Method for pushing material, electronic device, storage medium and program product

Abstract

Embodiments of the present application disclose a method for pushing material, an electronic device, a storage medium and a computer program product. The method for pushing material comprises inputting user information of a user into a business line preference prediction model, performing user preference probability distribution calculation for each business line respectively, and determining user preference probability distribution values of each business line; wherein the user information of the user comprises user behavior information of the user; inputting the user behavior information of the user into a material proportion prediction model, performing proportion calculation of different types of materials of each business line, and obtaining different type material proportion values of the user in each business line; and from a set of alternative materials, according to the determined user preference probability distribution values of each business line and the different type material proportion values of the user in the corresponding business line, screening out materials and pushing the materials to the user.

Description

Technical Field

[0001] This application relates to the field of computer network technology, and in particular to a method for pushing materials, an electronic device, a storage medium, and a computer program product. Background Technology

[0002] With the development of computer networks, more and more online platforms are hosted on these networks, providing various services to users. To better serve users, these platforms often need to display materials that users are interested in on their web pages. Currently, online platforms have evolved from tag-based push notifications of materials under a single business line to personalized feed-based push notifications of materials under personalized business lines. This has resulted in the push notification materials expanding from a single business line to multiple business lines, and the types of materials increasing from a single type, such as housing information, to various types, such as housing information, ranking information, live broadcast information, or shopping guide articles. How to identify the materials that users are interested in from multiple business lines and various types of alternative materials, and then accurately push them, is a core issue for online platforms in this push notification scenario. The more accurately the online platform predicts the business lines that users prefer, the more it can increase the depth of user behavior on the platform's web pages and improve user retention rates; the more accurately the platform predicts the types of materials that users prefer, the more it can improve the conversion rate of user behavior on the platform's web pages, such as click-through rates and whether a transaction is completed.

[0003] Currently, how to accurately predict and push materials that users are interested in, and improve the conversion rate and retention rate of user behavior on online platforms, is an urgent problem to be solved when pushing materials to users on online platforms. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method and system for pushing materials, which can accurately predict and push materials that users are interested in, thereby improving the conversion rate and retention rate of user behavior on the network platform.

[0005] The embodiments of this application are implemented as follows:

[0006] A method for pushing materials, the method comprising:

[0007] The user's information is input into the business line preference prediction model. For each business line, the user preference probability distribution is calculated to determine the user preference probability distribution value for each business line. The user's information includes the user's user behavior information.

[0008] The user's behavior information is input into the material ratio prediction model to calculate the ratio of different types of materials for each business line, and the ratio of different types of materials for each user in each business line is obtained.

[0009] From the set of candidate materials, materials are selected and pushed to the user based on the determined probability distribution value of user preferences for each business line and the ratio of different types of materials for the user in the corresponding business line.

[0010] Preferably, the business line preference prediction model includes: a personalized business line preference sub-model, a user profile-based business line preference sub-model, and a regional statistical business line preference sub-model;

[0011] The step of inputting user information into the business line preference prediction model, and calculating the probability distribution of user preferences for each business line to determine the probability distribution value of user preferences for each business line includes:

[0012] If the user preference probability distribution value of each business line can be calculated using the personalized business line preference sub-model, the user information of the user is input into the personalized business line preference sub-model, and the calculated personalized business line preference probability distribution value is used as the corresponding business line user preference probability distribution value.

[0013] In cases where it is impossible to calculate the user preference probability distribution value for each business line using a personalized business line preference sub-model, and where the user information also includes profile information, the profile information from the user information is input into a business line preference sub-model based on user profiles. The calculated probability distribution value of the user's transition behavior from the current behavior to the next adjacent behavior within the corresponding business line is used as the user preference probability distribution value for that business line.

[0014] When it is not possible to calculate the probability distribution value of user preferences for each business line using a business line preference sub-model based on user profiles, a regional statistical business preference sub-model is used to calculate the distribution of the proportion of materials of a set type under each business line in the region to which the user belongs, which is then used as the probability distribution value of user preferences for each business line in each region.

[0015] Preferably, the business line preference prediction model is trained using a tree model, wherein the personalized business line preference sub-model is trained using a tree model based on the extreme gradient boosting (XGBoost) algorithm, and the user profile-based business line preference sub-model is trained using a tree model based on the XGBoost algorithm.

[0016] Preferably, the material proportioning prediction model is a material proportioning prediction model with constraints, and the step of inputting the user's user behavior information into the material proportioning prediction model to calculate the proportioning values ​​of different types of materials for each business line includes:

[0017] Based on the corresponding business line, the material ratio prediction model with constraints calculates the baseline ratio values ​​of different types of materials.

[0018] Based on the constraints, the calculated baseline proportions of the different types of materials are adjusted to meet the constraints.

[0019] Preferably, the baseline proportion values ​​of different types of materials calculated by the constrained material proportion prediction model include:

[0020] When a user is determined to be a new user, the ratio between the amount of user behavior on a certain type of material and the number of times the material is exposed within a set time period is averaged and normalized within the set time period and then used as the baseline ratio value for that type of material.

[0021] When a user is determined to be a non-new user, the ratio between the user's behavior volume on a certain type of material and the number of times the material is exposed within the set time period is used as the baseline ratio value for that type of material.

[0022] Preferably, the adjustment to meet the constraints for different types of material proportions includes:

[0023] 1) For each type of material, determine whether the baseline proportion value of the material type is less than the set minimum constraint value of the material type. If so, use the minimum constraint value of the material type as the proportion value of the material type.

