Training method and device of data pre-estimation model and information pushing method and device

Abstract

The embodiment of the invention discloses a data pre-estimation model training method and device and an information pushing method and device. A specific embodiment of the training method comprises the steps of generating model training data representing value attribute distribution of each preset-level category article interacted by a user according to interaction data of the user and the article in a specified time period; based on the interaction data of the user, the number of tasks in a data estimation model and the model output dimension are set, and the data estimation model adopts a multi-task learning model structure; and inputting the sample data into the data prediction model, outputting an interaction value attribute prediction value of the user in each preset-level category, and adjusting model parameters according to a comparison result of the interaction value attribute prediction value and the sample label data so as to continue training. The implementation mode is related to an e-commerce platform optimization technology, the price preference of the user can be comprehensively described, and the accuracy and diversity of a model prediction result are improved.

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the technical field of e-commerce platform optimization, and specifically to a method and device for training a data prediction model, and a method and device for information push. Background Art

[0002] With the in-depth research and application of big data technology, the focus of e-commerce platform merchants has increasingly focused on big data precision marketing. How to match suitable products to users in need, so as to maximize the marketing effect in delivery and reach, is a problem that every merchant is very concerned about. A very important dimension to consider in precision marketing is price. Therefore, mining the price band preferences shown by users in different categories is a very important basis for precision marketing of each category.

[0003] The above information disclosed in this background art section is only used to enhance the understanding of the background of the inventive concept, and thus, it may include information that does not form the prior art known to those of ordinary skill in the art in this country. Summary of the Invention

[0004] The content part of the present disclosure is used to briefly introduce the concepts, which will be described in detail in the following detailed implementation part. The content part of the present disclosure is not intended to identify the key features or essential features of the claimed technical solution, nor is it intended to be used to limit the scope of the claimed technical solution.

[0005] Some embodiments of the present disclosure propose a method for training a data prediction model, a training device, a method for information push, an information push device, an electronic device, a computer-readable medium, and a computer program product to solve one or more of the technical problems mentioned in the above background art section.

[0006] In a first aspect, some embodiments of the present disclosure provide a method for training a data prediction model, including: generating model training data representing the value attribute distribution of each preset-level category item interacted by a user according to the interaction data between the user and the item within a specified time period, where the model training data includes sample data and corresponding sample label data; setting the number of tasks and the model output dimension in the data prediction model based on the user's interaction data, where the data prediction model adopts a multi-task learning model structure; inputting the sample data into the data prediction model, outputting the predicted values of the interaction value attributes of the user in each preset-level category, and adjusting the model parameters for continuous training according to the comparison result between the predicted values of the interaction value attributes and the sample label data.

[0007] In some embodiments, model training data representing the value attribute distribution of each preset-level category of items interacted by a user is generated based on the interaction data between the user and the items within a specified time period, including: for each type of item of a preset-level category interacted by the user within the specified time period, determining the range of unit value attributes after value reduction of the items under this preset-level category, dividing the range of unit value attributes into a first number of value attribute bands, and performing data statistics on the value attribute bands to which the items under this preset-level category belong; generating model training data based on the statistical data of the items under various preset-level categories interacted by the user.

[0008] In some embodiments, model training data is generated based on the statistical data of the items under various preset-level categories interacted by the user, including: for the order data of the user, taking at least one of the order quantity, distribution frequency, mean value, and mode of the order items in the value attribute bands of each preset-level category of items as the order sample data, and taking the average value of the value attribute bands to which the order items of the user under each preset-level category belong as the order sample label data; for the browsing data of the user, taking at least one of the number of times, distribution frequency, mean value, and mode of browsing the value attribute bands of each preset-level category of items as the browsing sample data, and taking the average value of the value attribute bands to which the items browsed by the user under each preset-level category belong as the browsing sample label data.

[0009] In some embodiments, based on the interaction data of the user, the number of tasks and the model output dimensions in the data prediction model are set, including: setting one task in the data prediction model to learn the score of the order value attribute band of the user based on the order sample data; and setting another task in the data prediction model to learn the score of the browsing value attribute band of the user based on the browsing sample data; setting the data prediction model to output the predicted scores of the order value attribute band and the browsing value attribute band of each preset-level category of items for the same user.

[0010] In some embodiments, generating model training data based on the statistical data of items under various preset category levels of user interaction further includes: dividing the first number of value attribute bands of each preset category level item into a second number of value attribute band intervals, where the second number is less than the first number; for the user's order data, taking at least one of the order quantity, distribution frequency, mean, and mode of the order items in the value attribute band of each preset category level item as the order sample data, and taking the proportion of the order items of the user under each preset category level belonging to each value attribute band interval as the order sample label data; for the user's browsing data, taking at least one of the number of times, distribution frequency, mean, and mode of browsing the value attribute band of each preset category level item as the browsing sample data, and taking the proportion of the items browsed by the user under each preset category level belonging to each value attribute band interval as the browsing sample label data.

[0011] In some embodiments, setting the number of tasks and the model output dimension in the data prediction model based on the user's interaction data further includes: setting a second number of tasks in the data prediction model to learn the probability that the user belongs to each order value attribute band interval according to the order sample data; and setting another second number of tasks in the data prediction model to learn the probability that the user belongs to each browsing value attribute band interval according to the browsing sample data; setting the data prediction model to output the predicted distribution of the same user in the order value attribute band interval and the predicted distribution in the browsing value attribute band interval for each preset category level item.

[0012] In some embodiments, inputting the sample data into the data prediction model includes: inputting the sample data belonging to the same preset category level into the data prediction model corresponding to that preset category level to output the predicted value of the value attribute of the user interacting with the items of that preset category level; or inputting the sample data and category identifiers belonging to various preset category levels into the same data prediction model to output the predicted value of the value attribute of the user interacting with the items of each preset category level, where the category identifiers of various preset category levels are set in one-hot encoding mode.

[0013] In a second aspect, some embodiments of the present disclosure provide a training apparatus for a data prediction model, including: a training data generation unit configured to generate model training data representing the value attribute distribution of each preset-level category of items interacted by a user based on the interaction data between the user and the items within a specified time period, where the model training data includes sample data and corresponding sample label data; a model setting unit configured to set the number of tasks and the model output dimension in the data prediction model based on the interaction data of the user, where the data prediction model adopts a multi-task learning model structure; and a parameter adjustment unit configured to input the sample data into the data prediction model, output the predicted values of the interaction value attributes of the user for each preset-level category, and adjust the model parameters for continuous training according to the comparison result between the predicted values of the interaction value attributes and the sample label data.

[0014] In some embodiments, the training data generation unit includes: a value division sub-unit configured to, for each preset-level category of items interacted by the user within a specified time period, determine the range of unit value attributes after value reduction of the items under the preset-level category, divide the range of unit value attributes into a first number of value attribute bands, and perform data statistics on the value attribute bands to which the items under the preset-level category belong; and a generation sub-unit configured to generate model training data according to the statistical data of the items under various preset-level categories interacted by the user.

[0015] In some embodiments, the generation sub-unit is further configured to, for the order data of the user, use at least one of the order quantity, distribution frequency, mean value, and mode of the order items in the value attribute bands of each preset-level category of items as the order sample data, and use the average value of the value attribute bands to which the order items of the user under each preset-level category belong as the order sample label data; for the browsing data of the user, use at least one of the number of times, distribution frequency, mean value, and mode of browsing the value attribute bands of each preset-level category of items as the browsing sample data, and use the average value of the value attribute bands to which the items browsed by the user under each preset-level category belong as the browsing sample label data.

[0016] In some embodiments, the model setting unit is further configured to set one task in the data prediction model to learn the score of the order value attribute band of the user according to the order sample data; and set another task in the data prediction model to learn the score of the browsing value attribute band of the user according to the browsing sample data; and set the data prediction model to output the predicted scores of the order value attribute band and the browsing value attribute band of each preset-level category of items for the same user.

