and sample user operation information, to train the ranking model and obtain predicted revenue information output by the model. Based on this, a target loss value can be calculated using this prior business knowledge, predicted revenue information, and revenue tags. The parameters of the ranking model can then be adjusted according to the target loss value, resulting in a ranking model with higher accuracy and better application performance. In other words, this method fully utilizes prior business knowledge to train the model, using supervised information from the prior business knowledge to guide feature dimensionality reduction, improving model training efficiency, thereby enhancing model accuracy and application performance, and ultimately increasing platform revenue.

[0134] Figure 3 This is a flowchart illustrating an information push method according to an exemplary embodiment of this disclosure. (Refer to...) Figure 3 The information push method includes:

[0135] Step S301: Obtain the target advertising merchant information and the target non-advertising merchant information to be pushed.

[0136] Step S302: Input the target advertising merchant information and the target non-advertising merchant information into the sorting model to obtain the combined push order of the target merchant information and the target non-advertising merchant information.

[0137] The ranking model was trained using the model training method described above.

[0138] Step S303: Based on the combined push order, push target advertising merchant information and target non-advertising merchant information.

[0139] It is easy to understand that the ranking model trained by the above model training method has a higher accuracy. By applying this ranking model, it is possible to more accurately predict the revenue that the business platform can obtain when pushing advertising merchant information and merchant information in different states according to different combination push orders.

[0140] Building upon this foundation, for the target advertising merchant information and target non-advertising merchant information to be pushed, a ranking model can be used to predict the revenue information of the target advertising merchant information and target non-advertising merchant information under different combined push orders, and determine the target combination push order that yields the highest revenue. In this way, the business platform can obtain the maximum revenue when pushing target advertising merchant information and target non-advertising merchant information according to the target combination push order.

[0141] The technical solution provided in this disclosure uses sample merchant information labeled with feature tags, sample user operation information of the sample merchants, and preset order information labeled with revenue tags as sample data. This sample data is then input into a ranking model for training. This process fully utilizes prior business knowledge, including sample merchant information and sample user operation information, and guides feature dimensionality reduction through supervised information within this prior knowledge. This avoids the difficulty ranking models face in learning high-dimensional state-action sequences, improving model training efficiency, accuracy, and application effectiveness, thereby increasing platform revenue.

[0142] Figure 4 This is a block diagram illustrating a model training apparatus according to an exemplary embodiment of the present disclosure. (Refer to...) Figure 4 The model training device 100 includes:

[0143] The first acquisition module 101 is used to acquire sample data for training. The sample data includes sample merchant information labeled with feature tags, sample user operation information of sample merchant information, and preset order information labeled with revenue tags. The sample merchant information includes advertising merchant information and non-advertising merchant information. The preset order information is used to represent the combined push order of advertising merchant information and non-advertising merchant information.

[0144] The first input module 102 is used to input sample data into the sorting model and perform the following operations through the sorting model: based on the sample data, generate multiple candidate lists to characterize the push order of each combination of merchant information in the sample merchant information, and for each candidate list, determine the predicted user operation information, the predicted feature information of the merchant information in the candidate list, and the predicted revenue information corresponding to the merchant information in the candidate list;

[0145] The calculation module 103 is used to calculate the target loss value based on the sample user operation information, the predicted user operation information, the feature label, the predicted feature information, the profit label, and the predicted profit information.

[0146] Adjustment module 104 is used to adjust the parameters of the ranking model according to the target loss value.

[0147] This embodiment of the disclosure uses sample merchant information labeled with feature tags, sample user operation information of the sample merchants, and preset order information labeled with revenue tags as sample data, and inputs the sample data into a ranking model to train the ranking model. In this process, it fully utilizes prior business knowledge, including sample merchant information and sample user operation information, and guides feature dimensionality reduction through supervised information in the prior business knowledge. This avoids the problem of the ranking model's difficulty in learning high-dimensional state-action sequences, improves model training efficiency, and enhances the accuracy and application effect of the trained ranking model, thereby increasing platform revenue.

[0148] Optionally, the sample user operation information of the sample merchant information includes the sample click pass rate of users on the advertising merchant information in the sample merchant information and the sample pull-down probability of users on the combined push order of various merchant information in the sample merchant information. The first input module 102 is used for:

[0149] Determine the predicted click-through rate for advertising merchant information in the candidate list and the predicted pull-down probability for the combined push order of various merchant information in the candidate list;

[0150] Calculation module 103 is used for:

[0151] The target loss value is calculated based on the sample click pass rate, sample drop-down probability, predicted click pass rate, predicted drop-down probability, feature labels, predicted feature information, revenue labels, and predicted revenue information.

