Resource delivery methods, devices, electronic equipment and storage media

By selecting a user set and fine-tuning the training of a predictive model, the click-through rate and conversion rate of users are determined, and SMS resources are accurately pushed, solving the problem of SMS push mismatch and improving conversion rate and user satisfaction.

CN121262273BActive Publication Date: 2026-11-17CHINA MOBILE INTERNET CO LTD +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202511442182.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-10-10
Publication Date
2026-11-17
Estimated Expiration
2045-10-10

AI Technical Summary

Technical Problem

In existing technologies, SMS resource push lacks accurate identification of user needs, resulting in mismatched resource recommendations, reduced conversion efficiency, and information interference and overload.

Method used

By selecting a second user set from the user set, using a pre-trained prediction model to determine the click-through rate and conversion rate of users, fine-tuning the training to obtain the second prediction model, and determining the comprehensive prediction probability based on the click-through rate and conversion rate, resources are accurately pushed to the users with the highest probability.

Benefits of technology

It enables precise delivery of resources, improves conversion rates, reduces unnecessary disturbances to users with low willingness, and enhances user acceptance and satisfaction with SMS services.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121262273B_ABST
    Figure CN121262273B_ABST
Patent Text Reader

Abstract

This disclosure provides a resource push method, apparatus, electronic device, and storage medium, relating to the field of mobile communication technology. The method includes: after selecting a first user set for a first target resource, fine-tuning a pre-trained first prediction model using attributes of some users in the first user set and click and conversion tags of some users for the first target resource to obtain a second prediction model; determining a second predicted click-through rate and a second predicted conversion rate for each user in the first user set for the first target resource using the second prediction model; determining a comprehensive prediction probability for each user regarding the first target resource based on the second predicted click-through rate and the second predicted conversion rate; and pushing the first target resource to multiple users with the highest comprehensive prediction probability via SMS. This achieves accurate resource push.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This disclosure relates to the field of mobile communication technology, and in particular to a resource push method, apparatus, electronic device, and storage medium. Background Technology

[0002] With the development of mobile communication technology, SMS, as an efficient way of transmitting information, has been widely used to push various resources (such as membership benefits, event information, etc.) to users.

[0003] In related technologies, the process of pushing resources to users via SMS often suffers from a lack of accurate identification of user needs, leading to a mismatch between pushed resources and user requirements. Inaccurate resource recommendations not only reduce conversion efficiency but also result in users receiving a large number of SMS messages unrelated to their interests or actual needs, causing information overload and interference. Therefore, how to push targeted resources to users who genuinely need them has become a pressing technical problem to be solved in this field. Summary of the Invention

[0004] This disclosure provides a resource push method, apparatus, electronic device, and storage medium.

[0005] In a first aspect, this disclosure provides a resource push method, the method comprising: determining a second user set from a first user set selected for a first target resource, wherein the number of users in the second user set is less than the number of users in the first user set; for each first user in the second user set, inputting the attribute information of the first user and the attribute information of the first target resource into a pre-trained first prediction model to obtain a first predicted click-through rate and a first predicted conversion rate for the first user regarding the first target resource; fine-tuning the first prediction model based on the first predicted click-through rate, the first predicted conversion rate, the click tags and conversion tags of the first user regarding the first target resource to obtain a second prediction model; for each second user in the first user set, inputting the attribute information of the second user and the attribute information of the first target resource into the second prediction model to obtain a second predicted click-through rate and a second predicted conversion rate for the second user regarding the first target resource; determining a comprehensive prediction probability for the second user regarding the first target resource based on the second predicted click-through rate and the second predicted conversion rate, and determining a number of third users with the highest comprehensive prediction probability from the first user set; and pushing the first target resource to the third users via SMS.

[0006] Secondly, this disclosure provides a resource push device, the device comprising: a first determining module, configured to determine a second user set from a first user set selected for a first target resource, wherein the number of users in the second user set is less than the number of users in the first user set; a first prediction module, configured to input the attribute information of the first user and the attribute information of the first target resource into a pre-trained first prediction model for each first user in the second user set, to obtain a first predicted click-through rate and a first predicted conversion rate for the first user regarding the first target resource; and a fine-tuning training module, configured to, based on the first predicted click-through rate, the first predicted conversion rate, and the click tags of the first user regarding the first target resource, ... The system includes a conversion tag and a fine-tuning training module for the first prediction model to obtain a second prediction model. A second prediction module is used to input the attribute information of each second user in the first user set and the attribute information of the first target resource into the second prediction model to obtain the second predicted click-through rate and the second predicted conversion rate for each second user regarding the first target resource. A second determination module is used to determine the comprehensive prediction probability of the second user regarding the first target resource based on the second predicted click-through rate and the second predicted conversion rate, and to determine the number of third users with the highest comprehensive prediction probability from the first user set. A push module is used to push the first target resource to the third users via SMS.

[0007] Thirdly, this disclosure provides an electronic device, including: 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 resource push method disclosed in the embodiments of this disclosure.

[0008] Fourthly, this disclosure provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the resource push method disclosed in the embodiments of this disclosure.

[0009] Fifthly, this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the resource push method disclosed in the embodiments of this disclosure.

[0010] The technical solutions provided by the embodiments of this disclosure bring at least the following beneficial effects: From the first user set selected for the first target resource, a second user set is determined. For each first user in the second user set, the attribute information of the first user and the attribute information of the first target resource are input into a pre-trained first prediction model to obtain the first predicted click-through rate and the first predicted conversion rate of the first user for the first target resource. Based on the first predicted click-through rate, the first predicted conversion rate, the click tags and conversion tags of the first user for the first target resource, the first prediction model is fine-tuned to obtain a second prediction model. The attribute information of the second users in the first user set and the attribute information of the first target resource are input into the second prediction model to obtain the second predicted click-through rate and the second predicted conversion rate of the second user for the first target resource. Based on the second predicted click-through rate and the second predicted conversion rate, the comprehensive prediction probability of the second user for the first target resource is determined. Then, the number of third users with the highest comprehensive prediction probability is determined from the first user set, and the first target resource is pushed to the third users via SMS. Therefore, by fine-tuning the second prediction model, the conversion rate and click-through rate of the user for the first target resource can be accurately determined. Based on the click-through rate and conversion rate, the comprehensive prediction probability of the user for the first target resource can be determined. The first target resource can then be sent to the user with the highest comprehensive prediction probability via SMS. This enables precise delivery of resources, which not only improves the resource conversion rate but also effectively reduces ineffective disturbance to users with low willingness, avoids information overload, and enhances user acceptance and satisfaction with SMS services. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0012] Figure 1 This is a flowchart illustrating a resource push method according to an exemplary embodiment; Figure 2 This is a schematic diagram illustrating the process of constructing a score matrix according to an exemplary embodiment; Figure 3 This is a flowchart illustrating a resource push method according to another exemplary embodiment; Figure 4 This is a schematic diagram of the model structure for the initial prediction model; Figure 5 This is a schematic diagram illustrating the structure of a resource push device according to an exemplary embodiment; Figure 6 This is a structural block diagram of an electronic device according to an exemplary embodiment. Detailed Implementation

[0013] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0014] The technical solutions of this disclosure and how they solve the aforementioned technical problems will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of this disclosure will now be described with reference to the accompanying drawings.

