Information push method, device, equipment and storage medium

Abstract

The present application provides an information push method, device, equipment and storage medium, wherein the method includes: obtaining a scenario model corresponding to the recommendation scenario currently displayed by the client, estimating the user's preferred products in the recommendation scenario according to the scenario model corresponding to the recommendation scenario, and recommending information about the user's preferred products in the recommendation scenario. In this technical solution, by using the scenario model obtained by joint data of multiple recommendation scenarios and scenario task training, the user's preferred products in each recommendation scenario are estimated, data sharing of different recommendation scenarios can be achieved, the correlation between each scenario task in different recommendation scenarios is fully considered, and it is more in line with the actual situation of user feedback data, and the user's product preferences can be more accurately estimated, thereby improving the recommendation effect of information.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to an information push method, device, equipment and storage medium. Background Art

[0002] With the development of big data and machine learning technology, it has been applied in the field of personalized recommendation technology. Personalized recommendation refers to analyzing user preferences by mining user historical behavior characteristics, so as to recommend preferred information and products to users through the client's recommendation page.

[0003] In the prior art, click-through rate (CTR) prediction models are widely used in personalized recommendation systems. CTR prediction models are used to learn and predict user feedback. By collecting a large amount of training expectations and conducting learning training, they can predict the information and products that users prefer.

[0004] However, in multi-recommendation scenarios, there are correlations between different scenario tasks. Since the CTR prediction model uses a single-task learning method, it does not consider the mutual connection between scenario tasks and ignores the user's behavior data in different recommendation scenarios. As a result, in multi-recommendation scenarios, the CTR prediction model cannot accurately push user-preferred information and products. Summary of the invention

[0005] The present application provides an information push method, apparatus, device and storage medium, which are used to solve the problem that the existing estimation model cannot accurately push information and products preferred by users in multiple recommendation scenarios.

[0006] In a first aspect, an embodiment of the present application provides an information push method, including:

[0007] Obtaining a scenario model corresponding to a recommendation scenario currently displayed on the client, wherein the recommendation scenario is used to characterize a product recommendation page of the client, and the scenario model is obtained by training with joint data of multiple recommendation scenarios and scenario tasks, wherein the scenario tasks include a click task and an order task;

[0008] According to the scenario model corresponding to the recommendation scenario, estimating the product preferred by the user in the recommendation scenario;

[0009] Recommend information about products preferred by the user in the recommendation scenario.

[0010] In a second aspect, an embodiment of the present application provides an information push device, including:

[0011] An acquisition module is used to acquire a scenario model corresponding to a recommendation scenario currently displayed on the client, wherein the recommendation scenario is used to characterize a product recommendation page of the client, and the scenario model is obtained by training the joint data of multiple recommendation scenarios and scenario tasks, wherein the scenario tasks include a click task and an order task;

[0012] An estimation module, used to estimate the product preferred by the user in the recommendation scenario according to the scenario model corresponding to the recommendation scenario;

[0013] The push module is used to recommend information about products preferred by users in the recommendation scenario.

[0014] In a third aspect, an embodiment of the present application provides a processing device, including a memory and at least one processor;

[0015] The memory stores computer-executable instructions;

[0016] The at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor performs the above method.

[0017] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer-readable storage medium has computer instructions stored therein, and the computer instructions are used to implement the above-mentioned method when executed by a processor.

[0018] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program / instruction, which implements the above method when executed by a processor.

[0019] The information push method, apparatus, device and storage medium provided in the embodiments of the present application use the scenario model obtained by joint data of multiple recommendation scenarios and scenario task training to estimate the products preferred by users in each recommendation scenario, thereby realizing data sharing of different recommendation scenarios, fully considering the correlation between each scenario task in different recommendation scenarios, being more in line with the actual situation of user feedback data, being able to more accurately estimate the user's product preferences and improving the information recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic diagram of an application scenario of the information push method provided in an embodiment of the present application;

[0021] Figure 2 A flowchart of the information push method provided in the embodiment of the present application;

[0022] Figure 3 A schematic diagram of the structure of a preset model provided in an embodiment of the present application;

[0023] Figure 4 A schematic diagram of the structure of an information push device provided in an embodiment of the present application;

[0024] Figure 5 A schematic diagram of the structure of a processing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0026] Below, some terms in the embodiments of the present application are explained to facilitate understanding by those skilled in the art:

[0027] Recommendation scenario: Users can install the client on their mobile terminals and call up different product recommendation pages in the client. Each product recommendation page can recommend products and information for users to view. The recommendation scenario is used to characterize the product recommendation page.