[0152] Optionally, the calculation module 103 is used for:

[0153] Calculate the click pass rate loss value based on the sample click pass rate and the predicted click pass rate;

[0154] Calculate the drop-down loss value based on the sample drop-down probability and the predicted drop-down probability;

[0155] Calculate the key information loss value based on the feature labels and predicted feature information;

[0156] The target predicted revenue value is determined based on the click pass rate loss value, drop-down loss value, key information loss value, and predicted revenue information.

[0157] Calculate the target loss value based on the target predicted return value and the return label.

[0158] Optionally, the first input module 102 is used for:

[0159] Determine the pre-design fee vector corresponding to the advertiser information in the candidate list, and multiply the predicted click-through rate corresponding to the advertiser information with the pre-design fee vector to obtain the predicted revenue information.

[0160] Optionally, the ranking model includes deep sub-networks and shallow sub-networks, and the first input module 102 is used for:

[0161] Input the sample user operation information and sample merchant information into the deep sub-network to obtain the merchant information vector output by the deep sub-network;

[0162] The preset order information is converted into an order information vector, and the order information vector and the merchant information vector are input into a shallow sub-network to generate multiple candidate lists to represent the push order of each combination of merchant information in the sample merchant information.

[0163] Optionally, the deep sub-network is a remotely running network model, and the shallow sub-network is a locally running network model.

[0164] Optionally, the first input module 102 is used for:

[0165] Based on the sample user operation information and non-advertising merchant information in the sample merchant information, generate a non-advertising merchant information vector, and based on the sample user operation information and advertising merchant information in the sample merchant information, generate an advertising merchant information vector.

[0166] Based on the non-advertising merchant information vector, the advertising merchant information vector, and the preset order information, a candidate list is generated to represent the push order of each combination of merchant information in the sample merchant information.

[0167] Optionally, the first input module 102 is used for:

[0168] Invert the preset sequence information to obtain the target reverse sequence information;

[0169] The preset sequence information is converted into a first action matrix, and the target reverse sequence information is converted into a second action matrix;

[0170] The first action matrix is ​​multiplied by the advertising merchant information vector to obtain the first state action matrix, and the second action matrix is ​​multiplied by the non-advertising merchant information vector to obtain the second state action matrix.

[0171] Summing the first-state action matrix and the second-state action matrix yields the target-state action matrix;

[0172] Based on the target state action matrix, a candidate list is generated to characterize the push order of each combination of merchant information in the sample merchant information.

[0173] Optionally, the feature tags of the sample merchant information include at least one of the following tags: merchant delivery fee tag, merchant delivery distance tag, merchant sales volume tag, merchant rating tag, and brand tag used to characterize whether the merchant associated with the sample merchant information is a brand merchant.

[0174] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0175] Figure 5 This is a block diagram illustrating an information push device according to an exemplary embodiment of the present disclosure. (Refer to...) Figure 4 The information push device 300 includes:

[0176] The second acquisition module 301 is used to acquire information on the target advertising merchants and the target non-advertising merchants to be pushed;

[0177] The second input module 302 is used to input the target advertising merchant information and the target non-advertising merchant information into the ranking model to obtain the combined push order of the target merchant information and the target non-advertising merchant information, wherein the ranking model is trained by the above-mentioned model training method.

[0178] The push module 303 is used to push target advertising merchant information and target non-advertising merchant information based on the combined push order.

[0179] This embodiment of the disclosure uses sample merchant information labeled with feature tags, sample user operation information of the sample merchants, and preset order information labeled with revenue tags as sample data, and inputs the sample data into a ranking model to train the ranking model. In this process, it fully utilizes prior business knowledge, including sample merchant information and sample user operation information, and guides feature dimensionality reduction through supervised information in the prior business knowledge. This avoids the problem of the ranking model's difficulty in learning high-dimensional state-action sequences, improves model training efficiency, and enhances the accuracy and application effect of the trained ranking model, thereby increasing platform revenue.

[0180] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0181] Figure 6 This is a block diagram illustrating an electronic device 400 according to an exemplary embodiment. Figure 6 As shown, the electronic device 400 may include a processor 401 and a memory 402. The electronic device 400 may also include one or more of a multimedia component 403, an input / output (I / O) interface 404, and a communication component 405.