[0015] First, combine Figure 1 The resource push method provided in the embodiments of this disclosure will be described by way of example.

[0016] Figure 1 This is a flowchart illustrating a resource push method according to an exemplary embodiment.

[0017] It should be noted that the resource push method provided in this embodiment can be executed by a resource push device, which can be implemented by software and / or hardware. The resource push device can be an electronic device, or it can be configured in an electronic device to enable the electronic device to have resource push functionality. The electronic device can be any device with computing capabilities, such as a terminal device, a server, etc., and this embodiment does not specifically limit the electronic device.

[0018] like Figure 1 As shown, this resource push method includes the following steps: Step 101: Determine a second user set from the first user set selected for the first target resource, wherein the number of users in the second user set is less than the number of users in the first user set.

[0019] The first target resource mentioned above can be any resource to be pushed. For example, the first target resource may include, but is not limited to, membership benefits, marketing activity links, data packages, call duration, etc. This embodiment does not specifically limit the first target resource.

[0020] In some embodiments, filtering conditions set for a first target resource can be obtained. Correspondingly, users that meet the filtering conditions can be filtered from a preset user set, and a first user set can be formed based on the filtered users.

[0021] The filtering criteria may include, but are not limited to, user region, age range, activity level, and device type, etc., but this embodiment does not impose specific limitations on them.

[0022] The second set of users includes: users who have clicked on the first target resource, and / or users who have converted to the first target resource, and users who have neither clicked on nor converted to the first target resource.

[0023] In other embodiments, another possible way to obtain the first user set is to: obtain the business line to which the target resource belongs and the target region to which the target resource is to be deployed; obtain users located in the target region and belonging to the business line from the preset user set to obtain the first user set.

[0024] In some embodiments, the target region for which the target resource is to be delivered may be a region that the corresponding user pre-configures for the first target resource on the corresponding configuration page according to the push requirements.

[0025] In some embodiments, the aforementioned business lines can also be pre-configured for the first target resource in the configuration page. This satisfies the personalized needs of users to set the business line to which the first target resource belongs.

[0026] Step 102: For each first user in the second user set, input the attribute information of the first user and the attribute information of the first target resource into the pre-trained first prediction model to obtain the first predicted click-through rate and the first predicted conversion rate of the first user for the first target resource.

[0027] Step 103: Based on the first predicted click-through rate, the first predicted conversion rate, the first user's click tags and conversion tags for the first target resource, fine-tune the first prediction model to obtain the second prediction model.

[0028] In some embodiments, the click tags and conversion tags of the first user for the first target resource can be determined, and the total loss function value of the first prediction model can be determined based on the first predicted click-through rate, the first predicted conversion rate, the click tags and conversion tags of the first user for the first target resource, and the first prediction model can be fine-tuned and trained based on the total loss function value to obtain the second prediction model.

[0029] In some embodiments, the model parameters of the first prediction model can be adjusted according to the total loss function value, and the adjusted first prediction model can continue to be trained until a preset number of training times is reached or the total loss function meets a preset condition, so as to obtain the second prediction model.

[0030] The preset conditions are the conditions for the model training to end. These preset conditions can be configured according to actual needs. For example, satisfying the preset conditions could mean the total loss function value is less than a preset value, or that the change in the total loss function value approaches stability, meaning the difference between the total loss function values ​​of two or more adjacent training iterations is less than a set value, essentially indicating that the total loss function value no longer changes.

[0031] The preset number of training sessions is set in advance according to actual needs, and this embodiment does not specifically limit the value of the preset number of training sessions.

[0032] In some embodiments, one possible implementation for determining the total loss function value of the first prediction model based on the first predicted click-through rate, the first predicted conversion rate, the click tags of the first user for the first target resource, and the conversion tags is as follows: Determine the click loss function value of the first prediction model based on the first predicted click-through rate and click tags; determine the conversion loss function value of the first prediction model based on the first predicted conversion rate and conversion tags; and perform a weighted summation of the click loss function value and the conversion loss function value to obtain the total loss function value of the first prediction model. Thus, by performing a weighted summation of the click loss function value and the conversion loss value, the total loss function value of the first prediction model is accurately determined.

[0033] In some embodiments, to meet the personalized needs of users setting expected click weights and expected conversion weights for the first target resource, the expected click weights and expected conversion weights set for the first target resource can also be obtained. Correspondingly, a possible implementation of weighted summation of the click loss function value and the conversion loss function value to obtain the total loss function value of the first prediction model can be as follows: Based on the expected click weights and expected conversion weights, determine a first weight and a second weight, where the sum of the first weight and the second weight equals 1; weight the click loss function value according to the first weight, and weight the conversion loss function value according to the second weight; sum the weighted click loss function value and the conversion loss function value to obtain the total loss function value.

[0034] In some embodiments, the expected click weight and the expected conversion weight can be summed to obtain a summed result, and the expected click weight can be compared with the summed result to obtain a first weight.

[0035] In some embodiments, it is used to obtain the first weight. The formula can be:

[0036] Among them, in the formula Indicates the expected click weight; This indicates the expected conversion weight. For example, the expected click weight can be 3, and the expected conversion weight can be 7.

[0037] In some embodiments, after obtaining the first weight, 1 can be subtracted from the first weight to obtain the second weight. Wherein, the method for obtaining the second weight is... The formula can be expressed as: .

[0038] Among them, the value of the total loss function is obtained. The formula can be expressed as:

[0039] Among them, in the formula This represents the click loss function value; This represents the value of the transformation loss function.

[0040] As an example, the expected click weight and expected conversion weight can be set by the corresponding user in the corresponding configuration interface. This facilitates the personalized setting of the expected click weight and expected conversion weight of the primary target resource for each user.

[0041] The expected click weight represents the expected click weight of the first target resource after it has been pushed to the corresponding audience. A higher weight value indicates a higher probability that users within the audience will click on the first target resource.

[0042] The expected conversion weight represents the expected conversion weight of the first target resource after it is pushed to the corresponding audience. A larger weight value indicates a higher probability that the audience will convert to the first target resource; for example, in the case of a purchase conversion, a larger weight value indicates a higher probability that users in that audience will purchase the first target resource.

[0043] As an example, the expected click weight can range from 0 to 10. Correspondingly, the expected conversion weight can also range from 0 to 10.

[0044] In some embodiments, the click tag of the first user on the first target resource is determined based on whether the first user has clicked on the first target resource. For example, if the first user has clicked on the first target resource, the click tag of the first user on the first target resource is determined to be 1; if the first user has not clicked on the first target resource, the click tag of the first user on the first target resource is determined to be 0.

[0045] In some embodiments, the conversion tag of the first user for the first target resource can be determined based on whether the first user has a conversion behavior for the first target resource. For example, if the first user has a conversion behavior for the first target resource, the conversion tag of the first user for the first target resource is determined to be 1; if the first user has no conversion behavior for the first target resource, the click tag of the first user for the first target resource is determined to be 1.

