Task recommendation method

Abstract

Embodiments of the present specification provide a task recommendation method, which comprises: receiving a task set corresponding to a target user, wherein the task set contains a to-be-processed task; processing the task set based on a first task processing unit in a task recommendation model to obtain a set representation of the task set; processing a user parameter of the target user based on a parameter processing unit in a preference prediction model to obtain a user representation corresponding to the target user; processing the set representation and the user representation based on a second task processing unit in the task recommendation model to obtain a task feature of the to-be-processed task; and inputting the task feature into a task prediction unit in the task recommendation model to obtain a target task corresponding to the target user, thereby accurately recommending a marketing activity meeting the preference of the user to the user.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of computer technology, and in particular to a task recommendation method. Background Technology

[0002] With the continuous development of internet technology, many internet companies use various marketing methods to achieve marketing goals such as user activation, user acquisition, and user retention. In traditional marketing scenarios, marketing activities are recommended to users through marketing models. For example, marketing activities that offer rewards for completing specific tasks.

[0003] However, traditional marketing models do not consider user preferences when recommending marketing campaigns, resulting in a low contribution of recommended campaigns to marketing objectives such as user activation, new user acquisition, and user retention. Therefore, there is an urgent need for a method that can accurately recommend marketing campaigns that match user preferences. Summary of the Invention

[0004] In view of this, embodiments of this specification provide a task recommendation method. One or more embodiments of this specification also relate to a task recommendation model training method, a task recommendation apparatus, a task recommendation model training apparatus, a computing device, a computer-readable storage medium, and a computer program, to address the technical deficiencies existing in the prior art.

[0005] According to a first aspect of the embodiments of this specification, a task recommendation method is provided, comprising:

[0006] Receive a task set corresponding to the target user, wherein the task set contains tasks to be processed;

[0007] The task set is processed by the first task processing unit in the task recommendation model to obtain a set representation of the task set.

[0008] Based on the parameter processing unit in the preference prediction model, the user parameters of the target user are processed to obtain the user representation corresponding to the target user;

[0009] Based on the second task processing unit in the task recommendation model, the set representation and the user representation are processed to obtain the task features of the task to be processed.

[0010] The task features are input into the task prediction unit in the task recommendation model to obtain the target task corresponding to the target user.

[0011] According to a second aspect of the embodiments of this specification, a task recommendation apparatus is provided, comprising:

[0012] The receiving module is configured to receive a task set corresponding to the target user, wherein the task set contains tasks to be processed;

[0013] The first processing module is configured to process the task set based on the first task processing unit in the task recommendation model to obtain a set representation of the task set.

[0014] The second processing module is configured to process the user parameters of the target user based on the parameter processing unit in the preference prediction model, and obtain the user representation corresponding to the target user.

[0015] The third processing module is configured to process the set representation and the user representation based on the second task processing unit in the task recommendation model to obtain the task features of the task to be processed.

[0016] The task acquisition module is configured to input the task features into the task prediction unit in the task recommendation model to obtain the target task corresponding to the target user.

[0017] According to a third aspect of the embodiments of this specification, a method for training a task recommendation model is provided, comprising:

[0018] Determine the training samples and sample labels for the task recommendation model, wherein the training samples are a set of sample tasks corresponding to the sample users, the set of sample tasks contains sample tasks, and the sample labels are the preferred tasks of the sample users;

[0019] The sample task set is processed by the first task processing unit in the task recommendation model to obtain a sample set representation of the sample task set.

[0020] Obtain the sample user representation corresponding to the sample user, wherein the sample user representation is obtained by processing the sample user parameters of the sample user in the preference prediction model.

[0021] Based on the second task processing unit in the task recommendation model, the sample set representation and the sample user representation are processed to obtain the sample task features of the sample task.

[0022] The sample task features are input into the task prediction unit of the task recommendation model to obtain the task recommendation result;

[0023] Based on the task recommendation results and the preferred tasks of the sample users, the task recommendation model is trained until the training stops, thus obtaining the trained task recommendation model.

[0024] According to a fourth aspect of the embodiments of this specification, a task recommendation model training apparatus is provided, comprising:

[0025] The sample determination module is configured to determine training samples and sample labels for the task recommendation model, wherein the training samples are a set of sample tasks corresponding to the sample user, the set of sample tasks contains sample tasks, and the sample labels are the preferred tasks of the sample user.

[0026] The first processing module is configured to process the sample task set based on the first task processing unit in the task recommendation model to obtain a sample set representation of the sample task set.

[0027] The second processing module is configured to obtain the sample user representation corresponding to the sample user, wherein the sample user representation is obtained by the parameter processing unit in the preference prediction model by processing the sample user parameters of the sample user.

[0028] The second processing module is configured to process the sample set representation and the sample user representation based on the second task processing unit in the task recommendation model to obtain the sample task features of the sample task.

[0029] The result acquisition module is configured to input the sample task features into the task prediction unit in the task recommendation model to obtain the task recommendation result.

[0030] The model training module is configured to train the task recommendation model based on the task recommendation results and the sample users' preferred tasks until the training stops, thereby obtaining the trained task recommendation model.

[0031] According to a fifth aspect of the embodiments of this specification, a computing device is provided, comprising:

[0032] Memory and processor;

[0033] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the task recommendation method and the task recommendation model training method.

[0034] According to a sixth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the task recommendation method and the task recommendation model training method.

[0035] According to a seventh aspect of the embodiments of this specification, a computer program is provided, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the task recommendation method and the task recommendation model training method.

[0036] The task recommendation method provided in this specification includes: receiving a task set corresponding to a target user, wherein the task set contains tasks to be processed; processing the task set based on a first task processing unit in a task recommendation model to obtain a set representation of the task set; processing the user parameters of the target user based on a parameter processing unit in a preference prediction model to obtain a user representation corresponding to the target user; processing the set representation and the user representation based on a second task processing unit in the task recommendation model to obtain task features of the tasks to be processed; and inputting the task features into a task prediction unit in the task recommendation model to obtain the target task corresponding to the target user.

[0037] Specifically, in the process of recommending target tasks to target users through the task recommendation model, this method obtains the user representation corresponding to the target user based on the parameter processing unit in the preference prediction model, and inputs the user representation and the task representation into the task prediction unit in the task recommendation model. The task prediction unit then predicts the target task corresponding to the target user, thereby accurately recommending tasks that match the user's preferences to the user. Attached Figure Description

[0038] Figure 1 This is a schematic diagram illustrating an application scenario of a task recommendation method provided in one embodiment of this specification;

[0039] Figure 2 This is a flowchart of a task recommendation method provided in one embodiment of this specification;

[0040] Figure 3 This is a flowchart illustrating a task recommendation model training method provided in one embodiment of this specification;

[0041] Figure 4A This is a flowchart illustrating the training process of a user preference prediction model in a task recommendation method provided in one embodiment of this specification.

[0042] Figure 4B This is a flowchart illustrating the training process of a task recommendation model in a task recommendation method provided in one embodiment of this specification.

[0043] Figure 5 This is a schematic diagram of the model structure in a task recommendation method provided in one embodiment of this specification;

[0044] Figure 6This is a schematic diagram of the distribution of user-perceived representations in a task recommendation method provided in one embodiment of this specification;

[0045] Figure 7 This is a schematic diagram of the distribution of scene perception representation in a task recommendation method provided in one embodiment of this specification;

[0046] Figure 8 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0047] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0048] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0049] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0050] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0051] Representation Learning: In the field of machine learning, representation learning is a method of transforming raw data into a form that can be effectively developed by machine learning. It can remove invalid or redundant information from the raw data and extract the effective information to form features. It enables machines not only to learn the features of the data, but also to use these features to complete a specific task.