[0182] The processor 401 controls the overall operation of the electronic device 400 to complete all or part of the steps in the aforementioned model training method or information push method. The memory 402 stores various types of data to support the operation of the electronic device 400. This data may include, for example, instructions for any application or method operating on the electronic device 400, and application-related data such as contact data, sent and received messages, pictures, audio, video, etc. The memory 402 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. Multimedia component 403 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in memory 402 or transmitted via communication component 405. The audio component also includes at least one speaker for outputting audio signals. I / O interface 404 provides an interface between processor 401 and other interface modules, such as a keyboard, mouse, buttons, etc. These buttons may be virtual or physical buttons. Communication component 405 is used for wired or wireless communication between the electronic device 400 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G, NB-IoT, eMTC, or other 5G technologies, or combinations thereof, is not limited here. Therefore, the corresponding communication component 405 may include: a Wi-Fi module, a Bluetooth module, an NFC module, etc.

[0183] In an exemplary embodiment, the electronic device 400 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the model training method or information push method described above.

[0184] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, which, when executed by a processor, implement the steps of the model training method or information push method described above. For example, the computer-readable storage medium may be the memory 402 including the program instructions described above, which may be executed by the processor 401 of the electronic device 400 to complete the model training method or information push method described above.

[0185] In another exemplary embodiment, a computer program product is also provided, the computer program product comprising a computer program executable by a programmable device, the computer program having a code portion for performing the above-described model training method or information push method when executed by the programmable device.

[0186] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0187] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0188] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. A model training method, characterized in that, The method comprises: obtaining sample data for training, the sample data comprising sample merchant information labeled with a feature label, sample user operation information of the sample merchant information, and preset sequence information labeled with a revenue label, the sample merchant information comprising advertising merchant information and non-advertising merchant information, and the preset sequence information being used to represent a combined push sequence of the advertising merchant information and the non-advertising merchant information; inputting the sample data into a ranking model, and performing the following operations by using the ranking model: generating, according to the sample data, a plurality of candidate lists used to represent a combined push sequence of each merchant information in the sample merchant information, and determining, for each candidate list, predicted user operation information of the merchant information in the candidate list, predicted feature information of the merchant information in the candidate list, and predicted revenue information corresponding to the merchant information in the candidate list; calculating a target loss value according to the sample user operation information, the predicted user operation information, the feature label, the predicted feature information, the revenue label, and the predicted revenue information; adjusting parameters of the ranking model according to the target loss value; the ranking model comprises a deep subnetwork and a shallow subnetwork, and the generating, according to the sample data, of the plurality of candidate lists used to represent the combined push sequence of each merchant information in the sample merchant information comprises: inputting the sample user operation information and the sample merchant information into the deep subnetwork to obtain a merchant information vector output by the deep subnetwork; converting the preset sequence information into a sequence information vector, inputting the sequence information vector and the merchant information vector into the shallow subnetwork, and generating the plurality of candidate lists used to represent the combined push sequence of each merchant information in the sample merchant information.

2. The method of claim 1, wherein, the sample user operation information of the sample merchant information comprises a sample click-through rate of a user to advertising merchant information in the sample merchant information and a sample pull-down probability of the user to a combined push sequence of each merchant information in the sample merchant information, and the determining of the predicted user operation information of the merchant information in the candidate list comprises: determining a predicted click-through rate of the advertising merchant information in the candidate list and a predicted pull-down probability of the combined push sequence of each merchant information in the candidate list; the calculating of the target loss value according to the sample user operation information, the predicted user operation information, the feature label, the predicted feature information, the revenue label, and the predicted revenue information comprises: calculating the target loss value according to the sample click-through rate, the sample pull-down probability, the predicted click-through rate, the predicted pull-down probability, the feature label, the predicted feature information, the revenue label, and the predicted revenue information.

3. The method of claim 2, wherein, the calculating of the target loss value according to the sample click-through rate, the sample pull-down probability, the predicted click-through rate, the predicted pull-down probability, the feature label, the predicted feature information, the revenue label, and the predicted revenue information comprises: calculating a click-through rate loss value according to the sample click-through rate and the predicted click-through rate; According to the sample pull-down probability and the predicted pull-down probability, a pull-down loss value is calculated; According to the feature label and the predicted feature information, a key information loss value is calculated; According to the click-through rate loss value, the pull-down loss value, the key information loss value and the predicted revenue information, a target predicted revenue value is determined; According to the target predicted revenue value and the revenue label, a target loss value is calculated.

4. The method according to claim 2 or 3, characterized in that, The determination of the predicted revenue information corresponding to the merchant information in the candidate list comprises: A predicted click-through rate corresponding to the target advertisement merchant information is determined, and a predicted design fee vector corresponding to the target advertisement merchant information is determined, and the predicted click-through rate corresponding to the target advertisement merchant information is multiplied by the predicted design fee vector to obtain the predicted revenue information.