[0017] In some embodiments, the generation subunit is further configured to further divide the first number of value-attribute bands of each preset-level category item into a second number of value-attribute band intervals, where the second number is less than the first number; for the order data of a user, use at least one of the order quantity, distribution frequency, mean, and mode of the order items in the value-attribute bands of each preset-level category item as order sample data, and use the proportion of the order items of the user under each preset-level category belonging to each value-attribute band interval as order sample label data; for the browsing data of the user, use at least one of the number of times, distribution frequency, mean, and mode of browsing the value-attribute bands of each preset-level category item as browsing sample data, and use the proportion of the items browsed by the user under each preset-level category belonging to each value-attribute band interval as browsing sample label data.

[0018] In some embodiments, the model setting unit is further configured to set a second number of tasks in the data prediction model to learn the probability that the user belongs to each order value-attribute band interval according to the order sample data; and set another second number of tasks in the data prediction model to learn the probability that the user belongs to each browsing value-attribute band interval according to the browsing sample data; set the data prediction model to output the predicted distributions of the same user in the order value-attribute band interval and in the browsing value-attribute band interval for each preset-level category item.

[0019] In some embodiments, the parameter adjustment unit is further configured to input the sample data belonging to the same preset-level category into the data prediction model corresponding to this preset-level category to output the predicted value of the value attribute of the user interacting with the items of this preset-level category; or input the sample data and category identifiers belonging to various preset-level categories into the same data prediction model to output the predicted value of the value attribute of the user interacting with the items of each preset-level category, where the category identifiers of various preset-level categories are set in a one-hot encoding manner.

[0020] In a third aspect, some embodiments of the present disclosure provide an information push method, including: inputting the value-attribute distribution data of each preset-level category item interacted by a target user within a specified time period and the attribute information of the user into a data prediction model to output the predicted value of the interaction value attribute of the target user for each preset-level category, where the data prediction model is obtained by using the training method of the data prediction model described in any implementation manner of the first aspect above; performing aggregation analysis on the output predicted values of the interaction value attributes of each preset-level category to determine the predicted value of the interaction value attribute of the target user for each parent category, where the parent category is the upper-level category to which the preset-level category belongs; determining the target items that match the target user according to the predicted values of the interaction value attributes of the target user at all levels and in all categories, and pushing the information of the target items to the target user.

[0021] In some embodiments, aggregate analysis is performed on the predicted values of the interaction value attributes of each preset-level category in the output to determine the predicted values of the interaction value attributes of the target user in each parent category, including: for multiple preset-level categories belonging to the same parent category, determining the product of the predicted value of the interaction value attribute of the target user on each preset-level category and the number of items interacted by the target user in this preset-level category, and determining the total product sum of the multiple preset-level categories; determining the total number of items interacted by the target user in the multiple preset-level categories; and determining the ratio of the product sum to the total number of items as the predicted value of the interaction value attribute of the target user in the parent category to which the multiple preset-level categories belong.

[0022] In some embodiments, the method further includes: in response to determining that the predicted value of the interaction value attribute of the target user in the target category is empty, determining the target parent category to which the target category belongs; and determining the predicted value of the interaction value attribute of the target user in the target parent category as the predicted value of the interaction value attribute of the target user in the target category.

[0023] In some embodiments, the method further includes: in response to determining that the predicted value of the interaction value attribute of the target user in the target category is empty, determining the correlation coefficient of the target category and the target parent category to which it belongs in the predicted value of the interaction value attribute of the user as a benchmark threshold; determining the correlation coefficient of the target category and its peer categories in the predicted value of the interaction value attribute of the user; and determining the predicted value of the interaction value attribute of the target user in the target category based on the predicted values of the interaction value attributes of each user in the candidate categories, where the candidate categories are peer categories whose correlation coefficient with the target category is greater than the benchmark threshold.

[0024] In a fourth aspect, some embodiments of the present disclosure provide an information push device, including: a prediction unit configured to input the value attribute distribution data of each preset-level category item interacted by the target user within a specified time period and the attribute information of the user into a data estimation model, and output the predicted value of the interaction value attribute of the target user in each preset-level category, where the data estimation model is obtained by using the training method of the data estimation model described in any implementation manner of the first aspect above; an aggregation unit configured to perform aggregate analysis on the predicted values of the interaction value attributes of each preset-level category output, and determine the predicted value of the interaction value attribute of the target user in each parent category, where the parent category is the upper-level category to which the preset-level category belongs; and a push unit configured to determine a target item matching the target user according to the predicted values of the interaction value attributes of the target user in each level and each category, and push the information of the target item to the target user.

[0025] In some embodiments, the aggregation unit is further configured to, for a plurality of preset category levels belonging to the same parent category, determine the product of the predicted value of the interaction value attribute of the target user on each preset category level and the number of items with which the target user interacts on that preset category level, and determine the total product sum of the plurality of preset category levels; determine the total number of items with which the target user interacts on the plurality of preset category levels; and determine the ratio of the product sum to the total number of items as the predicted value of the interaction value attribute of the target user on the parent category to which the plurality of preset category levels belong.

[0026] In some embodiments, the information push device further includes a category generalization unit, which is configured to, in response to determining that the predicted value of the interaction value attribute of the target user on the target category is empty, determine the target parent category to which the target category belongs; and determine the predicted value of the interaction value attribute of the target user on the target parent category as the predicted value of the interaction value attribute of the target user on the target category.

[0027] In some embodiments, the category generalization unit is further configured to, in response to determining that the predicted value of the interaction value attribute of the target user on the target category is empty, determine the correlation coefficient between the target category and the target parent category to which it belongs in terms of the predicted value of the interaction value attribute of the user as a reference threshold; determine the correlation coefficient between the target category and the same-level categories in terms of the predicted value of the interaction value attribute of the user; and determine the predicted value of the interaction value attribute of the target user on the target category based on the predicted values of the interaction value attributes of each user on the candidate categories, where the candidate categories are the same-level categories whose correlation coefficient with the target category is greater than the reference threshold.

[0028] In a fifth aspect, some embodiments of the present disclosure provide an electronic device, including: one or more processors; a storage device having stored thereon one or more programs, which when executed by the one or more processors cause the one or more processors to implement the method described in any implementation manner of the first aspect or the third aspect above.

[0029] In a sixth aspect, some embodiments of the present disclosure provide a computer-readable medium having stored thereon a computer program, where the computer program, when executed by a processor, implements the method described in any implementation manner of the first aspect or the third aspect above.

[0030] In a seventh aspect, some embodiments of the present disclosure provide a computer program product, including a computer program, which when executed by a processor, implements the method described in any implementation manner of the first aspect or the third aspect above.

[0031] The above - mentioned various embodiments of the present disclosure have the following beneficial effects: The training method of the data prediction model in some embodiments of the present disclosure can obtain a model that can more comprehensively analyze the interaction value attributes of users. Specifically, in the existing related methods and technologies of user price - band preferences, the method of analyzing price - band preferences at the user granularity generally has difficulty in characterizing the preference differences of users in each category. Generally speaking, users may have different price - band preferences in different categories, and it is not suitable to generalize all categories. This will affect the matching degree between different category items and users. For example, in the case of electronic products, users prefer high prices. While in the case of snacks, users may prefer low prices. If the products of each category to be pushed to this user are determined accordingly, it will surely affect the probability of interaction between the user and the pushed products.

[0032] Based on this, the training method of the data prediction model in some embodiments of the present disclosure, through the multi - task learning algorithm, can mine the historical behaviors of users in multiple aspects, learn their price - band preferences in each preset - level category, and thus can predict the interaction value attribute situation of users in each preset - level category. In this way, it can comprehensively characterize the price preferences of users, improve the accuracy and diversity of the model prediction results. By pushing products that meet the expected price levels of users in each category, the accuracy of product pushing can be improved. Furthermore, it can greatly increase the probability of interaction between the user and the pushed products, thereby increasing the page views and order quantities of the products. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Combined with the accompanying drawings and with reference to the following specific embodiments, the above - mentioned and other features, advantages and aspects of each embodiment of the present disclosure will become more obvious. Throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic, and the elements and elements are not necessarily drawn to scale.