[0052] Multi-task Learning (MTL): MTL is an important component of machine learning and a type of transfer learning. As a branch of transfer learning, MTL has its own unique learning background and applications. Given input and output data for several related tasks, MTL can leverage the relationships between tasks and learn multiple models simultaneously.

[0053] Marketing mindset: A marketing technique that involves studying, understanding, and mastering the psychological patterns of consumers, leveraging their mature minds and transforming them into purchasing motives to achieve the goal of marketing products.

[0054] CTR (Click-Through-Rate): This is a commonly used term in internet advertising, referring to the click-through rate of online advertisements (image ads / text ads / keyword ads / ranking ads / video ads, etc.), that is, the actual number of clicks on the ad.

[0055] CVR (Conversion Rate): The conversion rate is the rate at which a user clicks on an ad and becomes a valid activated, registered, or even paying user.

[0056] "Turn to 100" or "Raise to 100": This means increasing a user's holdings on the financial service platform from less than 100 yuan (including 0 yuan) to 100 yuan or more.

[0057] User acquisition: attracting new users, which means recommending tasks to users so that they can click, convert, and invest.

[0058] PLE: Multi-task learning model.

[0059] DNN (Deep Neural Networks): Deep neural networks.

[0060] MLP (Multilayer Perceptron): A multilayer perceptron is a type of feedforward artificial neural network model.

[0061] With the continuous development of internet technology, many internet companies use different marketing methods to achieve marketing goals such as activation, user acquisition, and user retention. For example, some internet finance companies provide users with a one-stop wealth management service platform, where users can purchase different types of wealth management products.

[0062] However, within the marketing ecosystem of this wealth management service platform, the marketing mindset plays a unique role in platform activation, traffic exchange, and new user / entry point acquisition for large-scale wealth management. Users can earn rewards by completing tasks recommended by the corresponding task system within specific marketing scenarios (such as purchasing funds or experiencing the platform's investment services), and these rewards can be exchanged for certain wealth management assets. It's worth noting that this one-stop wealth management service platform offers users a variety of services, such as lending, insurance, consumer loans, and wealth management products.

[0063] Furthermore, compared to other scenarios (reward-oriented and decision-making-oriented scenarios), the value of marketing-oriented scenarios for various businesses is primarily influenced by traffic and its distribution efficiency. Currently, marketing-oriented scenarios mainly include cash rewards (where users can receive cash upon completing tasks), discount days (offering discounts on products on specific dates), and financial trial funds, etc. The task recommendation modules within each scenario cover a variety of business domains, including the aforementioned lending services, insurance services, consumer loan services, and financial services, and the task formats displayed for these various products are rich and diverse.

[0064] For example, taking wealth management services as an example, there are tasks such as experiencing fund fixed investment, using red envelopes, experiencing wealth management trial funds, and experiencing high-end wealth management, as well as more complex tasks such as word-of-mouth marketing. These tasks cover the user's life cycle stages in different business domains, thereby promoting the achievement of business goals such as user acquisition, user activation, and user retention.

[0065] However, for one-stop wealth management service platforms that can provide users with a variety of wealth management services, the marketing scenarios for these services differ from traditional natural recommendation scenarios. The recommendation models in these marketing scenarios have their own unique challenges, specifically in the following three aspects:

[0066] (1) Rich user status.

[0067] In marketing scenarios covering a one-stop wealth management service platform, users throughout their entire lifecycle are represented. Some business metrics are strongly correlated with user segments, and the data distribution varies significantly across different user segments, with different business objectives. For example, in wealth marketing, the business objective for new customers is to achieve click-conversion-investment (i.e., new customer acquisition), while the business objective for users with low holdings is to increase their holdings to over 100 yuan (i.e., attracting 100 yuan).

[0068] (2) The products are heterogeneous and multi-sourced.

[0069] Marketing scenarios cover a wide range of business domains and diverse product forms. Tasks in heterogeneous business scenarios naturally exhibit differences in completion rates, making it challenging to use a unified model to uniformly sort these heterogeneous tasks. For example, the task recommendation module in a wealth marketing scenario covers various products including lending services, insurance services, consumer loan services, and wealth management services. These tasks cover different stages of the user lifecycle within different business domains, driving the achievement of different business goals such as new user acquisition, user activation, and user retention.

[0070] (3) Multiple objectives.

[0071] The operational goals of multiple business segments are strongly dependent on each other, with some dependencies between the goals and the customer base and business operations. For example, in a wealth marketing scenario, in addition to optimizing the CTR and CVR of each business domain task, it is also necessary to optimize the new user acquisition rate and the conversion rate of platform users. These metrics have some dependencies: Click--(fully dependent)--> Completion--(partially dependent)--> Conversion / New User Acquisition--(partially dependent)--> Conversion to 100% / Conversion to 100%

[0072] To address the aforementioned issues, this specification provides a multi-model solution based on heterogeneous business segmentation. This solution divides the business domains covered by the tasks deployed in business operations into three main categories: CTR model, CVR model, and GMV model. Each model is then used separately to model the relevance of different business domains to operational objectives. However, this solution has drawbacks. It uses a unified model at the user level, failing to capture user preferences for heterogeneous services across different timeframes. Furthermore, the segmentation based on heterogeneous businesses prevents the sharing of underlying task data and the identification of commonalities between different services, leading to biased modeling and hindering the improvement of overall traffic distribution efficiency.

[0073] Based on this, a task recommendation method is provided in this specification. One or more embodiments of this specification also relate to a task recommendation model training method, a task recommendation apparatus, a task recommendation model training apparatus, a computing device, a computer-readable storage medium, and a computer program, which will be described in detail in the following embodiments.

[0074] Figure 1 This diagram illustrates an application scenario of a task recommendation method according to an embodiment of this specification. (See attached diagram.) Figure 1 As can be seen, the task recommendation method provided in this specification can receive a task set corresponding to a target user, which contains tasks to be processed; then, the task set is input into the first task processing unit in the task recommendation model, and the task set is processed based on the first task processing unit in the task recommendation model to obtain a set representation of the task set.

[0075] Next, based on the parameter processing unit in the preference prediction model, the user parameters of the target user are processed to obtain the user representation corresponding to the target user; secondly, based on the second task processing unit in the task recommendation model, the set representation and user representation are processed to obtain the task features of the task to be processed; finally, the task features are input into the task prediction unit in the task recommendation model to obtain the target task corresponding to the target user.

[0076] Based on this, the task recommendation method provided in this specification proposes a layered awareness multi-task recommendation scheme for layered users and heterogeneous businesses. Through the optimization of the model structure, a single model can cover users throughout the entire lifecycle of wealth management, while being compatible with recommendations from different business domains. This ensures that the model can simultaneously optimize the aforementioned multiple operational objectives, thereby improving the traffic utilization rate in the entire marketing scenario.

[0077] Figure 2 A flowchart of a task recommendation method according to an embodiment of this specification is shown, specifically including the following steps. It should be noted that the task recommendation method provided in this specification can be applied to a multi-task recommendation model that includes a preference prediction model and a task recommendation model. This multi-task recommendation model enables hierarchical awareness of multi-task recommendations for tiered users and heterogeneous services.

[0078] Step 202: Receive the task set corresponding to the target user, wherein the task set contains tasks to be processed.

[0079] In this context, the target user can be understood as the user who needs task recommendations; the task set can be understood as a collection of multiple tasks with the same characteristics. In practical applications, each business scenario has multiple tasks; therefore, a business scenario can be a task set. It should be noted that the task recommendation method provided in this specification can be applied to different scenarios; and in different scenarios, the task set and the tasks to be processed are different. For example, when this task recommendation method is applied to a financial management scenario, the task set can be the lending business in a financial management platform, which has tasks such as new user acquisition (e.g., cash back for new user registration) and user activation (e.g., cash back for daily check-in). As another example, when this task recommendation method is applied to a shopping scenario, the task set can be the shopping business in a shopping platform, which has tasks such as new user acquisition (e.g., giving product coupons to new users upon registration) and promotional tasks (e.g., giving product coupons for completing check-ins), etc.

