Method and device for determining recommendation channel, equipment and storage medium
By generating multiple channels for users, combining feature vectors and using linear gradient interpolation method to train prediction models, the problem of insufficient accuracy of prediction models in the existing technology is solved, the accuracy and reliability of recommendation channels are improved, and the user experience and Cost control.
Patent Information
- Application Number
- CN202410997325.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-24
- Publication Date
- 2024-11-12
AI Technical Summary
When the user information is recommended, the existing prediction model cannot obtain the conversion data tags of the same user under the combination of different recommendation channels at the same time, resulting in limited sample data, which in turn affects the accuracy of the model and makes it difficult to provide accurate and reliable recommendations. Channel strategy.
By generating multiple channel combination feature vectors for each user, inputting the prediction model to obtain the prediction conversion results under each channel combination, and determining the optimal recommended channel combination based on the maximum value of the objective function, the prediction model is trained using linear gradient interpolation method to improve accuracy.
Through sufficient user combination feature vector training, the accuracy of the prediction model and the accuracy and reliability of the recommendation channel are improved, excessive user disturbance is avoided, recommendation costs are controlled, and the rationality of the recommendation strategy is improved.
Smart Images

Figure CN118939877A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of computer technology, and in particular to a method, device, equipment and storage medium for determining a recommendation channel. Background Art
[0002] When recommending information to users, currently common recommendation channels or means include phone calls, text messages, and message push notifications within applications. After the recommendation is completed, the recommendation results are recorded to form sample data, which is then used to train a prediction model. The user conversion rate output by the prediction model is used as the basis for formulating user recommendation channels or strategies. It can be seen that the accuracy of the prediction model is closely related to the accuracy and reliability of the recommendation channel.
[0003] However, the accuracy of existing prediction models is not ideal, because it is impossible to simultaneously obtain the conversion data labels of the same user under different recommendation channel combinations at the same time. In other words, the sample data used to train the model is limited, and the model cannot be fully trained, which in turn affects the accuracy of the model.
[0004] Therefore, how to improve the accuracy of the prediction model and provide accurate and reliable basis for user recommendation channels is an urgent problem to be solved. Summary of the invention
[0005] The present application provides a method, apparatus, device and storage medium for determining a recommendation channel, so as to improve the accuracy and reliability of determining the recommendation channel.
[0006] In a first aspect, the present application provides a method for determining a recommendation channel, the method comprising:
[0007] For each of the N users, based on the feature vector of each user under a single channel and the M channel combinations, M combined feature vectors of each user are formed; any channel combination represents whether at least one of the C channels is used to recommend information to the user; N, M, and C are all positive integers;
[0008] Inputting the M combined feature vectors of each user among the N users into the prediction model to obtain M predicted conversion results corresponding to each user under the M channel combinations;
[0009] According to the M predicted conversion results of each user among the N users, determine a recommendation channel combination for each user when the objective function satisfies the maximum value, thereby obtaining a final recommendation channel for the N users;
[0010] Among them, the objective function represents the comprehensive result of the predicted conversion results corresponding to the channel combination used when recommending information to the N users; the prediction model is obtained by training with the negative number corresponding to the maximum value of the objective function as the loss value.
[0011] The method for determining the recommended channel provided by the present application first forms M combined feature vectors for each user based on the feature vector of the user under a single channel and M channel combinations, wherein any channel combination in the M channel combinations represents whether at least one of the C channels is used to recommend information to the user. In this way, the M combined feature vectors for each user solve the problem of limited conversion data labels for each user in the existing scheme; then the M combined feature vectors for each user are input into the prediction model to obtain M predicted conversion results corresponding to each user under the M channel combinations; and then the final recommended channel for each user is determined based on the M predicted conversion results of each user. The present application can fully train the prediction model based on sufficient and complete user combination feature vectors, thereby improving the accuracy of the prediction model. Based on a more accurate prediction model, the accuracy of the M predicted conversion results corresponding to each user under the M channel combinations can be improved, further improving the accuracy and reliability of the final recommended channel for each user.
[0012] In a possible design, the objective function is used to represent the sum of predicted conversion results corresponding to the channel combination used when recommending information to the N users;
[0013] The objective function satisfies the maximum value when the sum of the predicted conversion results of the N users under the recommended channel combination is the largest and the recommended channel combination of the N users satisfies the implementation constraint condition.
