Promotion strategy generation and model training method, computing device, medium and product

Abstract

Embodiments of the present application provide a promotion strategy generation and model training method, a computing device, a medium and a product. In the method, a service end responds to a strategy generation event, obtains multi-dimensional target data corresponding to a target account, and a rule portrait label and a rule confidence corresponding to a matched first rule; a portrait generation model is used to identify the basic features of the multi-dimensional target data and determine the association strength thereof with the first rule; the rule confidence is combined to identify the rule features of the first rule; the weighted features of the basic features and the rule features are calculated based on the association strength, the target portrait label is generated based on the weighted features and the prior features of the rule portrait; the decision maker information and the portrait summary information of the target account are determined based on the target portrait label, the promotion object is determined based on the portrait summary information and the decision maker information as a guide, and the promotion strategy for the promotion object is generated. The present application can improve the generation efficiency and accuracy of the promotion strategy.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, computing device, medium and product for generating promotion strategies and training models. Background Technology

[0002] With the development of computer technology, current online service systems typically generate targeted promotion strategies based on the profile tags of target accounts, which are adapted to the business scenarios of the target accounts, in order to promote the target to users and improve the service conversion rate of the promoted targets.

[0003] The profile tags mentioned in this article are structured tags constructed based on relevant data of the corresponding target account, used to characterize the target account's behavioral characteristics, preferences, and needs in specific dimensions. For example, if the target account is an organization with recruitment needs in an online service system, its profile tags may include tags such as the organization's strength, the organization's activity level in the online service system, and the recruitment service conversion rate.

[0004] Currently, the method of manually integrating and analyzing target account data is commonly used to build profile tags. However, due to the large amount of data associated with target accounts, this method of building profile tags is time-consuming and prone to missing key information, which in turn affects the efficiency and accuracy of promotion strategy generation and urgently needs improvement. Summary of the Invention

[0005] This application provides a method, computing device, medium, and product for generating and training an information promotion strategy, in order to solve the problems of low efficiency and poor accuracy in the process of generating promotion strategies based on profile tags due to the long time consumption and inaccuracy of profile tag construction in the prior art.

[0006] Firstly, this application provides a method for generating a promotion strategy, applied to a server, the method comprising: In response to the strategy generation event, obtain multi-dimensional target data corresponding to the target account; Obtain the rule profile label and rule confidence level corresponding to the first profile rule matching the multi-dimensional target data; Using a profile generation model, the basic features of the multi-dimensional target data are identified, and the correlation strength between the first profile rule and the basic features is determined; combined with the rule confidence, the rule features of the first profile rule are identified; based on the correlation strength, the weighted features of the basic features and the rule features are calculated; and the target profile label is generated by combining the prior features of the rule profile label and the weighted features; the profile generation model is trained based on multi-dimensional sample data of sample accounts and their matched second profile rules and labeled profile labels, as well as the sample confidence and sample profile labels corresponding to the second profile rule. Based on the target profile tags, determine the decision-maker information of the target account and generate profile summary information; Based on the profile summary information and guided by the decision-maker information, the target audience is identified, and a promotion strategy is generated for the target audience; the promotion strategy is used to promote the target audience to the decision-maker's client.

[0007] Secondly, embodiments of this application provide a model training method, including: Obtain multi-dimensional sample data of the sample accounts, and the labeled profile tags of the multi-dimensional sample data; Obtain the sample profile labels and sample confidence scores corresponding to the second profile rule matched by the multi-dimensional sample data; The portrait generation model is trained using the multi-dimensional sample data, the second portrait rule and its corresponding sample confidence and sample portrait label as input data, and the labeled portrait label and the sample portrait label as training labels. The profile generation model is used to identify the basic features of multi-dimensional target data corresponding to the target account, and determine the correlation strength between the first profile rule matching the multi-dimensional target data and the basic features. Combining the rule confidence of the first profile rule, the model identifies the rule features of the first profile rule. Based on the correlation strength, the model calculates the weighted features of the basic features and the rule features. The model generates target profile tags by combining the prior features of the rule profile tags and the weighted features. The target profile tags are used to determine the decision-maker information of the target account and generate profile summary information. The decision-maker information and the profile summary information are used to determine the target audience and generate a promotion strategy for the target audience. The promotion strategy is used to promote the target audience to the decision-maker's client.

[0008] Thirdly, this application provides a computing device, including a processing component and a storage component; the storage component stores a computing program; the computer program is invoked and executed by the processing component to implement the promotion strategy generation method of the first aspect or the model training method of the second aspect.

[0009] Fourthly, this application provides a computer storage medium storing a computer program thereon. When the computer program is executed by a processing component, it implements the promotion strategy generation method of the first aspect or the model training method of the second aspect.

[0010] Fifthly, this application provides a computer program product, including a computer program or instructions, which, when executed by a processing component, implement the promotion strategy generation method of the first aspect or the model training method of the second aspect.

[0011] In this embodiment, the server responds to the strategy generation event by obtaining multi-dimensional target data corresponding to the target account, obtaining the rule profile label and rule confidence level corresponding to the first profile rule matching the multi-dimensional target data; using the profile generation model to identify the basic features of the multi-dimensional target data, and determining the correlation strength between the first profile rule and the basic features; combining the rule confidence level to identify the rule features of the first profile rule; then, based on the correlation strength, calculating the weighted features of the basic features and the rule features; combining the prior features and weighted features of the rule profile label to generate the target profile label; determining the decision-maker information of the target account based on the target profile label; generating profile summary information; using the profile summary information as a basis and the decision-maker information as a guide to determine the promotion target and generate a promotion strategy for the promotion target.

[0012] This embodiment provides a scheme for automatically and accurately constructing portrait tags based on portrait rules and a portrait generation model. The extraction of rule features is determined by rule confidence. Based on the correlation strength between the matched first portrait rule and the basic features of multi-dimensional target data, a weighted fusion of rule features and basic features is driven. Finally, the prior features of the corresponding rule portrait tag of the first portrait rule are fused with the weighted features to generate the target portrait tag. It is evident that this embodiment is not simply using rules first and then the model, but rather organically integrating business logic and intelligent parsing within the model. This achieves deep collaboration between rules and the model rather than mechanical splicing, preserving the interpretability and controllability of the rules while leveraging the model's context awareness and generalization capabilities. Since the initial portrait tags are obtained through rule matching, there is no need to wait for complex reasoning; and the portrait generation model only needs to extract and fuse features in the feature space, significantly improving the efficiency of portrait generation and the real-time nature of promotion decisions. Furthermore, this embodiment introduces correlation strength as a dynamic adjustment signal to dynamically adjust the weights of feature weighting processing, effectively suppressing the interference of low-correlation features on the target portrait tag, significantly improving the robustness and accuracy of the target portrait tag.