[0024] 2) If not, multiply the baseline proportion value of the material of the type by the calculated proportion value to obtain the updated proportion value of the material of the type. The calculated proportion value is the ratio between the remaining value of the sum of the proportion values ​​of all other materials of the type for which the minimum constraint value of other types is set, and the sum of the baseline proportion values ​​of all other materials of the type for which the minimum constraint value of other types is not set.

[0025] 3) Determine whether the updated proportion value of the material of the type is greater than the set maximum constraint value of the material of the type. If so, use the maximum constraint value of the material of the type as the proportion value of the material of the type.

[0026] 4) If not, the proportion value of the material of the aforementioned type is averaged with the proportion value of other types of materials that have not been set with the maximum constraint value corresponding to other types, and the maximum proportion difference value is obtained by distributing the maximum proportion difference value equally among the proportion values ​​of the material of the aforementioned type. The maximum proportion difference value is the difference between the proportion value of other types of materials that have been set with the maximum constraint value corresponding to other types and the maximum constraint value corresponding to other types of materials before the update.

[0027] 5) Determine whether the proportion of all types of materials under the business line meets the constraint condition, which is that the proportion of all types of materials is between the set minimum and maximum constraint values. If not, return to step 4) to continue execution; if yes, end this process.

[0028] Preferably, the step of selecting materials for push notification based on the determined user preference probability distribution value for each business line and the user's allocation of different types of materials in the corresponding business line includes:

[0029] From the set of candidate materials, select the business line to which the material belongs as the candidate material for the current business line, and the proportion of the candidate material quantity for the current business line conforms to the user preference probability distribution value of the current business line;

[0030] For each business line, select different types of materials that match the user's material ratio for the corresponding business line from the candidate materials.

[0031] Preferably, before selecting the different types of materials that match the user's different material ratios in the corresponding business line, the method further includes:

[0032] The user's behavior information is determined and updated. The updated user behavior information is then input into the material ratio prediction model to calculate the ratio of different types of materials for each business line, thereby obtaining the updated ratio of different types of materials for each business line.

[0033] The selection of different types of materials that meet the user's different material ratios in the corresponding business line is based on the user's updated material ratio values ​​for each business line.

[0034] Preferably, the user information further includes at least one of the following:

[0035] User characteristics; and

[0036] The intersection of user characteristics and user behavior information.

[0037] An electronic device, comprising:

[0038] processor;

[0039] A memory storing a program configured to implement, when executed by the processor, a method for pushing materials as described in any of the preceding claims.

[0040] A non-transient storage medium storing a program configured to, when executed by a processor, cause the processor to perform the steps of the method for pushing material as described in any of the preceding claims.

[0041] A computer program product, comprising a computer program or instructions, characterized in that, when executed by a processor, the computer program or instructions implement the steps of the method for pushing materials as described in any of the preceding claims.

[0042] As seen above, in this embodiment of the application, a business line preference prediction model and a corresponding business line material ratio prediction model are pre-trained. When pushing materials to a user, the user's information is input into the business line preference prediction model, and the user preference probability distribution is calculated for each business line to determine the user preference probability distribution value for each business line. The user's information includes the user's user behavior information. The user's user behavior information is input into the corresponding business line material ratio prediction model to calculate the ratio of different types of materials for each business line, thus obtaining the user's material ratio for each business line. From the candidate material set, materials are selected and pushed to the user based on the determined user preference probability distribution value for each business line and the user's material ratio value for each business line. Because this application adjusts the proportion of different types of materials for user preferences in each business line based on the behavioral information of different users, and the prediction model for the proportion of different types of materials in each business line considers the differences in the quantity of different types of materials during training and application to improve the prediction accuracy, it can accurately predict and push materials that users are interested in, thereby improving the conversion rate and retention rate of user behavior on the network platform. Attached Figure Description

[0043] Figure 1 A schematic diagram illustrating the process framework for determining the materials to be pushed in this application;

[0044] Figure 2 A flowchart illustrating a method for pushing materials, as provided in an embodiment of this application;

[0045] Figure 3 A schematic diagram illustrating the process framework for determining the delivery of materials on a network platform, provided in an embodiment of this application;

[0046] Figure 4 A flowchart illustrating a method for determining the probability distribution value of user preferences for a business line using a business line preference prediction model, as provided in an embodiment of this application.

[0047] Figure 5 This is a schematic diagram illustrating the process of inputting user behavior information into a constrained material proportion prediction model for a corresponding business line, and calculating the proportion of different types of materials based on the constraints, as provided in this application embodiment.

[0048] Figure 6A schematic diagram illustrating the process of updating the baseline proportion of different types of materials based on the constraints of the model, as provided in this embodiment of the application;

[0049] Figure 7 This application provides a schematic diagram of a system structure for pushing materials.

[0050] Figure 8 This is a schematic diagram of an electronic device provided for another embodiment of this application. Detailed Implementation

[0051] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0052] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0053] The technical solution of this application will be described in detail below with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0054] Currently, when pushing materials to users on online platforms, the following methods can be used: Figure 1 The process shown, Figure 1This is a schematic diagram illustrating the process framework for determining the materials to be pushed on the network platform according to an embodiment of this application. As shown in the figure, the network platform trains different types of material ratio prediction models for each single business line and a business line preference prediction model. When pushing materials to users, firstly, the different types of material ratio prediction models for each single business line are used to determine the different types of material ratios for each business line, and then different types of materials for each single business line are selected from the set of spare materials. Secondly, user information is input into the business line preference prediction model, and probability distribution calculations are performed for each business line. The probability values ​​of business lines with probability values ​​are used as the user preference probability distribution values ​​for that business line. Finally, different types of materials for each business line are selected from the different types of materials for each single business line, and based on the user preference probability distribution values ​​for that business line, different types of materials are selected as the materials to be pushed to the user. Among them, the prediction model for the different material ratios of each single business line uses the materials exposed on the webpage of the network platform as samples and is trained with the conversion rate of the different types of materials of each single business line as the target to obtain the neural network model of the different material ratios of each business line; the business line preference prediction model uses user information including user characteristics and user behavior as samples and is trained with the conversion rate of the behavior (click behavior or transaction behavior, etc.) of different business lines as the target to obtain the neural network model of the probability distribution of business line preferences.