[0046] In other embodiments, the conversion tag of the first user for the first target resource may also be determined based on whether the first user has conversion behavior for the first target resource and whether the first user has conversion behavior for the K second target resources most similar to the first target resource.

[0047] In some embodiments, the conversion tag of the first user for the first target resource can be determined as follows: from a preset resource library, obtain K second target resources that are most similar to the first target resource, where K is an integer greater than 1; determine the conversion tag of the first user for the first target resource based on whether the first user has converted for each second target resource in the past preset time period, whether the first user has converted for the first target resource in the past preset time period, and the similarity between the first target resource and each second target resource.

[0048] In some embodiments, the similarity value ranges from 0 to 1.

[0049] Wherein, K is a value preset according to actual needs. For example, K can be 10, 6, 8 or 9, etc. This embodiment does not specifically limit the value of K.

[0050] For example, if K is 10, after obtaining the 10 second target resources that are most similar to the first target resource, a similarity vector of length 11 can be formed, where the similarity vector J = [j0, j1, j2, ..., j...]. k ], where j0 represents the similarity between the first target resource and the first target resource, and the value of this similarity is equal to 1, where j i This represents the similarity score ranked i-th in the similarity ranking results. The similarity ranking results are obtained by sorting the similarity scores of the first target resource with each of the 10 target resources in descending (or ascending) order, where i ranges from 1 to 10. Additionally, after obtaining whether the first user has converted on each of the second target resources within a preset time period, and whether the first user has converted on the first target resource within a preset time period, a conversion score vector of length 11 can be formed. In the conversion score vector, 0 indicates no conversion behavior, and 1 indicates conversion behavior. The conversion score vector Z = [z0, z1, ... z2]. k], where z0 represents whether the first user has performed a conversion action on the first target resource within a preset time period. If there is a conversion action, the value of z0 is 1; otherwise, the value of z0 is 0. i This indicates whether the first user has made a conversion to the i-th second target resource within a preset time period. If so, then z... i The value of z is equal to 1. If there is no transformation behavior, then z i The value of is equal to 0; where the similarity between the i-th second target resource and the first target resource is located in the i-th position in the similarity ranking result. Correspondingly, after obtaining the similarity vector and the conversion score vector, the cosine similarity can be calculated between the similarity vector and the conversion score vector to obtain the conversion tag of the first user for the first target resource in the past preset time period. Wherein, the conversion tag of the first user for the first target resource in the past preset time period It can be represented as:

[0051] in, Indicates to Transpose.

[0052] The preset time is set according to actual needs. For example, the preset time can be 1 month, 2 months, or 3 months. This embodiment does not specifically limit the preset time.

[0053] Step 104: For each second user in the first user set, input the attribute information of the second user and the attribute information of the first target resource into the second prediction model to obtain the second predicted click-through rate and the second predicted conversion rate of the second user for the first target resource.

[0054] Step 105: Determine the comprehensive predicted probability of the second user for the first target resource based on the second predicted click-through rate and the second predicted conversion rate, and determine the number of third users with the highest comprehensive predicted probability from the first user set.

[0055] In some embodiments, the expected click weight and expected conversion weight set for the first target resource can be obtained. Based on the expected click weight and expected conversion weight, a first weight and a second weight are determined, wherein the sum of the first weight and the second weight is equal to 1. The second predicted click rate is weighted according to the first weight, the second predicted conversion rate is weighted according to the second weight, and the weighted second predicted click rate and predicted conversion rate are summed to obtain the comprehensive predicted probability of the second user for the first target resource.

[0056] The target number can be preset based on actual needs.

[0057] As an example, the target number of users can be pre-configured by the corresponding user for the first target resource on the configuration page. In other words, the corresponding user can pre-set the target number of users to be pushed to the first target resource on the corresponding configuration page. For example, the target number of users can be 100,000, 3 million, or 3.5 million, etc. This embodiment does not specifically limit the target number of users.

[0058] Step 106: Push the first target resource to the third user via SMS.

[0059] The resource push method provided in this embodiment of the present disclosure determines a second user set from a first user set selected for a first target resource. For each first user in the second user set, the attribute information of the first user and the attribute information of the first target resource are input into a pre-trained first prediction model to obtain a first predicted click-through rate and a first predicted conversion rate for the first user for the first target resource. Based on the first predicted click-through rate, the first predicted conversion rate, the click tags and conversion tags of the first user for the first target resource, the first prediction model is fine-tuned to obtain a second prediction model. The attribute information of the second users in the first user set and the attribute information of the first target resource are input into the second prediction model to obtain a second predicted click-through rate and a second predicted conversion rate for the second user for the first target resource. Based on the second predicted click-through rate and the second predicted conversion rate, the comprehensive prediction probability of the second user for the first target resource is determined. The number of third users with the highest comprehensive prediction probability is determined from the first user set. Finally, the first target resource is pushed to the third users via SMS. Therefore, by fine-tuning the second prediction model, the conversion rate and click-through rate of the user for the first target resource can be accurately determined. Based on the click-through rate and conversion rate, the comprehensive prediction probability of the user for the first target resource can be determined. The first target resource can then be sent to the user with the highest comprehensive prediction probability via SMS. This enables precise delivery of resources, which not only improves the resource conversion rate but also effectively reduces ineffective disturbance to users with low willingness, avoids information overload, and enhances user acceptance and satisfaction with SMS services.

[0060] Based on the above embodiments, the similarity between the first target resource and each of the second target resources can be determined as follows: From a pre-constructed scoring matrix, obtain a first column vector corresponding to the first target resource, wherein the first column vector includes: a first comprehensive score for each user in a pre-constructed user set regarding the first target resource within a past pre-constructed time period, the first comprehensive score being determined based on the user's first historical click-through rate and first historical conversion rate regarding the first target resource within the past pre-constructed time period; From the pre-constructed scoring matrix, obtain a second column vector corresponding to the second target resource, wherein the second column vector includes: a second comprehensive score for each user in a pre-constructed user set regarding the second target resource within a past pre-constructed time period, the second comprehensive score being determined based on the user's second historical click-through rate and second historical conversion rate regarding the second target resource within the past pre-constructed time period; Perform cosine similarity calculation on the first column vector and the second column vector to obtain the similarity between the first target resource and each of the second target resources.

[0061] It should be noted that both the first target resource and the second target resource are resources in a preset resource library. That is, both the first target resource and the second target resource belong to the preset resource library.

[0062] In some embodiments, an exemplary process for obtaining a first comprehensive score may be: determining the historical click weight of the first target resource within a preset time period; determining the historical conversion weight of the first target resource within a preset time period; and performing a weighted summation of the first historical click rate and the first historical conversion rate to obtain a first comprehensive score, wherein the weight of the first historical click rate is the historical click weight, and the weight of the first historical conversion rate is the historical conversion weight.

[0063] In other embodiments, an exemplary process for obtaining the second comprehensive score may be: determining the historical click weight of the first target resource within a preset time period; determining the historical conversion weight of the first target resource within a preset time period; and weighting and summing the second historical click rate and the second historical conversion rate to obtain the second comprehensive score, wherein the weight of the second historical click rate is the historical click weight, and the weight of the second historical conversion rate is the historical conversion weight.