[0013] It is important to note that this solution is not a simple superposition of rules and models. Instead, it is a collaborative design that uses a feature weighting mechanism driven by confidence and relevance, along with a priori fusion architecture of rule profile tags, to quickly and accurately construct target profile tags. This results in a promotion strategy that is more precise due to the profile tags, significantly improving the conversion rate of the target audience.

[0014] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 The system architecture diagram of the online service system provided in this application is shown; Figure 2 A flowchart illustrating an embodiment of the promotion strategy generation method provided in this application is shown; Figure 3 This is a schematic diagram illustrating the display effect of the portrait summary information provided in this application; Figure 4 This is a schematic diagram illustrating the display effect of the decision-maker information provided in this application; Figure 5 A flowchart of an embodiment of the model training method provided in this application is shown; Figure 6 A flowchart illustrating the promotion strategy generation method in a practical application scenario of this application is shown; Figure 7 A schematic diagram of the structure of an embodiment of the promotion strategy generation apparatus provided in this application is shown; Figure 8 A schematic diagram of the structure of one embodiment of the model training apparatus provided in this application is shown; Figure 9 A schematic diagram of the structure of the computing device provided in this application is shown. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] It should be noted that, in the cases involving user account data in this application's embodiments, the account data (including but not limited to attribute data, behavioral data, interaction data, etc.) and other data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application's embodiments are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse. In addition, the various models involved in this application (including but not limited to language models or large models) comply with relevant laws and standards.

[0018] It should be noted that the technical solutions in this application are applicable to virtual network environments, and the users described generally refer to "virtual users." Real users can register user accounts on the server through registration to obtain user identities in the network environment. The same user account can log in to the server through different types of clients, enabling the server to identify the same user.

[0019] Interactions between the server and the user can be based on user accounts. The data received or sent by the server to the user is also based on the user account; in reality, the client corresponding to the user account receives or sends data to the server. Furthermore, users can also communicate with each other through their user accounts. Here, "user" can refer to an individual or an organization, such as a company; this application does not impose specific restrictions.

[0020] As described in the background section, servers typically generate targeted promotional strategies for target accounts based on their profile tags, tailored to their specific business scenarios. However, current profile tag analysis relies heavily on manual statistics, which is time-consuming and prone to missing key information, resulting in poor efficiency and accuracy in tag generation. To address this, the inventors conceived of a solution for automatically generating profile tags based on account data. For example, profile rules for tags could be pre-defined manually, and the tags could be determined by matching user account data in the online service system with these rules. However, this method suffers from poor flexibility and can only analyze profile tags based on structured interaction data (such as transaction data or behavioral data), limiting the amount of account data available for analysis. With the development of artificial intelligence technology, the inventors also considered using AI models to analyze large amounts of account data to determine profile tags. However, considering the poor interpretability of AI models and the initial sparseness of account data, resulting in poor model output, simply relying on AI models to generate profile tags cannot guarantee the accuracy of the generated results.

[0021] Therefore, facing the contradictions between the flexibility of profile rules in profile tag construction, the cold start defect and poor interpretability of the model, and the poor performance when data is sparse, the inventors conducted a series of studies and proposed a profile tag analysis method that combines profile rules and intelligent models (i.e., profile generation models) for application in the generation process of promotion strategies. The basic idea is as follows: The server responds to the strategy generation event, obtains multi-dimensional target data corresponding to the target account, obtains the rule profile tag and rule confidence of the first profile rule matching the multi-dimensional target data; uses the profile generation model to identify the basic features of the multi-dimensional target data, and determines the correlation strength between the first profile rule and the basic features; combines the rule confidence to identify the rule features of the first profile rule; then, based on the correlation strength, calculates the weighted features of the basic features and rule features; combines the prior features and weighted features of the rule profile tag to generate target profile tags; based on the target profile tags, determines the decision-maker information of the target account and generates profile summary information; based on the profile summary information and guided by the decision-maker information, determines the promotion target and generates a promotion strategy for the promotion target.

[0022] This embodiment provides a scheme for automatically and accurately constructing portrait tags based on portrait rules and a portrait generation model. The extraction of rule features is determined by rule confidence. Based on the correlation strength between the matched first portrait rule and the basic features of multi-dimensional target data, a weighted fusion of rule features and basic features is driven. Finally, the prior features of the corresponding rule portrait tag of the first portrait rule are fused with the weighted features to generate the target portrait tag. It is evident that this embodiment is not simply using rules first and then the model, but rather organically integrating business logic and intelligent parsing within the model. This achieves deep collaboration between rules and the model rather than mechanical splicing, preserving the interpretability and controllability of the rules while leveraging the model's context awareness and generalization capabilities. Since the initial portrait tags are obtained through rule matching, there is no need to wait for complex reasoning; and the portrait generation model only needs to extract and fuse features in the feature space, significantly improving the efficiency of portrait generation and the real-time nature of promotion decisions. Furthermore, this embodiment introduces correlation strength as a dynamic adjustment signal to dynamically adjust the weights of feature weighting processing, effectively suppressing the interference of low-correlation features on the target portrait tag, significantly improving the robustness and accuracy of the target portrait tag.

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] Figure 1This diagram illustrates a system architecture of an online service system according to an embodiment of this application. The system architecture may include a client 101 and a server 102. The client 101 can be a client corresponding to a service personnel within the service system, or a client corresponding to a user providing services within the service system (such as an agency providing recruitment services). The client 101 and server 102 can establish a connection via a network. The network provides a communication link between the client 101 and server 102. The network can include various connection types, such as wired, wireless, or fiber optic cables. The client 101 can interact with the server 102 via the network to receive or send messages, etc.

[0025] The client 101 can be a browser, an app (application), a web application such as an H5 (HyperText Markup Language 5) application, a lightweight application (also known as a mini-program), or a cloud application. The client 101 can be deployed on electronic devices and depends on the device or certain apps on the device to run. Electronic devices can have displays and support information browsing, such as personal mobile terminals like mobile phones, tablets, personal computers, desktop computers, smart speakers, smartwatches, etc. For ease of understanding, Figure 1 The client is primarily represented by the image of a device. Various other types of applications can also be configured in electronic devices, such as human-computer interaction applications, model training applications, text processing applications, web browser applications, shopping applications, search applications, instant messaging tools, email clients, and social media platform software. An electronic device can refer to a user-used device with the computing, internet access, and communication functions required by the user, such as a mobile phone, tablet computer, personal computer, or wearable device. An electronic device typically includes at least one processing component and at least one storage component. Electronic devices may also include basic configurations such as network interface card (NIC) chips, I / O (input / output) buses, and audio / video components; this application does not limit the scope of these components. Optionally, depending on the implementation of the electronic device, it may also include some peripheral devices, such as a keyboard, mouse, input pen, and printer; this application does not limit the scope of these peripheral devices.

[0026] Server 102 may include servers that provide various services, such as servers that provide background training for the profile generation model, or servers that analyze profile tag generation and promotion strategies.