[0028] Scenario task: Scenario task is to predict the user's interactive behavior in the recommendation scenario. For example, scenario tasks include click tasks and order tasks. The click task means predicting whether the user will click on the product recommended in the recommendation scenario, and the order task means predicting whether the user will place an order to purchase the product recommended in the recommendation scenario.

[0029] Figure 1 A schematic diagram of an application scenario of the information push method provided in the embodiment of the present application, such as Figure 1 As shown, in personalized recommendations, users can install applications (APP) on mobile terminals 11 (such as mobile phones, computers and tablet computers). Taking APP as a client as an example, APP includes product recommendation pages such as APP homepage, order page, and product details page. These product recommendation pages generally display products and information for users to browse, click, add to cart, place orders, etc. Among them, users' browsing, clicking, adding to cart and placing orders can be monitored by the platform server as historical behavior characteristics of users.

[0030] Exemplarily, the historical behavior characteristics of the user may be the historical behavior characteristics generated by the user in the past seven days or the historical behavior characteristics generated by the user in the past month.

[0031] The server 12 performs learning and training based on the user's historical behavior characteristics, optimizes the scenario tasks, estimates the products and information that the user is interested in, and feeds back to the client to recommend these products and information of interest to different product recommendation pages.

[0032] Currently, there are two main methods for personalized recommendation:

[0033] 1. The click-through rate (CTR) prediction model is used for learning and prediction. The CTR prediction model uses a single-task learning method. It learns and predicts whether a user will click on a certain product based on the user's click behavior. This method does not consider the relationship between scenario tasks. In addition, the CTR prediction model requires a large amount of training expectations during learning and training. The prediction effect requires a large amount of training expectations to support it, and it cannot be applied to some recommendation scenarios where training expectations are insufficient.

[0034] 2. Using multi-task learning, the personalized recommendation system not only pays attention to the CTR indicator, but also needs to optimize indicators from multiple angles such as conversion rate (CVR), browsing depth, and nurturing. Multi-task learning can train multiple scenario tasks at the same time, and link multiple mutually influencing scenario tasks together by sharing results, so that multiple goals can reach the overall optimal, and the generalization ability is better than that of a single task. However, in the multi-task learning method, since the user's browsing, adding to cart, and placing orders may be completed in different recommendation scenarios, this multi-task learning method does not consider the connection between different scenarios. Even for the same scenario task (such as click task, browsing task), there are differences in user feedback effects in different recommendation scenarios. The similarity and correlation of different recommendation scenarios are not weighed, resulting in unsatisfactory recommendation effects.

[0035] In summary, there is no method in the prior art for accurately estimating user product preferences and pushing information in multiple recommendation scenarios.

[0036] In response to the above problems, the present application provides an information push method, apparatus, device and storage medium, which utilizes the joint data and scenario tasks of multiple recommendation scenarios to train scenario models corresponding to different recommendation scenarios, thereby increasing the training data and allowing the data under different recommendation scenarios to influence each other, effectively weighing the similarities and correlations of different recommendation scenarios. At the same time, it is also possible to iteratively generate scenario models corresponding to multiple recommendation scenarios through one training, and the model training efficiency is higher. When a user browses a recommendation scenario, the scenario model corresponding to the recommendation scenario is used to predict the user's product preference in the recommendation scenario, and recommend product information to the recommendation scenario, ultimately achieving better recommendation effects and improving product conversion rates.

[0037] The technical solution of the present application is described in detail below through specific embodiments. It should be noted that the following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0038] Figure 2 The flowchart of the information push method provided in the embodiment of the present application is as follows. The method can be applied to a server in a network platform or a local processing device, such as Figure 2 As shown, the method comprises the following steps:

[0039] S201: Obtain a scene model corresponding to a recommended scene currently displayed on the client.