[0064] In some embodiments, one possible implementation of determining the historical click weight of the first target resource within a preset time period is as follows: averaging the historical click rates of each resource in the preset resource library within the preset time period to obtain the average click rate; summing the first historical click rates of each user in the preset user set for the first target resource within the preset time period to obtain the first target click rate of the first target resource within the preset time period; and determining the historical click weight of the first target resource within the preset time period based on the first target click rate and the average click rate.

[0065] In some embodiments, one possible implementation of determining the historical conversion weight of the first target resource over a preset period of time is as follows: averaging the historical conversion rates of each resource in the preset resource library over the preset period of time to obtain the average conversion rate; summing the second historical conversion rates of each user in the preset user set for the second target resource over the preset period of time to obtain the first target conversion rate of the first target resource over the preset period of time; and determining the historical conversion weight of the second target resource over the preset period of time based on the first target conversion rate and the average conversion rate.

[0066] The pre-constructed score matrix includes the comprehensive score of any user in the preset user set for any resource in the preset resource library within the past preset time period.

[0067] In some embodiments, an exemplary process for constructing the score matrix is ​​as follows: Figure 2 As shown.

[0068] Figure 2 This is a schematic diagram illustrating the process of constructing a score matrix according to an exemplary embodiment.

[0069] like Figure 2 As shown, the process of constructing the score matrix may include: Step 201: Obtain resource P from the preset resource library within the past preset time period. j For user U in the preset user set i Exposure times E ij User U i Resource P within the past preset time period j Click count C ij and the number of conversions Z ij .

[0070] It should be noted that the value of i ranges from 1 to N, and the value of j ranges from 1 to M, where N represents the total number of users in the preset user set, and M represents the total number of resources in the preset resource library.

[0071] Step 202, based on exposure data E ij And click data C ij Determine user U i For resource P j Click-through rate α ij And based on click data C ij And the conversion data Z ij Determine user U i For resource P j Conversion rate β ij .

[0072] In some embodiments, for obtaining user U i For resource P j Click-through rate The formula can be expressed as: .

[0073] In some embodiments, for obtaining user U i For resource P j conversion rate The formula can be expressed as: .

[0074] Step 203, determine the resource P j Click-through rate α j And determine the resource P j Conversion rate β j .

[0075] In some embodiments, the resource P is used to obtain the resource P j Click-through rate The formula can be expressed as:

[0076] In some embodiments, the resource P is used to obtain the resource P j conversion rate The formula can be expressed as:

[0077] Among them, in the formula This indicates that resource P j Total number of exposures, of which ; This indicates that resource P j The total number of conversions, of which, ; This indicates that resource P j The total number of conversions, of which, .

[0078] Step 204: Average the click-through rate of each resource in the preset resource library to obtain the average click-through rate α, and average the conversion rate of each resource in the preset resource library to obtain the average conversion rate β.

[0079] In some embodiments, the average click-through rate is obtained. The formula can be expressed as:

[0080] In some embodiments, for obtaining average conversion rate The formula can be expressed as:

[0081] Step 205: Determine the historical click weight of resource Pj based on its click-through rate αj and average click-through rate α. Based on the conversion rate βj and average conversion rate β of resource Pj, the resource P is determined. j Historical conversion weight .

[0082] In some embodiments, for obtaining resource P j Historical click weight The formula can be expressed as:

[0083] In some embodiments, for obtaining resource P j Historical transformation weight The formula can be expressed as:

[0084] Step 206, for resource P j Click-through rate α j and conversion rate β j Perform weighted processing to obtain user U i Regarding resource P j The overall score, in which resource P j Click-through rate α j The weight is the historical click weight ω cj Among them, the resource P j Conversion rate β j The weight is the historical conversion weight. .

[0085] In some embodiments, for user U i Regarding resource P j Overall score The formula can be expressed as:

[0086] Step 207: Construct a score matrix based on the comprehensive scores of each user for each resource.

[0087] Among them, in the score matrix , indicating user U i Regarding resource P j The overall score. It's understandable that the rows in the score matrix represent users, and the columns represent resources.

[0088] In some embodiments, after constructing the score matrix, resource P can also be determined based on the score matrix. iand P j The similarity between them.

[0089] In some embodiments, the resource P can be obtained from the scoring matrix. i The corresponding column vector And obtain resource P j column vectors Then, for resource P i The corresponding column vector and resources P j column vectors Perform cosine similarity calculation to obtain resource P. i and P j The similarity between them.

[0090] In some embodiments, for obtaining resource P i and P j similarity between The formula can be expressed as:

[0091] in, For column vectors score i The L2 norm; column vector The L2 norm.

[0092] Among them, column vectors Let be the i-th column vector in the score matrix; column vector Let be the vector in the j-th column of the score matrix.

[0093] In some embodiments, the similarity between any user in a preset user set and each resource in a preset resource library is stored.

[0094] In some embodiments, in order to accurately obtain the similarity between the first target resource and the second target resource, the SMS push date preset for the first target resource can be obtained, and the similarity between any user in the stored preset user set and each resource in the preset resource library at the target date can be obtained. The similarity between the first target resource and each second target resource can be determined from the obtained similarity. The stored similarity is determined based on the target score matrix calculated at the target date, where the target date is the day before the SMS push date. The SMS push date refers to the date on which an SMS message containing the first target resource is sent to the corresponding user. In other words, an SMS message containing the first target resource is sent to the corresponding user on the SMS push date.

[0095] In some embodiments, in order to clearly understand this disclosure, the following references are made to... Figure 3 The method of this embodiment is described by way of example.

[0096] Figure 3 This is a flowchart illustrating a resource push method according to another exemplary embodiment.

[0097] like Figure 3 As shown, the method may include: Step 301: In response to the resource push request, display the corresponding configuration page.

[0098] Step 302: Obtain the first target resource configured on the configuration page, the business line of the first target resource, the target region to be delivered to the first target resource, the SMS push date of the first target resource, the target number of people to be delivered to the first target resource, and the expected click weight and expected conversion weight of the first target resource.

[0099] For example, the first target resource is a traffic package, the target region to be delivered to the first target resource can be region 1, the SMS push date of the first target resource can be September 2, 2024, the target number of people to be delivered to the first target resource can be 3 million, the expected click weight of the first target resource can be 3, and the expected conversion weight can be 7.

[0100] Step 303: Obtain the first set of users located in the target region and belonging to the business line from the preset user set.

[0101] Step 304: Determine a second user set from the first user set, wherein the number of users in the second user set is less than the number of users in the first user set.

[0102] Step 305: Obtain the first prediction model corresponding to the target date, where the target date is the day before the SMS push date.

[0103] For example, if the SMS push date is September 2, 2024, the corresponding target date could be September 1, 2024.