[0080] This task can be understood as a marketing tactic that enables users to perform specific actions, such as new user acquisition tasks, promotional tasks, etc.

[0081] Step 204: Process the task set based on the first task processing unit in the task recommendation model to obtain a set representation of the task set.

[0082] The task recommendation model can be understood as a model capable of recommending corresponding tasks to users. For example, the task recommendation model can be composed of DNN, MLP, or PLE. The first task processing unit can be understood as a neural network model capable of processing the task set to obtain the corresponding representation; for example, the first task processing unit can be PLE.

[0083] Specifically, after receiving the task set, the first task processing unit in the task recommendation model is used to extract the representation of the task set, thereby obtaining the set representation of the task set.

[0084] Specifically, the first task processing unit in the task recommendation model processes the task set to obtain a set representation of the task set, including:

[0085] Feature extraction is performed on the tasks to be processed contained in the task set to obtain the task features of the tasks to be processed;

[0086] The task features are input into the first task processing unit in the task recommendation model to obtain the task representation of the task to be processed.

[0087] The set representation of the task set is determined based on the task representation.

[0088] This set of representations can be understood as scene-aware representations that represent the different features of different scenes.

[0089] Specifically, after obtaining the task set, a multi-task recommendation model is used to extract features from the tasks to be processed within the task set using any feature extraction method. This yields the task features corresponding to the tasks to be processed; this specification does not impose specific limitations on this. The task features are then input into the first task processing unit of the task recommendation model for processing to obtain the task features corresponding to each task to be processed. Based on these task features, a set feature of the task set is constructed. In other words, the first task processing unit can be a task feature processing model.

[0090] Continuing with the previous example, this multi-task recommendation model receives the business scenarios involved by the user. Each business scenario contains multiple tasks. Then, it extracts features from each task using any feature extraction method to obtain task features. The PLE module then models these task features to obtain a task representation for each task. Based on this task representation, a business scenario representation is constructed. This facilitates subsequent task recommendations for the user based on the business scenario representation.

[0091] It should be noted that the task features can include various types, including but not limited to sparse task features, dense task features, and multimodal task features. Subsequently, the first task processing unit (i.e., the PLE module) processes the various types of task features to obtain the task representation of the task to be processed.

[0092] In one embodiment provided in this specification, determining the set representation of the task set based on the task representation includes:

[0093] The initial set representation of the task set is determined based on the task representation;

[0094] The initial set representation is encoded and calculated based on the first encoding module to obtain the set weight parameters corresponding to the initial set representation;

[0095] Based on the set weight parameters and the initial set representation, the target set representation of the task set is obtained.

[0096] The determination of the initial set representation of the task set based on the task representation can be understood as constructing the initial set representation of the task set based on the task representation.

[0097] The first encoding module can be understood as the encoding module that calculates the corresponding weights for the scenario representation. For example, this first encoding module can be a scenario-aware encoding module. This module can determine the corresponding weights for the business scenario representation through different encoding methods. For instance, hard-segmentation encoding, i.e., using one-hot encoding, selects only one of the representations from a set each time as the final representation, thereby making the business meaning of different business scenarios clearer. Alternatively, soft-segmentation encoding can be used, i.e., allocating weights based on prior business knowledge, or the model learns a vector 'a' and assigns it to multiple sets of representations, finally weighting them to obtain the final representation. Its advantage is that it can weaken the segmentation boundary, and the model can learn the parameters itself. It should be noted that the set of weight parameters can be a vector.

[0098] The process of obtaining the target set representation of the task set based on the set weight parameters and the initial set representation can be understood as multiplying the set weight parameters and the initial set representation to obtain scenario-aware representations for different business scenarios.

[0099] Using the previous example, this multi-task recommendation model calculates the weight of each scene representation through the encoding module, and then multiplies the scene representation V with the weight a to obtain the final scene-aware representation.

[0100] Step 206: Based on the parameter processing unit in the preference prediction model, process the user parameters of the target user to obtain the user representation corresponding to the target user.

[0101] The preference prediction model can be understood as a model that can analyze user preferences. For example, the preference prediction model can be composed of DNN, MLP or PLE.

[0102] This parameter processing unit can be understood as being able to perform characterization processing on the user parameters to obtain a user-represented model. For example, this parameter processing unit can be a PLE.

[0103] This user representation can be understood as representing the characteristics of users at different levels; for example, this user representation can be a user-aware representation.

[0104] Specifically, the parameter processing unit in the preference prediction model processes the user parameters of the target user to obtain the user representation corresponding to the target user, including:

[0105] Receive user parameters of the target user, and perform feature extraction on the user parameters to obtain user features corresponding to the user parameters;

[0106] The user characteristics are input into the parameter processing unit of the preference prediction model to obtain the user representation corresponding to the target user.

[0107] The user parameters include, but are not limited to, user attribute information and user behavior data. User attribute information includes, but is not limited to, user name and user location. User behavior data can include tasks clicked by the user, tasks experienced by the user, task advertisements viewed by the user, etc.

[0108] Specifically, the preference prediction model is used to perform feature extraction on the received user parameters of the target user, thereby obtaining the user features corresponding to the user parameters. It should be noted that in the process of obtaining user features, any feature extraction method can be used to extract features from the user parameters, and this specification does not impose any specific restrictions on this.

[0109] Then, the task features are input into the parameter processing unit of the preference prediction model, and the parameter processing unit processes the user features to obtain the user representation corresponding to the target user. In other words, the parameter processing unit can be a user feature processing model.

[0110] It should be noted that the user features can include various types, including but not limited to user sparse features and user dense features. The subsequent parameter processing unit (i.e., the PLE module) processes these various types of user features to obtain the user representation corresponding to the target user.

[0111] This specification provides an embodiment in which the step of inputting the user features into the parameter processing unit of the preference prediction model to obtain the user representation corresponding to the target user includes:

[0112] The user features are input into the parameter processing unit of the preference prediction model to obtain the initial user representation corresponding to the target user;

[0113] The initial user representation is encoded and calculated based on the second encoding module to obtain the user weight parameters corresponding to the initial user representation;

[0114] Based on the initial user representation and the user weight parameters, the target user representation corresponding to the target user is determined.

[0115] The second encoding module can be understood as an encoding module that calculates the corresponding weights for the user representation. For example, this second encoding module can be a user-aware encoding module. It should be noted that this user encoding module can determine the corresponding weights for the user representation using different encoding methods. For example, hard-segment encoding, i.e., using one-hot encoding, selects only one of the representations from a set each time as the final representation, thus making the meaning of the user features clearer. Alternatively, soft-segment encoding can be used, i.e., allocating weights based on prior business knowledge, or the model learns a vector 'a' and assigns it to multiple sets of representations, finally weighting them to obtain the final representation. Its advantage is that it weakens the segmentation boundary, and the model can learn the parameters itself. It should be noted that the user weight parameter can be a vector.

[0116] The process of determining the target user representation based on the initial user representation and user weight parameters can be understood as multiplying the initial user representation and user weight parameters to obtain the user-aware representation.

[0117] Using the previous example, this multi-task recommendation model calculates the weight of each user representation through the encoding module, and then multiplies the weight by the user representation to obtain the final user perception representation.

[0118] Furthermore, after the parameter processing unit in the preference prediction model processes the user parameters of the target user to obtain the user representation corresponding to the target user, it also includes:

[0119] The user representation is input into the preference prediction unit in the preference prediction model to obtain user preference parameters.