[0014] In one possible design, the implementation constraint includes at least one of the following:
[0015] The recommended channel combination used when recommending information to each user is at most one of the M channel combinations;
[0016] The sum of the costs of the recommendation channel combinations used when recommending information to the N users is less than or equal to the preset cost.
[0017] The above real-time constraints can prevent users from being disturbed too much and improve user experience; they can also control the total cost required for recommendation within the expected range and improve the rationality of determining the recommendation channel.
[0018] In a possible design, the prediction model is obtained by training using a linear gradient interpolation method by taking the negative number corresponding to the maximum value of the objective function as a loss value.
[0019] In a possible design, the prediction model is obtained by training using a linear gradient interpolation method by taking a negative number corresponding to the maximum value of the objective function as a loss value, including:
[0020] Input the combined feature vector of the sample users into the prediction model to obtain the sample prediction value;
[0021] Based on the sample prediction value, determining a first output value of the objective function when the maximum value of the objective function is satisfied;
[0022] Derivative the first output value based on the loss value to determine a disturbance prediction value corresponding to the sample prediction value;
[0023] Based on the disturbance prediction value, determining a second output value of the objective function when the maximum value of the objective function is satisfied;
[0024] Based on the first output value and the second output value, a gradient value corresponding to the sample prediction value is determined; the gradient value is used to adjust the parameters of the prediction model during back propagation.
[0025] The use of linear gradient interpolation method can solve the gradient blocking problem encountered during prediction model training, ensure that the prediction model can be effectively adjusted, ensure the training effect of the prediction model, and thus improve the accuracy of the prediction model.
[0026] In a possible design, it also includes: when the difference between the first gradient value and the second gradient value corresponding to two consecutive trainings of the prediction model is within a preset range, the prediction model training ends.
[0027] In a second aspect, the present application further provides a device for determining a recommended channel, the device comprising: a transceiver unit and a processing unit;
[0028] The processing unit is used to form, for each user among the N users, M combined feature vectors of each user based on the feature vector of each user under a single channel and the M channel combinations; any channel combination represents whether at least one channel among the C channels is used to recommend information to the user; N, M, and C are all positive integers;
[0029] The transceiver unit is used to input the M combined feature vectors of each user among the N users into the prediction model to obtain the M predicted conversion results corresponding to each user under the M channel combinations;
[0030] The processing unit is further configured to determine, based on the M predicted conversion results of each user among the N users, a combination of recommended channels for each user when the objective function satisfies a maximum value, thereby obtaining a final recommended channel for the N users;
[0031] Among them, the objective function represents the comprehensive result of the predicted conversion results corresponding to the channel combination used when recommending information to the N users; the prediction model is obtained by training with the negative number corresponding to the maximum value of the objective function as the loss value.
[0032] In a possible design, the objective function is used to represent the sum of predicted conversion results corresponding to the channel combination used when recommending information to the N users;
[0033] The objective function satisfies the maximum value when the sum of the predicted conversion results of the N users under the recommended channel combination is the largest and the recommended channel combination of the N users satisfies the implementation constraint condition.
[0034] In one possible design, the implementation constraint includes at least one of the following:
[0035] The recommended channel combination used when recommending information to each user is at most one of the M channel combinations;
[0036] The sum of the costs of the recommendation channel combinations used when recommending information to the N users is less than or equal to the preset cost.
[0037] In a possible design, the prediction model is obtained by training using a linear gradient interpolation method by taking the negative number corresponding to the maximum value of the objective function as a loss value.
[0038] In a possible design, the prediction model is obtained by training using a linear gradient interpolation method by taking a negative number corresponding to the maximum value of the objective function as a loss value, including:
[0039] The transceiver unit is used to input the combined feature vector of the sample user into the prediction model to obtain the sample prediction value;
[0040] The processing unit is used to determine, based on the sample prediction value, a first output value of the objective function when the maximum value of the objective function is met; based on the loss value, derive the first output value to determine a disturbance prediction value corresponding to the sample prediction value; based on the disturbance prediction value, determine a second output value of the objective function when the maximum value of the objective function is met; based on the first output value and the second output value, determine a gradient value corresponding to the sample prediction value; the gradient value is used to adjust the parameters of the prediction model during back propagation.
[0041] In a possible design, it also includes: when the difference between the first gradient value and the second gradient value corresponding to two consecutive trainings of the prediction model is within a preset range, the prediction model training ends.