[0104] One possible implementation of the first prediction model is as follows: For each user in the preset user set, determine the click tag of the user on each resource in the preset resource library within the past preset time period of the target date, based on whether the user clicked on each resource in the preset resource library within the past preset time period of the target date; for each resource in the preset resource library, determine the K second resources with the highest similarity to the resource from the similarity between the resources stored at the target date, and determine the user's conversion tag for the resource based on whether the user converted on the second resource within the past preset time period of the target date, whether the user converted on the second resource within the past preset time period of the target date, and the similarity between the resource and each second resource. The user's attribute information and the resource's attribute information are input into the initial prediction model to obtain the user's predicted click-through rate and predicted conversion rate for the resource. Based on the click tags and predicted click-through rate, the click loss function value of the initial prediction model is determined. Based on the conversion tags and predicted conversion rate, the conversion loss function value of the initial prediction model is determined. Based on the click loss function value and conversion loss function value, the total loss function value of the initial prediction model is determined. Based on the total loss function value, the initial prediction model is pre-trained to obtain the first prediction model.

[0105] For example, if K is 10, when retrieving resource P from the preset resource library... i After identifying the 10 most similar secondary resources, a similarity vector of length 11 can be formed, where the similarity vector J = [j i0 j i1 j i2 , ..., j ik ], where j i0 Represents resource P i With resource P i The similarity between them, where the similarity value is equal to 1, and j ih This represents the similarity score ranked h-th in the similarity ranking results. The similarity ranking results are obtained by sorting the similarity scores of the first target resource with each of the top 10 target resources in descending (or ascending) order, where h ranges from 1 to 10. Additionally, it involves obtaining information on whether the user has converted data for each of the second resources within a preset time period, and the user's interactions with resource P within the preset time period. i After determining whether a transformation behavior exists, a transformation score vector of length 11 can be generated, where 0 in the transformation score vector indicates no transformation behavior and 1 indicates transformation behavior. The transformation score vector Z = [z...]. i0 , z i1 , ...z ik ], where z i0 Used to indicate user activity on resource P within a preset time period.i Does conversion behavior exist? If so, then z i0 The value of z is equal to 1. If there is no transformation behavior, then z i0 The value of z is equal to 0. ih This indicates whether the user has made a conversion to the i-th second resource within the past preset time period. If so, then z... ih The value of z is equal to 1. If there is no transformation behavior, then z ih The value of is equal to 0; where the i-th second resource is related to resource P. i The similarity between them is located in the i-th position of the similarity ranking results. Correspondingly, after obtaining the similarity vector and the conversion score vector, a cosine similarity calculation can be performed on the similarity vector and the conversion score vector to obtain the user's conversion tag for the first target resource within a preset time period. Here, the user's conversion tag for resource P within the preset time period... i conversion tags It can be represented as:

[0106] in, Indicates to Transpose.

[0107] The preset time is set according to actual needs. For example, the preset time can be 1 month, 2 months, or 3 months. This embodiment does not specifically limit the preset time.

[0108] In some embodiments, the click loss function of the initial prediction model can be the binary cross-entropy loss function.

[0109] For example, if the click loss function of the initial prediction model is the binary cross-entropy loss function, and the training data used to train the initial prediction model includes n samples and their corresponding click labels and conversion labels, then the click loss function of the initial prediction model... It can be represented as:

[0110] Among them, in the formula This represents the click tag corresponding to the i-th sample. This represents the predicted click-through rate corresponding to the i-th sample, where the i-th sample includes the attribute information of the sample user and the attribute information of the sample resource.

[0111] In some embodiments, the transformation loss function of the initial prediction model can be the variance loss function.

[0112] For example, if the conversion loss function of the initial prediction model is the variance loss function, and the training data used to train the initial prediction model includes n samples and their corresponding click and conversion labels, then the conversion loss function of the initial prediction model is... It can be represented as:

[0113] Among them, in the formula This represents the conversion label corresponding to the i-th sample. This represents the predicted conversion rate corresponding to the i-th sample, where the i-th sample includes the attribute information of the sample user and the attribute information of the sample resource.

[0114] It should be noted that the initial prediction model can be an untrained neural network model or a pre-trained prediction model. For example, the initial prediction model can be the prediction model corresponding to the first target date, where the first target date is the day before the target date.

[0115] In some embodiments, a schematic diagram of the model structure of the initial prediction model is shown, such as... Figure 4 As shown, through Figure 4 As can be seen, the initial prediction model can include an embedding representation layer, a discretization encoding layer, a feature processing layer, and click prediction and conversion prediction networks connected to the feature processing layer, respectively. The feature processing layer is connected to both the embedding representation layer and the discretization encoding layer. It should be noted that the embedding representation layer is used to embed the discrete features in the samples, while the discretization encoding layer is used to discretize and encode the continuous features in the samples. For example, the continuous features can be binned, and the binning results can be one-hot encoded; or the binning results can be normalized, and the normalized binning results can be one-hot encoded. The feature processing layer processes the output features of the embedding representation layer and the discretization encoding layer; the click prediction network predicts clicks based on the output features of the feature processing layer to output the predicted click-through rate; and the conversion prediction network predicts conversions based on the output features of the feature processing layer to output the predicted conversion rate.

[0116] In some embodiments, after the first prediction model is trained, it can be combined with the target verification user set to determine whether the first prediction model passes verification.

[0117] In some embodiments, the attribute information of users in the target verification user set and the attribute information of sample resources can also be input into the first prediction model to obtain the target predicted click-through rate and the target conversion rate. Based on the actual click tags of users in the target verification user set for the sample resources and the target predicted click-through rate, the click-through rate score of the target verification user set is determined. Based on the actual conversion tags of users in the target verification user set for the sample resources and the target predicted conversion rate, the conversion rate score of the target verification user set is determined. The click-through rate score and the conversion rate score are averaged and weighted to obtain an average weighted result. If the average weighted result meets the first target condition, the first prediction model is determined to have passed the verification.

[0118] In some embodiments, the actual click tags and the target predicted click-through rate (CTR) can be input into a calculation formula for calculating the CTR score, thereby obtaining the actual click tags and the target predicted CTR through the formula. In other embodiments, a first target number of target users whose actual click tags and target predicted CTRs match can be determined in the target verified user set, and the CTR score of the target verified user set can be determined based on the first target number of users. The first target number of users and the CTR score are positively correlated. In some embodiments, the actual conversion tag and the target predicted conversion rate can be input into a calculation formula for calculating the conversion rate score, so that the actual conversion tag and the target predicted conversion rate can be obtained through the calculation formula. In other embodiments, a second target number of target users whose actual conversion tags and target predicted conversion rates match can be determined, and the conversion rate score of the target verified user set can be determined based on the second target number of users. The second target number of users and the conversion rate score are positively correlated.

[0119] The first target condition is preset according to actual needs. For example, the first target condition can be that the average weighted result is greater than a preset threshold, or that the improvement ratio of the average weighted result for 5 consecutive times is less than 1%. This embodiment does not specifically limit the first target condition.

[0120] Step 306: Divide the second user set into a fine-tuning user set and a verification user set according to a preset ratio.

[0121] The preset ratio is set in advance according to actual needs. For example, the preset ratio can be 7:3. This embodiment does not specifically limit the value of the preset ratio.

[0122] Step 307: For each user in the fine-tuning user set, input the user's attribute information and the attribute information of the first target resource into the first prediction model to obtain the user's first predicted click-through rate and first predicted conversion rate for the first target resource.