[0040] Among them, the recommendation scenario is used to characterize the client's product recommendation page. The scenario model is trained through the joint data and scenario tasks of multiple recommendation scenarios. The joint data includes the scenario identifier, task identifier and the user's historical behavior characteristics. The scenario tasks include click tasks and order tasks.

[0041] Exemplarily, a user can install a client, such as an APP, on a mobile terminal. The APP may include product recommendation pages such as the APP homepage, the Guess You Like page, the product details page, the shopping cart, and the order page. When the user opens the APP and enters the APP homepage, the recommendation scene currently displayed on the client is the APP homepage. Some products and information can be recommended on the APP homepage, such as "Mobile Phone-Discount Information". The user can also click the "Shopping Cart" icon on the homepage to enter the shopping cart. At this time, the recommendation scene currently displayed on the client is the shopping cart page.

[0042] For example, the user's historical behavior characteristics may be different in different recommendation scenarios. For example, the user's historical behavior on the shopping page mainly includes ordering behavior, the historical behavior on the product details page mainly includes adding to cart behavior, and the historical behavior on the APP homepage mainly includes browsing and clicking behavior.

[0043] In this embodiment, when training the scenario model, the input training prediction integrates the historical behavior characteristics, scenario identifiers, and task identifiers of each recommendation scenario, wherein the scenario identifiers are used to identify different recommendation scenarios, and the task identifiers are used to identify each scenario task in different recommendation scenarios. Exemplarily, the scenario tasks of different recommendation scenarios can be the same, for example, different recommendation scenarios have the same click task and browse task.

[0044] Optionally, the scenario model can learn the importance of different recommendation scenarios and the importance of different scenario tasks in each recommendation scenario based on the joint data of multiple recommendation scenarios, that is, determine the weight of each recommendation scenario and the weight of each scenario task, and train scenario models corresponding to different recommendation scenarios.

[0045] S202: Estimate the user's preferred products in the recommended scenario based on the scenario model corresponding to the recommended scenario.

[0046] Specifically, each recommendation scenario corresponds to a trained scenario model. The scenario tasks of each recommendation scenario include a click task and an order task. The click task is used to predict whether the user will click on the product recommended to the recommendation scenario, and the order task predicts whether the user will place an order to purchase the product recommended to the recommendation scenario.

[0047] In this embodiment, the trained scenario model is combined with the weight of each scenario task to determine the product preferred by the user in the recommendation scenario.

[0048] For example, the scenario model can be used to output the user's preference score for a certain product, and each product can be ranked according to the preference scores of different products. Based on the ranking relationship, the user's preferred product in the recommendation scenario can be determined.

[0049] S203: Recommend information about products preferred by the user in the recommendation scenario.

[0050] Exemplarily, the recommendation scenario may recommend information about products preferred by multiple users. Based on the user's preference for each product, N (N is a positive integer) products with higher preference levels are selected, and information about N products is recommended in the recommendation scenario.

[0051] The embodiment of the present application obtains a scenario model corresponding to each recommendation scenario by using the joint data and scenario tasks of multiple recommendation scenarios for learning and training in advance. The scenario model obtained by combining the joint data training of multiple recommendation scenarios can more accurately predict the products preferred by users, thereby improving the product recommendation effect.

[0052] In some embodiments, before the above step S201, the above method further includes the following steps:

[0053] Obtain scenario tasks and joint data of multiple recommendation scenarios submitted by the client;

[0054] According to the preset loss function, scenario tasks and joint data, the preset model is trained to obtain the scenario model corresponding to each recommended scenario.

[0055] Specifically, the joint data includes the scenario identifier of the recommendation scenario, the task identifier and the user's historical behavior characteristics. The scenario identifier can be used to distinguish different recommendation scenarios, and the task identifier can be used to distinguish different scenario tasks. Users can browse, click, add to cart and place orders in different recommendation scenarios. These can be used as the user's historical behavior characteristics. The client can upload the user's historical behavior characteristics to the server at a regular interval.

[0056] Exemplarily, the preset loss function may be:

[0057]

[0058] In the above formula, Indicates whether sample i belongs to the recommended scene S corresponding to scene task t t ,ω t represents the weight of scene task t, is the weighted cross entropy.

[0059] In this embodiment, weighted cross entropy is a common loss function, where is the model prediction value of sample i, is the true data value of sample i, Represents the coefficient of positive samples when calculating weighted cross entropy.