[0027] It should be noted that server 102 can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. The server can also be a server in a distributed system, or a server integrated with blockchain. The server can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms, or an intelligent cloud computing server or intelligent cloud host with artificial intelligence technology.

[0028] It should be noted that the promotion strategy generation method provided in this application embodiment is generally executed by the server 102, and the corresponding promotion strategy generation device is generally located in the server 102. However, in other embodiments of this application, the promotion strategy generation method provided in this application embodiment can also be jointly executed by the client 101 and the server 102. It should be understood that... Figure 1 The number of clients and servers shown is merely illustrative. Depending on implementation needs, there can be any number of clients and servers.

[0029] The implementation details of the technical solutions in the embodiments of this application are described in detail below.

[0030] Figure 2 A flowchart illustrating an embodiment of a method for generating a promotion strategy provided in this application. The technical solution of this embodiment can be derived from... Figure 1 The server 102 in the middle executes, Figure 2 The method for generating the promotion strategy shown may include the following steps: S201, in response to the policy generation event, retrieves multi-dimensional target data corresponding to the target account.

[0031] The strategy generation event can be a promotion strategy generation event. In this embodiment, the strategy generation event can be triggered at a set time or when the promotion target is updated.

[0032] In this embodiment, the "account" can be a virtual account registered by a user (such as an individual or organization) in the online service system. For example, it could be an account registered by an organization with recruitment needs in the online service system. Optionally, this embodiment can use each account in the online service system sequentially as the target account, or it can use accounts that meet the screening requirements (such as level requirements, attribute requirements, etc.) as the target account; there is no limitation on this.

[0033] Multi-dimensional target data can be multi-dimensional data generated by the target account in the online service system, including but not limited to: basic account data (e.g., if the target account is an organization, the basic data could be the organization's size, industry, city, etc.), behavioral data (e.g., login data, search data, platform posting data, etc.), transaction data (e.g., recharge data, consumption data, order data, etc.), and interaction data (e.g., likes data, message reply data, outbound call connection rate, etc.). To further improve the adaptability of profile tags to business scenarios, the multi-dimensional target data in this embodiment can further include at least one service characteristic tag of the service field to which the target account belongs. For example, if the service field to which the target account belongs is the recruitment field, then the service characteristic tag can include recruitment peak season tags, urgent job tags, etc. This service characteristic tag can be labeled by the user or determined by the server based on analysis of basic account data, behavioral data, transaction data, interaction data, etc.

[0034] In this embodiment, after the server responds to the policy generation event, it determines the target account to be analyzed and then obtains multi-dimensional target data of the target account from the database (such as a customer relationship management system) where account data is stored in the online service system.

[0035] In practical applications, multi-dimensional account data of each account (including the target account) in the online service system can be pre-processed and stored in a database to quickly retrieve multi-dimensional target data in response to policy generation events. Optionally, before executing the method of this embodiment, multi-dimensional raw data of the target account is obtained; at least one service characteristic tag is determined based on the multi-dimensional raw data; the multi-dimensional raw data and at least one service characteristic tag are cleaned and / or standardized, and the processed data is associated with the account identifier of the target account and stored in the database.

[0036] The multi-dimensional raw data can be directly obtained from online service systems, including but not limited to: unprocessed basic data, behavioral data, transaction data, and interaction data. In this embodiment, after obtaining the multi-dimensional raw data for the target account, business scenario characteristic analysis can be performed on the raw data to determine at least one service characteristic tag. Then, the multi-dimensional raw data and at least one service characteristic tag are cleaned and / or standardized. Cleaning includes, but is not limited to: removing abnormal data and completing missing data. Standardization includes, but is not limited to: using standardized scores (such as Z-scores) for numerical data (such as login counts and outbound call connection rates) to eliminate the influence of unit of measurement. For non-numerical data, such as industry, city, or other categorical data, or text data such as reply messages, one-hot encoding is performed to adapt to the input format of the profile generation model. The processed data (including the processed multi-dimensional raw data and service characteristic tags) is then associated with the account identifier of the corresponding target account and stored in the database. Accordingly, when performing this step, in response to a policy generation event, the account identifier of the target account can be determined; data associated with the account identifier can then be retrieved from the database as the multi-dimensional target data of the target account. Specifically, in response to a policy generation event, the account identifier of the target account is determined, such as obtaining an account that meets the recommendation level requirements as the target account, and then obtaining the account identifier of the target account. Using this account identifier as an index, data associated with it can be retrieved from the database as the multi-dimensional target data of the target account. This embodiment integrates and stores the scattered and inconsistent multi-dimensional raw data to improve the efficiency and accuracy of subsequent multi-dimensional target data acquisition. In addition, the multi-dimensional target data in this embodiment includes not only account-related data but also service characteristic tags of the service domain to which the account belongs. These service characteristic tags help to make the constructed profile tags business-adaptable to the service scenario, thereby further improving the accuracy of target profile tag determination.

[0037] S202, obtain the rule profile label and rule confidence score corresponding to the first profile rule of multi-dimensional target data matching.

[0038] In this embodiment, the profiling rules are the judgment rules for profiling tags. These rules can be interpretable rules composed of profiling tags and account characteristics. For example, the profiling rule corresponding to the profiling tag "high conversion potential" could be "real estate industry + monthly recharge ≥ 5000 yuan". The profiling rule corresponding to the profiling tag "high churn risk" could be "login frequency decreased by 50% in the past 7 days + no job postings in the past 7 days".

[0039] The profiling rules in this embodiment can be extracted from multi-dimensional historical data of a large number of accounts in the online service system. The extraction method can be to obtain multi-dimensional historical data of sample accounts; for any profiling category, based on the multi-dimensional historical data, extract at least one set of associations between the profiling tags corresponding to the profiling category and the account features, as at least one profiling rule and its corresponding profiling tag corresponding to the profiling category; verify and optimize the extracted at least one profiling rule and its corresponding profiling tag to obtain the final profiling rule and its corresponding profiling tag.

[0040] Specifically, accounts with complete follow-up and clear results of order completion or churn can be selected as sample accounts. Similar to obtaining multi-dimensional target data in S201, multi-dimensional historical data of these sample accounts is obtained. This multi-dimensional historical data is then divided into rule extraction data and rule verification data at a certain ratio (e.g., 9:1). Account feature values, such as the number of times information is posted, industry type, and message response rate in the past 7 days, are extracted from the rule extraction data. Since the sample accounts are those with complete follow-up, their true profile tags, such as order completion probability, churn risk, and platform activity, are determinable. Therefore, the account feature values ​​extracted from sample accounts with the same profile tags can be statistically analyzed to extract preliminary profile rules for those tags. Finally, the accuracy of the preliminary profile rules is verified using rule verification data. High-accuracy profile rules are retained, and duplicate rules can be removed from the retained rules to reduce redundancy.