[0120] In this context, the preference prediction unit can be understood as the model in the preference prediction model that can perform user preference prediction. That is to say, the preference prediction unit can be a preference prediction sub-model in the preference prediction model that performs user preference prediction. The preference prediction model can be a DNN or an MLP.

[0121] This user preference parameter can be understood as a parameter that characterizes user preferences, such as the user's preference level.

[0122] Using the previous example, by inputting the user's representation into the MLP module of the preference prediction model, we can obtain the user's preference for the financial service platform.

[0123] Step 208: Based on the second task processing unit in the task recommendation model, process the set representation and the user representation to obtain the task features of the task to be processed.

[0124] The second task processing unit can be understood as a model capable of processing the set representation and the user representation; for example, the second task processing unit can be a PLE.

[0125] Specifically, the second task processing unit in the task recommendation model processes the set representation and the user representation to obtain the task features of the task to be processed, including:

[0126] Determine the task representation corresponding to the task to be processed from the set representation;

[0127] Based on the task representation and the user representation, the second task processing unit is input to obtain the task features corresponding to the task to be processed.

[0128] Following the previous example, the user perception representation shared by the user preference prediction model is concatenated with the scene perception representation, and the concatenated representation is input into the PLE module. By using the PLE module to model the concatenated representation, feature representations for multiple tasks are obtained.

[0129] Step 210: Input the task features into the task prediction unit in the task recommendation model to obtain the target task corresponding to the target user.

[0130] Specifically, the step of inputting the task features into the task prediction unit of the task recommendation model to obtain the target task corresponding to the target user includes:

[0131] The task features are input into the first task prediction unit of the task recommendation model to obtain a first candidate task; and / or

[0132] The task features are input into the second task prediction unit in the task recommendation model to obtain the second candidate task;

[0133] The first candidate task and / or the second candidate task are taken as the target task.

[0134] The first task prediction unit can be understood as a general task recommendation model in the task recommendation model; for example, it can be a DNN or MLP for general task recommendation. The second task prediction unit can be understood as a special task recommendation model in the task recommendation model; for example, it can be a DNN or MLP for special task recommendation. The first candidate task is a general task, and the second candidate task is a special task preferred by the user.

[0135] Furthermore, the step of inputting the task features into the first task prediction unit in the task recommendation model to obtain the first candidate task includes:

[0136] The task features are input into the first task prediction unit in the task recommendation model to obtain the task prediction result;

[0137] Based on the task prediction results, a first candidate task is determined for the target user from the tasks to be processed.

[0138] The task prediction result can be the task identifier of the task that needs to be recommended to the user. Subsequently, the corresponding task can be recommended to the user based on the task identifier. The task identifier can be understood as information that uniquely identifies a task, such as task name, task ID, task sequence number, etc.

[0139] Following the previous example, the task feature representation of each task is input into the first task prediction unit to obtain the task name of the general task that needs to be recommended to the user (such as the experience of financial management experience fund task), and the task named experience of financial management experience fund task is recommended to the user.

[0140] Furthermore, the step of inputting the task features into the second task prediction unit in the task recommendation model to obtain the second candidate task includes:

[0141] User preference parameters are determined based on the preference prediction model.

[0142] The user preference parameters and the task features are input into the second task prediction unit in the task recommendation model to obtain the task prediction result.

[0143] Based on the task prediction results, a second candidate task is determined for the target user from the tasks to be processed.

[0144] Continuing with the previous example, a special task recommendation model combines task features and user preferences to obtain the user's preferred tasks. Specifically, this task recommendation model obtains the user preference features input from the user preference model, concatenates these user preference features with the task features, and inputs the concatenated features into the special task recommendation model within the task recommendation model. The special task recommendation model processes the concatenated features to obtain the identifier of the feature task that matches the user's preferences, for example, the identifier can be the task name (such as a "red envelope task"). Then, the task with the name "red envelope task" is recommended to the user.

[0145] It should be noted that the training steps of the preference prediction model include:

[0146] Determine the training samples and sample labels for the preference prediction model, wherein the training samples are the sample user parameters of the sample users, and the sample labels are the sample user preference parameters of the sample users;

[0147] The sample user type of the sample user is determined based on the sample user parameters;

[0148] From the parameter processing units of the preference prediction model, determine the target parameter processing unit corresponding to the sample user type;

[0149] The sample user parameters are input into the target parameter processing unit to obtain the sample user representation corresponding to the sample user.

[0150] The sample user representation is input into the preference prediction unit in the preference prediction model to obtain the preference prediction result;

[0151] Based on the preference prediction results and the sample user preference parameters, the preference prediction model is trained until the training stops, thus obtaining the trained preference prediction model.

[0152] Here, the sample user parameters can be understood as the user parameters used as samples. It should be noted that these sample user parameters are different from the user parameters mentioned above.

[0153] The training stopping condition can be set according to the actual application scenario. For example, the training stopping condition can be that the preference prediction model has reached convergence, or that the preference prediction model has completed a preset number of training rounds, etc. This specification does not limit this. Furthermore, it should be noted that the processing steps of the preference prediction model on the training samples can be found in the above embodiments regarding the processing of user parameters by the preference prediction model, which will not be repeated in this specification.

[0154] Furthermore, the steps for training the preference prediction model based on the preference prediction results and the sample user preference parameters can refer to any method of training a model based on prediction results and sample labels; this manual does not impose any specific limitations on this.

[0155] In one embodiment provided in this specification, the training steps of the task recommendation model include:

[0156] Determine the training samples and sample labels for the task recommendation model, wherein the training samples are a set of sample tasks corresponding to the sample users, the set of sample tasks contains sample tasks, and the sample labels are the preferred tasks of the sample users;

[0157] The sample task set is processed by the first task processing unit in the task recommendation model to obtain a sample set representation of the sample task set.

[0158] Obtain the sample user representation corresponding to the sample user, wherein the sample user representation is obtained by processing the sample user parameters of the sample user in the preference prediction model.

[0159] Based on the second task processing unit in the task recommendation model, the sample set representation and the sample user representation are processed to obtain the sample task features of the sample task.

[0160] The sample task features are input into the task prediction unit of the task recommendation model to obtain the task recommendation result;

[0161] Based on the task recommendation results and the preferred tasks of the sample users, the task recommendation model is trained until the training stops, thus obtaining the trained task recommendation model.

[0162] Here, the sample task set can be understood as the set of tasks that serve as samples.

[0163] The training stopping condition can be set according to the actual application scenario. For example, the training stopping condition can be that the task recommendation model has reached convergence, or that the task recommendation model has completed a preset number of training rounds, etc. This specification does not limit this. Furthermore, it should be noted that the processing steps of the task recommendation model for the training samples can be found in the above embodiments regarding the processing of the task set by the task recommendation model, and this specification will not elaborate on this.

[0164] Furthermore, the steps for training the task recommendation model based on the task recommendation results and the preferences of sample users can refer to any method of training the model based on prediction results and sample labels; this manual does not impose any specific limitations on this.

[0165] The task recommendation method provided in this specification, in the process of recommending target tasks to target users through the task recommendation model, obtains the user representation corresponding to the target user based on the parameter processing unit in the preference prediction model, and inputs the user representation and the task representation into the task prediction unit in the task recommendation model. The task prediction unit then predicts the target task corresponding to the target user, thereby accurately recommending marketing activities that match the user's preferences.

[0166] Figure 3 A flowchart of a task recommendation model training method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0167] Step 302: Determine the training samples and sample labels for the task recommendation model, wherein the training samples are the sample task set corresponding to the sample user, the sample task set contains sample tasks, and the sample labels are the preferred tasks of the sample user.