[0060] The embodiment of the present application takes the scenario tasks and joint data of multiple recommendation scenarios as input, trains the preset model through a preset loss function, and can iteratively generate scenario models corresponding to multiple recommendation scenarios through one training, thereby improving the training efficiency of the model. In addition, the training process can weight different recommendation scenarios and scenario tasks, learn the similarities and correlations of different recommendation scenarios, and improve the recommendation effects of different recommendation scenarios.

[0061] Figure 3 A schematic diagram of the structure of the preset model provided in the embodiment of the present application, such as Figure 3 As shown, the preset model includes an input layer 31, a sharing layer 32 and a tower layer 33. The above-mentioned "training the preset model according to the preset loss function, scene task and joint data to obtain a scene model corresponding to each recommended scene" can be specifically implemented by the following steps:

[0062] According to the scene identifier and the task identifier, the recommended scene and scene task are identified;

[0063] According to the preset loss function and joint data, the shared layer is trained to obtain the adaptive weights of the scenario tasks for each recommended scenario;

[0064] According to the adaptive weights, the scenario tasks of each recommended scenario are weighted and summed, and input into the tower layer corresponding to each scenario task respectively;

[0065] According to the tower layers, the scenario tasks of each recommended scenario are decoupled to obtain the scenario model corresponding to each recommended scenario.

[0066] Exemplarily, the input layer 31 is used to input the joint data of the recommendation scene, the shared layer 32 includes an expert network expert and a gate network gate, and the tower layer 33 includes a plurality of tower networks TOWER, as shown in FIG. Figure 3 As shown, the recommendation positions A, B, C, and D refer to four recommendation scenarios. Each recommendation scenario includes a click task and an order task. The expert network includes expert1 to expert8. Each scenario task corresponds to a separate gate network gate, and each scenario task corresponds to a separate tower network TOWER learning. For example, Figure 3 In the above figure, gate recommendation position A click represents the gated network corresponding to the click task of recommendation position A, and TOWER recommendation position A click represents the tower network TOWER corresponding to the click task of recommendation position A.

[0067] In this embodiment, when training the shared layer, the inputs of different scene tasks will obtain different adaptive weights through the gated network gate to achieve selective use of the expert network expert, so that different scene tasks can learn different combined expert models, and the effect is similar to weighted summation. Then each tower network TOWER in the tower layer learns separately, decouples the scene tasks that need to be optimized under each recommendation position, and outputs the scene model.

[0068] By training the shared layer, the embodiment of the present application can learn the integration method of the shared layer, integrate the mutual relationship of the scenario tasks of each recommended scenario, improve the generalization ability, and through one training, it is possible to obtain the scenario models corresponding to multiple recommended scenarios, and the model training efficiency is higher.

[0069] Furthermore, in some embodiments, when training the shared layer, the joint data can be input into the expert network expert and the gated network gate, and the expert network expert and the gated network gate are trained according to a preset loss function to obtain adaptive weights for tasks in different scenarios.

[0070] Specifically, the expert network Experts learns a vector sequence (embedding sequence) based on the joint data, and the gate network gate learns adaptive weights and performs weighted summation on the outputs of the expert network Experts. After obtaining the corresponding results, they are input into the tower network TOWER corresponding to each scene task respectively. The tower network TOWER decouples and outputs the scene model.

[0071] The embodiment of the present application learns similarity and correlation models of different recommendation scenarios through an expert network, and learns the integration of shared layers through a gated network, so as to learn scenario tasks of different scales and types and adapt to scenario tasks of various recommendation scenarios.

[0072] Optionally, in some embodiments, before training the shared layer according to the loss function and the joint data, the above information push method may further include the following steps:

[0073] Get the number of positive samples of scene task t and the number of negative samples of scene task t;

[0074] According to the preset weight strategy, the number of positive samples and the number of negative samples, the preset loss function is weighted.

[0075] In this embodiment, the scale of the loss function may vary greatly for scenario tasks in different recommendation scenarios. For example, there may be a large gap in the magnitude of the joint data, i.e., the training corpus, in different recommendation scenarios. For scenario tasks of different magnitudes, it is necessary to assign weights to the loss function during the learning process so that scenario tasks of different magnitudes can be reasonably considered during the learning process.