[0034] Figure 1 is a flowchart of some embodiments of the training method of the data prediction model of the present disclosure;

[0035] Figure 2 is a schematic structural diagram of some embodiments of the data prediction model of the present disclosure;

[0036] Figure 3 is a schematic structural diagram of some embodiments of the training device of the data prediction model of the present disclosure;

[0037] Figure 4 is a flowchart of some embodiments of the information - pushing method of the present disclosure;

[0038] Figure 5 is a schematic structural diagram of some embodiments of the information - pushing device of the present disclosure;

[0039] Figure 6ASchematic diagrams of some embodiments of the category generalization method;

[0040] Figure 6B Schematic diagrams of application scenarios of some embodiments of the present disclosure;

[0041] Figure 7 Schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed implementation manners

[0042] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although some embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. On the contrary, these embodiments are provided to more thoroughly and completely understand the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not used to limit the protection scope of the present disclosure.

[0043] In addition, it should be noted that for the convenience of description, only parts related to the relevant invention are shown in the drawings. Without conflict, the embodiments in the present disclosure and the features in the embodiments can be combined with each other.

[0044] It should be noted that the concepts such as "first" and "second" mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence relationship of the functions performed by these devices, modules or units.

[0045] It should be noted that the modifications of "one" and "multiple" mentioned in the present disclosure are illustrative rather than restrictive. Those skilled in the art should understand that unless otherwise clearly specified in the context, it should be understood as "one or more".

[0046] For operations such as collection, storage, and use of user personal information (such as user attribute information and user historical behavior data) involved in the present disclosure, before performing the corresponding operations, relevant organizations or individuals shall fulfill obligations including conducting personal information security impact assessments, fulfilling the obligation of informing the personal information subject, and obtaining the prior authorization and consent of the personal information subject.

[0047] The present disclosure will be described in detail below with reference to the drawings and in combination with embodiments.

[0048] Figure 1 Flow 100 of some embodiments of the training method of the data prediction model according to the present disclosure is shown. The training method includes the following steps:

[0049] Step 101, generating model training data representing the value attribute distribution of each preset-level category item interacted by the user according to the interaction data between the user and the item within a specified time period.

[0050] In some embodiments, the execution subject of the training method of the data prediction model (such as a model training server) can communicate with other electronic devices through a wired connection method or a wireless connection method. For example, an application person can send the interaction data between a user and an item within a specified time period to the execution subject through a terminal device. For another example, when the execution subject receives a model training instruction sent by the application person, it can obtain the interaction data between the user and the item within the specified time period from the application database. The specified time period here can be set according to the actual situation, such as one week, one month, etc. To ensure sufficient data volume, the specified time period here can be set longer, such as half a year or one year, etc. The user interaction behavior indicated by the interaction data can be any behavior indicating that the user has the intention to obtain a certain item, and can include but are not limited to at least one of the following: placing an order to purchase, browsing, collecting, adding to the shopping cart, etc.

[0051] In some embodiments, the execution subject can generate model training data representing the value attribute distribution of each preset-level category item interacted by the user according to the interaction data between the user and the item within the specified time period. Among them, the model training data can include sample data and corresponding sample label data. The above-mentioned preset-level category can also be set according to the actual situation. Here, to ensure the sufficiency of the model training data and the accuracy of the model prediction result, the preset-level category is usually a relatively detailed category level, such as the third-level category in an e-commerce application platform.

[0052] As an example, the execution subject can generate model training data according to the unit value attribute (original unit price) of each preset-level category item interacted by the user within the specified time period. It can be understood that the price of an item often directly affects the user's interaction behavior. Therefore, to improve the accuracy of the model training result, the execution subject can generate model training data according to the unit value attribute after the item value is reduced, that is, the unit price after discount.

[0053] Specifically, for each preset-level category item interacted by the user within the specified time period, first, the range of the unit value attribute after the item value in this preset-level category is reduced can be determined. Then, this range of the unit value attribute can be divided into the first number of value attribute bands. After that, data statistics are performed on the value attribute bands to which the items in this preset-level category belong. Furthermore, model training data is generated according to the statistical data of the items in various preset-level categories interacted by the user.

[0054] As an example, the unit price of the item after discount of the product sku (Stock Keeping Unit) can be used as the price basis for dividing the price range. Since the amount after discount for different orders of the same product sku may vary, the average value within a certain time window can be used to calculate the unit price of the item after discount. Taking one week as an example, the unit price of the item after discount can be:

[0055] Unit price of the item after discount = sum(amount after discount in the recent week) / sum(sales quantity);

[0056] Within each range of each third - level category, sort the product skus according to the unit price of the item after discount. At this time, the 1% - 99% quantiles can be taken as thresholds to divide the prices of the third - level category products into 100 price ranges. That is, the 1% of the products with the lowest unit price belong to price range 1, the products with unit prices in the 1% - 2% quantile range belong to price range 2, and so on. The products with unit prices greater than the 99% quantile belong to price range 100. That is to say, for each product in each third - level category, the range formed by the maximum and minimum unit prices of the item after discount is divided into 100 sub - ranges. This division method can make the number of products in each price range evenly distributed. In this way, we have constructed the price range division for each third - level category respectively.

[0057] In some embodiments, in order to reduce the amount of data to be processed, for various interaction behaviors of users, the order - placing behavior and browsing behavior can be mainly considered. In this way, both the training efficiency of the model can be improved and the training effect of the model can be ensured. It should be noted that since the functions of the data prediction model are different, that is, the content of analysis and prediction is different, the model training data required are often also different.

[0058] In some embodiments, the data prediction model can be used for predictive analysis of the interaction value attribute band scores of users under each preset - level category, that is, the price - band preference score prediction model. At this time, for the order - placing data of users, at least one of the order quantity, distribution frequency, mean, and mode of the value attribute band of the order items in each preset - level category item can be used as the order sample data. And the average value of the value attribute band to which the order items of the user belong under each preset - level category can be used as the order sample label data. For the browsing data of users, at least one of the number of times, distribution frequency, mean, and mode of browsing the value attribute band of each preset - level category item can be used as the browsing sample data. And the average value of the value attribute band to which the items browsed by the user belong under each preset - level category can be used as the browsing sample label data.

[0059] As an example, such as Figure 2The schematic diagram of the model architecture shown mainly uses the user's interaction data for feature processing at the feature layer. That is, the commodity SKU price is mapped to the price band. For example: the order volume, distribution frequency, mean, and mode of the user's orders in each category of commodity price bands in the past year / half year / one month. And the number of times the user browses each category of commodity price bands in the past 30 days / 7 days / 3 days, distribution frequency, mean, and mode. During the processing, first, the order and the browsed commodity sku are matched to the corresponding price band number according to the unit price per piece. Then, statistics such as the quantity, frequency, mean (mean of price band numbers), and mode of each price band are calculated as features. These features can describe the user's interest in price bands shown in different past time periods and are also an important basis for predicting future price band preferences.

[0060] Step 102: Based on the user's interaction data, set the number of tasks and the model output dimension in the data prediction model.

[0061] In some embodiments, based on the user's interaction data, the execution entity can set the number of tasks in the data prediction model and the model output dimension. Among them, the data prediction model in the embodiments of the present disclosure can adopt a multi-task learning model structure. It can be understood that according to the different prediction functions of the model, the network structure in the model is usually different.

[0062] As an example, for Figure 2 the price band preference score prediction model shown on the left, at the model layer, the execution entity can set the number of tasks (Tasks) in the data prediction model to two. Among them, one task (Task1) can be used to learn the score of the user's order value attribute band according to the order sample data. Another task (Task2) can be used to learn the score of the user's browsing value attribute band according to the browsing sample data. Additionally, at the output layer, the output of the data prediction model can be set as: the predicted score of the order value attribute band of each preset category of items for the same user, and the predicted score of the browsing value attribute band.

[0063] In some embodiments, on this basis, the first number of value attribute bands of each preset category item can be further divided into the second number of value attribute band intervals. Among them, the second number is less than the first number. At this time, the execution entity can also set the model output layer to output: the predicted stratification of the order value attribute band of each preset category item for the same user, and the predicted stratification of the browsing value attribute band. For example, for the above 100 price bands, they can be divided into 5 price band intervals. That is to say, the execution entity can divide the preference score into 5 intervals to obtain the preference stratification. Among them, the order price band preference stratification = ceiling(order price band preference score / 20). The browsing price band preference stratification = ceiling(browsing price band preference score / 20). Here, ceiling means rounding up.