[0041] It should be noted that each profiling rule in this embodiment has its unique corresponding profiling label. A single profiling label can correspond to multiple profiling rules. For each profiling rule, this embodiment can also set a corresponding confidence level based on the accuracy of its corresponding profiling label. This confidence level can be set when refining the profiling rule, for example, using the accuracy determined during the verification phase as its confidence level. Alternatively, it can be determined by real-time or periodic statistical analysis of the consistency between the target profiling label generated by the profiling generation model and the rule profiling labels associated with the target profiling label, using this as a feedback signal to adjust the confidence level of the profiling rule.

[0042] This embodiment involves determining whether the multi-dimensional target data matches the account characteristics of each extracted profiling rule. If they match, the profiling rule is considered the first profiling rule for matching the multi-dimensional target data. The corresponding profiling tag for this first profiling rule is then obtained as the rule profiling tag, and the corresponding confidence score for this first profiling rule is obtained as the rule confidence score. It should be noted that in this embodiment, the number of first profiling rules for matching the multi-dimensional target data can be one or more. If there are multiple first profiling tags, the corresponding rule profiling tag and rule confidence score can be determined separately for each first profiling rule.

[0043] In practical applications, user profile tags are typically set based on user profile categories. For example, when the user profile category is "resource investment," the corresponding user profile tags include: high resource investment, medium resource investment, and low resource investment. To improve the accuracy of user profile tag positioning, the account characteristic items in the user profile rules for the same user profile category are usually fixed. The only difference is that the characteristic values ​​corresponding to each account characteristic item differ in the user profile rules for different user profile tags. For example, when the user profile category is "resource investment," one account characteristic item for this category can be the ranking range of the platform recharge amount in the past 7 days within the corresponding customer group (i.e., a group from the same city, size, and industry). In this case, the ranking range corresponding to the "high resource investment" tag under this account characteristic item is the top 20%; the ranking range corresponding to the "medium resource investment" tag is 20%-50%; and the ranking range corresponding to the "low resource investment" tag is 51%-80%.

[0044] Therefore, when there are multiple profile categories, and each profile category corresponds to multiple profile tags and multiple account feature items, matching profile rules one by one is inefficient and prone to errors. Therefore, this step can also be implemented through the following sub-steps: Sub-step 1: For any given profile category, determine at least one account feature item corresponding to that profile category. The account feature items for each profile category can be pre-set. To accurately determine profile tags, there are usually multiple account feature items for each profile category. This embodiment can directly obtain at least one pre-set account feature item for each profile category.

[0045] Sub-step 2: Based on the multi-dimensional target data, determine the sub-feature values ​​corresponding to at least one account feature item. Specifically, for each account feature item, relevant data for that feature item can be found in the multi-dimensional target data and directly used as the corresponding sub-feature value (such as city, organization size, industry, etc.). Alternatively, the relevant data can be further analyzed and processed to obtain the corresponding sub-feature value. For example, after obtaining the account's platform recharge amount in the past 7 days (i.e., relevant data), the ranking range of its customer group can be calculated, and the calculation result can be used as the sub-feature value.

[0046] Sub-step 3: Perform weighted fusion processing on the sub-feature values ​​corresponding to at least one account feature item to obtain the total feature value of the portrait category. Specifically, for each portrait category, its corresponding multiple account feature items are assigned corresponding weight values. In this embodiment, the sub-feature values ​​can be weighted and summed (or averaged) according to the weight values ​​corresponding to at least one account feature item to obtain the total feature value corresponding to the portrait category.

[0047] Sub-step 4: Among the at least one portrait rule corresponding to the portrait category, the portrait rule that matches at least one sub-feature value and the total feature value is taken as the first portrait rule, and the rule portrait label and rule confidence corresponding to the first portrait rule are obtained. In this embodiment, each portrait rule corresponding to the same portrait category has its own corresponding feature value range and total feature value range for each account feature item. In this embodiment, for each portrait category, it can be determined whether there is a portrait rule among the at least one portrait rule corresponding to it that matches the corresponding sub-feature value and total feature value range. If so, it is taken as the first portrait rule, and the portrait label corresponding to the first portrait rule is obtained as the rule portrait label, and the confidence of the first portrait rule is obtained as the rule confidence.

[0048] S203. Using a portrait generation model, identify the basic features of multi-dimensional target data and determine the correlation strength between the first portrait rule and the basic features; combine the rule confidence to identify the rule features of the first portrait rule; calculate the weighted features of the basic features and rule features based on the correlation strength; and generate target portrait labels by combining the prior features and weighted features of the rule portrait labels.

[0049] The portrait generation model in this paper, as well as the strategy recognition model, conversation summarization model, and intelligent generation model discussed later, can refer to a large-parameter model (i.e., a large model) trained using large-scale data and powerful computing capabilities. These are machine learning models with complex structures, capable of processing massive amounts of data and completing various complex tasks, such as natural language processing, computer vision, and speech recognition. They can include Large Language Models (LLMs) or Multimodal Large Models (MLMs). The portrait generation model discussed in this paper can be a pre-trained model, which is retrained through model fine-tuning to adapt to different processing tasks. This utilizes the powerful capabilities of the pre-trained model while also adapting to new data distributions, thus ensuring the model's generalization ability and reducing overfitting. The portrait generation model in this embodiment is trained based on multi-dimensional sample data of sample accounts, their matched second portrait rules and labeled portrait tags, as well as the sample confidence and sample portrait tags corresponding to the second portrait rules. The specific training process will be described in detail in subsequent embodiments.

[0050] Specifically, the portrait generation model in this embodiment includes a dual-channel feature fusion network and an attention mapping matrix. The first channel is the rule data input channel, used to identify the features of the input multi-dimensional target data in multiple spatial dimensions, i.e., basic features. Combining the confidence level of each first portrait rule, it identifies the rule features corresponding to the first portrait rule, as well as the prior features of the rule portrait label in multiple spatial dimensions. The attention mapping matrix is ​​used to determine the association strength between each first portrait rule and the basic features of the multi-dimensional target data. The second channel is the data-driven channel, used to use the association scheduling corresponding to each first portrait rule as the weight of the rule features of that first portrait rule, weighting the basic features and rule features to obtain weighted features, and concatenating or weighting the prior features and weighted features of the rule portrait label, transforming the processed features into corresponding portrait labels, i.e., target portrait labels.

[0051] Optionally, this embodiment uses various methods to determine the association strength between each first profile rule and the basic features of the multi-dimensional target data using the attention mapping matrix. One approach is to calculate the association strength between each first profile rule and the basic features of the multi-dimensional target data based on the rule confidence level. For example, the higher the rule confidence level of a first profile rule, the stronger its association with the basic features of the multi-dimensional target data. Another approach is to identify the matching rate of the multi-dimensional target data for each first profile rule. For example, the matching degree between the account feature values ​​in the multi-dimensional target data and the account feature values ​​corresponding to the first profile rule is calculated as the matching rate of the multi-dimensional target data for the first profile rule. Then, based on the matching rate, the association strength between the first profile rule and the basic features is determined; for example, the higher the matching rate, the stronger the corresponding association.