[0055] use Figure 1 The method shown has limitations in determining the materials to be pushed to users. It cannot adjust the proportion of different types of materials in each individual business line according to the user information of different users. Furthermore, the prediction model for the material proportion of each individual business line does not take into account the difference in the number of samples of different types during training, resulting in low prediction accuracy. Consequently, the materials to be pushed may not be the materials that users are interested in, leading to a low conversion rate of user behavior on the online platform.

[0056] from Figure 1As can be seen from the described scheme, determining whether the materials pushed to users are of interest to them is closely related to the prediction accuracy of the two trained models. Specifically, 1) A prediction model for the proportion of different types of materials under each single business line. Typically, different types of materials under each business line are modeled with conversion rate (click / sales, etc.) as the objective. This model merges and sorts different types of materials within the same business line, selecting the appropriate materials. In this process, different types of materials are treated uniformly in the model, leading to an imbalance in the number of different types of materials: materials with less exposure on the webpage are not reflected in the model training process due to their small sample size, and the model training is influenced by the mainstream material types. Furthermore, since different types of materials are modeled uniformly using the same model, the trained model obtains a probability distribution of materials with high conversion rates, failing to reflect the probability of the proportion of each type of material and unable to control the quantity of different types of materials. 2) A business line preference prediction model. Typically, user information such as historical behavior and characteristics is used as samples. The model is trained based on whether users generate business line behaviors such as clicks or transactions across different business lines, obtaining the probability distribution of user behavior across each business line. The probability distribution of user preferences for each business line is then determined based on this distribution. However, this model suffers from a cold start problem when applying it to new users.

[0057] From the above descriptions of the two models obtained from the training, we can conclude that: 1) The sample types of different types of materials in the prediction models for different material ratios under each single business line are unbalanced during training: For different business lines, there are low-frequency and high-frequency material behaviors, and the quantity of materials also differs, resulting in differences in the number of different types of materials as samples. Materials with a small proportion in the same model cannot be fully trained, and for some low-conversion-rate materials, training with this model may reduce its prediction quantity, exacerbate the long-tail effect of subsequent predictions of this type of material, and reduce the diversity of the predicted material types of the model; 2) For different users, the quantity of different types of materials under the same business line cannot be controlled. The prediction models for different types of materials under each single business line are trained with conversion rate as the target, treating different types of materials equally. When the trained model is used for subsequent predictions, the prediction results cannot guarantee the minimum and maximum quantities of each type of material, and the ratio of each type of material cannot be adjusted as needed.

[0058] In summary, the background technology's solution, when determining the materials to be pushed to users, suffers from low prediction accuracy because it cannot adjust the proportion of different types of materials in each individual business line according to the user information of different users, and the prediction model for the material proportion of each individual business line does not take into account the differences in the quantity of different types of materials during training and prediction. As a result, the materials to be pushed may not be the materials that users are interested in, leading to low conversion rates and low retention rates of users on the online platform.

[0059] When pushing promotional materials to users, specifically the distribution of various types of materials, the main issue to be addressed is: what proportion of different types of materials from different business lines should be selected from the recalled candidate set (i.e., the set of alternative materials) to push to users in order to maximize user conversion and retention rates on the online platform? Therefore, the method provided in this application focuses on the following two points: 1) the proportion of materials from different business lines pushed to the current user; and 2) the proportion of different types of materials pushed to the user within the same business line. For example, if a user prefers secondhand housing materials, occasionally browses rental materials, and does not browse new housing materials, the materials determined by the solution provided in this application should primarily focus on the secondhand housing business line, followed by the rental business line, with secondhand housing listings being the primary type of material within the secondhand housing business line. The solution provided in this application accurately predicts the probability distribution of user preferences for business lines and the proportion of different types of materials within the corresponding business lines. This allows for the selection of different types of materials from the candidate set (i.e., the set of alternative materials) that match the probability distribution of user preferences for business lines and the proportion of different types of materials within the corresponding business lines, thereby improving user retention and conversion rates.

[0060] To achieve the aforementioned objectives, this application remodels the business line preference prediction model and the corresponding business line's material ratio prediction model. Specifically, the corresponding business line's material ratio prediction model is set as a constrained model, allowing adjustment of the material ratios of different types. This ensures that different material types receive a certain proportion of exposure during training, mitigating sample imbalance and avoiding low prediction accuracy in the trained model. It also guarantees the diversity of material types subsequently determined based on the trained model. To enable the constrained material ratio prediction model to adjust the material ratios based on real-time user behavior, this application first uses the business line preference prediction model to predict the probability distribution of user preferences for each business line. Then, based on user behavior information, the constrained material ratio prediction model predicts the user's material ratios across different business lines, and finally, a comprehensive selection of candidate materials is made. In this way, business needs can be met, and the current user interests can be perceived in real time to shift the push of materials. This allows for a more accurate inference of user interests and real-time adjustment of the proportion of different types of materials pushed to users, thereby improving user retention and conversion rates.