[0123] Step 308: Determine the click loss function value of the first prediction model based on the click tags and the first predicted click rate of each user in the fine-tuned user set within the past preset time period, and determine the conversion loss function value of the first prediction model based on the conversion tags and the first predicted conversion rate of each user in the fine-tuned user set within the past preset time period.

[0124] For details on the specific implementation of determining click tags and conversion tags, please refer to the relevant descriptions in other embodiments, which will not be repeated here.

[0125] Step 309: Determine the first weight and the second weight based on the expected click weight and the expected conversion weight. Weight the click loss function value according to the first weight and the conversion loss function value according to the second weight. Sum the weighted click loss function value and the conversion loss function value to obtain the total loss function value of the first prediction model.

[0126] In some embodiments, the total loss function value of the first prediction model ,in It is the first weight. Indicates the second weight, where, Where 'a' represents the expected click weight entered on the configuration page, and 'b' represents the expected conversion weight entered on the configuration page. For example, if 'a' is 3 and 'b' is 7, then the total loss function value of the first prediction model is... ,in, This represents the click loss function value. This represents the value of the transformation loss function.

[0127] Step 310: Based on the total loss function value, fine-tune the first prediction model to obtain the second prediction model.

[0128] Step 311: Validate the second prediction model based on the set of validating users to obtain the validation results.

[0129] In some embodiments, a possible implementation of validating the second prediction model based on the validating user set to obtain the validation result is as follows: Calculate the ratio of the target number of users to the number of users in the first user set to obtain a first ratio value; for each user in the validating user set, input the user's attribute information and the attribute information of the first target resource into the second prediction model to obtain the user's second predicted click-through rate and second predicted conversion rate for the first target resource; weight the second predicted click-through rate according to a first weight, and weight the second predicted conversion rate according to a second weight; sum the weighted second predicted click-through rate and second predicted conversion rate to obtain a weighted prediction probability; sort the users in the validating user set in descending order according to the weighted prediction probability to obtain a ranking result; extract the users with the first ratio value before ranking from the ranking result to form a target user subset; determine the first number L of users in the validating user set who actually clicked on the first target resource; determine the second number Q of users in the target user subset who actually clicked on the first target resource; and determine the clicking user score S based on the first number L and the second number Q. ctr Determine the third number W of users in the verification user set who have actually converted for the first target resource, and the fourth number E of users in the target user subset who have actually converted for the first target resource. Based on the third number W and the fourth number E, determine the conversion user score. ;Score the clicking user based on the first weight. ctr Weighted scores are applied, and converted users are scored based on the second weight. Weighting is performed, and the weighted user score S is calculated. ctr and conversion user score The results are summed to obtain a comprehensive score S; if the comprehensive score S satisfies the second objective condition, the second prediction model is verified.

[0130] The second objective condition is preset according to actual needs. For example, the second objective condition can be that the comprehensive score S is greater than a preset score threshold, or that the comprehensive score S increases by less than 1% for 5 consecutive times. This embodiment does not specifically limit the second objective condition.

[0131] In some embodiments, the click user score S is determined based on a first quantity L and a second quantity Q. ctr One possible implementation is to compare the first quantity L with the second quantity Q to obtain the click user score S. ctr .

[0132] In some embodiments, the click user score S is obtained. ctr The formula can be expressed as:

[0133] In some embodiments, the conversion user score is determined based on the third quantity W and the fourth quantity E. One possible implementation is to compare the third quantity W with the fourth quantity E to obtain the converted user score. .

[0134] In some embodiments, a conversion user score is obtained. The formula can be expressed as:

[0135] In some embodiments, the formula for obtaining the overall score S can be expressed as:

[0136] For example, if the first weight is set to 0.3 and the second weight to 0.7, the formula for obtaining the overall score S can be expressed as: .

[0137] Step 312: If the second prediction model passes the verification, for each user in the first user set, send the user's attribute information and the attribute information of the first target resource to the second prediction model to obtain the user's second predicted click-through rate and second predicted conversion rate for the second target resource.

[0138] Step 313: Weight the second predicted click-through rate according to the first weight, and weight the second predicted conversion rate according to the second weight. Then sum the weighted second predicted click-through rate and the second predicted conversion rate to obtain the comprehensive predicted probability of the user for the first target resource.

[0139] Step 314: Determine the number of third target users with the highest comprehensive prediction probability from the first user set.

[0140] Step 315: If the current date is the SMS push date, send the first target resource to the third user via SMS.

[0141] The resource push method provided in this disclosure, after obtaining a set of users located in the target region and belonging to the business line, fine-tunes a pre-trained first prediction model using attribute information of some users in the user set and attribute information of the first target resource to obtain a second prediction model. The second prediction model obtained through fine-tuning accurately determines the conversion rate and click-through rate of the user for the first target resource, and determines the comprehensive prediction probability of the user for the first target resource based on the click-through rate and conversion rate. The first target resource is then sent to the user with the highest comprehensive prediction probability via SMS. This enables precise resource push, improves resource conversion rate, effectively reduces invalid disturbance to low-interest users, avoids information overload, and enhances user acceptance and satisfaction with SMS services.

[0142] Figure 5 This is a schematic diagram illustrating the structure of a resource push device according to an exemplary embodiment.

[0143] like Figure 5 As shown, the resource push device 500 includes: a first determination module 501, a first prediction module 502, a fine-tuning training module 503, a second prediction module 504, a second determination module 505, and a push module 506, wherein: The first determining module 501 is used to determine a second user set from the first user set selected for the first target resource, wherein the number of users in the second user set is less than the number of users in the first user set.

[0144] The first prediction module 502 is used to input the attribute information of the first user and the attribute information of the first target resource into the pre-trained first prediction model for each first user in the second user set, so as to obtain the first predicted click-through rate and the first predicted conversion rate of the first user for the first target resource.

[0145] The fine-tuning training module 503 is used to fine-tune the first prediction model based on the first predicted click-through rate, the first predicted conversion rate, the first user's click tags and conversion tags for the first target resource, and to obtain the second prediction model.

[0146] The second prediction module 504 is used to input the attribute information of the second user and the attribute information of the first target resource into the second prediction model for each second user in the first user set, so as to obtain the second predicted click-through rate and the second predicted conversion rate of the second user for the first target resource.

[0147] The second determining module 505 is used to determine the comprehensive predicted probability of the second user for the first target resource based on the second predicted click-through rate and the second predicted conversion rate, and to determine the number of third users with the highest comprehensive predicted probability from the first user set.

[0148] The push module 506 is used to push the first target resource to a third user via SMS.

[0149] In one embodiment of this disclosure, the fine-tuning training module 503 is specifically configured to: determine the click tags and conversion tags of the first user for the first target resource; determine the click loss function value of the first prediction model based on the first predicted click-through rate and click tags; determine the conversion loss function value of the first prediction model based on the first predicted conversion rate and conversion tags; perform a weighted summation of the click loss function value and the conversion loss function value to obtain the total loss function value of the first prediction model; and fine-tune the first prediction model based on the total loss function value to obtain a second prediction model.