[0064] Step 103: Input the sample data into the data prediction model, output the predicted values of the interaction value attributes of the user in each preset category, and adjust the model parameters for continuous training according to the comparison results between the predicted values of the interaction value attributes and the sample label data.

[0065] In some embodiments, the execution entity can input the sample data obtained in step 101 into the data prediction model set in step 102. Thus, the predicted values of the interaction value attributes of the user in each preset category can be output through the model. And according to the comparison results between the predicted values of the interaction value attributes and the sample label data, the execution entity can determine the loss function, and then adjust the model parameters for continuous training.

[0066] As an example, for the price band preference score prediction model, users who have placed orders in the category in the most recent month are used as samples, and a multi-task learning model is adopted. A total of two tasks are set. The first task is used to learn the order price band preference score. The label of the sample is the average value of the price bands where the products ordered by the user in this category in the most recent month are located, which can be regarded as the actual order price band preference score. The second task is used to learn the browsing price band preference score. The label of the sample is the average value of the price bands where the products browsed by the user in this category in the most recent month are located, which can be regarded as the actual browsing price band preference score. The specific calculation method is: Denote (Y i , Z i ) as the label value of the i-th user. Denote as the price band number corresponding to the M i skus ordered by the i-th user in a certain category, then

[0067]

[0068] Denote as the price band number corresponding to the L i sub-skus browsed by the i-th user in a certain category, then

[0069]

[0070] Among them, the number of views can be calculated according to PV (PageView, the number of accesses).

[0071] It should be noted that the model can adopt an NN (neural network) neural network or other network models, or tree models such as random forest and xgboost. Among them, xgboost (Exterme Gradient Boosting) is usually Extreme Gradient Boosting, which is an ensemble machine learning algorithm based on decision trees. Taking the NN model as an example, as Figure 2 shown in the model layer of the architecture diagram, two tasks can share the underlying layer, and the output layer is 2-dimensional, corresponding to the labels of the two tasks respectively. The model is not limited to the examples in the figure, and attention (attention mechanism), convolutional layers, etc. can also be added, or personalized layers for each task can be constructed after sharing several underlying layers. Denote (Y′ i , Z′ i ) as the predicted values of the two tasks respectively, and the loss function of the model can be:

[0072]

[0073] That is, a linear combination of the squared loss functions of the two tasks. Among them, λ represents the weight of the loss function of the browsing price band task, which can take 1, or the weight can be appropriately reduced according to the actual scenario requirements.

[0074] It can be understood that the sample data belonging to the same preset-level category can be input into the data prediction model corresponding to this preset-level category, so as to output the predicted value of the value attribute of the user interacting with the items of this preset-level category. Or, the sample data and category identifiers belonging to various preset-level categories can be input into the same data prediction model, so as to output the predicted value of the value attribute of the user interacting with each preset-level category item. Among them, the category identifiers of various preset-level categories can be set in the one-hot encoding method.

[0075] That is to say, for the processing of categories, in the case of rich computing resources, a multi-task learning model can be independently trained for each category. The category id (identity identifier) can also be added to the model in the one-hot (one-hot encoding) form, and the data of all categories can be aggregated together to train a general model. If a general model is selected for training, attention should be paid to the distribution of the training set data volume among various categories. If the distribution difference is large, sampling needs to be considered to balance the data volume to ensure that each category can be fully learned.

[0076] As can be seen from the above description, in the training method of the data prediction model according to some embodiments of the present disclosure, through the multi-task learning algorithm, the historical behaviors of users can be mined from multiple aspects, and the price band preferences of users in each preset category can be learned, so that the interaction value attributes of users in each preset category can be predicted. In this way, the price preferences of users can be comprehensively characterized, and the accuracy and diversity of the model prediction results can be improved. Thereby, the products that meet the expected price levels of users in each category can be pushed, and the accuracy of product pushing can be improved. Furthermore, the probability of users interacting with the pushed products can be greatly increased, thereby increasing the view volume and order quantity of products.

[0077] It can be understood that in some actual scenarios, there may be various types of preferences of users in the price bands of each category. Some users' preferred price bands in some categories are relatively concentrated, and they are likely to purchase products within a certain fixed price range. For example, most users have their own fixed preferred prices for mobile phones. Some other users' preferred price bands in some other categories are relatively dispersed. For example, many users' price preferences for snacks are relatively dispersed, and they will buy snacks at various prices. There are also some situations where some users have two relatively concentrated preferred price bands in some categories. For example, the scenario where a user buys a high-price mobile phone for himself and a medium- and low-price elderly mobile phone for his parents. Considering the above actual scenarios, using only one value to estimate the price band preference score or stratification of a user in a certain category is not sufficient to fully characterize various complex preference situations. Therefore, the present invention proposes a model to estimate the price band preference distribution probability and correspondingly give the probabilities of each price band interval.

[0078] That is to say, in some embodiments, the above data prediction model can also be used to predict the distribution of the interaction value attribute band intervals of users in each preset category, that is, the price band preference distribution probability prediction model. In this case, when the execution entity generates model training data based on the second number of value attribute band intervals divided previously. Specifically, for the order placement behavior of users, at least one of the order quantity, distribution frequency, mean, and mode of the value attribute band of the order in each preset category of items can be used as the order sample data. And the proportion of the order items of users in each preset category belonging to each value attribute band interval can be used as the order sample label data. For the browsing behavior of users, at least one of the number of times, distribution frequency, mean, and mode of the value attribute band of browsing each preset category of items can be used as the browsing sample data. And the proportion of the items browsed by users in each preset category belonging to each value attribute band interval can be used as the browsing sample label data.

[0079] At this time, when setting the model parameters, such as Figure 2The price band preference distribution prediction model shown on the right side in the figure can set a second number of tasks (such as Task1-5) in the data prediction model to learn the probability that a user belongs to each order value attribute band interval. And set a second number of tasks (such as Task6-10) in the data prediction model to learn the probability that a user belongs to each browsing value attribute band interval. Moreover, the output of the data prediction model can be set as: for each preset category item of the same user, the prediction distribution in the order value attribute band interval and the prediction distribution in the browsing value attribute band interval.

[0080] As can be seen from Figure 2 , in terms of features, the two models can share features. To further improve the accuracy of the model preset results, some portrait features of the user, such as age, gender, etc., can also be added. That is to say, the samples of the price band preference distribution prediction model can be aligned with the samples of the price band preference score prediction model. Also, the users who have placed orders in the category in the recent month are used as samples. The model also adopts multi-task learning and a total of 10 tasks are set. The 1st to 5th tasks are used to learn the preference probability on the order price band intervals 1-5. The label of the sample is the quantity of the goods ordered by the user in this category in the recent month, and the proportion of the price belonging to the price band interval 1-5 in the total quantity of the goods ordered by the user in this category. The 6th to 10th tasks are used to learn the preference probability on the browsing price band intervals 1-5. The label of the sample is the pv of the goods browsed by the user in this category in the recent month, and the proportion of the price belonging to the price band interval 1-5 in the total pv of the user in this category.

[0081] The specific calculation method is: Denote (Y i1 , Y i2 , …, Y i5 , Z i1 , Z i2 , …, Z i5 ) as the label value of the i-th user. Denote as the price band numbers corresponding to the M i skus ordered by the i-th user in a certain category, then:

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] In addition, denote For the \(L\) i price band numbers corresponding to the sku browsed by the \(i\)-th user in a certain category, then:

[0088]

[0089]

[0090]

[0091]

[0092]

[0093] where \(i = 1, 2, \ldots, n\); \(I\{\}\) is an indicator function that takes the value 1 when the condition in the curly braces is satisfied and 0 otherwise.

[0094] Similarly, taking the NN model as an example, as Figure 2 shown in the model layer of the architecture diagram, the two tasks share the underlying layer, and the output layer is 10-dimensional, corresponding to the labels of 10 tasks respectively. Denote \((Y' i1 , Y' i2 , \ldots, Y' i5 , Z' i1 , Z' i2 , \ldots, Z' i5 ) as the predicted values of 10 tasks respectively, and the loss function of the model is:

[0095]

[0096] That is, a linear combination of the squared loss functions of 10 tasks. \(\lambda\) represents the weight of the loss function for the browsing price band task, which can take the value 1, or be appropriately reduced according to the actual scenario requirements. In addition, the processing of the category can be similar to the above price band preference score estimation model, and can be trained independently, or the category id can be added to the model in one-hot form and trained together to form a general model.