5. The method of claim 1, wherein, The deep sub-network is a remotely running network model, and the shallow sub-network is a locally running network model.

6. The method according to any one of claims 1 to 3, characterized in that, The method comprises: According to the sample data, a plurality of candidate lists for representing the push sequence of each combination of merchant information in the sample merchant information are generated, comprising: According to the sample user operation information of non-advertisement merchant information in the sample merchant information and the non-advertisement merchant information, a non-advertisement merchant information vector is generated, and according to the sample user operation information of advertisement merchant information in the sample merchant information and the advertisement merchant information, an advertisement merchant information vector is generated; 7. The method of claim 6, wherein, According to the non-advertisement merchant information vector, the advertisement merchant information vector and the preset sequence information, a candidate list for representing the push sequence of each combination of merchant information in the sample merchant information is generated. The method comprises: The preset sequence information is inverted to obtain target reverse sequence information; The preset sequence information is converted into a first action matrix, and the target reverse sequence information is converted into a second action matrix; The first action matrix is multiplied by the advertisement merchant information vector to obtain a first state action matrix, and the second action matrix is multiplied by the non-advertisement merchant information vector to obtain a second state action matrix; The first state action matrix and the second state action matrix are summed to obtain a target state action matrix; 8. The method according to any one of claims 1 to 3, characterized in that, Based on the target state action matrix, a candidate list for representing the push sequence of each combination of merchant information in the sample merchant information is generated.

9. An information push method characterized by comprising: The feature label of the sample merchant information comprises at least one of the following labels: a merchant delivery fee label, a merchant delivery distance label, a merchant sales label, a merchant rating label and a brand label for representing whether the merchant associated with the sample merchant information is a brand merchant. The method comprises: Obtaining target advertisement merchant information and target non-advertisement merchant information to be pushed; Inputting the target advertisement merchant information and the target non-advertisement merchant information into a sorting model to obtain the combination push sequence of the target advertisement merchant information and the target non-advertisement merchant information, wherein the sorting model is trained by the model training method of any one of claims 1-8. Push the target advertiser merchant information and the target non-advertiser merchant information based on the combination push sequence.

10. A model training apparatus, comprising: The device comprises: The first obtaining module is configured to obtain sample data for training, the sample data comprising sample merchant information labeled with a feature label, sample user operation information of the sample merchant information, and preset sequence information labeled with a benefit label, the sample merchant information comprising advertiser merchant information and non-advertiser merchant information, and the preset sequence information being used to represent a combination push sequence of the advertiser merchant information and the non-advertiser merchant information. The first input module is configured to input the sample data into a ranking model and perform the following operations on the ranking model: generating a plurality of candidate lists representing the combination push sequence of each merchant information in the sample merchant information according to the sample data, and determining, for each candidate list, predicted user operation information of the merchant information in the candidate list, predicted feature information of the merchant information in the candidate list, and predicted benefit information corresponding to the merchant information in the candidate list; the ranking model comprises a deep subnetwork and a shallow subnetwork, and the generating of the plurality of candidate lists representing the combination push sequence of each merchant information in the sample merchant information according to the sample data comprises: inputting the sample user operation information and the sample merchant information into the deep subnetwork to obtain a merchant information vector output by the deep subnetwork; converting the preset sequence information into a sequence information vector, and inputting the sequence information vector and the merchant information vector into the shallow subnetwork to generate the plurality of candidate lists representing the combination push sequence of each merchant information in the sample merchant information. The computing module is configured to calculate a target loss value according to the sample user operation information, the predicted user operation information, the feature label, the predicted feature information, the benefit label, and the predicted benefit information. The adjusting module is configured to adjust parameters of the ranking model according to the target loss value.

11. An information push apparatus characterized by comprising: The device comprises: The second obtaining module is configured to obtain target advertiser merchant information and target non-advertiser merchant information to be pushed. The second input module is configured to input the target advertiser merchant information and the target non-advertiser merchant information into a ranking model to obtain a combination push sequence of the target advertiser merchant information and the target non-advertiser merchant information, wherein the ranking model is obtained by training the model training method according to any one of claims 1-8. The pushing module is configured to push the target advertiser merchant information and the target non-advertiser merchant information based on the combination push sequence.

12. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by a processor to implement the steps of the method of any one of claims 1-8 or claim 9.

13. An electronic device, comprising: comprise: a memory having a computer program stored thereon; a processor configured to execute the computer program in the memory to implement the steps of the method of any one of claims 1-8 or claim 9.

Images