[0052] It should be noted that the portrait generation model in this embodiment can play a leading role in the case of sparse multi-dimensional target data, with high-confidence rule features and prior features of rule portrait labels playing a dominant role, ensuring the basic rationality of portrait labels and solving the cold start problem. As multi-dimensional target data of target accounts accumulates, the contribution of basic features is adaptively enhanced, and by introducing correlation strength, the interpretability of weighted features can be enhanced, thereby improving the efficiency and accuracy of target portrait label generation.

[0053] S204. Based on the target profile tags, determine the decision-maker information of the target account and generate profile summary information.

[0054] In this context, the decision-maker for the target account can be anyone with the authority to decide whether or not to accept promotional offers. This person may or may not be the one operating the target account. For example, if the target account is a personal account, the decision-maker can be the person operating the account. If the target account is an institutional account, the decision-maker can be the account operator or the institution's manager (such as a director). There can be one or more decision-makers for the target account. The decision-maker information in this embodiment may include, but is not limited to, the decision-maker's contact information, job title, and contact time.

[0055] Since the target profile tags in this embodiment include multi-dimensional profile tags corresponding to multiple profile categories such as user (individual or organization) strength, platform activity, resource investment, and conversion effect, organizational complexity analysis can be performed on the target account based on these target profile tags. For example, it can be determined whether the target account has team-based operation characteristics. If it does, its organizational complexity is high; otherwise, its organizational complexity is low. Then, a decision-making mode is determined based on the organizational complexity. For example, if the organizational complexity is low, the operator of the target account can be considered the decision-maker, and the decision-maker information can be obtained. If the organizational complexity is high, the organizational structure data corresponding to the target account can be obtained, and a pre-trained decision recognition model can be used to identify the decision-maker features of each person in the organizational structure data to determine and output one or more target decision-makers in the organizational structure, thereby obtaining their corresponding decision-maker information.

[0056] This embodiment can also utilize a conversation summary model to extract and summarize the core information of the target profile tags, add explanations corresponding to the target profile tags, and finally output the profile summary information in natural language. To facilitate service personnel's intuitive understanding of the profile tag information, this embodiment can also add corresponding level labels (such as high, medium, and low levels), color labels, and prompts to the profile summary information for profile tags that require special attention from service personnel.

[0057] S205, based on the profile summary information and guided by decision-maker information, identifies the target audience and generates promotional strategies for the target audience.

[0058] In this embodiment, the target of promotion can be a product or service that the online service system needs to promote to decision-makers. For example, it could be a membership product or intelligent customer service provided by the online service system.

[0059] This embodiment can search for preliminary matching objects that match the profile summary information in the target pool. Then, combined with the decision-maker information, it can further filter the preliminary matching objects to select the objects that the decision-maker has the authority to decide as the final target objects. Then, for the target objects, it can generate a promotion strategy for the target objects by combining the profile summary information and the decision-maker information.

[0060] Optionally, this embodiment can employ an intelligent generation model in conjunction with prompts to generate promotional strategies. The prompts may include input data, generation instructions, role information, generation requirements, thought chain information, script templates, and / or sample data, etc. The content items included in the prompts can be set according to actual circumstances, and this application does not limit them.

[0061] The input data provides contextual information for model reasoning, such as profile summary information, decision-maker information, and target audience. Generation instructions explicitly tell the model what to do, such as "Generate corresponding promotion strategies for the target audience based on the input data." Role information instructs the model to assume a specific role, changing its professional focus, such as "You are a professional promotion and operations expert." Generation requirements specify constraints, such as generating promotion strategies based on communication priorities, script templates, and resource support. Communication priorities are determined by the urgency of the need; for example, high-demand accounts should prioritize service efficiency. Script templates should align with the decision-maker's position; for example, if the decision-maker is the HR director of a recruitment agency, the script template should emphasize improving resume screening efficiency for the target audience. Resource support should be linked to profile summary information; for example, high-performing accounts can receive complimentary top-ranking promotional offers. Thought chain information guides the model's reasoning process, and example data provides learning samples to help the model understand the operations performed. This embodiment leverages the powerful semantic understanding and intelligent reasoning capabilities of the intelligent generative model to improve the accuracy of promotion decisions. This embodiment generates a promotion strategy from three dimensions: communication focus, script templates, and resource support. It forms a progressive promotion strategy loop from awareness to positioning to execution, which helps to improve the conversion rate of the target audience.

[0062] It should be noted that the promotion strategy generated in this embodiment can be used to promote the target audience to the decision-maker's client. One implementation method is to generate promotional information for the target audience based on the promotion strategy. For example, using a large language model, the promotional information for the target audience can be output in natural language description based on the promotion strategy, and then sent to the decision-maker's client for the decision-maker to view.

[0063] Another implementation method is to send the promotion strategy to the service personnel's client, so that the service personnel's client can respond to the information input operation for the promotion strategy, determine the promotion information of the target audience, and send the promotion information to the decision-maker's client. Specifically, after the server sends the promotion strategy to the service personnel's client, the service personnel can determine the promotion information based on the communication points, script templates, resource support, etc. listed in the promotion strategy, and then perform the input operation of the promotion information on their client's chat page with the decision-maker. The service personnel's client obtains the promotion information corresponding to the input operation and sends it to the decision-maker's client. Optionally, after viewing the promotion strategy through their client, the service personnel can also promote the target audience to the decision-maker by establishing a call connection with the decision-maker's terminal; there is no limitation on this. This enables service personnel to accurately target promotion to each user.

[0064] In practical applications, this embodiment can also involve sending the profile summary information and / or decision-maker information to the service personnel's client after determining the decision-maker information and profile summary information. In this embodiment, the service personnel can be staff responsible for target promotion within the online service system. For example, Figure 3 and Figure 4 The diagrams illustrate the sending of profile summary information and decision-maker information to the client. This embodiment sends profile summary information to the service personnel's client, helping them efficiently grasp key information, such as favorable profile tags for the target audience (e.g., high demand, high capability, high conversion rate) and unfavorable profile tags (e.g., low platform activity), thus solving the problem of difficulty in extracting tag information. Sending decision-maker information to the service personnel's client addresses the issue of service personnel being unable to find the contact person when they have target audience promotion needs.

[0065] In this embodiment, the server responds to the strategy generation event by obtaining multi-dimensional target data corresponding to the target account, obtaining the rule profile label and rule confidence level corresponding to the first profile rule matching the multi-dimensional target data, using the profile generation model to identify the basic features of the multi-dimensional target data, and determining the correlation strength between the first profile rule and the basic features. Combining the rule confidence level, the server identifies the rule features of the first profile rule, and then calculates the weighted features of the basic features and rule features based on the correlation strength. Combining the prior features and weighted features of the rule profile label, the server generates the target profile label, determines the decision-maker information of the target account based on the target profile label, and generates profile summary information. Based on the profile summary information and guided by the decision-maker information, the server determines the target audience and generates a promotion strategy for the target audience.