[0150] In one embodiment of this disclosure, the device may further include: The first acquisition module is used to acquire the expected click weight and expected conversion weight set for the first target resource; Specifically, the fine-tuning training module 503 performs a weighted sum of the click loss function value and the conversion loss function value to obtain the total loss function value of the first prediction model in the following way: Based on the expected click weight and the expected conversion weight, a first weight and a second weight are determined, wherein the sum of the first weight and the second weight is equal to 1; the click loss function value is weighted according to the first weight, and the conversion loss function value is weighted according to the second weight; the weighted click loss function value and the conversion loss function value are summed to obtain the total loss function value.

[0151] In one embodiment of this disclosure, the fine-tuning training module 503 determines the comprehensive predicted probability of the second user for the first target resource based on the second predicted click-through rate and the second predicted conversion rate in the following manner: the second predicted click-through rate is weighted according to the first weight, and the second predicted conversion rate is weighted according to the second weight; the weighted second predicted click-through rate and the second predicted conversion rate are summed to obtain the comprehensive predicted probability of the second user for the first target resource.

[0152] In one embodiment of this disclosure, the conversion tag of the first user for the first target resource is determined as follows: from a preset resource library, K second target resources that are most similar to the first target resource are obtained, where K is an integer greater than 1; based on whether the first user has converted each of the second target resources in the past preset time period, whether the first user has converted the first target resource in the past preset time period, and the similarity between the first target resource and each of the second target resources, the conversion tag of the first user for the first target resource is determined.

[0153] In one embodiment of this disclosure, the similarity between the first target resource and each of the second target resources is determined as follows: A first column vector corresponding to the first target resource is obtained from a pre-constructed scoring matrix. The first column vector includes: a first comprehensive score for each user in a pre-constructed user set regarding the first target resource within a pre-constructed time period. The first comprehensive score is determined based on the user's first historical click-through rate and first historical conversion rate for the first target resource within the pre-constructed time period. A second column vector corresponding to the second target resource is obtained from the pre-constructed scoring matrix. The second column vector includes: a second comprehensive score for each user in the pre-constructed user set regarding the second target resource within a pre-constructed time period. The second comprehensive score is determined based on the user's second historical click-through rate and second historical conversion rate for the second target resource within the pre-constructed time period. Cosine similarity is calculated between the first column vector and the second column vector to obtain the similarity between the first target resource and each of the second target resources.

[0154] In one embodiment of this disclosure, the apparatus further includes: a preprocessing module, configured to: determine the historical click weight of the first target resource within a preset time period; determine the historical conversion weight of the first target resource within a preset time period; perform a weighted summation of the first historical click rate and the first historical conversion rate to obtain a first comprehensive score, wherein the weight of the first historical click rate is the historical click weight, and the weight of the first historical conversion rate is the historical conversion weight; and perform a weighted summation of the second historical click rate and the second historical conversion rate to obtain a second comprehensive score, wherein the weight of the second historical click rate is the historical click weight, and the weight of the second historical conversion rate is the historical conversion weight.

[0155] In one embodiment of this disclosure, the specific implementation of determining the historical click weight of the first target resource within a preset time period can be as follows: averaging the historical click rates of each resource in the preset resource library within the preset time period to obtain the average click rate; summing the first historical click rates of each user in the preset user set for the first target resource within the preset time period to obtain the first target click rate of the first target resource within the preset time period; and determining the historical click weight of the first target resource within the preset time period based on the first target click rate and the average click rate.

[0156] In one embodiment of this disclosure, the specific method for determining the historical conversion weight of the first target resource within a preset time period may be as follows: averaging the historical conversion rates of each resource in the preset resource library within the preset time period to obtain the average conversion rate; summing the second historical conversion rates of each user in the preset user set for the second target resource within the preset time period to obtain the first target conversion rate of the first target resource within the preset time period; and determining the historical conversion weight of the second target resource within the preset time period based on the first target conversion rate and the average conversion rate.

[0157] In one embodiment of this disclosure, the device may further include: The second acquisition module is used to acquire the business line to which the target resource belongs and the target region where the target resource is to be deployed. The third acquisition module is used to acquire users located in the target region and belonging to the business line from the preset user set to obtain the first user set.

[0158] It should be noted that the foregoing description of the resource push method embodiment also applies to the resource push device of this embodiment, and will not be repeated here.

[0159] It should be noted that the apparatus provided in this disclosure corresponds to the aforementioned method embodiments. Its technical effects are the same as those in the aforementioned method embodiments; please refer to the description in the method section for details.

[0160] According to embodiments of this disclosure, an electronic device is also provided, including: a processor; and a memory for storing processor-executable instructions, wherein the processor is configured to: implement the resource push method disclosed in embodiments of this disclosure.

[0161] To implement the above embodiments, this disclosure also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the resource push method disclosed in this disclosure.

[0162] To implement the above embodiments, this disclosure also provides a computer program product. This computer program product includes a computer program that, when executed by a processor, implements the resource push method disclosed in this disclosure.

[0163] Figure 6 This is a structural block diagram of an electronic device according to an exemplary embodiment. Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0164] like Figure 6As shown, the electronic device 1000 includes a processor 111, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 112 or a program loaded from memory 116 into random access memory (RAM) 113. The RAM 113 also stores various programs and data required for the operation of the electronic device 1000. The processor 111, ROM 112, and RAM 113 are interconnected via a bus 114. An input / output (I / O) interface 115 is also connected to the bus 114.

[0165] The following components are connected to I / O interface 115: memory 116 including hard disks, etc.; and communication section 117 including network interface cards such as local area network (LAN) cards, modems, etc., communication section 117 performs communication processing via a network such as the Internet; and driver 118 is also connected to I / O interface 115 as needed.

[0166] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 117. When the computer program is executed by processor 111, it performs the functions defined in the methods of this disclosure.

[0167] In an exemplary embodiment, a storage medium including instructions is also provided, such as a memory including instructions, which can be executed by the processor 111 of the electronic device 1000 to perform the above-described method. Optionally, the storage medium may be a non-transitory computer-readable storage medium, such as a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device.

[0168] In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wireline, optical fiber, RF, etc., or any suitable combination thereof.