[0061] Figure 2 A flowchart of a method for pushing materials provided in this application embodiment includes the following specific steps:

[0062] Step 201: Input the user's information into the business line preference prediction model. For each business line, calculate the user preference probability distribution and determine the user preference probability distribution value for each business line. The user's information includes the user's user behavior information.

[0063] Step 202: Input the user's behavior information into the material ratio prediction model, calculate the ratio of different types of materials for each business line, and obtain the ratio value of different types of materials for each business line for the user.

[0064] Step 203: From the candidate material set, based on the determined user preference probability distribution value for each business line and the user's material ratio for different types in the corresponding business line, select materials and push them to the user.

[0065] Before this method is executed, the process includes: training a business line preference prediction model and training a prediction model for the proportion of different types of materials corresponding to the business line. Here, the prediction model for the proportion of different types of materials is actually a prediction model for the proportion of different types of materials with constraints. In this way, in subsequent steps, the business line preference prediction model and the prediction model for the proportion of different types of materials with constraints corresponding to the business line can be used.

[0066] In this method, the user's user behavior information is actually a type of information within the user's user information, which is used to describe user behavior.

[0067] The above method can adjust the proportion of different types of materials in each business line based on the behavioral information of different users. The prediction model for the proportion of different types of materials in each business line is improved by considering the difference in the quantity of different types of materials during training and subsequent application. This can accurately predict the materials that users are interested in and push them to them, thereby improving the conversion rate and retention rate of users on the network platform.

[0068] like Figure 3 As shown, Figure 3 This is a schematic diagram illustrating the process framework for determining the materials to be pushed on a network platform, as provided in the embodiments of this application. When pushing materials to users, the above scheme first uses a business line preference prediction model to predict the probability distribution value of user preferences for each business line. Then, based on the user's behavior information, it uses a constrained prediction model for the proportion of different types of materials in the corresponding business line to predict the proportion of different types of materials the user prefers in that business line. The advantage of this approach is that it eliminates the need to merge and sort different types of materials from each individual business line in the first step to select materials, thus preserving all the backup material sets. Subsequently, the proportion of different types of materials preferred by the user in each business line can be adjusted in real time based on business needs and real-time user behavior information. After adjustment, materials matching the user's preferred proportion of different types of materials in the business line are selected from the backup material set, ultimately determining the list of materials displayed to the user.

[0069] In the above method, the business line preference prediction model and the prediction model for different types of material proportions corresponding to the business line are models obtained through training.

[0070] In the above method, the business line preference prediction model consists of three business line preference prediction sub-models with different priorities: a personalized business line preference sub-model, a user profile-based business line preference sub-model, and a regional statistical business line preference sub-model. The personalized business line preference sub-model has the highest priority, the user profile-based sub-model has the second highest priority, and the regional statistical business line preference sub-model has the lowest priority. The personalized business line preference sub-model is trained using a neural network, specifically a tree model using the Extreme Gradient Boosting (XGBoost) algorithm. The user profile-based business line preference sub-model is trained using a neural network, specifically a tree model using the XGBoost algorithm. The regional statistical business line preference sub-model calculates the probability distribution value for each business line under each region based on the proportion of a specified material type within each business line.

[0071] 1) Personalized Business Line Preference Sub-model. This sub-model targets the conversion probability of user behavior across different business lines, such as the probability distribution of clicks on actual property listings. It uses user information, including user characteristics, user behavior, and the cross-features between these characteristics and behavior. The sub-model is then trained to determine the conversion probability of user behavior across different business lines, and this probability is used as the user preference probability distribution value for the corresponding business line. Because this sub-model relies on historical user information, it can present a cold start problem for new users. Therefore, a regional statistical business line preference sub-model is added to the business line preference prediction model, using the probability distribution of each business line within the user's region as the user preference probability distribution value for the corresponding business line.

[0072] 2) Business Line Preference Sub-model Based on User Profile. a) For users with a single business line, directly use that business line as the user's preferred business line; b) For users with multiple business lines, determine the depth of user behavior conversion, and train the sub-model using the conversion probability of behaviors at adjacent depths to behaviors in each business line as the objective. When applied, this sub-model predicts the probability distribution of user conversion from the current behavior to the next adjacent behavior in each business line, and uses this probability distribution as the user preference probability distribution value for that business line. For example, from click behavior to opportunity behavior, this sub-model predicts the probability distribution of opportunity behavior that a user who has already clicked will experience in each business line, and uses this probability distribution as the user preference probability distribution value for that business line.

[0073] 3) Regional Statistical Business Line Preference Sub-model. Regions can be set as city regions. The percentage distribution of material types across different business lines within each region is used as the preference probability distribution value for each business line in that region. When a user belongs to that region, the regional statistical business line preference sub-model is used to calculate the preference probability distribution value for that region's business lines. This preference probability distribution value is then used as the user preference probability distribution value for that business line. For example, the percentage of property listing visits across different business lines in each city region can be used as the preference probability distribution value for that city region's business lines.

[0074] Figure 4 The flowchart of the method for determining the probability distribution value of user preferences for a business line using a business line preference prediction model provided in this application embodiment includes the following specific steps:

[0075] Step 401: Train to obtain the personalized business line preference sub-model, the user profile-based business line preference sub-model, and the regional statistical business line preference sub-model;

[0076] In this step, the personalized business line preference sub-model is provided to step 402 for execution, the user profile-based business line preference sub-model is provided to step 405 for execution, and the regional statistics business line preference sub-model is provided to step 407 for execution.