[0097] Through the above model, the output result is the probability score (order and browsing) of the user in each price band interval of each category. Further, the probability scores of the order and browsing can be respectively normalized to be used as the probability estimates of each price band interval:

[0098]

[0099]

[0100] where \(Y'' ij , Z'' ijrespectively represent the estimated order probability and the estimated browsing probability of the \(i\)-th user in the \(j\)-th price band interval of each category, after normalization; \(Y'\) ij and \(Z'\) ij respectively represent the order probability score and the browsing probability score of the \(i\)-th user in the \(j\)-th price band interval of each category.

[0101] As can be seen from the above description, in addition to estimating the price band preference stratification and price band preference scores of users for each category, the embodiments of the present disclosure can also estimate the preference probabilities of users for each price band, so as to more comprehensively characterize the price band preferences of users. If a user has relatively high probabilities for two price bands, it indicates that this user has two significant price band preferences.

[0102] Next, refer to Figure 3 as an implementation of the above Figure 1 shown training method, some embodiments of the present disclosure provide a training device for a data estimation model. These device embodiments correspond to Figure 1 the training method embodiments shown. The training device for the data estimation model can be specifically applied to various electronic devices.

[0103] As Figure 3 shown, in some embodiments, the training device 300 for the data estimation model may include: a training data generation unit 301, configured to generate model training data representing the value attribute distribution of the items of each preset-level category interacted by the user according to the interaction data between the user and the items within a specified time period, where the model training data includes sample data and corresponding sample label data; a model setting unit 302, configured to set the number of tasks and the model output dimension in the data estimation model based on the interaction data of the user, where the data estimation model adopts a multi-task learning model structure; a parameter adjustment unit 303, configured to input the sample data into the data estimation model, output the predicted values of the interaction value attributes of the user for each preset-level category, and adjust the model parameters for continuous training according to the comparison result between the predicted values of the interaction value attributes and the sample label data.

[0104] In some embodiments, the training data generation unit 301 may include: a value division sub-unit (not shown in the figure), configured to determine the range of the unit value attribute after value reduction of the items of each preset-level category interacted by the user within a specified time period, divide the range of the unit value attribute into the first number of value attribute bands, and perform data statistics on the value attribute bands to which the items of the preset-level category belong; a generation sub-unit (not shown in the figure), configured to generate model training data according to the statistical data of the items of various preset-level categories interacted by the user.

[0105] In some embodiments, the generation subunit may be further configured to, for the user's order placement behavior, use at least one of the order quantity, distribution frequency, mean, and mode of the value attribute bands of items of each preset-level category as order sample data, and use the average value of the value attribute bands to which the order items under each preset-level category belong as order sample label data; for the user's browsing behavior, use at least one of the number of times, distribution frequency, mean, and mode of browsing the value attribute bands of items of each preset-level category as browsing sample data, and use the average value of the value attribute bands to which the browsed items under each preset-level category belong as browsing sample label data.

[0106] In some embodiments, the model setting unit 302 may be further configured to set a task in the data prediction model to learn the score of the value attribute band of the user's order; and set another task in the data prediction model to learn the score of the value attribute band of the user's browsing; set the data prediction model to output the predicted scores of the order value attribute band and the browsing value attribute band of the same user for items of each preset-level category.

[0107] In some embodiments, the generation subunit may also be further configured to further divide the first number of value attribute bands of items of each preset-level category into the second number of value attribute band intervals, where the second number is less than the first number; for the user's order placement behavior, use at least one of the order quantity, distribution frequency, mean, and mode of the value attribute bands of items of each preset-level category as order sample data, and use the proportion of the order items under each preset-level category belonging to each value attribute band interval as order sample label data; for the user's browsing behavior, use at least one of the number of times, distribution frequency, mean, and mode of browsing the value attribute bands of items of each preset-level category as browsing sample data, and use the proportion of the browsed items under each preset-level category belonging to each value attribute band interval as browsing sample label data.

[0108] In some embodiments, the model setting unit 302 may also be further configured to set the second number of tasks in the data prediction model to learn the probability of the user belonging to each order value attribute band interval; and set another second number of tasks in the data prediction model to learn the probability of the user belonging to each browsing value attribute band interval; set the data prediction model to output the predicted distribution of the same user in the order value attribute band interval and the browsing value attribute band interval for items of each preset-level category.

[0109] In some embodiments, the parameter adjustment unit 303 may further be configured to input sample data belonging to the same preset category into the data prediction model corresponding to the preset category, and output the predicted value of the value attribute of the user's interaction with the item of the preset category; or input the sample data and category identifiers belonging to various preset categories into the same data prediction model, and output the predicted value of the value attribute of the user's interaction with each item of the preset category, where the category identifiers of various preset categories are set in a one-hot encoding manner.

[0110] It can be understood that the various units described in the training device 300 of the data prediction model correspond to the respective steps in the training method described in the reference Figure 1 Therefore, the operations, features, and beneficial effects described above for the method also apply to the training device 300 of the data prediction model and the units included therein, and will not be elaborated here.

[0111] Continuing to refer to Figure 4 , which shows a process 400 of some embodiments of the information push method according to the present disclosure. The method may further include the following steps:

[0112] Step 401, input the value attribute distribution data of each preset category item interacted by the target user within a specified time period, and the user's attribute information, into the data prediction model, and output the predicted value of the interaction value attribute of the target user for each preset category.

[0113] In some embodiments, the execution entity (such as an application server) of the information push method of the present disclosure may use the data prediction model in the above embodiments to predict the interaction value attribute of users on the e-commerce application platform. That is, the execution entity may input the value attribute distribution data of each preset category item interacted by the target user within a specified time period, and the user's attribute information, into the data prediction model. Among them, the data prediction model may be obtained by using the training method of the data prediction model described in any implementation manner in the above Figure 1 embodiments. Furthermore, the predicted value of the interaction value attribute of the target user for each preset category is obtained through the model output.

[0114] The target user here may be any user on the e-commerce application platform. The specified time period here may be the same as or different from the Figure 1 specified time period in the embodiments. The user's attribute information may be relevant information characterizing the user's personal attributes, such as the above portrait features. In addition, the predicted value of the interaction value attribute may be the predicted value of the value attribute of the target user's interaction with each preset category item. The predicted value of the interaction value attribute here may include at least one of the following: the predicted score of the interaction value attribute band, the stratification of the interaction value attribute band, the predicted distribution of the interaction value attribute band interval, etc.

[0115] It should be noted that the execution entity in the embodiments of the present disclosure may be the same as or different from the execution entity in the Figure 1 embodiment.

[0116] Step 402: Aggregate and analyze the predicted values of the interaction value attributes of each preset-level category output, and determine the predicted values of the interaction value attributes of the target user in each parent-level category.

[0117] In some embodiments, the execution entity may aggregate and analyze the predicted values of the interaction value attributes of the target user in each preset-level category output by the model, so as to determine the predicted values of the interaction value attributes of the target user in each parent-level category. Wherein, the parent-level category is the upper-level category to which the preset-level category belongs. For example, the parent of the third-level category is the second-level category, and the parent of the second-level category is the first-level category. The aggregation analysis method is not limited here.

[0118] As an example, for multiple preset-level categories belonging to the same parent-level category, first, the product of the predicted value of the interaction value attribute of the target user on each preset-level category and the number of items interacted by the target user in this preset-level category can be determined. Then, the sum of the products of multiple preset-level categories can be determined. After that, the total number of items interacted by the target user in multiple preset-level categories is determined. Finally, the ratio of the sum of the products to the total number of items can be determined as the predicted value of the interaction value attribute of the target user in the parent-level category to which multiple preset-level categories belong.