[0168] Step 304: Process the sample task set based on the first task processing unit in the task recommendation model to obtain the sample set representation of the sample task set.

[0169] Step 306: Obtain the sample user representation corresponding to the sample user, wherein the sample user representation is obtained by processing the sample user parameters of the sample user in the preference prediction model.

[0170] Step 308: Based on the second task processing unit in the task recommendation model, process the sample set representation and the sample user representation to obtain the sample task features of the sample task.

[0171] Step 310: Input the sample task features into the task prediction unit in the task recommendation model to obtain the task recommendation result.

[0172] Step 312: Based on the task recommendation results and the sample users' preferred tasks, train the task recommendation model until the training stops, and obtain the trained task recommendation model.

[0173] For a detailed explanation of the above steps, please refer to the corresponding or relevant content in the embodiments of the above task recommendation method, which will not be elaborated further here.

[0174] In the embodiments provided in this specification, the training steps of the preference prediction model include:

[0175] Determine the training samples and sample labels for the preference prediction model, wherein the training samples are the sample user parameters of the sample users, and the sample labels are the sample user preference parameters of the sample users;

[0176] The sample user type of the sample user is determined based on the sample user parameters;

[0177] From the parameter processing units of the preference prediction model, determine the target parameter processing unit corresponding to the sample user type;

[0178] The sample user parameters are input into the target parameter processing unit to obtain the sample user representation corresponding to the sample user.

[0179] The sample user representation is input into the preference prediction unit in the preference prediction model to obtain the preference prediction result;

[0180] Based on the preference prediction results and the sample user preference parameters, the preference prediction model is trained until the training stops, thus obtaining the trained preference prediction model.

[0181] For a detailed explanation of the above steps, please refer to the corresponding or relevant content in the embodiments of the above task recommendation method, which will not be elaborated further here.

[0182] The task prediction model training method provided in this manual can refer to the sample user representation determined by the preference prediction model during the training process of the task recommendation model, thereby obtaining a task recommendation model that can accurately recommend marketing activities that match users' preferences. This facilitates the subsequent accurate recommendation of marketing activities that match users' preferences based on this task recommendation model.

[0183] The following is in conjunction with the appendix Figure 4A Appendix Figure 4B Taking the task recommendation method provided in this specification as an example in a marketing scenario, the task recommendation method will be further explained. Figure 4A This document illustrates a flowchart of the training process for a user preference prediction model in a task recommendation method according to an embodiment of this specification. Figure 4B This document illustrates a flowchart of the training process for a task recommendation model in an embodiment of this specification. It should be noted that the task recommendation method described herein provides a layered, multi-task-oriented recommendation solution for tiered users and heterogeneous businesses in a marketing scenario. This solution provides a neural network model, the overall design of which is as follows: Figure 5 As shown. Based on Figure 5 It can be seen that the neural network model is based on a user bias learning model and a task recommendation model. The task recommendation model includes a common task learning sub-model and a user-specific task learning sub-model.

[0184] Among them, user bias learning is mainly responsible for learning tasks that are only related to users. Its purpose is to obtain user preferences (user bias). In practical applications, when the task recommendation method is applied to a one-stop financial service platform, this module (i.e., user bias learning) is mainly used to learn the degree of user preference for the one-stop financial service platform.

[0185] The user-specific task learning is mainly responsible for tasks related to user preferences (user bias) and recommending tasks that match user preferences to users, such as new user acquisition and conversion tasks. Moreover, this part of the tasks is only strongly correlated with the business domains of some financial service platforms, and tasks in other business domains will not trigger the conversion of this operational goal.

[0186] Among them, common task learning is responsible for predicting CTR and CVR for common tasks, thereby recommending common tasks to users, namely clicks and completions.

[0187] Furthermore, it should be noted that the neural network model provided in this manual employs a general joint training framework, see [link to documentation]. Figure 5 It can be seen that, from a horizontal perspective, this neural network model can be divided into: bottom embedding sharing, middle-layer sharing, and high-level sharing.

[0188] Bottom embedding sharing mainly refers to the sharing of underlying model parameters. There is no longer a distinction between customer groups or business scenarios. All sub-models share the same set of underlying embedding parameters.

[0189] Middle-layer sharing refers to the sharing of middle-layer representations after the model maps low-level features to middle-layer representations.

[0190] It should be noted that the intermediate representation includes user-aware representation on the user side (user-specific task learning) and scenario-aware representation on the item side (common task learning, user-specific task learning).

[0191] High-level sharing is mainly based on the partial dependencies between prediction tasks, introducing the user's high-level representation (user bias) into the user-specific task to share the high-level representation.

[0192] Therefore, the task processing method provided in this specification is based on Figure 4A , Figure 4B This section explains the training and application of neural network models. See also... Figure 4A , Figure 4A The training process of a user bias learning model is shown, which includes the following steps.

[0193] Step 402: Use user parameters as training samples and user preferences as sample labels.

[0194] The user parameters include, but are not limited to, user attribute information and user behavior data. User attribute information includes, but is not limited to, user name and user location. User behavior data can include tasks clicked by the user, tasks experienced by the user, task advertisements viewed by the user, etc.

[0195] Step 404: Extract features from user parameters to obtain user features.

[0196] Specifically, the task recommendation method provided in this specification can use any feature extraction method to extract features from user parameters and obtain the features corresponding to those user parameters. Based on this, by extracting features from the user parameters, the user features corresponding to those user parameters are obtained.

[0197] The user features include dense user features and sparse user features. It's important to note that the sparse user feature can be an unstructured feature, such as a user's location being Shanghai. To ensure model performance, it is trained using structured vectors. Therefore, when it's necessary to transform Shanghai into a usable structured vector for the model, a mapping relationship is used to transform the unstructured vector into a denser (i.e., structured) vector.

[0198] Based on the above method, user features can be obtained for training the user preference prediction model.

[0199] Step 406: Input user features into the PLE module to obtain user representation.

[0200] Specifically, the dense user features and the dense user features obtained through mapping relationship transformation are concatenated, and the concatenated user features are input into the PLE module in the preference prediction model.

[0201] The PLE module in this preference prediction model learns multiple different representations V based on user characteristics. These representations correspond to different user strata, meaning users at different lifecycle stages. Figure 5 The U in the text corresponds to the number of user tiers.

[0202] Step 408: Calculate the weights of different user representations through the encoding module.

[0203] Specifically, after determining the user representations of different layers, the weight 'a' of different representations is obtained through the User-aware Encoding module. There are multiple encoding methods to choose from; for example, hard-segment encoding, which uses one-hot encoding, selects only one of the representations in a set as the final representation each time. The advantage is that the business meaning is relatively clear.

[0204] Alternatively, soft-segmentation encoding can be used, where quotas are allocated based on prior business knowledge, or the model learns a vector 'a' and assigns it to multiple representations, finally weighting them to obtain the final representation. The advantages are: it weakens the segmentation boundaries, and the model can learn the parameters itself. It should be noted that this user encoding module calculates corresponding weights for user representations based on the user's investment information on the financial platform. This investment information includes, but is not limited to: the platform's historical new user acquisition, platform liquidation, platform holdings of 0-100, platform holdings of 100-1K, platform holdings of 1K-1W, and platform holdings of 1W+.

[0205] Step 410: Multiply the representation V and the weight a to obtain the user-aware representation.

[0206] In practical applications, to improve the recommendation performance of the task recommendation model, the user-perceived representation can be shared with the task recommendation model to enhance its performance.

[0207] Step 412: Input the user-perceived representation into the DNN to obtain user bias.

[0208] In practical applications, to avoid the problem that task recommendation models cannot perceive the preferences of users at different time periods for heterogeneous services, these user preferences can be shared with the task recommendation model. This allows the task recommendation model to make task recommendations based on the preferences of users at different time periods for heterogeneous services.