[0076] Exemplarily, the preset weight strategy is:

[0077]

[0078]

[0079] In the above formula, represents the number of positive samples of scene task t, represents the number of negative samples of scene task t, is the coefficient of positive samples when calculating weighted cross entropy.

[0080] The embodiment of the present application assigns weights to the loss function through a preset weight strategy. During subsequent learning and training, the differences between scenario tasks of different magnitudes can be fully considered to avoid the loss function performing well on some scenario tasks but poorly on other scenario tasks.

[0081] For example, in some embodiments, the above step S202 may be implemented by the following steps:

[0082] According to the scenario model, an output of a scenario task of the recommended scenario is obtained, where the output of the scenario task is used to indicate an importance score of the scenario task;

[0083] According to the preset fusion strategy, the scenario tasks of the recommendation scenario are fused to obtain the output of the recommendation scenario. The output of the recommendation scenario is used to indicate the user's preference for different products in the recommendation scenario.

[0084] According to the output of the recommendation scenario, the products preferred by the user in the recommendation scenario are obtained.

[0085] In this embodiment, each recommended scenario may correspond to multiple scenario tasks, and the importance scores of different scenario tasks may be different. For example, the importance score of a click task of a recommended scenario may be 0.8, while the importance score of an order task of the recommended scenario may be 0.2.

[0086] Exemplarily, the preset fusion strategy is:

[0087]

[0088] In the above formula, represents the output of the recommended scene s, T s represents all scenario tasks for the recommended scenario s, represents the weight of the output of scene task t, Represents the output of scene task t.

[0089] For example, It can be adjusted by the operator according to the actual situation. The output of the recommendation scenario represents the predicted ranking of the user's preference for all commodities. From the ranking, the top N products can be selected as the user's preferred products in the recommendation scenario, where N is a positive integer. For example, N can be 10.

[0090] The embodiment of the present application integrates the scenario tasks of the recommendation scenario so that the output of the recommendation scenario integrates the performance of multiple recommendation tasks, thereby ultimately achieving the best overall recommendation effect.

[0091] Optionally, in some embodiments, if the product recommendation page includes at least one of a non-empty shopping cart, an empty shopping cart, an order details page, and an order tracking page, the above step S203 may be specifically implemented by the following steps:

[0092] Recommend information about products preferred by the user in at least one of a non-empty shopping cart, an empty shopping cart, an order details page, and an order tracking page.

[0093] In this embodiment, the user can operate on the client to add the products recommended in a certain product recommendation page to the shopping cart, and can also delete the products added to the shopping cart in the shopping cart product recommendation page, and can also place an order to purchase the products added to the shopping cart, and view the order information on the order details page, and view the logistics status of the order on the order tracking page.

[0094] Exemplarily, when a user views order information on an order details page, a product recommendation area may be preset on the order details page, and user-preferred products may be recommended in the product recommendation area.

[0095] The embodiment of the present application sets multiple recommendation scenarios, namely, product recommendation pages, and can integrate the joint data of multiple recommendation scenarios to obtain more training corpora, so that the low-traffic recommendation scenario with smaller training corpora can draw on the training corpora of other recommendation scenarios, thereby improving the recommendation effect of each recommendation scenario.

[0096] Based on the above embodiment, the above step of "recommending information about products preferred by the user in at least one of a non-empty shopping cart, an empty shopping cart, an order details page, and an order tracking page" may specifically include the following steps:

[0097] Get product price information, picture information and parameter information;

[0098] The price information, picture information and parameter information of the product are recommended in at least one of a non-empty shopping cart, an empty shopping cart, an order details page and an order tracking page.

[0099] Exemplarily, the price information of a product includes the current price and historical prices of the product, the picture information includes a display picture of the product, and the parameter information includes the model, specification, quantity, color, etc. of the product.

[0100] Optionally, two virtual areas may be divided in the recommendation area (ie, the product recommendation page), one of which is used to display product image information, and the other virtual area is used to display product parameter information and price information.

[0101] The embodiment of the present application recommends product price information, image information, and parameter information to corresponding recommendation scenarios, making it convenient for users to intuitively and quickly understand relevant product information and improving recommendation effects.