[0119] It can be understood that Figure 1 the model prediction in the embodiment is the price band preference score, stratification and distribution probability of the user in each third-level category. In this embodiment, a method for predicting the price band preference score, stratification and distribution probability of the user in the second-level category and the first-level category is proposed. Since there is a subordinate relationship among the first-level, second-level, and third-level categories, the price band preference of the second-level category can be aggregated from the price band preferences of the third-level categories it contains, and the price band preference of the first-level category can be aggregated from the price band preferences of the second-level categories it contains. Here, the weighted average method can be used to calculate the price band preferences of the first-level and second-level categories. The specific calculation method is as follows:

[0120] Denote that a certain second-level category j contains n j third-level categories, then the order (browse) price band preference score of user i in the second-level category j is:

[0121]

[0122] Denote that a certain first-level category k contains m k second-level categories, then the order (browse) price band preference score of user i in the first-level category k is:

[0123]

[0124] For the preference stratification of the order (browse) price range of the first- and second-level categories, the mapping method from the price range preference score of the third-level category to the preference stratification is still adopted, which will not be elaborated here. For the distribution probability of the order (browse) price range preference of the first- and second-level categories, a weighted average process is carried out in a similar way to the price range preference score. That is:

[0125] The preference probability score of user i for the order (browse) price range q of the second-level category j is:

[0126]

[0127] The preference probability score of user i for the order (browse) price range q of the first-level category k is:

[0128]

[0129] where q = 1, 2, 3, 4, 5. Here, the probability scores of orders and browsing can also be normalized respectively to obtain the corresponding probability estimates.

[0130] Step 403: Determine the target item that matches the target user according to the predicted values of the interaction value attributes of the target user at all levels and in all categories, and push the information of the target item to the target user.

[0131] In some embodiments, based on the aggregation analysis in step 402, the execution subject can determine the target item that matches the target user according to the predicted values of the interaction value attributes of the target user at all levels and in all categories. That is, the item whose unit value attribute (especially the unit price per piece after discount) matches the predicted value of the interaction value attribute. For example, the unit price of the target item is within the predicted price range (especially the order price range). At this time, the execution subject can push the information of the target item to the target user. For example, the execution subject can generate a notification message of the application according to the information of the target item, and then send the notification message to the terminal used by the target user. Another example is that when the target user opens the application, the information of the target item can be displayed on the home page of the application.

[0132] The information push method of the embodiments of the present disclosure can use a data prediction model to predict the predicted values of the interaction value attributes of the user in each preset-level category. In this way, the price preferences of the user in each category can be comprehensively analyzed, and at the same time, the accuracy of the analysis results can be ensured. Then, the predicted values of these preset-level categories can be aggregated and analyzed to obtain the predicted values of the interaction value attributes of the user in the parent category. In this way, the price preferences of the user in each level of category can be obtained. Furthermore, with reference to the price preferences of the user in each level and each category, the target item that better meets the user's expectations can be determined and pushed to the user. This helps to increase the probability of the user performing interaction operations with the target item, achieve precision marketing, and thus increase the sales volume and / or the number of views of the target item.

[0133] It should be noted that in Figure 1 and Figure 3 the embodiments, the recall scope of the user is all users who have placed orders and browsing behaviors in the target category in the past year. However, even those users who are not within the recall scope should also have price band preferences. Because price band preferences are generally a reflection of the user's consumption concept and do not completely depend on specific order placement and browsing behaviors. From another perspective, in marketing activities, it is also desirable to expand the scope of operable users as much as possible, not limited to users who have placed orders and browsing behaviors in the target category in the past year. In the embodiments of the present disclosure, the way of category generalization can be adopted to expand the user coverage. Here, generalization generally means that when the price band preference of the user in the target category is not estimated, the price band preferences of other categories that can approximately represent the target category are regarded as the price band preferences of the target category. Here, category generalization is not restricted, for example, it can include up-roll generalization and similarity generalization.

[0134] In some embodiments, if it is determined that the predicted value of the interaction value attribute of the target user in the target category is empty, first, the target parent category to which the target category belongs can be determined. Then, the predicted value of the interaction value attribute of the target user in the target parent category can be determined as the predicted value of the interaction value attribute of the target user in the target category.

[0135] As an example, if the price band preference of the user in a certain category is not estimated, the price band preference of the parent category where the user is located can be generalized to this category. For example Figure 6A in the up-roll generalization shown in, the user has no behavior in the third-level category "dried meat and dried fruits", so the price band preference cannot be estimated. At this time, the price band preference of its second-level category "snacks" can be used as the price band preference of the third-level category "dried meat and dried fruits". Similarly, for the second-level category "beverage brewing", the price band preference of its first-level category "food and beverage" can be used as the price band preference of this second-level category.

[0136] It should be noted that usually only one up-roll is considered, that is, only the results of the parent category are used for generalization. If the price band preference of the parent category has not been estimated yet, the parent of the parent category will not be used for generalization. The reason for this is to consider the rationality of generalization. The price band preference of the same user for the target category and the parent category is usually highly correlated. However, the correlation with the price band preference of the parent of the parent category is greatly reduced. Therefore, only the results of the parent category are used for generalization. This can ensure the accuracy of the price band preference generalization result, thus guaranteeing the accuracy of the item information recommendation.

[0137] Optionally, if it is determined that the predicted value of the interaction value attribute of the target user for the target category is empty, the execution entity can also adopt a similar generalization method. That is, first, the correlation coefficient of the target category and its target parent category in the predicted value of the interaction value attribute of the user can be determined as a benchmark threshold (benchmark). Then, the correlation coefficient of the target category and its sibling categories in the predicted value of the interaction value attribute of the user can be determined. After that, based on the predicted values of the interaction value attributes of each user for the candidate categories, the predicted value of the interaction value attribute of the target user for the target category can be determined. Among them, the candidate categories are the sibling categories whose correlation coefficient with the target category is greater than the benchmark threshold.

[0138] As an example, first, calculate the Pearson correlation coefficient of the price band preference scores of the target category and its parent category as the benchmark score. Among them, the Pearson Correlation Coefficient is usually used to measure whether two data sets are on the same line. Also, calculate the Pearson correlation coefficient of the price band preference scores of the target category and other sibling categories. Generalize the price band preference scores of the categories with results greater than the benchmark score to the target category. As Figure 6A shown in the similarity generalization, the price band preference of the third-level category "dried meat and meat shops" can be determined by using the price band preferences of the third-level categories "coffee" and "cereal beverages", and so on. When the Pearson correlation coefficient results of multiple categories are all greater than the benchmark score, the average value of the price band preferences of these categories can be taken as the price band preference of the target category. Or, when the price band preferences of these categories include the target user, in order to improve the accuracy of the generalization result, the average value of the price band preferences of the target user for these categories can be taken as the price band preference of the target category.

[0139] It should be noted that in the generalization method, roll-up generalization can be preferentially adopted. If the user has no predicted price band preference for a certain category and has not been roll-up generalized to a price band preference, similarity generalization can be adopted. In this way, both the accuracy of the price band preference can be ensured, and the user coverage can be greatly expanded, and the price band preference of users without behavior can be mined. Thus, the target user group can be expanded for precision marketing and category new customer acquisition, the coverage of marketing activities can be increased, and higher benefits can be obtained.

[0140] Refer to the following Figure 5 , as an implementation of the above Figure 4 shown method, the present disclosure provides some embodiments of an information push device. These device embodiments correspond to Figure 4 the method embodiments shown. The information push device can be specifically applied to various electronic devices.

[0141] As Figure 5 shown, an information push device 500 in some embodiments may include: a prediction unit 501 configured to input the value attribute distribution data of each preset-level category item interacted by a target user within a specified time period, and the attribute information of the user into a data prediction model, and output the predicted value of the interaction value attribute of the target user for each preset-level category, where the data prediction model is obtained by using the training method of the data prediction model described in any implementation manner in the above Figure 1 embodiment; an aggregation unit 502 configured to perform aggregation analysis on the predicted values of the interaction value attributes of each preset-level category output, and determine the predicted value of the interaction value attribute of the target user for each parent category, where the parent category is the upper-level category to which the preset-level category belongs; a push unit 503 configured to determine a target item matching the target user according to the predicted values of the interaction value attributes of the target user for each category at each level, and push the information of the target item to the target user.

[0142] In some embodiments, the aggregation unit 502 may be further configured to, for multiple preset-level categories belonging to the same parent category, determine the product of the predicted value of the interaction value attribute of the target user on each preset-level category and the number of items interacted by the target user on this preset-level category, and determine the sum of the products of the multiple preset-level categories; determine the total number of items interacted by the target user on the multiple preset-level categories; and determine the ratio of the sum of the products to the total number of items as the predicted value of the interaction value attribute of the target user for the parent category to which the multiple preset-level categories belong.