[0042] In a third aspect, the present application further provides a device for determining a recommendation channel, the device comprising: a processor, and a memory communicatively connected to the processor;
[0043] The memory stores computer-executable instructions;
[0044] The processor executes the computer-executable instructions stored in the memory to implement the method described in the first aspect above.
[0045] In a fourth aspect, the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium includes a program, and when the program is executed on a device, the device executes the method as described in any one of the above-mentioned first aspects.
[0046] In a fifth aspect, the present application further provides a computer program product, wherein the computer program product comprises a computer program, and when the computer program is executed by a processor, the method described in the first aspect is implemented. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0048] Figure 1 A flowchart of a method for determining a recommendation channel provided in an embodiment of the present application;
[0049] Figure 2 An interactive schematic diagram of a method for determining a recommendation channel provided in an embodiment of the present application;
[0050] Figure 3 A schematic diagram of the structure of the device for determining the recommended channel provided in the embodiment of the present application Figure 1 ;
[0051] Figure 4 A schematic diagram of the structure of the device for determining the recommended channel provided in the embodiment of the present application Figure 2 . DETAILED DESCRIPTION
[0052] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.
[0053] The application scenarios described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. It is known to those skilled in the art that with the emergence of new application scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems. In the description of the present application, unless otherwise specified, the meaning of "multiple" is two or more.
[0054] With the rapid development of the times, information recommendations to users are now mostly made through Internet marketing. The advantages of Internet marketing include precise positioning, rapid dissemination, flexibility and strong interactivity. Enterprises can adjust their marketing strategies in a timely manner based on marketing feedback to achieve accurate push.
[0055] As mentioned in the background technology section, the accuracy of existing prediction models is poor. The fundamental reason is that it is impossible to obtain the conversion data labels of the same user for different channel combinations at the same time, because a user cannot make multiple different decisions at the same time. This is an inherent problem in Internet marketing and poses a great challenge to the training of prediction models. To solve this problem, this application proposes the following Figure 1 The method shown in the figure can be executed in the server or in a functional module of the server, and this application does not limit this. The method includes:
[0056] Step 100: For each user among the N users, based on the feature vector of each user under a single recommendation channel and the M channel combinations, M combined feature vectors of each user are formed.
[0057] For example, currently, information recommendation to users is mostly done through channels such as phone calls, text messages, or in-app push messages, and the channel combination is determined based on the above channels. In other words, assuming there are C types of channels, the channel combination is composed of at least one of the C types of channels. Specifically, the channel combination can be composed of one of the C types of channels, or two of the C types of channels, ..., or composed of C types of channels, and any channel combination of M channel combinations represents whether at least one of the C types of channels is used to recommend information to users. M and C are both positive integers.
[0058] For example, assuming that C is 3, and the three channels are making phone calls, sending text messages, and pushing messages in apps, then the M channel combinations can be: ① making phone calls; ② sending text messages; ③ pushing messages in apps; ④ making phone calls + sending text messages; ⑤ making phone calls + pushing messages; ⑥ sending text messages + pushing messages in apps; ⑦ making phone calls + sending text messages + pushing messages in apps. In this case, the value of M is 7, among which channel combination ① means recommending information to users only through the channel of making phone calls; channel combination ② means recommending information to users only through the channel of sending text messages; channel combination ③ means recommending information to users only through the channel of pushing messages in apps; channel combination ④ means recommending information to users through both channels of making phone calls and sending text messages; channel combination ⑤ means recommending information to users through both channels of making phone calls and pushing messages in apps; channel combination ⑥ means recommending information to users through both channels of sending text messages and pushing messages in apps; channel combination ⑦ means recommending information to users through three channels: making phone calls, sending text messages, and pushing messages in apps.
[0059] Exemplarily, if the above three channels are taken as an example, the feature vector of each user among N users in a single channel can be understood as the feature vector of the user in the calling channel, the feature vector of the user in the texting channel, and the feature vector of the user in the in-app push message channel, among which the feature vector of the user in the calling channel can be determined based on whether the call is answered, the duration of the call, and other features; the feature vector of the user in the texting channel can be determined based on whether the text message is replied, the content length of the reply text message, and other features; the feature vector of the user in the in-app push message channel can be determined based on whether the message is read, whether the message is replied, and other features. It should be noted that the feature vector of the user in any single channel can also be determined based on the user's age, gender, work unit, industry, and other personal characteristics, and this application does not limit this. For example, the feature vector of the user in the texting channel can also be determined based on the user's age and work unit. N is a positive integer.