[0102] The following are device embodiments of the present application, which can be used to execute the method embodiments of the present application. For details not disclosed in the device embodiments of the present application, please refer to the method embodiments of the present application.

[0103] Figure 4This is a schematic diagram of the structure of the information push device provided in the embodiment of the present application. Optionally, the information push device 40 can be applied to a server, or can be an independent server and work in conjunction with the server to implement the technical solution of the present application, such as Figure 4 As shown, the information pushing device includes: an acquisition module 41, an estimation module 42 and a pushing module 43.

[0104] Among them, the acquisition module 41 is used to obtain the scene model corresponding to the recommended scene currently displayed by the client; the estimation module 42 is used to estimate the products preferred by the user in the recommended scene based on the scene model corresponding to the recommended scene; the push module 43 is used to recommend information about the products preferred by the user in the recommended scene.

[0105] Among them, the recommendation scenario is used to characterize the client's product recommendation page. The scenario model is trained through the joint data of multiple recommendation scenarios and scenario tasks. The scenario tasks include click tasks and order tasks.

[0106] In some embodiments, the information push device 40 further includes a training module for:

[0107] Obtain scenario tasks and joint data of multiple recommendation scenarios submitted by the client;

[0108] According to the preset loss function, scenario tasks and joint data, the preset model is trained to obtain the scenario model corresponding to each recommended scenario.

[0109] Based on the above embodiments, in some embodiments, if the preset model includes a shared layer and a tower layer corresponding to each scenario task, the above training module is specifically used for:

[0110] According to the scene identifier and the task identifier, the recommended scene and scene task are identified;

[0111] According to the preset loss function and joint data, the shared layer is trained to obtain the adaptive weights of the scenario tasks for each recommended scenario;

[0112] According to the adaptive weights, the scenario tasks of each recommended scenario are weighted and summed, and input into the tower layer corresponding to each scenario task respectively;

[0113] According to the tower layers, the scenario tasks of each recommended scenario are decoupled to obtain the scenario model corresponding to each recommended scenario.

[0114] Optionally, in some embodiments, if the shared layer includes an expert network and a gating network, the training module is specifically used for:

[0115] The joint data is fed into the expert network and the gating network;

[0116] According to the preset loss function, the expert network and the gating network are trained to obtain adaptive weights for tasks in different scenarios.

[0117] In some embodiments, the above-mentioned preset loss function may be:

[0118]

[0119] In the above formula, Indicates whether sample i belongs to the recommended scene S corresponding to scene task t t ,ω t represents the weight of scene task t, is the weighted cross entropy.

[0120] In some embodiments, the information recommendation device 40 further includes an allocation module for:

[0121] Get the number of positive samples of scene task t and the number of negative samples of scene task t;

[0122] According to the preset weight strategy, the number of positive samples and the number of negative samples, the preset loss function is weighted.

[0123] In some embodiments, the above-mentioned preset weight strategy may be:

[0124]

[0125]

[0126] In the above formula, represents the number of positive samples of scene task t, represents the number of negative samples of scene task t, is the coefficient of positive samples when calculating weighted cross entropy.

[0127] For example, It can be input by the operator and adjusted later according to the actual situation.

[0128] In some embodiments, the estimation module 42 is specifically used to:

[0129] According to the scenario model, obtain the output of the scenario task of the recommended scenario;

[0130] According to the preset fusion strategy, the scene tasks of the recommended scene are fused to obtain the output of the recommended scene;

[0131] According to the output of the recommendation scenario, the products preferred by the user in the recommendation scenario are obtained.

[0132] Among them, the output of the scenario task is used to indicate the importance score of the scenario task, and the output of the recommendation scenario is used to indicate the user's preference for different products in the recommendation scenario.

[0133] Exemplarily, in some embodiments, the preset fusion strategy may be:

[0134]

[0135] In the above formula, represents the output of the recommended scene s, T s represents all scenario tasks for the recommended scenario s, represents the weight of the output of scene task t, Represents the output of scene task t.

[0136] In some embodiments, if the product recommendation page includes at least one of a non-empty shopping cart, an empty shopping cart, an order details page, and an order tracking page, the push module 43 is specifically used to:

[0137] Recommend information about products preferred by the user in at least one of a non-empty shopping cart, an empty shopping cart, an order details page, and an order tracking page.