[0143] In some embodiments, the information push device 500 may further include a category generalization unit (not shown in the figure), which is configured to, in response to determining that the predicted value of the interaction value attribute of the target user for the target category is empty, determine the target parent category to which the target category belongs; and determine the predicted value of the interaction value attribute of the target user for the target parent category as the predicted value of the interaction value attribute of the target user for the target category.

[0144] In some embodiments, the category generalization unit may further be configured to, in response to determining that the predicted value of the interaction value attribute of the target user for the target category is empty, determine the correlation coefficient between the target category and the target parent category to which it belongs in terms of the predicted value of the interaction value attribute of the user as a reference threshold; determine the correlation coefficient between the target category and the peer categories in terms of the predicted value of the interaction value attribute of the user; and determine the predicted value of the interaction value attribute of the target user for the target category based on the predicted values of the interaction value attributes of each user for the candidate categories, where the candidate categories are the peer categories whose correlation coefficient with the target category is greater than the reference threshold.

[0145] It can be understood that the various units described in the information push device 500 correspond to the respective steps in the method described in the reference Figure 4 Therefore, the operations, features, and beneficial effects described above for the method also apply to the information push device 500 and the units included therein, and will not be elaborated herein.

[0146] Further referring to Figure 6B , which shows a schematic diagram of an application scenario combining the training method and the information push method of the embodiments of the present disclosure. As can be seen from Figure 6B , through the historical item interaction (such as placing an order, browsing, etc.) data of the user on the e-commerce application platform, the corresponding product sku price band can be obtained. Then, using these price bands, model estimation can be performed to obtain the price band preference of the third-level category of the user's pin (Personal identification number). Through model aggregation, the price band preference of the second-level category and the first-level category of the user can be obtained. For the categories where the user has no behavior, up-roll generalization and similarity generalization can be used to obtain the price band preference of the user for these categories. Finally, the price band preference of the user at various granularities can be output.

[0147] It should be noted that in the related methods and technologies regarding the user price band preference, the price band preference analysis method at the user-category granularity usually can cover a limited number of users. Whether it is a statistical method or a machine learning method, it can only cover the users who have behaviors on the target category in the recent period. The coverage is limited, especially in relatively niche categories, and fewer users with preferences can be mined.

[0148] In addition, it is impossible to characterize the diversity distribution of user preferences. Whether it is a statistical method or a machine learning method, the output is the unique price band preference of the user for each category. However, in fact, many users have preferences for multiple price bands. For example, a user has a preference for high-end mobile phones for personal use and also has a preference for relatively low-priced senior mobile phones for family members. Existing methods are difficult to support such cases of multiple preferences.

[0149] The embodiments of the present disclosure can improve the defects of the above-mentioned existing technologies. Specifically, by using a multi-task learning algorithm, it is possible to estimate the price band preference stratification, price band preference score, and price band preference distribution probability of the user under each category. It can characterize the different preferences of the user for each category and also has the ability to predict the future. At the same time, it also solves the problem that it is difficult to characterize the multiple preference price bands of users. In addition, through the analysis of the similarity of category price bands, the present invention generalizes the price band preferences of similar categories to users without behavior, solving the problem of limited user coverage. Furthermore, it helps to further increase the number of commodity orders.

[0150] Reference is made below to Figure 7 , which shows a schematic structural diagram of an electronic device 700 suitable for use in implementing some embodiments of the present disclosure. Figure 7 The electronic device shown is merely an example and should not impose any limitation on the functions and usage scope of the embodiments of the present disclosure.

[0151] As Figure 7 shown, the electronic device 700 may include a processing device 701 (such as a central processing unit, a graphics processing unit, etc.), which may perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage device 708 into a random access memory (RAM) 703. In the RAM 703, various programs and data required for the operation of the terminal device 700 are also stored. The processing device 701, the ROM 702, and the RAM 703 are connected to each other through a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0152] Generally, the following devices may be connected to the I / O interface 705: an input device 706 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 707 including, for example, a speaker, a vibrator, etc.; a storage device 708 including, for example, a magnetic disk, a hard disk, etc.; and a communication device 709. The communication device 709 may allow the electronic device 700 to communicate with other devices wirelessly or wiredly to exchange data. Although Figure 7 shows an electronic device 700 having various devices, it should be understood that it is not required to implement or include all the shown devices. Instead, more or fewer devices may be implemented or included.Figure 7 Each block shown may represent one device or, as required, multiple devices.

[0153] In particular, according to some embodiments of the present disclosure, the processes described above with reference to the flowcharts may be implemented as computer software programs. For example, some embodiments of the present disclosure include a computer program product that includes a computer program carried on a computer-readable medium, the computer program including program code for performing the methods shown in the flowcharts. In such some embodiments, the computer program may be downloaded and installed from the network via a communication device 709, or installed from a storage device 708, or installed from a ROM 702. When the computer program is executed by a processing device 701, the above functions defined in the methods of some embodiments of the present disclosure are performed.

[0154] It should be noted that the computer-readable medium described in some embodiments of the present disclosure may be a computer-readable signal medium or a computer-readable storage medium or any combination of the two. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of the computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In some embodiments of the present disclosure, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. And in some embodiments of the present disclosure, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which computer-readable program code is carried. Such a propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0155] In some embodiments, the client and the server can communicate using any currently known or future-developed network protocol such as HTTP (Hyper Text Transfer Protocol), and can be interconnected with digital data communication in any form or medium (e.g., communication network). Examples of communication networks include local area networks ("LAN"), wide area networks ("WAN"), the Internet (e.g., the Internet), and end-to-end networks (e.g., ad hoc end-to-end networks), as well as any currently known or future-developed networks.

[0156] The above computer-readable medium can be included in the above electronic device; or it can exist separately without being assembled into the electronic device. The above computer-readable medium carries one or more programs. When the above one or more programs are executed by the electronic device, the electronic device is caused to: generate model training data representing the value attribute distribution of each preset-level category of items interacted by the user based on the interaction data between the user and the items within a specified time period, where the model training data includes sample data and corresponding sample label data; set the number of tasks and the model output dimension in the data prediction model based on the user's interaction data, where the data prediction model adopts a multi-task learning model structure; input the sample data into the data prediction model, output the predicted values of the interaction value attributes of the user in each preset-level category, and adjust the model parameters for continuous training according to the comparison results between the predicted values of the interaction value attributes and the sample label data.

[0157] Alternatively, the electronic device is caused to: input the value attribute distribution data of each preset-level category of items interacted by the target user within a specified time period and the user's attribute information into the data prediction model, and output the predicted values of the interaction value attributes of the target user in each preset-level category, where the data prediction model is obtained by Figure 1 In the embodiment, it is obtained by the training method of the data prediction model described in any implementation manner; perform aggregation analysis on the predicted values of the interaction value attributes of each preset-level category output, and determine the predicted values of the interaction value attributes of the target user in each parent category, where the parent category is the upper-level category to which the preset-level category belongs; determine the target items that match the target user according to the predicted values of the interaction value attributes of the target user at all levels and in all categories, and push the information of the target items to the target user.

[0158] In addition, computer program code for performing the operations of some embodiments of the present disclosure can be written in one or more programming languages or combinations thereof. The programming languages include object-oriented programming languages such as Java, Smalltalk, C++, and also include conventional procedural programming languages such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, executed as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, by connecting through the Internet using an Internet service provider).

[0159] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagram may represent a module, a program segment, or a part of code that contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that marked in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and the combinations of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0160] The units described in some embodiments of the present disclosure can be implemented in software or in hardware. The described units can also be provided in a processor. For example, it can be described as: a processor includes a training data generation unit, a model setting unit, and a parameter adjustment unit; or includes a prediction unit, an aggregation unit, and a push unit. Among them, the names of these units do not constitute a limitation to the unit itself in some cases. For example, the training data generation unit can also be described as "a unit that generates model training data representing the value attribute distributions of various preset-level category items interacted by the user".