[0209] Step 414: Based on the user preference and sample labels, train the user preference prediction model until the running stop condition is met, and obtain the trained user preference prediction model.

[0210] It should be noted that, in practical applications, the steps for processing user parameters of the trained user preference prediction model can be found in the steps described above. The steps for processing the training samples will not be elaborated upon in this manual.

[0211] Furthermore, the task processing methods provided in this specification... Figure 4B The training process of the task recommendation model is shown, which includes the following steps.

[0212] Step 416: Determine the training samples and sample labels for training the task recommendation model.

[0213] The training samples can be tasks that users have used in the past in different scenarios. These different scenarios can be understood as different financial services. The sample labels can be tasks that users prefer.

[0214] Step 418: Extract features from the training samples to obtain task features.

[0215] The task features include sparse features, dense features, and multimodal features.

[0216] The multimodal features of this task can be understood as converting the textual information in the task into an NLP (word vector).

[0217] Step 420: Input the task features into the PLE module in the task recommendation model to obtain the scene representation.

[0218] Specifically, the PLE module learns multiple different representations V based on user characteristics. These representations correspond to different scenarios, among which... Figure 5 The S in the text corresponds to the number of business scenarios for the task product.

[0219] Step 422: Calculate the weights of different scene representations through the encoding module.

[0220] The encoding module can be a scenario-aware encoding module.

[0221] Step 424: Multiply the scene representation V and the weight a to obtain the scene-aware representation.

[0222] Step 426: Concatenate the user-perceived representation shared by the user preference prediction model with the scene-perceived representation, and input the concatenated representation into the PLE module to obtain the task feature representation.

[0223] Specifically, by using the PLE module to model the concatenated representation, feature representations of multiple tasks are obtained, where T corresponds to the number of tasks.

[0224] Step 428: Concatenate the task feature representation with the user preferences shared by the user preference prediction model to obtain a task feature representation that includes user preferences.

[0225] Step 430: Construct a corresponding DNN tower for each task feature representation, predict the target, and obtain the prediction result.

[0226] This DNN tower can be understood as an MLP (Multilayer Perceptron).

[0227] It should be noted that, due to some dependencies between tasks, MLP is divided into user-specific and common task.

[0228] The "common task" refers to some general-purpose tasks, such as CTR and CVR, with CTR primarily serving as an auxiliary task to assist in the learning of CVR. Each of these general-purpose tasks builds a simple, independent DNN tower, entirely based on a feature representation from the PLE module.

[0229] User-specific tasks are relatively deeper-level tasks that require user preference features for assistance and are strongly related to certain business scenarios, such as conversion and transfer tasks in wealth management. These tasks depend not only on the output of the PLE module but also on the output of user bias.

[0230] Step 432: Based on the prediction result and sample labels, train the task recommendation model until the training stopping condition is met, and obtain the trained task recommendation model.

[0231] It should be noted that, in practical applications, the processing steps for tasks that users have used in the past and in different scenarios, after the task recommendation model has been trained, can be found in the steps described above. The processing steps for the training sample will not be repeated in this manual.

[0232] Based on this, the training process for the user preference prediction model and the task recommendation model has been demonstrated through the above steps. It should be noted that the User bias Learning module is primarily used for learning user preferences for specific tasks. In the context of financial marketing, this module is mainly used to learn the degree of user preference for financial services.

[0233] Furthermore, considering the dependencies between certain tasks, such as how users' preferences for wealth management services affect the new user acquisition and conversion rates, the model needs to incorporate high-level features of user preferences into the learning modules for conversion and conversion rates. This will help learn deep funnel tasks like conversion and conversion rates.

[0234] This User-specific and Common Task Learning module is mainly used for learning general / specific business tasks, learning the commonalities between different tasks and the unique characteristics of different scenarios.

[0235] Based on this, after completing the above training, a neural network model capable of hierarchical awareness multi-task recommendation for hierarchical users and heterogeneous businesses can be obtained. Based on this neural network model, while predicting user preferences, it can recommend tasks that match the user's preferences.

[0236] In practical applications, by incorporating user hierarchical information and scenario-aware information from heterogeneous businesses into the hierarchical awareness encoding module, combined with the aforementioned model structure, the model can automatically identify users at different lifecycles and content delivered to different business domains during the training phase. This allows users at the same lifecycle and tasks within the same business domain to be trained and optimized within the same PLE model, while sharing underlying features, increasing the model's generalization ability, and learning user representations at different lifecycles and task representations for different businesses. The distribution of the user-aware representation and scenario-aware representation outputs from the model, after unsupervised dimensionality reduction using T-SNE, can be found in [reference needed]. Figure 6 , Figure 7 . Figure 6 This is a schematic diagram illustrating the distribution of user-perceived representations in a task recommendation method provided in one embodiment of this specification. Figure 7 This is a schematic diagram illustrating the distribution of scene-aware representations in a task recommendation method provided in one embodiment of this specification. See also... Figure 6 Based on the distribution of user groups at different lifecycle stages, users with the same lifecycle are distributed in a relatively close sample space. This demonstrates that the model can automatically identify and optimize different user groups. For example, "new" refers to new users of the wealth management platform, "low" refers to users with low holdings (>0 & <100), and "high" refers to users with holdings greater than or equal to 100 yuan. See also... Figure 7 Based on the distribution of task representations for different business operations, tasks related to the same business are distributed in relatively similar sample spaces. This demonstrates that the model can automatically identify and optimize different types of business domains. Here, "yeb" refers to balance-enhancing service products, "active" refers to active users, "wealth" refers to wealth management products, "insurance" refers to insurance products, "shares" refers to securities products, and "ecology" refers to the platform ecosystem. It should be noted that the meanings of the above English terms are different from those of... Figure 5 The nouns yeb, active, wealth, insurance, shares, and ecology have the same meaning.

[0237] Based on this, the task recommendation method provided in this specification proposes a layered awareness multi-task recommendation scheme for layered users and heterogeneous services. This scheme can achieve the following three beneficial effects:

[0238] (1) Tiered users.

[0239] Based on the characteristics of different customer groups, a dedicated network is introduced to model mixed customer groups. At the same time, a soft-encoding module is used to achieve soft isolation between homogeneous groups, enabling differentiated modeling for different customer groups.

[0240] (2) Heterogeneous business.

[0241] Add dedicated networks for heterogeneous business scenarios, and achieve hard isolation of heterogeneous scenarios through the Hard-Encoding module to learn the differences between different scenarios.

[0242] (3) Multiple objectives.

[0243] To address different types of optimization objectives, User-Bias, Common-Task, and Specific-Task modules were added, each covering different user groups. The model also considers the dependencies between multiple objectives and provides auxiliary modeling and learning for tasks with dependencies.

[0244] Corresponding to the above method embodiments, this specification also provides a task recommendation device embodiment, which includes:

[0245] The receiving module is configured to receive a task set corresponding to the target user, wherein the task set contains tasks to be processed;

[0246] The first processing module is configured to process the task set based on the first task processing unit in the task recommendation model to obtain a set representation of the task set.

[0247] The second processing module is configured to process the user parameters of the target user based on the parameter processing unit in the preference prediction model, and obtain the user representation corresponding to the target user.

[0248] The third processing module is configured to process the set representation and the user representation based on the second task processing unit in the task recommendation model to obtain the task features of the task to be processed.

[0249] The task acquisition module is configured to input the task features into the task prediction unit in the task recommendation model to obtain the target task corresponding to the target user.