[0060] Further exemplarily, based on the feature vector of each user in a single channel and M channel combinations, M combined feature vectors of each user are formed, and the M combined feature vectors correspond one-to-one to the M channel combinations. Among them, if the number of channels included in the channel combination is greater than 1, for example, any one of the channel combinations ④ to ⑦, then the user combination feature vector corresponding to the channel combination is obtained by splicing the feature vector of the user in a single recommended channel.
[0061] For example, the C recommendation channels are still based on calling, texting, and pushing messages. Assume that the corresponding feature vector of user A in the calling channel is x 1 , the corresponding feature vector of user A in the SMS channel is x 2, the corresponding feature vector of the user in the push message channel is x 3 , then based on the feature vector of user A in a single channel and the above channel combinations ① to ⑦, 7 combined feature vectors of user A can be formed, which are: Combined feature vector ① formed based on user feature vector of user A and channel combination ①: x 1 ; Combined feature vector ② formed based on user feature vector of user A and channel combination ②: x 2 ; Combined feature vector ③ formed based on user feature vector of user A and channel combination ③: x 3 ; Combined feature vector ④ formed based on user feature vector of user A and channel combination ④: x 1 and x 2 The vector obtained by splicing; the combined feature vector ⑤ formed based on the user feature vector of user A and the channel combination ⑤: x 1 and x 3 The vector obtained by splicing; the combined feature vector ⑥ formed based on the user feature vector of user A and the channel combination ⑥: x 2 and x 3 The vector obtained by splicing; the combined feature vector ⑦ formed based on the user feature vector of user A and the channel combination ⑦: x 1 、x 2 and x 3 Concatenate the obtained vectors.
[0062] Step 101: inputting M combined feature vectors of each user among N users into a prediction model to obtain M predicted conversion results corresponding to each user under M channel combinations.
[0063] Exemplarily, the prediction model can be obtained through machine learning, for example, through a neural network architecture; it can also be obtained through deep learning, for example, through a multi-layer neural network architecture; this application does not limit this. The prediction model of the present application can be expressed by the following formula:
[0064] a (i+1) =σ(W (i) a (i) +b (i) )
[0065] Among them, σ is the activation function, W (i) and b (i) is the parameter of the prediction model, W (i) and b (i) With an initial value, mainly by adjusting W (i) and b (i) To optimize the prediction model, the value of i ranges from 1 to K-1, K is the number of layers of the prediction model, K is generally 4, a (1)is the input of the prediction model, is the M combined feature vectors of each user in the N users in step 100, if the output of the prediction model is recorded as but It consists of M predicted conversion results corresponding to each user under M channel combinations.
[0066] Through step 100, M combined feature vectors of each user can be obtained. By taking the M combined feature vectors of each user as the input of the prediction model, M predicted conversion results corresponding to each user under the M channel combinations can be obtained, which solves the problem in the existing solution that it is impossible to obtain the conversion data labels of the same user for different channel combinations at the same time, ensures that there is sufficient and complete sample data when training the prediction model, improves the accuracy of the preset model, and improves the accuracy of the M predicted conversion results corresponding to each user under the M channel combinations.
[0067] Step 102: According to the M predicted conversion results of each user among the N users, determine the recommendation channel combination for each user when the objective function satisfies the maximum value, so as to obtain the final recommendation channels of the N users.
[0068] Exemplarily, after obtaining the M predicted conversion results corresponding to each user under the M channel combinations through step 101, the final recommended channel for N users is determined according to the M predicted conversion results of each user and the objective function. Among them, the objective function represents the comprehensive result of the predicted conversion results corresponding to the channel combination used when recommending information to N users, and is used to represent the sum of the predicted conversion results corresponding to the channel combination used when recommending information to N users. When recommending information to N users, if you want to obtain a better overall recommendation effect, the sum of the predicted conversion results of each user in the N users should be maximized. Therefore, the objective function needs to satisfy the maximum value. The objective function satisfying the maximum value means that the sum of the predicted conversion results of N users under the recommended channel combination is the largest. The objective function formula is as follows:
[0069]
[0070] Where N is the number of users, M is the number of channel combinations, and y ij represents the predicted conversion result corresponding to the i-th user under the j-th channel combination, obtained through step 101; z ij The value of z is 0 or 1. ij If z is 0, it means that the jth channel combination is not used to recommend information to the i-th user; ij If is 1, it means that the jth channel combination is used to recommend information to the ith user. The input of the objective function is the M predicted conversion results of each user, and the output value of the objective function is z, which contains the final recommended channel for each user.