[0161] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that can be used include: Field Programmable Gate Arrays (FPGAs), Application Specific Integrated Circuits (ASICs), Application Specific Standard Products (ASSPs), Systems on Chip (SOCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0162] Some embodiments of the present disclosure also provide a computer program product, including a computer program, which when executed by a processor, implements any of the above-described data prediction model training methods or information pushing methods.

[0163] The above description is only some preferred embodiments of the present disclosure and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, technical solutions formed by mutually replacing the above features with technical features having similar functions (but not limited to) disclosed in the embodiments of the present disclosure.

Claims

1. A method for training a data estimation model, comprising: Generate model training data representing the value attribute distribution of items of each preset level category interacted by the user according to the interaction data between the user and the item in a specified time period, wherein the model training data includes sample data and corresponding sample label data; Based on the user's interaction data, setting the number of tasks and model output dimensions in a data estimation model, wherein the data estimation model adopts a multi-task learning model structure; The sample data is input into the data estimation model, the predicted values ​​of the interaction value attributes of the user in each preset level category are output, and according to the comparison result between the predicted values ​​of the interaction value attributes and the sample label data, the model parameters are adjusted to continue training.

2. The method for training a data prediction model according to claim 1, wherein: The generating of model training data representing the value attribute distribution of items of each preset level category interacted by the user according to the interaction data between the user and the item within the specified time period includes: For each item of a preset category interacted by the user within a specified time period, determine a unit value attribute range after the value of the item under the preset category is reduced, divide the unit value attribute range into a first number of value attribute bands, and perform data statistics on the value attribute bands to which the items under the preset category belong; Model training data is generated based on the statistical data of items under various preset levels of categories interacted by the user.

3. The training method of the data prediction model according to claim 2, wherein: The generating of model training data according to the statistical data of items under various preset level categories interacted by the user includes: For the order data of the user, at least one of the order quantity, distribution frequency, mean, and mode of the value attribute band of the order items in each preset level category is used as the order sample data, and the average value of the value attribute band to which the order items of the user in each preset level category belong is used as the order sample label data; For the user's browsing data, at least one of the number of times the value attribute band of each preset category item is browsed, the distribution frequency, the mean and the mode is used as browsing sample data, and the average value of the value attribute band of the items browsed by the user under each preset category is used as browsing sample label data.

4. The method for training a data prediction model according to claim 3, wherein: The step of setting the number of tasks and the model output dimension in the data estimation model based on the user's interaction data includes: Setting a task in the data estimation model to learn the score of the user's order value attribute band based on the order sample data; and Setting another task in the data estimation model to learn the score of the user's browsing value attribute band based on the browsing sample data; The data estimation model is set to output the predicted scores of the order value attribute band and the browsing value attribute band for each preset level category item of the same user.

5. The method for training a data prediction model according to claim 2, wherein: The generating of model training data according to the statistical data of items under various preset level categories interacted by the user also includes: Further dividing the first number of value attribute bands of each preset level category item into a second number of value attribute band intervals, wherein the second number is smaller than the first number; For the order data of the user, at least one of the order quantity, distribution frequency, mean, and mode of the order items in each preset level category is used as the order sample data, and the proportion of the order items of the user in each preset level category that belong to each value attribute band interval is used as the order sample label data; For the browsing data of the user, at least one of the number of times the value attribute band of each preset category item is browsed, the distribution frequency, the mean and the mode is used as the browsing sample data, and the proportion of the items browsed by the user under each preset category belonging to each value attribute band interval is used as the browsing sample label data.

6. The method for training a data prediction model according to claim 5, wherein: The step of setting the number of tasks and the model output dimension in the data estimation model based on the user's interaction data further includes: The second number of tasks is set in the data estimation model to learn the probability that the user belongs to each order value attribute band interval based on the order sample data; and The second number of tasks is further set in the data estimation model to learn the probability that the user belongs to each browsing value attribute band interval according to the browsing sample data; The data estimation model is set to output the predicted distribution of the same user in the order value attribute band interval and the predicted distribution in the browsing value attribute band interval for each preset level category item.

7. The method for training a data estimation model according to any one of claims 1 to 6, wherein: The step of inputting the sample data into the data estimation model comprises: Input sample data belonging to the same preset category into a data estimation model corresponding to the preset category, and output a predicted value attribute of the user's interaction with the item of the preset category; or The sample data and category identifiers belonging to various preset categories are input into the same data estimation model to output the predicted value attribute of each preset category item interacted by the user, wherein the category identifiers of various preset categories are set using a unique hot encoding method.

8. A training device for a data prediction model, comprising: A training data generating unit is configured to generate model training data representing the value attribute distribution of items of each preset level category interacted by the user according to the interaction data between the user and the item in a specified time period, wherein the model training data includes sample data and corresponding sample label data; A model setting unit, configured to set the number of tasks and model output dimensions in a data estimation model based on the user's interaction data, wherein the data estimation model adopts a multi-task learning model structure; The parameter adjustment unit is configured to input the sample data into the data estimation model, output the predicted value of the interaction value attribute of the user in each preset level category, and adjust the model parameters to continue training based on the comparison result between the predicted value of the interaction value attribute and the sample label data.

9. An information push method, comprising: Inputting the value attribute distribution data of the items of each preset level category interacted by the target user within a specified time period and the attribute information of the user into a data estimation model, and outputting the predicted value attribute of the interaction value of the target user in each preset level category, wherein the data estimation model is obtained by using the training method of the data estimation model according to any one of claims 1 to 7; Performing aggregate analysis on the output predicted values ​​of the interaction value attributes of each preset category to determine the predicted values ​​of the interaction value attributes of the target user in each parent category, wherein the parent category is the upper category to which the preset category belongs; According to the predicted values ​​of the interaction value attributes of the target user at all levels and categories, a target item matching the target user is determined, and information of the target item is pushed to the target user.

10. The information push method according to claim 9, wherein: The step of performing aggregate analysis on the output predicted values ​​of the interaction value attributes of each preset level category to determine the predicted values ​​of the interaction value attributes of the target user in each parent level category includes: For multiple preset categories belonging to the same parent category, determine the product of the target user's interaction value attribute prediction value on each preset category and the number of items interacted by the target user in the preset category, and determine the sum of the products of the multiple preset categories; Determine the total number of items interacted by the target user in the multiple preset level categories; The ratio of the sum of the products to the total number of items is determined as the predicted value of the interaction value attribute of the target user in the parent category to which the multiple preset categories belong.

11. The information push method according to claim 9, wherein: The method further comprises: In response to determining that the predicted value of the interaction value attribute of the target user in the target category is null, determining a target parent category to which the target category belongs; The predicted value of the interaction value attribute of the target user in the target parent category is determined as the predicted value of the interaction value attribute of the target user in the target category.

12. The information push method according to any one of claims 9 to 11, wherein: The method further comprises: In response to determining that the predicted value of the interaction value attribute of the target user in the target category is empty, determining a correlation coefficient between the target category and its corresponding target parent category on the predicted value of the interaction value attribute of the user as a reference threshold; Determine the correlation coefficient between the target category and the same level category in the predicted value of the user's interactive value attribute; Based on the predicted values ​​of the interaction value attributes of each user in the candidate categories, the predicted values ​​of the interaction value attributes of the target user in the target category are determined, wherein the candidate categories are categories of the same level whose correlation coefficient with the target category is greater than the benchmark threshold.

13. An information push device, comprising: The prediction unit is configured to input the value attribute distribution data of the preset-level category items interacted by the target user within a specified time period and the attribute information of the user into a data estimation model, and output the predicted value attribute of the interaction value of the target user in each preset-level category, wherein the data estimation model is obtained by the training method of the data estimation model according to any one of claims 1 to 7; an aggregation unit configured to perform aggregation analysis on the output interaction value attribute prediction values ​​of each preset category to determine the interaction value attribute prediction value of the target user in each parent category, wherein the parent category is the upper category to which the preset category belongs; The push unit is configured to determine a target item that matches the target user according to the predicted values ​​of the interaction value attributes of the target user at all levels and categories, and push the information of the target item to the target user.

14. An electronic device comprising: one or more processors; a storage device having one or more programs stored thereon, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1-7 or 9-12.

15. A computer readable medium having a computer program stored thereon, wherein: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 or 9 to 12 is implemented.

16. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the method according to any one of claims 1 to 7 or 9 to 12.

Images