[0250] Optionally, the task acquisition module is further configured to:

[0251] The task features are input into the first task prediction unit of the task recommendation model to obtain a first candidate task; and / or

[0252] The task features are input into the second task prediction unit in the task recommendation model to obtain the second candidate task;

[0253] The first candidate task and / or the second candidate task are taken as the target task.

[0254] Optionally, the task acquisition module is further configured to: input the task features into the first task prediction unit in the task recommendation model to obtain the task prediction result;

[0255] Based on the task prediction results, a first candidate task is determined for the target user from the tasks to be processed.

[0256] Optionally, the task acquisition module is further configured to: determine user preference parameters based on the preference prediction model;

[0257] The user preference parameters and the task features are input into the second task prediction unit in the task recommendation model to obtain the task prediction result.

[0258] Based on the task prediction results, a second candidate task is determined for the target user from the tasks to be processed.

[0259] Optionally, the first processing module is further configured to:

[0260] Feature extraction is performed on the tasks to be processed contained in the task set to obtain the task features of the tasks to be processed;

[0261] The task features are input into the first task processing unit in the task recommendation model to obtain the task representation of the task to be processed.

[0262] The set representation of the task set is determined based on the task representation.

[0263] Optionally, the first processing module is further configured to:

[0264] The initial set representation of the task set is determined based on the task representation;

[0265] The initial set representation is encoded and calculated based on the first encoding module to obtain the set weight parameters corresponding to the initial set representation;

[0266] Based on the set weight parameters and the initial set representation, the target set representation of the task set is obtained.

[0267] Optionally, the third processing module is further configured to:

[0268] Determine the task representation corresponding to the task to be processed from the set representation;

[0269] Based on the task representation and the user representation, the second task processing unit is input to obtain the task features corresponding to the task to be processed.

[0270] Optionally, the second processing module is further configured to:

[0271] Receive user parameters of the target user, and perform feature extraction on the user parameters to obtain user features corresponding to the user parameters;

[0272] The user characteristics are input into the parameter processing unit of the preference prediction model to obtain the user representation corresponding to the target user.

[0273] Optionally, the second processing module is further configured to:

[0274] The user features are input into the parameter processing unit of the preference prediction model to obtain the initial user representation corresponding to the target user;

[0275] The initial user representation is encoded and calculated based on the second encoding module to obtain the user weight parameters corresponding to the initial user representation;

[0276] Based on the initial user representation and the user weight parameters, the target user representation corresponding to the target user is determined.

[0277] Optionally, the task recommendation device further includes a preference prediction module, configured to:

[0278] The user representation is input into the preference prediction unit in the preference prediction model to obtain user preference parameters.

[0279] Optionally, the training steps of the preference prediction model include:

[0280] Determine the training samples and sample labels for the preference prediction model, wherein the training samples are the sample user parameters of the sample users, and the sample labels are the sample user preference parameters of the sample users;

[0281] The sample user type of the sample user is determined based on the sample user parameters;

[0282] From the parameter processing units of the preference prediction model, determine the target parameter processing unit corresponding to the sample user type;

[0283] The sample user parameters are input into the target parameter processing unit to obtain the sample user representation corresponding to the sample user.

[0284] The sample user representation is input into the preference prediction unit in the preference prediction model to obtain the preference prediction result;

[0285] Based on the preference prediction results and the sample user preference parameters, the preference prediction model is trained until the training stops, thus obtaining the trained preference prediction model.

[0286] Optionally, the training steps of the task recommendation model include:

[0287] Determine the training samples and sample labels for the task recommendation model, wherein the training samples are a set of sample tasks corresponding to the sample users, the set of sample tasks contains sample tasks, and the sample labels are the preferred tasks of the sample users;

[0288] The sample task set is processed by the first task processing unit in the task recommendation model to obtain a sample set representation of the sample task set.

[0289] Obtain the sample user representation corresponding to the sample user, wherein the sample user representation is obtained by processing the sample user parameters of the sample user in the preference prediction model.

[0290] Based on the second task processing unit in the task recommendation model, the sample set representation and the sample user representation are processed to obtain the sample task features of the sample task.

[0291] The sample task features are input into the task prediction unit of the task recommendation model to obtain the task recommendation result;

[0292] Based on the task recommendation results and the preferred tasks of the sample users, the task recommendation model is trained until the training stops, thus obtaining the trained task recommendation model.

[0293] The above is an illustrative scheme of a task recommendation device according to this embodiment. It should be noted that the technical solution of this task recommendation device and the technical solution of the task recommendation method described above belong to the same concept. For details not described in detail in the technical solution of the task recommendation device, please refer to the description of the technical solution of the task recommendation method described above.

[0294] Corresponding to the above method embodiments, this specification also provides an embodiment of a task recommendation model training device, which includes:

[0295] The sample determination module is configured to determine training samples and sample labels for the task recommendation model, wherein the training samples are a set of sample tasks corresponding to the sample user, the set of sample tasks contains sample tasks, and the sample labels are the preferred tasks of the sample user.

[0296] The first processing module is configured to process the sample task set based on the first task processing unit in the task recommendation model to obtain a sample set representation of the sample task set.

[0297] The second processing module is configured to obtain the sample user representation corresponding to the sample user, wherein the sample user representation is obtained by the parameter processing unit in the preference prediction model by processing the sample user parameters of the sample user.

[0298] The second processing module is configured to process the sample set representation and the sample user representation based on the second task processing unit in the task recommendation model to obtain the sample task features of the sample task.

[0299] The result acquisition module is configured to input the sample task features into the task prediction unit in the task recommendation model to obtain the task recommendation result.

[0300] The model training module is configured to train the task recommendation model based on the task recommendation results and the sample users' preferred tasks until the training stops, thereby obtaining the trained task recommendation model.

[0301] Optionally, the training steps of the preference prediction model include:

[0302] Determine the training samples and sample labels for the preference prediction model, wherein the training samples are the sample user parameters of the sample users, and the sample labels are the sample user preference parameters of the sample users;

[0303] The sample user type of the sample user is determined based on the sample user parameters;

[0304] From the parameter processing units of the preference prediction model, determine the target parameter processing unit corresponding to the sample user type;

[0305] The sample user parameters are input into the target parameter processing unit to obtain the sample user representation corresponding to the sample user.

[0306] The sample user representation is input into the preference prediction unit in the preference prediction model to obtain the preference prediction result;

[0307] Based on the preference prediction results and the sample user preference parameters, the preference prediction model is trained until the training stops, thus obtaining the trained preference prediction model.

[0308] The above is an illustrative scheme of a task recommendation model training device according to this embodiment. It should be noted that the technical solution of this task recommendation model training device and the technical solution of the task recommendation model training method described above belong to the same concept. Details not described in detail in the technical solution of the task recommendation model training device can be found in the description of the technical solution of the task recommendation model training method described above.

[0309] Figure 8 A structural block diagram of a computing device 800 according to one embodiment of this specification is shown. The components of the computing device 800 include, but are not limited to, a memory 810 and a processor 820. The processor 820 is connected to the memory 810 via a bus 830, and a database 850 is used to store data.

[0310] The computing device 800 also includes an access device 840, which enables the computing device 800 to communicate via one or more networks 860. Examples of these networks include a Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 840 may include one or more of any type of wired or wireless network interface (e.g., a Network Interface Card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) interface, a Wi-MAX interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and so on.

[0311] In one embodiment of this specification, the above-described components of the computing device 800 and Figure 8 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 8 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0312] The computing device 800 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs. The computing device 800 can also be a mobile or stationary server.

[0313] The processor 820 is configured to execute the following computer-executable instructions, which, when executed by the processor 820, implement the steps of the task recommendation method and the task recommendation model method described above.

[0314] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device belongs to the same concept as the technical solutions of the task recommendation method and the task recommendation model method described above. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solutions of the task recommendation method and the task recommendation model method described above.