[0138] In some embodiments, the push module 43 is specifically used for:

[0139] Get product price information, picture information and parameter information;

[0140] The price information, picture information and parameter information of the product are recommended in at least one of a non-empty shopping cart, an empty shopping cart, an order details page and an order tracking page.

[0141] The device provided in the embodiments of the present application can be used to execute the method in the above embodiments. Its implementation principle and technical effects are similar and will not be repeated here.

[0142] It should be noted that it should be understood that the division of the various modules of the above device is only a division of logical functions. In actual implementation, they can be fully or partially integrated into one physical entity, or they can be physically separated. And these modules can all be implemented in the form of software called by processing elements; they can also be all implemented in the form of hardware; some modules can also be implemented in the form of software called by processing elements, and some modules can be implemented in the form of hardware. For example, the estimation module can be a separately established processing element, or it can be integrated in a chip of the above device. In addition, it can also be stored in the memory of the above device in the form of program code, and called and executed by a processing element of the above device. The implementation of other modules is similar. In addition, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each module above can be completed by an integrated logic circuit of hardware in the processor element or instructions in the form of software.

[0143] Figure 5 A schematic diagram of the structure of the processing device provided in the embodiment of the present application, such as Figure 5 As shown, the processing device includes a memory 51 and at least one processor 52;

[0144] The memory 51 stores computer-executable instructions. The processing device 50 further includes a bus 53 , wherein the memory 51 is connected to the processor 52 via the bus 53 .

[0145] In a specific implementation process, at least one processor 52 executes the computer-executable instructions stored in the memory 51, so that at least one processor 52 executes the method as described above.

[0146] Exemplarily, the memory 51 can also be used to store preset short-strategy risk control rules and long-strategy risk control rules, etc. The memory 51 can be cloud storage or local storage.

[0147] Exemplarily, the processing device may be a computer device or a server.

[0148] For example, the bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of the present application is not limited to only one bus or one type of bus.

[0149] The specific implementation process of the processor can be found in the above method embodiment, and its implementation principle and technical effect are similar, so this embodiment will not be repeated here.

[0150] Optionally, the present application also provides a readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the steps of the above method are implemented.

[0151] Optionally, an embodiment of the present application further provides a computer program product, including a computer program / instructions, which implements the steps of the above method when executed by a processor.

[0152] In the present application, "at least one" means one or more, and "plurality" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship; in the formula, the character " / " indicates that the previous and next associated objects are in a "division" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, c can be single or multiple.

[0153] It is to be understood that the various numerical numbers involved in the embodiments of the present application are only for the convenience of description and are not intended to limit the scope of the embodiments of the present application. In the embodiments of the present application, the size of the sequence number of each process does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An information push method, characterized in that: The method comprises: Obtaining a scenario model corresponding to the recommendation scenario currently displayed by the client, wherein the recommendation scenario is used to characterize the product recommendation page of the client, and the scenario model is obtained by training the joint data of multiple recommendation scenarios and scenario tasks, wherein the joint data includes a scenario identifier, a task identifier, and historical behavior characteristics of the user, and the scenario tasks include a click task and an order task; According to the scenario model corresponding to the recommendation scenario, estimating the product preferred by the user in the recommendation scenario; Recommend information about products preferred by the user in the recommendation scenario; The scene model training method comprises: Identifying the recommended scene and scene task according to the scene identifier and the task identifier; According to the preset loss function and the joint data, the shared layer of the preset model is trained to obtain the adaptive weight of the scene task of each recommended scene, wherein the preset model includes the shared layer and the tower layer corresponding to each scene task; According to the adaptive weights, weighted sums are performed on the scenario tasks of each recommended scenario, and the sums are input into the tower layer corresponding to each scenario task respectively; According to the tower layers, the scenario tasks of each recommended scenario are decoupled to obtain a scenario model corresponding to each recommended scenario.

2. The method according to claim 1, characterized in that Before obtaining the scene model corresponding to the recommended scene currently displayed by the client, the method further includes: Get the scenario tasks and joint data of multiple recommendation scenarios submitted by the client.