[0066] This embodiment provides a scheme for automatically and accurately constructing portrait tags based on portrait rules and a portrait generation model. The extraction of rule features is determined by rule confidence. Based on the correlation strength between the matched first portrait rule and the basic features of multi-dimensional target data, a weighted fusion of rule features and basic features is driven. Finally, the prior features of the corresponding rule portrait tag of the first portrait rule are fused with the weighted features to generate the target portrait tag. It is evident that this embodiment is not simply using rules first and then the model, but rather organically integrating business logic and intelligent parsing within the model. This achieves deep collaboration between rules and the model rather than mechanical splicing, preserving the interpretability and controllability of the rules while leveraging the model's context awareness and generalization capabilities. Since the initial portrait tags are obtained through rule matching, there is no need to wait for complex reasoning; and the portrait generation model only needs to extract and fuse features in the feature space, significantly improving the efficiency of portrait generation and the real-time nature of promotion decisions. Furthermore, this embodiment introduces correlation strength as a dynamic adjustment signal to dynamically adjust the weights of feature weighting processing, effectively suppressing the interference of low-correlation features on the target portrait tag, significantly improving the robustness and accuracy of the target portrait tag.

[0067] In practical applications, this embodiment can also obtain feedback behavior information regarding the promotion strategy after sending the promotion strategy to the client of the service personnel or decision-maker; and fine-tune the first profile rule and / or the profile generation model based on the feedback behavior information. The feedback behavior information is used to characterize whether the target account completes the expected operation after the promotion target is promoted to the decision-maker's client. For example, if the promotion target is a membership product, the feedback behavior information is whether the target account purchases the membership product after it is promoted to the decision-maker's client according to the promotion strategy. In this embodiment, if the feedback behavior information indicates that the target account has not completed the expected operation (e.g., has not purchased the promotion target), it indicates that the promotion strategy may be inaccurate. This is likely due to inaccurate determination of the target profile tag, and in this case, the first profile rule and / or profile generation model used to determine the target profile tag can be fine-tuned. Optionally, to further improve the accuracy of the fine-tuning results, when the feedback behavior information indicates that the target account has not completed the expected operation, the feedback behavior data in this embodiment may further include feedback information from service personnel, such as the accuracy rate of each target profile, the accuracy rate of decision-maker information, and the effectiveness of the promotion strategy adoption. This feedback information from service personnel can more accurately help identify the problems with inaccurate promotion decisions, thereby allowing for targeted adjustments to the first profile rules and / or the profile generation model to continuously improve the online target promotion effect.

[0068] Next, the training method for the image generation model in this embodiment will be described, which may include the following sub-steps: Sub-step A: Obtain multi-dimensional sample data of the sample accounts, and the labeled profile tags for the multi-dimensional sample data. Specifically, this step can adopt a similar method to obtaining multi-dimensional target data in S201 above to obtain multi-dimensional sample data of the sample accounts, and then determine the profile tags (i.e., label the profile tags) of the multi-dimensional sample data through manual annotation.

[0069] Sub-step B involves obtaining the sample profile labels and sample confidence scores corresponding to the second profile rule matched with multi-dimensional sample data. The specific implementation is similar to S202 and will not be elaborated here.

[0070] Sub-step C: Using multi-dimensional sample data, the second portrait rule and its corresponding sample confidence and sample portrait label as input data, and the labeled portrait label and sample portrait label as training labels, train the portrait generation model.

[0071] Specifically, this embodiment utilizes a portrait generation model, similar to S203, to generate predicted portrait labels based on multi-dimensional sample data, second portrait rules, their corresponding sample confidence scores, and sample portrait labels. For example, the portrait generation model identifies the basic features of the multi-dimensional sample data and determines the sample association strength between the second portrait rules and the basic features; combined with the sample rule confidence score, it identifies the sample rule features of the second portrait rules; based on the sample association strength, it calculates the sample weighted features of the basic features and the sample rule features; and it generates predicted portrait labels by combining the sample prior features and the sample weighted features of the sample portrait labels. Then, the consistency between the predicted portrait labels and the sample portrait labels is used as the first constraint, and the consistency between the predicted portrait labels and the labeled portrait labels is used as the second constraint to train the portrait generation model. In this embodiment, the training of the portrait generation model does not rely solely on labeled portrait labels, but rather employs a dual-source constraint mechanism of labeled portrait labels and sample portrait labels matched based on portrait rules, improving the accuracy and robustness of the model training results. By introducing sample image labels from image rule matching as soft constraints, the model can make reasonable inferences based on the prior features of the rule image labels even when the labeled image labels are missing or inaccurate, thereby improving the generalization ability of the image generation model in small sample scenarios.

[0072] In practical applications, the labeled portrait, predicted portrait, and sample portrait labels may all include portrait labels corresponding to multiple portrait categories. In this case, to further improve the model training efficiency and accuracy, sub-step C can be further implemented through the following sub-steps: Sub-step C1: Using the portrait generation model, based on the multi-dimensional sample data, the second portrait rule and its corresponding sample confidence and sample portrait label, determine the predicted portrait label. The specific implementation method has been described in the above embodiments and will not be repeated here.

[0073] Sub-step C2: For any image category, determine the first deviation value between the predicted image label and the sample image label corresponding to the image category. Specifically, during the model training phase, while the image generation model determines the predicted image label, it can further output the probability value corresponding to the predicted image label. If the predicted image label is different from the sample image label (e.g., one is high platform activity, and the other is low platform activity), then a default larger deviation value is set as the first deviation value. If the predicted image label is the same as the sample image label, then the probability difference between the predicted image label and the sample image label is further calculated as the first deviation value, where the probability value of the sample image label is 1.

[0074] Sub-step C3: Determine at least one deviation profile category whose first deviation value is greater than the deviation threshold, and determine the first loss value based on the first deviation value corresponding to each of the at least one deviation profile category and / or the number of deviation profile categories. Specifically, this can be achieved by fusing (e.g., summing, or weighted summing) the first deviation values ​​corresponding to each of the at least one deviation profile category to obtain the first loss value; alternatively, the first loss value can be increased according to the number of deviation profile categories, such as increasing the first loss value by 5 for each additional deviation profile category.

[0075] Sub-step C4: Determine the second loss value based on the second deviation value between the predicted image label and the labeled image label corresponding to different image categories. Specifically, when calculating the second loss value, the second deviation value between the predicted image label and the labeled image label for the same image category can be calculated as the second loss value. The method for determining the second deviation value is similar to that for determining the first deviation value, and will not be elaborated here.

[0076] Sub-step C5: Train the image generation model based on the first loss value and the second loss value. Specifically, in this embodiment, the image generation model can be trained sequentially based on the first loss value and the second loss value, with minimizing the first loss value and the second loss value as the training objective. Alternatively, the first loss value and the second loss value can be fused (e.g., summed or weighted summed) to obtain the target loss value, and the image generation model can be trained with minimizing the target loss value as the training objective.