[0071] Further exemplarily, each channel combination has a corresponding cost when recommending information to users. In order for the objective function to satisfy the maximum value, in addition to satisfying the maximum sum of the predicted conversion results of N users under the recommended channel combination, it must also satisfy implementation constraints, and the implementation constraints include at least one of the following: Constraint ①: The recommended channel combination used when recommending information to each user is at most one of the M channel combinations; this constraint is to avoid excessive disturbance of users; Constraint ②: The sum of the costs of the recommended channel combinations used when recommending information to N users is less than or equal to the preset cost; this constraint is to ensure the rationality of the output value of the objective function.
[0072] For example, the exhaustive list z ij possible values of z, find a set of ij The sum of the predicted conversion results of N users under the recommended channel combination is maximized and meets the implementation constraints. As the final recommended channel for N users, this group z ij This is the output of the objective function, denoted as set z. The elements in set z include z 1,1 、z 1,2 ,......,z 1,M 、z 2,1 、z 2,2 ,......,z N,M , where z 1,1 Indicates whether to use the first channel combination to recommend information to the first user, z 1,2 Indicates whether to use the second channel combination to recommend information to the first user, ..., z 1,M Indicates whether to use the Mth channel combination to recommend information to the first user, z 2,1 Indicates whether to use the first channel combination to recommend information to the second user, z 2,2 Indicates whether to use the second channel combination to recommend information to the second user, ..., z N,M Indicates whether to use the Mth channel combination to recommend information to the Nth user.
[0073] Steps 100 to 102 can also be referred to Figure 2 , where step 100 corresponds to Figure 2 In "User Conversion Rate Prediction", step 101 corresponds to Figure 2 In "Training prediction model based on linear gradient interpolation", step 102 corresponds to Figure 2 "Determine the user's final referral channel" in.
[0074] The following explains the training process of the prediction model:
[0075] Exemplarily, the prediction model is obtained by training the negative number corresponding to the maximum value of the objective function as the loss value. The loss value is the output of the loss function, which can also be called the loss function value. The formula of the loss function is as follows:
[0076]
[0077] The smaller the loss value, the larger the sum of the predicted conversion results of N users.
[0078] For example, in order to train the prediction model, it is necessary to obtain the gradient value at the output value of the prediction model to determine whether the prediction model training is completed. If the difference between the first gradient value and the second gradient value corresponding to two consecutive trainings of the prediction model is within a preset range, the prediction model training is completed; if the prediction model training is not completed, the parameters of the prediction model are adjusted according to the gradient value (W (i) and b (i) ).
[0079] Since the input of the prediction model is the user's combined feature vector, this will cause the problem of gradient blocking in the training process of the prediction model. To solve this problem, the prediction model of the present application is obtained by training with a linear gradient interpolation method by taking the negative number corresponding to the maximum value of the objective function as the loss value.
[0080] Further exemplary, the specific process of using the linear gradient interpolation method to train the prediction model is as follows: first, the combined feature vector of the sample user is input into the prediction model to obtain the sample prediction value, and the sample prediction value is recorded as Based on the sample prediction value, determine the first output value of the objective function when the maximum value of the objective function is met, assuming that based on Determine the first output value of the objective function when the maximum value of the objective function is met Based on the loss value, the first output value is derived to determine the disturbance prediction value corresponding to the sample prediction value, and the disturbance prediction value is recorded as y λ ,y λ It can be expressed by the following formula:
[0081]
[0082] Among them, λ is the parameter of the control disturbance; then based on the disturbance prediction value, the second output value of the objective function when the maximum value of the objective function is met is determined. Assuming that based on y λ , determine the second output value of the objective function when the maximum value of the objective function is met λ ; Finally, based on the first output value and the second output value, determine the gradient value corresponding to the sample prediction value:
[0083] because therefore,
[0084] Since the gradient value corresponding to the sample prediction value is difficult to obtain, the perturbation prediction value is obtained by perturbing the sample prediction value, and the change value between the output value of the objective function corresponding to the sample prediction value and the perturbation prediction value is calculated. The change value is approximated as the gradient value at the sample prediction value, which is used to adjust the parameters of the prediction model during back propagation until the difference between the gradient values corresponding to two adjacent trainings is within the preset range. The introduction of linear gradient interpolation solves the problem of gradient blocking during prediction model training.