3. The method according to claim 1, characterized in that The preset loss function is: In the above formula, Indicates whether sample i belongs to the recommended scene S corresponding to scene task t t ,ω t represents the weight of scene task t, is the weighted cross entropy.

4. The method according to claim 3, characterized in that Before training the shared layer according to the preset loss function and the joint data to obtain the adaptive weight of the scene task of each recommended scene, the method further includes: Get the number of positive samples of scene task t and the number of negative samples of scene task t; The preset loss function is weighted according to a preset weight strategy, the number of positive samples and the number of negative samples.

5. The method according to claim 4, characterized in that The preset weight strategy is: In the above formula, represents the number of positive samples of scene task t, represents the number of negative samples of scene task t, is the coefficient of positive samples when calculating weighted cross entropy.

6. The method according to claim 1, characterized in that The shared layer includes an expert network and a gating network. The shared layer is trained according to a preset loss function and joint data to obtain an adaptive weight of a scene task for each recommended scene, including: inputting the joint data into the expert network and the gating network; According to the preset loss function, the expert network and the gating network are trained to obtain adaptive weights for different scenario tasks.

7. The method according to claim 1, characterized in that The estimating the product preferred by the user in the recommendation scenario according to the scenario model corresponding to the recommendation scenario includes: According to the scenario model, obtaining an output of a scenario task of the recommended scenario, wherein the output of the scenario task is used to indicate an importance score of the scenario task; According to a preset fusion strategy, the scenario tasks of the recommendation scenario are fused to obtain the output of the recommendation scenario, where the output of the recommendation scenario is used to indicate the user's preference for different products in the recommendation scenario; According to the output of the recommendation scenario, the product preferred by the user in the recommendation scenario is obtained.

8. The method according to claim 7, characterized in that The preset fusion strategy is: In the above formula, represents the output of the recommended scene s, T s represents all scenario tasks for the recommended scenario s, represents the weight of the output of scene task t, Represents the output of scene task t.

9. The method according to any one of claims 1 to 8, characterized in that: The product recommendation page includes at least one of a non-empty shopping cart, an empty shopping cart, an order details page, and an order tracking page, and the information of recommending the user's preferred product in the recommendation scenario includes: Recommend information about products preferred by the user in at least one of the non-empty shopping cart, the empty shopping cart, the order details page, and the order tracking page.

10. The method according to claim 9, characterized in that The information of recommending the user's preferred product in at least one of the non-empty shopping cart, the empty shopping cart, the order details page, and the order tracking page includes: Obtain price information, image information and parameter information of the product; The price information, picture information and parameter information of the product are recommended in at least one of the non-empty shopping cart, the empty shopping cart, the order details page and the order tracking page.

11. An information push device, characterized in that: include: An acquisition module is used to acquire a scenario model corresponding to a recommendation scenario currently displayed by the client, wherein the recommendation scenario is used to characterize a product recommendation page of the client, and the scenario model is obtained by training the joint data of multiple recommendation scenarios and scenario tasks, wherein the joint data includes a scenario identifier, a task identifier, and historical behavior characteristics of the user, and the scenario tasks include a click task and an order task; An estimation module, used to estimate the product preferred by the user in the recommendation scenario according to the scenario model corresponding to the recommendation scenario; A push module, used to recommend information about products preferred by users in the recommendation scenario; The device also includes: a training module, The training module is specifically used for: Identifying the recommended scene and scene task according to the scene identifier and the task identifier; According to the preset loss function and the joint data, the shared layer of the preset model is trained to obtain the adaptive weight of the scene task of each recommended scene, wherein the preset model includes the shared layer and the tower layer corresponding to each scene task; According to the adaptive weights, weighted sums are performed on the scenario tasks of each recommended scenario, and the sums are input into the tower layer corresponding to each scenario task respectively; According to the tower layers, the scenario tasks of each recommended scenario are decoupled to obtain a scenario model corresponding to each recommended scenario.

12. A processing device, characterized in that: comprising a memory and at least one processor; The memory stores computer-executable instructions; The at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor performs the method according to any one of claims 1 to 10.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer-readable storage medium stores computer instructions, which are used to implement the method according to any one of claims 1 to 10 when executed by a processor.

14. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the method according to any one of claims 1 to 10 is implemented.

Images