[0077] In practical applications, this embodiment can freeze the underlying parameters of the portrait generation model and adjust the adaptation layer parameters of the portrait generation model based on the first loss value and the second loss value. The underlying parameters can be the parameters of the first few layers of the network used to extract general features in the portrait generation model, and the adaptation layer parameters can be the parameters of the top-level network added or fine-tuned for the portrait label generation task of this application. This method, similar to freezing the underlying parameters and adjusting the adaptation layer parameters, fine-tunes the portrait generation model to reduce the computational cost of the model training process, avoid model overfitting, and improve training stability.

[0078] Optionally, based on the above embodiments, this embodiment may further evaluate the trained portrait generation model to determine the prediction accuracy of the portrait generation model for multiple portrait label categories; determine whether the prediction accuracy of the multiple portrait label categories all meet the corresponding accuracy requirements; if yes, then complete the training operation of the portrait generation model; if no, then continue to execute the training operation of the portrait generation model.

[0079] Figure 5 This is a flowchart of an embodiment of a model training method provided in this application. The technical solution of this embodiment can be derived from... Figure 1 The server 102 in the middle executes the model trained in this embodiment, which is the portrait generation model described in the above embodiment. Figure 5 The model training method shown may include the following steps: S501, obtain multi-dimensional sample data of sample accounts, as well as the labeled profile tags of the multi-dimensional sample data.

[0080] S502, obtain the sample profile label and sample confidence level corresponding to the second profile rule matching the multi-dimensional sample data.

[0081] S503 uses multi-dimensional sample data, second profile rules and their corresponding sample confidence and sample profile labels as input data, and labeled profile labels and sample profile labels as training labels to train the profile generation model.

[0082] The profile generation model is used to identify the basic features of multi-dimensional target data corresponding to the target account, and determine the correlation strength between the first profile rule matching the multi-dimensional target data and the basic features. Combining the rule confidence of the first profile rule, the model identifies the rule features of the first profile rule. Based on the correlation strength, the model calculates the weighted features of the basic features and the rule features. The model generates target profile tags by combining the prior features of the rule profile tags and the weighted features. The target profile tags are used to determine the decision-maker information of the target account and generate profile summary information. The decision-maker information and the profile summary information are used to determine the target audience and generate a promotion strategy for the target audience. The promotion strategy is used to promote the target audience to the decision-maker's client.

[0083] In some embodiments, training the portrait generation model using the multi-dimensional sample data, the second portrait rule and its corresponding sample confidence and sample portrait label as input data, and the labeled portrait label and the sample portrait label as training labels includes: using the portrait generation model, determining a predicted portrait label based on the multi-dimensional sample data, the second portrait rule and its corresponding sample confidence and sample portrait label; when the labeled portrait label, the predicted portrait label, and the sample portrait label all include portrait labels corresponding to multiple portrait categories, determining a first deviation value between the predicted portrait label and the sample portrait label corresponding to any portrait category; determining at least one biased portrait category whose first deviation value is greater than a deviation threshold, and determining a first loss value based on the first deviation value and / or the number of biased portrait categories corresponding to the at least one biased portrait category; determining a second loss value based on the second deviation value between the predicted portrait label and the labeled portrait label corresponding to different portrait categories; and training the portrait generation model based on the first loss value and the second loss value.

[0084] In some embodiments, training the portrait generation model based on the first loss value and the second loss value includes: freezing the underlying parameters of the portrait generation model and adjusting the adaptation layer parameters of the portrait generation model based on the first loss value and the second loss value.

[0085] The detailed implementation methods and beneficial effects of each step in this embodiment have been described in detail in the foregoing embodiments, and will not be elaborated here.

[0086] In a practical application scenario, this embodiment of the application takes the registered account of a target enterprise with recruitment needs on an online service system as an example to introduce the promotion strategy generation process of this embodiment. For example... Figure 6As shown, this embodiment's solution is implemented based on a three-layer architecture: a data layer, a model layer, and an application layer. The data layer is primarily used for multi-dimensional data processing (such as providing, cleaning, standardizing, and fusion processing) and the extraction of profiling rules, including a data processing module and a rule extraction module. The model layer is used for model fine-tuning (i.e., training) and the model-based profiling process, including a model fine-tuning module and a profiling generation module. The application layer is used for promoting targets based on the generated target profiling tags, including a strategy analysis model and a sales execution module. Additionally, this embodiment's architecture includes an effect iteration feedback module, used to fine-tune the model or rules based on feedback information after sales promotion. The following will combine... Figure 6 This paper provides a detailed introduction to the entire process of how the above models work together to achieve multi-dimensional data input, profile generation, and business application.

[0087] To achieve accurate construction of user profile tags, this embodiment requires preparation operations including multi-dimensional data processing, profile rule extraction, and model fine-tuning. Specifically, the multi-dimensional data processing stage involves the data processing module cleaning, standardizing, and fusing the multi-dimensional data of each user's corresponding account (including target accounts and sample accounts) in the system, and then storing it in a database for use in rule extraction, model fine-tuning, and user profile tag generation. In the profile rule extraction stage, multi-dimensional historical data provided by the data processing module for the sample accounts is obtained, and then the rule extraction module extracts multiple profile rules from the multi-dimensional historical data to build a rule library.

[0088] In the model fine-tuning stage, the data processing module can obtain multi-dimensional historical data provided by the sample account, match the multi-dimensional historical data with each portrait rule in the rule base, determine the sample portrait label and sample confidence corresponding to the portrait rule (i.e. the second portrait rule) matched by the multi-dimensional sample data, and then combine the multi-dimensional sample data and its labeled portrait labels to fine-tune the portrait generation model through the model fine-tuning module.

[0089] After completing the above preparations, multi-dimensional target data of the target account provided by the data processing module can be obtained and matched with various profile rules in the rule base to determine the rule profile label and rule confidence level corresponding to the profile rule matching the multi-dimensional target data (i.e., the first profile rule). Then, through the profile generation module, the trained profile generation model is called to generate target profile labels based on the multi-dimensional target data, the first profile rule, and its corresponding rule profile label and rule confidence level. The target profile labels are input into the strategy analysis module, which generates the corresponding profile summary letter, decision-maker information, and promotion strategy for the target audience and sends them to the sales execution end. The service personnel at the sales execution end promote the target audience by sending messages to the decision-maker's client or making phone calls. This embodiment can also obtain feedback behavior information on the promotion strategy through the effect iteration module after the promotion operation is completed, so as to fine-tune the profile rules and / or profile generation model.

[0090] The detailed implementation methods and beneficial effects of each step in this embodiment have been described in detail in the foregoing embodiments, and will not be elaborated here.