[0315] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the task recommendation method and the task recommendation model method described above.

[0316] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solutions of the task recommendation method and the task recommendation model method described above. Details not described in detail in the technical solution of the storage medium can be found in the descriptions of the technical solutions of the task recommendation method and the task recommendation model method described above.

[0317] An embodiment of this specification also provides a computer program, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described task recommendation method and task recommendation model method.

[0318] The above is an illustrative example of a computer program according to this embodiment. It should be noted that the technical solution of this computer program belongs to the same concept as the technical solutions of the task recommendation method and task recommendation model method described above. Details not described in detail in the computer program's technical solution can be found in the descriptions of the technical solutions of the task recommendation method and task recommendation model method described above.

[0319] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0320] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0321] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0322] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0323] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A task recommendation method, comprising: Receive a task set corresponding to the target user, wherein the task set contains tasks to be processed; The task set is processed by the first task processing unit in the task recommendation model to obtain a set representation of the task set. Based on the parameter processing unit in the preference prediction model, the user parameters of the target user are processed to obtain the user representation corresponding to the target user; Based on the second task processing unit in the task recommendation model, the set representation and the user representation are processed to obtain the task features of the task to be processed. The task features are input into the task prediction unit in the task recommendation model to obtain the target task corresponding to the target user.

2. The task recommendation method according to claim 1, wherein inputting the task features into the task prediction unit in the task recommendation model to obtain the target task corresponding to the target user includes: The task features are input into the first task prediction unit in the task recommendation model to obtain the first candidate task; and / or The task features are input into the second task prediction unit in the task recommendation model to obtain the second candidate task; The first candidate task and / or the second candidate task are taken as the target task.

3. The task recommendation method according to claim 2, wherein inputting the task features into the first task prediction unit in the task recommendation model to obtain the first candidate task includes: The task features are input into the first task prediction unit in the task recommendation model to obtain the task prediction result; Based on the task prediction results, a first candidate task is determined for the target user from the tasks to be processed.

4. The task recommendation method according to claim 2, wherein inputting the task features into the second task prediction unit in the task recommendation model to obtain a second candidate task includes: User preference parameters are determined based on the preference prediction model. The user preference parameters and the task features are input into the second task prediction unit in the task recommendation model to obtain the task prediction result. Based on the task prediction results, a second candidate task is determined for the target user from the tasks to be processed.

5. The task recommendation method according to claim 1, wherein the first task processing unit in the task recommendation model processes the task set to obtain a set representation of the task set, comprising: Feature extraction is performed on the tasks to be processed contained in the task set to obtain the task features of the tasks to be processed; The task features are input into the first task processing unit in the task recommendation model to obtain the task representation of the task to be processed. The set representation of the task set is determined based on the task representation.

6. The task recommendation method according to claim 5, wherein determining the set representation of the task set based on the task representation includes: The initial set representation of the task set is determined based on the task representation; The initial set representation is encoded and calculated based on the first encoding module to obtain the set weight parameters corresponding to the initial set representation; Based on the set weight parameters and the initial set representation, the target set representation of the task set is obtained.

7. The task recommendation method according to claim 1, wherein the step of processing the set representation and the user representation based on the second task processing unit in the task recommendation model to obtain the task features of the task to be processed includes: Determine the task representation corresponding to the task to be processed from the set representation; Based on the task representation and the user representation, the second task processing unit is input to obtain the task features corresponding to the task to be processed.

8. The task recommendation method according to claim 1, wherein the parameter processing unit in the preference prediction model processes the user parameters of the target user to obtain the user representation corresponding to the target user, comprising: Receive user parameters of the target user, and perform feature extraction on the user parameters to obtain user features corresponding to the user parameters; The user characteristics are input into the parameter processing unit of the preference prediction model to obtain the user representation corresponding to the target user.

9. The task recommendation method according to claim 8, wherein inputting the user features into the parameter processing unit of the preference prediction model to obtain the user representation corresponding to the target user includes: The user features are input into the parameter processing unit of the preference prediction model to obtain the initial user representation corresponding to the target user; The initial user representation is encoded and calculated based on the second encoding module to obtain the user weight parameters corresponding to the initial user representation; Based on the initial user representation and the user weight parameters, the target user representation corresponding to the target user is determined.

10. The task recommendation method according to claim 1, further comprising, after the parameter processing unit in the preference prediction model processes the user parameters of the target user to obtain the user representation corresponding to the target user, the method further comprising: The user representation is input into the preference prediction unit in the preference prediction model to obtain user preference parameters.

11. The task recommendation method according to claim 1, wherein the training step of the preference prediction model includes: Determine the training samples and sample labels for the preference prediction model, wherein the training samples are the sample user parameters of the sample users, and the sample labels are the sample user preference parameters of the sample users; The sample user type of the sample user is determined based on the sample user parameters; From the parameter processing units of the preference prediction model, determine the target parameter processing unit corresponding to the sample user type; The sample user parameters are input into the target parameter processing unit to obtain the sample user representation corresponding to the sample user. The sample user representation is input into the preference prediction unit in the preference prediction model to obtain the preference prediction result; Based on the preference prediction results and the sample user preference parameters, the preference prediction model is trained until the training stops, thus obtaining the trained preference prediction model.

12. The task recommendation method according to claim 1 or 11, wherein the training step of the task recommendation model includes: Determine the training samples and sample labels for the task recommendation model, wherein the training samples are a set of sample tasks corresponding to the sample users, the set of sample tasks contains sample tasks, and the sample labels are the preferred tasks of the sample users; The sample task set is processed by the first task processing unit in the task recommendation model to obtain a sample set representation of the sample task set. Obtain the sample user representation corresponding to the sample user, wherein the sample user representation is obtained by processing the sample user parameters of the sample user in the preference prediction model. Based on the second task processing unit in the task recommendation model, the sample set representation and the sample user representation are processed to obtain the sample task features of the sample task. The sample task features are input into the task prediction unit of the task recommendation model to obtain the task recommendation result; Based on the task recommendation results and the preferred tasks of the sample users, the task recommendation model is trained until the training stops, thus obtaining the trained task recommendation model.

13. A method for training a task recommendation model, comprising: Determine the training samples and sample labels for the task recommendation model, wherein the training samples are a set of sample tasks corresponding to the sample users, the set of sample tasks contains sample tasks, and the sample labels are the preferred tasks of the sample users; The sample task set is processed by the first task processing unit in the task recommendation model to obtain a sample set representation of the sample task set. Obtain the sample user representation corresponding to the sample user, wherein the sample user representation is obtained by processing the sample user parameters of the sample user in the preference prediction model. Based on the second task processing unit in the task recommendation model, the sample set representation and the sample user representation are processed to obtain the sample task features of the sample task. The sample task features are input into the task prediction unit of the task recommendation model to obtain the task recommendation result; Based on the task recommendation results and the preferred tasks of the sample users, the task recommendation model is trained until the training stops, thus obtaining the trained task recommendation model.

14. The task recommendation model training method according to claim 13, wherein the training step of the preference prediction model includes: Determine the training samples and sample labels for the preference prediction model, wherein the training samples are the sample user parameters of the sample users, and the sample labels are the sample user preference parameters of the sample users; The sample user type of the sample user is determined based on the sample user parameters; From the parameter processing units of the preference prediction model, determine the target parameter processing unit corresponding to the sample user type; The sample user parameters are input into the target parameter processing unit to obtain the sample user representation corresponding to the sample user. The sample user representation is input into the preference prediction unit in the preference prediction model to obtain the preference prediction result; Based on the preference prediction results and the sample user preference parameters, the preference prediction model is trained until the training stops, thus obtaining the trained preference prediction model.

Images