[0077] Step 402: Determine whether the user information required for the personalized business line preference sub-model can be obtained. Calculate the user preference probability distribution value for each business line using the personalized business line preference sub-model. If yes, proceed to step 403; otherwise, proceed to step 404.

[0078] Step 403: Obtain user information and input it into the personalized business line preference sub-model. Use the obtained user behavior conversion probability in different business lines as the user preference probability distribution value of the corresponding business line.

[0079] Step 404: Obtain the user profile information of the user, input the user profile information into the business line preference sub-model based on the user profile, and then execute step 405.

[0080] Here, user profile information is a part of user information, used to describe user roles; when it is not possible to calculate the probability distribution value of user preferences for each business line using the personalized business line preference sub-model, it is confirmed whether user profile information in the user information can be obtained, and the business line preference sub-model of user profile is used for processing.

[0081] Step 405: Determine whether the user preference probability distribution value for each business line can be calculated using a business line preference sub-model based on user profiles. If yes, proceed to step 406; otherwise, proceed to step 407.

[0082] In this step, if user profile information can be obtained, then it can be determined that the user preference probability distribution value of each business line can be calculated using the business line preference sub-model based on user profile.

[0083] Step 406: The business line preference sub-model based on user profiles uses the probability distribution of the conversion behavior of users from the current behavior to the next adjacent behavior of each business line as the user preference probability distribution value of the corresponding business line.

[0084] Step 407: Use the regional statistical business line preference sub-model to calculate and obtain the preference probability distribution value of the business line in the region. Use the distribution of the proportion of set type materials of each business line in the user's region as the user preference probability distribution value of the business line in each region.

[0085] This step is implemented when it is not possible to calculate the probability distribution of user preferences for each business line using a business line preference sub-model based on user profiles.

[0086] from Figure 4 The process shows that the priority levels of the personalized business line preference sub-model, the user profile-based business line preference sub-model, and the regional statistical business line preference sub-model are from high to low. When determining the user preference probability distribution value for each business line, the process proceeds according to priority. If the user is an existing user and their user information is complete, the personalized business line preference information required by the personalized business line preference sub-model can be obtained from the user information. This personalized business line preference sub-model then processes the information to determine the user preference probability distribution value for each business line. If the user is an existing user and their user information is incomplete, only containing user profile information, the personalized business line preference sub-model cannot process it. Therefore, the user profile-based business line preference sub-model is used according to priority to process the user profile information and determine the user preference probability distribution value for each business line. If the user is a new user and their user information is empty, neither the personalized business line preference sub-model nor the user profile-based business line preference model can be used. Therefore, the regional statistical business line preference sub-model is used according to priority to determine the user preference probability distribution value for each business line.

[0087] exist Figure 2In step 201, the business line preference prediction model obtained through training includes: a personalized business line preference sub-model, a business line preference sub-model based on user profiles, and a regional statistical business line preference sub-model.

[0088] exist Figure 2 In step 202, the user information is input into the business line preference prediction model. For each business line, the probability distribution of user preferences is calculated separately, and the probability distribution value of user preferences for each business line is determined, including:

[0089] If the user preference probability distribution value of each business line can be calculated using the personalized business line preference sub-model, the user information of the user is input into the personalized business line preference sub-model, and the calculated personalized business line preference probability distribution value is used as the corresponding business line user preference probability distribution value.

[0090] In cases where it is impossible to calculate the user preference probability distribution value for each business line using a personalized business line preference sub-model, and where the user information also includes profile information, a user profile-based business line preference sub-model is used. The profile information from the user information is input into the user profile-based business line preference sub-model, and the calculated probability distribution value of the user's transition behavior from the current behavior to the next adjacent behavior within the corresponding business line is used as the user preference probability distribution value for that business line.

[0091] When it is not possible to calculate the probability distribution value of user preferences for each business line using a business line preference sub-model based on user profiles, a regional statistical business preference sub-model is used to calculate the distribution of the proportion of materials of a set type under each business line in the region to which the user belongs, which is then used as the probability distribution value of user preferences for each business line in each region.

[0092] Here, the business line preference prediction model is trained using a neural network. The personalized business line preference sub-model is trained using a tree model based on the extreme gradient boosting (XGBoost) algorithm, and the business line preference sub-model based on user profiles is trained using a tree model based on the XGBoost algorithm.

[0093] In this method, inputting the user's information into the business line preference prediction model includes:

[0094] The user information, including user characteristics, user behavior information, and / or the cross-features of user characteristics and user behavior information, is input into the business line preference prediction model.

[0095] The user information includes user behavior information that is identical to the user behavior information of the user in the input material proportioning prediction model. The user behavior information describes the user's operational behavior across various business lines.

[0096] exist Figure 2 In step 203, pre-trained material proportion prediction models with constraints for each business line are applied. Specifically, Figure 5 This illustration shows the process of inputting user behavior information into a constrained prediction model for the proportion of different types of materials in a corresponding business line, as provided in this application embodiment, and calculating the proportion of different types of materials based on the constraints. As shown, the model calculates the proportion of different types of materials for the user under their preferred business line based on the user's historical behavior data, and uses this as the baseline proportion of different types of materials. The prediction model can adjust the baseline proportion of different types of materials according to the set constraints to satisfy them, thus obtaining the proportion values ​​of different types of materials. Specifically, after obtaining the user's real-time behavior information, the model can adjust the baseline proportion of different types of materials in real time, and then dynamically adjust the proportion values ​​of different types of materials according to the constraints, iterating repeatedly to update the proportion values ​​of different types of materials for the user under their preferred business line in real time.