[0169] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0170] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A resource push method, characterized in that, The method includes: From the first set of users selected for the first target resource, a second set of users is determined, wherein the number of users in the second set of users is less than the number of users in the first set of users; For each first user in the second user set, the attribute information of the first user and the attribute information of the first target resource are input into the pre-trained first prediction model to obtain the first predicted click-through rate and the first predicted conversion rate of the first user for the first target resource. Based on the first predicted click-through rate, the first predicted conversion rate, the first user's click tags and conversion tags for the first target resource, the first prediction model is fine-tuned and trained to obtain a second prediction model. The click tags are determined based on whether the first user clicked on the first target resource. The conversion tags for the first user's first target resource are determined as follows: from a preset resource library, K second target resources most similar to the first target resource are obtained, where K is an integer greater than 1; based on whether the first user converted on each of the second target resources within a preset time period, whether the first user converted on the first target resource within a preset time period, and the similarity between the first target resource and each of the second target resources, the conversion tags for the first user's first target resource are determined. For each second user in the first user set, the attribute information of the second user and the attribute information of the first target resource are input into the second prediction model to obtain the second predicted click-through rate and the second predicted conversion rate of the second user for the first target resource. The comprehensive predicted probability of the second user for the first target resource is determined based on the second predicted click-through rate and the second predicted conversion rate, and the number of third users with the highest comprehensive predicted probability are determined from the first user set. The first target resource is pushed to the third user via SMS; The similarity between the first target resource and each of the second target resources is determined as follows: From the pre-constructed score matrix, obtain the first column vector corresponding to the first target resource, wherein the first column vector includes: the first comprehensive score of each user in the preset user set for the first target resource in the past preset time period, and the first comprehensive score is determined based on the first historical click rate and the first historical conversion rate of the user for the first target resource in the past preset time period; From the pre-constructed score matrix, obtain the second column vector corresponding to the second target resource, wherein the second column vector includes: the second comprehensive score of each user in the preset user set for the second target resource in the past preset time period, and the second comprehensive score is determined based on the user's second historical click rate and second historical conversion rate for the second target resource in the past preset time period; The cosine similarity is calculated on the first column vector and the second column vector to obtain the similarity between the first target resource and each of the second target resources.

2. The method as described in claim 1, characterized in that, The step of fine-tuning and training the first prediction model based on the first predicted click-through rate, the first predicted conversion rate, the first user's click tags and conversion tags for the first target resource, to obtain the second prediction model includes: Determine the click tags and conversion tags of the first user for the first target resource; Based on the first predicted click-through rate and the click tag, determine the click loss function value of the first prediction model; Based on the first predicted conversion rate and the conversion label, determine the conversion loss function value of the first prediction model; The total loss function value of the first prediction model is obtained by weighted summing of the click loss function value and the conversion loss function value. Based on the total loss function value, the first prediction model is fine-tuned and trained to obtain the second prediction model.

3. The method as described in claim 2, characterized in that, Before determining the second user set from the first user set selected for the first target resource, the method further includes: Obtain the expected click weight and expected conversion weight set for the first target resource; The step of weighted summing of the click loss function value and the conversion loss function value to obtain the total loss function value of the first prediction model includes: Based on the expected click weight and the expected conversion weight, a first weight and a second weight are determined, wherein the sum of the first weight and the second weight is equal to 1; The click loss function value is weighted according to the first weight, and the conversion loss function value is weighted according to the second weight; The total loss function value is obtained by summing the weighted click loss function value and the conversion loss function value.

4. The method as described in claim 3, characterized in that, The step of determining the comprehensive predicted probability of the second user for the first target resource based on the second predicted click-through rate and the second predicted conversion rate includes: The second predicted click-through rate is weighted according to the first weight, and the second predicted conversion rate is weighted according to the second weight. The weighted second predicted click-through rate and the second predicted conversion rate are summed to obtain the comprehensive predicted probability of the second user for the first target resource.

5. The method as described in claim 1, characterized in that, The method includes: Determine the historical click weight of the first target resource within a preset time period; Determine the historical conversion weight of the first target resource over a preset period of time; The first historical click-through rate and the first historical conversion rate are weighted and summed to obtain the first comprehensive score, wherein the weight of the first historical click-through rate is the historical click weight, and the weight of the first historical conversion rate is the historical conversion weight. The second historical click-through rate and the second historical conversion rate are weighted and summed to obtain the second comprehensive score, wherein the weight of the second historical click-through rate is the historical click weight, and the weight of the second historical conversion rate is the historical conversion weight.

6. The method as described in claim 5, characterized in that, Determining the historical click weight of the first target resource over a preset time period includes: The average click rate is obtained by averaging the historical click rates of each resource in the preset resource library over a preset time period. The first historical click-through rates of each user in the preset user set for the first target resource within a preset time period are summed to obtain the first target click-through rate of the first target resource within the preset time period. Based on the first target click-through rate and the average click-through rate, the historical click weight of the first target resource over a preset period of time is determined.

7. The method as described in claim 5, characterized in that, Determining the historical conversion weight of the first target resource over a preset period of time includes: The average conversion rate is obtained by averaging the historical conversion rates of each resource in the preset resource library over a preset period of time. The second historical conversion rate of each user in the preset user set for the second target resource within a preset time period is summed to obtain the first target conversion rate of the first target resource within the preset time period. Based on the first target conversion rate and the average conversion rate, the historical conversion weight of the second target resource over a preset period of time is determined.

8. The method according to any one of claims 1-7, characterized in that, Before determining the second user set from the first user set selected for the target resource, the method further includes: Obtain the business line to which the target resource belongs and the target region where the target resource is to be deployed; From the preset user set, users located in the target region and belonging to the business line are obtained to obtain the first user set.

9. A resource delivery device, characterized in that, The device includes: The first determining module is configured to determine a second user set from a first user set selected for the first target resource, wherein the number of users in the second user set is less than the number of users in the first user set; The first prediction module is used to input the attribute information of the first user and the attribute information of the first target resource into a pre-trained first prediction model for each first user in the second user set, so as to obtain the first predicted click-through rate and the first predicted conversion rate of the first user for the first target resource. The fine-tuning training module is used to fine-tune the first prediction model based on the first predicted click-through rate, the first predicted conversion rate, the first user's click tags for the first target resource, and conversion tags, to obtain a second prediction model. The click tags are determined based on whether the first user clicked on the first target resource. The conversion tags for the first user's first target resource are determined as follows: from a preset resource library, obtain the K most similar second target resources to the first target resource, where K is an integer greater than 1; determine the first user's conversion tags for the first target resource based on whether the first user converted on each of the second target resources within a preset time period, whether the first user converted on the first target resource within a preset time period, and the similarity between the first target resource and each of the second target resources. The second prediction module is used to input the attribute information of the second user and the attribute information of the first target resource into the second prediction model for each second user in the first user set, so as to obtain the second predicted click-through rate and the second predicted conversion rate of the second user for the first target resource. The second determining module is used to determine the comprehensive predicted probability of the second user for the first target resource based on the second predicted click-through rate and the second predicted conversion rate, and to determine the number of third users with the highest comprehensive predicted probability from the first user set. The push module is used to push the first target resource to the third user via SMS; The similarity between the first target resource and each of the second target resources is determined as follows: From the pre-constructed score matrix, obtain the first column vector corresponding to the first target resource, wherein the first column vector includes: the first comprehensive score of each user in the preset user set for the first target resource in the past preset time period, and the first comprehensive score is determined based on the first historical click rate and the first historical conversion rate of the user for the first target resource in the past preset time period; From the pre-constructed score matrix, obtain the second column vector corresponding to the second target resource, wherein the second column vector includes: the second comprehensive score of each user in the preset user set for the second target resource in the past preset time period, and the second comprehensive score is determined based on the user's second historical click rate and second historical conversion rate for the second target resource in the past preset time period; The cosine similarity is calculated on the first column vector and the second column vector to obtain the similarity between the first target resource and each of the second target resources.

10. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-8.

12. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method of any one of claims 1-8.

Citation Information

Patent Citations

  • Recommendation model training method, electronic equipment and storage medium

    CN113868523A

  • Internet online advertisement accurate putting method based on multi-task learning

    CN114298734A