[0085] The method for determining the recommended channel proposed in the present application first determines the M combined feature vectors of each user, takes the M combined feature vectors of each user as the input of the prediction model, and obtains the predicted conversion results corresponding to each user under the M channel combinations, thereby solving the problem in the existing scheme that it is impossible to obtain the conversion data labels of the same user for different channel combinations at the same time, obtains sufficient and complete sample data, and improves the accuracy of the prediction model; based on a more accurate prediction model, the predicted conversion results corresponding to each user under the M recommendation channels are more accurate, and further makes the user's final recommendation channel more accurate and reliable.
[0086] Figure 3 and Figure 4 A schematic diagram of a possible device for determining a recommendation channel provided in an embodiment of the present application. If the above method is executed in a server, these devices for determining a recommendation channel can be used to implement the functions of the server, thereby also achieving the beneficial effects of the above method embodiment.
[0087] like Figure 3 As shown, the device 300 for determining a recommended channel includes a transceiver unit 310 and a processing unit 320. When the device 300 for determining a recommended channel is used to implement the function of the server:
[0088] The processing unit 320 is used to form, for each user among the N users, M combined feature vectors of each user based on the feature vector of each user under a single channel and the M channel combinations; any channel combination represents whether at least one channel among the C channels is used to recommend information to the user; N, M, and C are all positive integers;
[0089] The transceiver unit 310 is used to input the M combined feature vectors of each user among the N users into the prediction model to obtain the M predicted conversion results corresponding to each user under the M channel combinations;
[0090] The processing unit 320 is further configured to determine, based on the M predicted conversion results of each user among the N users, a combination of recommended channels for each user when the objective function satisfies the maximum value, thereby obtaining a final recommended channel for the N users;
[0091] Among them, the objective function represents the comprehensive result of the predicted conversion results corresponding to the channel combination used when recommending information to the N users; the prediction model is obtained by training with the negative number corresponding to the maximum value of the objective function as the loss value.
[0092] In a possible design, the objective function is used to represent the sum of predicted conversion results corresponding to the channel combination used when recommending information to the N users;
[0093] The objective function satisfies the maximum value when the sum of the predicted conversion results of the N users under the recommended channel combination is the largest and the recommended channel combination of the N users satisfies the implementation constraint condition.
[0094] In one possible design, the implementation constraint includes at least one of the following:
[0095] The recommended channel combination used when recommending information to each user is at most one of the M channel combinations;
[0096] The sum of the costs of the recommendation channel combinations used when recommending information to the N users is less than or equal to the preset cost.
[0097] In a possible design, the prediction model is obtained by training using a linear gradient interpolation method by taking the negative number corresponding to the maximum value of the objective function as a loss value.
[0098] In a possible design, the prediction model is obtained by training using a linear gradient interpolation method by taking a negative number corresponding to the maximum value of the objective function as a loss value, including:
[0099] The transceiver unit 310 is used to input the combined feature vector of the sample user into the prediction model to obtain the sample prediction value;
[0100] The processing unit 320 is used to determine, based on the sample prediction value, a first output value of the objective function when the maximum value of the objective function is met; based on the loss value, derive the first output value to determine a disturbance prediction value corresponding to the sample prediction value; based on the disturbance prediction value, determine a second output value of the objective function when the maximum value of the objective function is met; based on the first output value and the second output value, determine a gradient value corresponding to the sample prediction value; the gradient value is used to adjust the parameters of the prediction model during back propagation.
[0101] In a possible design, it also includes: when the difference between the first gradient value and the second gradient value corresponding to two consecutive trainings of the prediction model is within a preset range, the prediction model training ends.
[0102] For more detailed description of the transceiver unit 310 and the processing unit 320, please refer to Figure 1 The relevant description in the method embodiment shown is directly obtained and will not be repeated here.