[0097] In other words, Figure 2 In step 203 of the method, the calculation of the proportioning values ​​of different types of materials based on constraints includes two steps:

[0098] Based on the corresponding business line, the material ratio prediction model with constraints calculates the baseline ratio values ​​of different types of materials.

[0099] Based on the constraints, the calculated baseline proportions of the different types of materials are adjusted to meet the constraints.

[0100] Here, because the benchmark proportions of different types of materials output by the model depend on the user's historical behavior, there is a cold start problem for new users. Therefore, this embodiment of the application uses the average normalization method of all user behavior information under the user's preferred business line to calculate the benchmark proportions of different types of materials as the benchmark proportions of different types of materials for new users.

[0101] When calculating the baseline proportions for different types of materials, the formula for new users is as follows:

[0102]

[0103] The time period can be set to 30 days. In each time period, the amount of behavior for each type of material can be the number of clicks for each type of material. After calculating the proportion of each type of material, the proportion of that type of material is averaged and normalized.

[0104] For existing users, the formula is:

[0105]

[0106] The time period can be set to 3 days, and the amount of action for each type of material can be the number of clicks for each type of material.

[0107] In other words, the baseline proportion values ​​for different types of materials calculated by the material proportion prediction model under constraints include:

[0108] When a user is identified as a new user, the ratio between the amount of user behavior on a certain type of material and the number of times the material is exposed within a set time period is averaged and normalized within that set time period and then used as the benchmark ratio value for that type of material.

[0109] When a user is determined to be a non-new user, the ratio between the user's activity level with a certain type of material and the number of times the material is exposed within a set time period will be used as the baseline ratio for that type of material.

[0110] The update strategy for the baseline proportions of different types of materials based on the constraints of this model is as follows.

[0111] Constraints: Minimum and maximum constraints are set for each material type. The minimum constraint (min) represents the minimum proportion of the material, and the maximum constraint (max) represents the maximum proportion of the material; N represents the total number of material types; p i This is to update the previous material proportioning values. For each type of material, the minimum constraint value can be set differently or the same, and the maximum constraint value can be set differently or the same. The minimum and maximum constraint values ​​can be set as needed, without restriction.

[0112] Update strategy:

[0113] The first step is to update the proportions of all materials that are less than the minimum constraint value for their respective material types, based on the user's baseline proportion values ​​for different material types, to the minimum constraint value (p′) for that material type. i =min i {p i <min i},p i =p′ iFor the remaining material types that do not undergo minimum constraint updates, the baseline proportions are updated proportionally, using the following formula:

[0114]

[0115] The second step is to determine whether the proportion value of the material with the non-constrained minimum value exceeds the constraint maximum value for the corresponding material type. If so, the proportion value of that material is updated to the constraint maximum value for the corresponding material type. For the remaining materials that do not require constraint maximum value updates, the maximum proportion difference value is allocated. The difference value is calculated as: (Proportion value before update of the material with the maximum updated proportion value) - (Maximum constraint maximum value set). The formula used is:

[0116]

[0117] The third step is to determine whether the updated proportions of all materials meet the constraints. If they do, the updated proportions of all materials are returned. If they do not meet the constraints, the second step is repeated iteratively.

[0118] The adjustments to meet the constraints for different types of material proportions include:

[0119] 1) For each type of material, determine whether the baseline proportion value of the material type is less than the set minimum constraint value of the material type. If so, use the minimum constraint value of the material type as the proportion value of the material type.

[0120] 2) If not, multiply the baseline proportion value of the material of the type by the calculated proportion value to obtain the updated proportion value of the material of the type. The calculated proportion value is the ratio between the remaining value of the sum of the proportion values ​​of all other materials of the type for which the minimum constraint value of other types is set, and the sum of the baseline proportion values ​​of all other materials of the type for which the minimum constraint value of other types is not set.

[0121] 3) Determine whether the updated proportion value of the material of the type is greater than the set maximum constraint value of the material of the type. If so, use the maximum constraint value of the material of the type as the proportion value of the material of the type.

[0122] 4) If not, the proportion value of the material of the aforementioned type is averaged with the proportion value of other types of materials that have not been set with the maximum constraint value corresponding to other types, and the maximum proportion difference value is obtained by distributing the maximum proportion difference value equally among the proportion values ​​of the material of the aforementioned type. The maximum proportion difference value is the difference between the proportion value of other types of materials that have been set with the maximum constraint value corresponding to other types and the maximum constraint value corresponding to other types of materials before the update.

[0123] 5) Determine whether the proportion of all types of materials under the business line meets the constraint condition, which is that the proportion of all types of materials is between the set minimum and maximum constraint values. If not, return to step 4) to continue execution; if yes, end this process.

[0124] In other words, the proportions of different types of materials under each business line are constrained to between the minimum and maximum constraints of their respective types, and are determined in relation to the proportions of other types of materials. This ensures that the proportions of materials that do not have the minimum or maximum constraints of their respective types share the maximum proportion difference on an average basis. This allows for appropriate adjustments to the proportions of each type of material under each business line.

[0125] A specific example will be used for illustration.

[0126] As shown in Table 1, Table 1 lists the constraints for different material types in this model.

[0127] Material type Minimum constraint Constraint maximum value live streaming 0.12 0.15 Salesperson 0.12 0.15 Rankings 0.1 0.15 Housing 0.2 0.7 Demand Card 0.12 0.15

[0128] Table 1

[0129] By inputting user behavior information into the model, the baseline proportions of different types of materials are obtained, as shown in Table 2.

[0130] Material type The baseline proportion of materials live streaming 0.1 Salesperson 0.1 Rankings 0.3 Housing 0.9 Demand Card 0.1

[0131] Table 2

[0132] For the baseline allocation value of live streaming, the minimum constraint value is updated, adjusting the material allocation value for live streaming to 0.12. Similarly, for the baseline allocation value of sales guide type, the minimum constraint value is updated, adjusting the material allocation value to 0.12. For the baseline allocation value of ranking list type, since it exceeds the minimum constraint value, an unconstrained minimum value adjustment is performed, calculated proportionally as 0.3 / 0.3 + 0.9*(1 - 0.12*3) = 0.16. Since the obtained 0.16 exceeds the maximum constraint value, it is updated to the maximum constraint value of 0.15, thus adjusting the baseline allocation value for ranking list type. The base allocation value for the housing type is adjusted to 0.16. Since it exceeds the minimum constraint value, an adjustment to the unconstrained minimum value is made proportionally: 0.9 / 0.3 + 0.9*(1-0.12*3) = 0.48. The base allocation value for the housing type is adjusted to 0.48. Since 0.48 is less than the maximum constraint value, an adjustment is made to 1-0.12*3-0.15 = 0.49. The material allocation value for the housing type is adjusted to 0.49. For the base allocation value for the demand card type, the minimum constraint value is updated, and the material allocation value for the demand card type is adjusted to 0.12. The entire process is as follows: Figure 6 As shown, Figure 6 This is a schematic diagram illustrating the process of updating the baseline proportion of different types of materials based on the constraints of the model, as provided in the embodiments of this application.

[0133] exist Figure 2 In step 204, the process of selecting materials for push notifications based on the determined user preference probability distribution value for each business line and the proportion of different types of materials used by users in the corresponding business line includes:

[0134] From the set of candidate materials, select the business line to which the material belongs as the candidate material for the current business line, and the proportion of the candidate material quantity for the current business line conforms to the user preference probability distribution value of the current business line;

[0135] For each business line, select different types of materials that match the user's material ratio for the corresponding business line from the candidate materials.

[0136] exist Figure 2 The method further includes, before selecting different types of materials that match the user's material ratios for different business lines:

[0137] The user's behavior information is determined and updated. The updated user behavior information is then input into the material ratio prediction model to calculate the ratio of different types of materials for each business line, thereby obtaining the updated ratio of different types of materials for each business line.

[0138] The selection of different material types that match the user's material allocation ratio in the corresponding business line is based on the user's updated material allocation ratio values ​​for each business line. In other words, when the user's behavior information is updated, the material allocation ratio values ​​for each business line can be adjusted in real time, thereby adjusting the material allocation ratio pushed to the user accordingly.

[0139] This application also provides a system for pushing materials, such as Figure 7 As shown, the system includes: an application unit and a filtering unit, wherein,

[0140] The application unit is used to input user information into the business line preference prediction model, calculate the user preference probability distribution for each business line, and determine the user preference probability distribution value for each business line.

[0141] The user's behavior information is input into the material ratio prediction model to calculate the ratio of different types of materials for each business line, and the ratio of different types of materials for each user in each business line is obtained.

[0142] The filtering unit is used to filter materials from the candidate material set according to the determined user preference probability distribution value of each business line and the user's material ratio of different types in the corresponding business line, and push them to the user.

[0143] In another embodiment of the present invention, an electronic device is also provided, including: a processor; and a memory storing a program configured to implement the above-described method for pushing materials when executed by the processor.

[0144] In another embodiment of this application, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing instructions, which, when executed by a processor, cause the processor to perform a method for pushing materials as described in the foregoing embodiments. Figure 8 This is a schematic diagram of an electronic device provided for another embodiment of this application. (See diagram below.) Figure 8 As shown, another embodiment of this application also provides an electronic device, which may include a processor 801, wherein the processor 801 is used to perform the steps of the above-described method for pushing materials. From Figure 8 It can also be seen that the electronic device provided in the above embodiments further includes a non-transitory computer-readable storage medium 802, on which a computer program is stored, and the computer program is executed by the processor 801 to perform the steps of the above-described method for pushing materials.

[0145] Specifically, the non-transitory computer-readable storage medium 802 can be a general-purpose storage medium, such as a mobile disk, hard disk, FLASH, read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), or portable compact disk read-only memory (CD-ROM), etc. When the computer program on the non-transitory computer-readable storage medium 802 is run by the processor 801, it can cause the processor 801 to execute the various steps of the above-described method for pushing materials.

[0146] In practical applications, the non-transitory computer-readable storage medium 802 may be included in the device / apparatus / system described in the above embodiments, or it may exist independently without being assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, can perform the various steps of the above-described method for pushing materials.

[0147] Another embodiment of this application provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method for pushing materials.

[0148] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments disclosed in this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings. For example, two blocks shown connectedly may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0149] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this disclosure can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, without departing from the spirit and teachings of this application, the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways, and all such combinations and / or combinations fall within the scope of this application.

[0150] This document uses specific embodiments to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and core ideas of this application, and are not intended to limit this application. For those skilled in the art, changes can be made to the specific implementation methods and application scope based on the ideas, spirit and principles of this application. Any modifications, equivalent substitutions, improvements, etc., made should be included within the scope of protection of this application.