[0103] like Figure 4 As shown, the device 400 for determining the recommended channel includes a processor 410 and an interface circuit 420. The processor 410 and the interface circuit 420 are coupled to each other. It is understood that the interface circuit 420 can be a transceiver or an input-output interface. Optionally, the device 400 for determining the recommended channel can also include a memory 430 for storing instructions executed by the processor 410 or storing input data required by the processor 410 to execute instructions or storing data generated after the processor 410 executes instructions.
[0104] When the device 400 for determining the recommended channel is used to implement Figure 1 When the method is shown, the processor 410 is used to implement the functions of the processing unit 320, and the interface circuit 420 is used to implement the functions of the transceiver unit 310.
[0105] The division of units in the embodiments of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional unit in each embodiment of the present application may be integrated into a processor, or may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0106] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.
[0107] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.
Claims
1. A method for determining referral channels, It is characterized in that The method includes: For each of the N users, based on the feature vector of each user under a single channel and the M channel combinations, M combined feature vectors of each user are formed; any channel combination represents whether at least one of the C channels is used to recommend information to the user; N, M, and C are all positive integers; Inputting the M combined feature vectors of each user among the N users into the prediction model to obtain M predicted conversion results corresponding to each user under the M channel combinations; According to the M predicted conversion results of each user among the N users, determine a recommendation channel combination for each user when the objective function satisfies the maximum value, thereby obtaining a final recommendation channel for the N users; Among them, the objective function represents the comprehensive result of the predicted conversion results corresponding to the channel combination used when recommending information to the N users; the prediction model is obtained by training with the negative number corresponding to the maximum value of the objective function as the loss value.
2. The method according to claim 1, It is characterized in that The objective function is used to represent the sum of predicted conversion results corresponding to the channel combination used when recommending information to the N users; The objective function satisfies the maximum value when the sum of the predicted conversion results of the N users under the recommended channel combination is the largest and the recommended channel combination of the N users satisfies the implementation constraint condition.
3. The method according to claim 2, It is characterized in that The implementation constraints include at least one of the following: The recommended channel combination used when recommending information to each user is at most one of the M channel combinations; The sum of the costs of the recommendation channel combinations used when recommending information to the N users is less than or equal to the preset cost.
4. The method according to any one of claims 1 to 3, It is characterized in that The prediction model is obtained by training using a linear gradient interpolation method by taking the negative number corresponding to the maximum value of the objective function as a loss value.
5. The method according to claim 4, It is characterized in that The prediction model is obtained by training using a linear gradient interpolation method using a negative number corresponding to the maximum value of the objective function as a loss value, including: Input the combined feature vector of the sample users into the prediction model to obtain the sample prediction value; Based on the sample prediction value, determining a first output value of the objective function when the maximum value of the objective function is satisfied; Derivative the first output value based on the loss value to determine a disturbance prediction value corresponding to the sample prediction value; Based on the disturbance prediction value, determining a second output value of the objective function when the maximum value of the objective function is satisfied; Based on the first output value and the second output value, a gradient value corresponding to the sample prediction value is determined; the gradient value is used to adjust the parameters of the prediction model during back propagation.
6. The method according to claim 5, It is characterized in that The method further includes: when the difference between the first gradient value and the second gradient value respectively corresponding to two adjacent trainings of the prediction model is within a preset range, the prediction model training ends.
7. A device for determining a recommendation channel, It is characterized in that The device comprises: a transceiver unit and a processing unit; The processing unit is used to form, for each user among the N users, M combined feature vectors of each user based on the feature vector of each user under a single channel and the M channel combinations; any channel combination represents whether at least one channel among the C channels is used to recommend information to the user; N, M, and C are all positive integers; The transceiver unit is used to input the M combined feature vectors of each user among the N users into the prediction model to obtain the M predicted conversion results corresponding to each user under the M channel combinations; The processing unit is further configured to determine, based on the M predicted conversion results of each user among the N users, a combination of recommended channels for each user when the objective function satisfies a maximum value, thereby obtaining a final recommended channel for the N users; Among them, the objective function represents the comprehensive result of the predicted conversion results corresponding to the channel combination used when recommending information to the N users; the prediction model is obtained by training with the negative number corresponding to the maximum value of the objective function as the loss value.
8. A device for determining a referral channel, It is characterized in that include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 1 to 6.
9. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 6 when executed by a processor.
10. A computer program product, It is characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 6 when being executed by a processor.