Claims

1. A method for pushing materials, characterized in that, The method includes: The user's information is input into the business line preference prediction model. For each business line, the user preference probability distribution is calculated to determine the user preference probability distribution value for each business line. The user's information includes the user's user behavior information. The user's behavior information is input into the material ratio prediction model to calculate the ratio of different types of materials for each business line, and the ratio of different types of materials for each user in each business line is obtained. From the set of candidate materials, materials are selected and pushed to the user based on the determined probability distribution value of user preferences for each business line and the ratio of different types of materials for the user in the corresponding business line. The material proportioning prediction model is a material proportioning prediction model with constraints. The step of inputting the user's behavior information into the material proportioning prediction model to calculate the proportions of different types of materials for each business line includes: Based on the corresponding business line, the material ratio prediction model with constraints calculates the baseline ratio values ​​of different types of materials. Based on the constraints, the calculated baseline proportions of the different types of materials are adjusted to meet the constraints. The constrained material proportion prediction model calculates the baseline proportion values ​​for different types of materials, including: When a user is determined to be a new user, the ratio between the amount of user behavior on a certain type of material and the number of times the material is exposed within a set time period is averaged and normalized within the set time period and then used as the baseline ratio value for that type of material. When a user is determined to be a non-new user, the ratio between the user's behavior volume on a certain type of material and the number of times the material is exposed within the set time period is used as the baseline ratio value for that type of material. The adjustments to meet the constraints for different types of material proportions include: 1) For each type of material, determine whether the baseline proportion value of the material type is less than the set minimum constraint value of the material type. If so, use the minimum constraint value of the material type as the proportion value of the material type. 2) If not, multiply the baseline proportion value of the material of the type by the calculated proportion value to obtain the updated proportion value of the material of the type. The calculated proportion value is the ratio between the remaining value of the sum of the proportion values ​​of all other materials of the type for which the minimum constraint value of other types is set, and the sum of the baseline proportion values ​​of all other materials of the type for which the minimum constraint value of other types is not set. 3) Determine whether the updated proportion value of the material of the type is greater than the set maximum constraint value of the material of the type. If so, use the maximum constraint value of the material of the type as the proportion value of the material of the type. 4) If not, the proportion value of the material of the aforementioned type is averaged with the proportion value of other types of materials that have not been set with the maximum constraint value corresponding to other types, and the maximum proportion difference value is obtained by distributing the maximum proportion difference value among them. The maximum proportion difference value is the difference between the proportion value of other types of materials before the update and the maximum constraint value corresponding to other types of materials. 5) Determine whether the proportion of all types of materials under the business line meets the constraint condition. The constraint condition is that the proportion of all types of materials is between the set minimum and maximum constraint values. If not, return to step 4) to continue execution; if yes, end this process.

2. The method for pushing materials as described in claim 1, characterized in that, The business line preference prediction model includes: a personalized business line preference sub-model, a user profile-based business line preference sub-model, and a regional statistical business line preference model; The step of inputting user information into the business line preference prediction model, and calculating the probability distribution of user preferences for each business line to determine the probability distribution value of user preferences for each business line includes: If the user preference probability distribution value of each business line can be calculated using the personalized business line preference sub-model, the user information of the user is input into the personalized business line preference sub-model, and the calculated personalized business line preference probability distribution value is used as the corresponding business line user preference probability distribution value. In cases where it is impossible to calculate the user preference probability distribution value for each business line using a personalized business line preference sub-model, and where the user information also includes profile information, the profile information from the user information is input into a business line preference sub-model based on user profiles. The calculated probability distribution value of the user's transition behavior from the current behavior to the next adjacent behavior within the corresponding business line is used as the user preference probability distribution value for that business line. When it is not possible to calculate the probability distribution value of user preferences for each business line using a business line preference sub-model based on user profiles, a regional statistical business preference sub-model is used to calculate the distribution of the proportion of materials of a set type under each business line in the region to which the user belongs, which is then used as the probability distribution value of user preferences for each business line in each region.

3. The method for pushing materials as described in claim 2, characterized in that, The business line preference prediction model is trained using a tree model. Specifically, the personalized business line preference sub-model is trained using a tree model based on the extreme gradient boosting (XGBoost) algorithm, and the business line preference sub-model based on user profiles is trained using a tree model based on the XGBoost algorithm.

4. The method for pushing materials as described in claim 1, characterized in that, The step of selecting materials for push notifications based on the determined user preference probability distribution value for each business line and the ratio of different types of materials for each user in the corresponding business line includes: From the set of candidate materials, select the business line to which the material belongs as the candidate material for the current business line, and the proportion of the candidate material quantity for the current business line conforms to the user preference probability distribution value of the current business line; For each business line, select different types of materials that match the user's material ratio for the corresponding business line from the candidate materials.

5. The method for pushing materials as described in claim 4, characterized in that, Before selecting the different types of materials that match the user's material ratios for the corresponding business line, the process also includes: The user's behavior information is determined and updated. The updated user behavior information is then input into the material ratio prediction model to calculate the ratio of different types of materials for each business line, thereby obtaining the updated ratio of different types of materials for each business line. The selection of different types of materials that meet the user's different material ratios in the corresponding business line is based on the user's updated material ratio values ​​for each business line.

6. The method for pushing materials according to any one of claims 1 to 5, characterized in that, The user information also includes at least one of the following: User characteristics; and The intersection of user characteristics and user behavior information.

7. An electronic device, characterized in that, include: processor; A memory storing a program configured to implement, when executed by the processor, the method of pushing material as described in any one of claims 1 to 6.

8. A non-transient storage medium storing a program, characterized in that, The program is configured to cause the processor to perform the steps of the method for pushing materials as described in any one of claims 1 to 6 when executed by the processor.

9. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method for pushing materials as described in any one of claims 1 to 6.

Images