[0091] It should be noted that some processes described in the above embodiments and accompanying drawings include multiple operations that appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order they appear in this document, or they may be executed in parallel. The sequence numbers of the operations are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should also be noted that the descriptions such as "first" and "second" in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0092] Figure 7 A schematic diagram of a promotion strategy generation apparatus provided for an exemplary embodiment of this application is shown. The apparatus includes: The first data processing module 701 is used to respond to the strategy generation event and obtain multi-dimensional target data corresponding to the target account; The first rule matching module 702 is used to obtain the rule profile label and rule confidence level corresponding to the first profile rule of the multi-dimensional target data matching; The portrait generation module 703 is used to identify the basic features of the multi-dimensional target data using a portrait generation model, and determine the correlation strength between the first portrait rule and the basic features; identify the rule features of the first portrait rule by combining the rule confidence; calculate the weighted features of the basic features and the rule features based on the correlation strength; and generate target portrait labels by combining the prior features of the rule portrait labels and the weighted features; the portrait generation model is trained based on multi-dimensional sample data of sample accounts and their matched second portrait rules and labeled portrait labels, as well as the sample confidence and sample portrait labels corresponding to the second portrait rule. The strategy analysis module 704 is used to determine the decision-maker information of the target account based on the target profile tags, and generate profile summary information; based on the profile summary information and guided by the decision-maker information, determine the target audience and generate a promotion strategy for the target audience; the promotion strategy is used to promote the target audience to the decision-maker's client.

[0093] In some embodiments, the apparatus further includes an effect iteration module, configured to acquire feedback behavior information for the promotion strategy; the feedback behavior information is used to characterize whether the target account completes the expected operation behavior after the promotion object is promoted to the decision-maker's client; and to fine-tune the first profile rule and / or the profile generation model based on the feedback behavior information.

[0094] In some embodiments, the first rule matching module 702 of the device is specifically used to determine at least one account feature item corresponding to any portrait category; determine sub-feature values ​​corresponding to the at least one account feature item according to the multi-dimensional target data; perform weighted fusion processing on the sub-feature values ​​corresponding to the at least one account feature item to obtain the total feature value of the portrait category; take the portrait rule that matches at least one sub-feature value and the total feature value among the at least one portrait rule corresponding to the portrait category as the first portrait rule, and obtain the rule portrait label and rule confidence degree corresponding to the first portrait rule.

[0095] In some embodiments, the device further includes a model fine-tuning module, configured to acquire multi-dimensional sample data of sample accounts and labeled profile tags of the multi-dimensional sample data; acquire sample profile tags and sample confidence scores corresponding to the second profile rule matched by the multi-dimensional sample data; and train the profile generation model using the multi-dimensional sample data, the second profile rule and its corresponding sample confidence scores and sample profile tags as input data, and the labeled profile tags and sample profile tags as training tags.

[0096] In some embodiments, the model fine-tuning module is specifically used to utilize the portrait generation model to determine the predicted portrait label based on the multi-dimensional sample data, the second portrait rule and its corresponding sample confidence and sample portrait label; when the labeled portrait label, the predicted portrait label, and the sample portrait label all include portrait labels corresponding to multiple portrait categories, for any portrait category, determine a first deviation value between the predicted portrait label and the sample portrait label corresponding to the portrait category; determine at least one deviated portrait category whose first deviation value is greater than a deviation threshold, and determine a first loss value based on the first deviation value and / or the number of deviated portrait categories corresponding to the at least one deviated portrait category; determine a second loss value based on the second deviation value between the predicted portrait label and the labeled portrait label corresponding to different portrait categories; and train the portrait generation model based on the first loss value and the second loss value.

[0097] In some embodiments, the model fine-tuning module is further configured to freeze the underlying parameters of the portrait generation model and adjust the adaptation layer parameters of the portrait generation model based on the first loss value and the second loss value.

[0098] In some embodiments, the first data processing module 701 is further configured to acquire multi-dimensional raw data corresponding to the target account; determine at least one service feature tag based on the multi-dimensional raw data; perform cleaning and / or standardization processing on the multi-dimensional raw data and the at least one service feature tag, and associate the processed data with the account identifier of the target account and store it in the database; determine the account identifier of the target account in response to a policy generation event; and search the database for data associated with the account identifier as the multi-dimensional target data corresponding to the target account.

[0099] In some embodiments, the apparatus further includes a sending module for sending the promotion strategy to a service personnel's client, so that the service personnel's client, in response to an information input operation for the promotion strategy, determines the promotion information of the target object and sends the promotion information to the decision-maker's client.

[0100] In some embodiments, the sending module is further configured to send the profile summary information and / or the decision-maker information to the service personnel's client.

[0101] Figure 7 The device for generating the promotion strategy can execute... Figure 2The implementation principle and technical effects of the promotion strategy generation method described in the illustrated embodiment will not be repeated here. The specific methods by which each module and unit of the promotion strategy generation device in the above embodiments performs its operations have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0102] Figure 8 A schematic diagram of a model training apparatus provided for an exemplary embodiment of this application is shown. The apparatus includes: The second data processing module 801 is used to acquire multi-dimensional sample data of the sample account, and the labeled profile tags of the multi-dimensional sample data. The second rule matching module 802 is used to obtain the sample profile label and sample confidence level corresponding to the second profile rule matched by the multi-dimensional sample data. The model training module 803 is used to train the portrait generation model with the multi-dimensional sample data, the second portrait rule and its corresponding sample confidence and sample portrait label as input data, and the labeled portrait label and the sample portrait label as training labels. The profile generation model is used to identify the basic features of multi-dimensional target data corresponding to the target account, and determine the correlation strength between the first profile rule matching the multi-dimensional target data and the basic features. Combining the rule confidence of the first profile rule, the model identifies the rule features of the first profile rule. Based on the correlation strength, the model calculates the weighted features of the basic features and the rule features. The model generates target profile tags by combining the prior features of the rule profile tags and the weighted features. The target profile tags are used to determine the decision-maker information of the target account and generate profile summary information. The decision-maker information and the profile summary information are used to determine the target audience and generate a promotion strategy for the target audience. The promotion strategy is used to promote the target audience to the decision-maker's client.

[0103] In some embodiments, the model training module 803 is specifically used to utilize the portrait generation model to determine the predicted portrait label based on the multi-dimensional sample data, the second portrait rule and its corresponding sample confidence and sample portrait label; when the labeled portrait label, the predicted portrait label, and the sample portrait label all include portrait labels corresponding to multiple portrait categories, for any portrait category, determine a first deviation value between the predicted portrait label and the sample portrait label corresponding to the portrait category; determine at least one biased portrait category whose first deviation value is greater than a deviation threshold, and determine a first loss value based on the first deviation value and / or the number of biased portrait categories corresponding to the at least one biased portrait category; determine a second loss value based on the second deviation value between the predicted portrait label and the labeled portrait label corresponding to different portrait categories; and train the portrait generation model based on the first loss value and the second loss value.

[0104] In some embodiments, the model training module 803 is further configured to freeze the underlying parameters of the portrait generation model and adjust the adaptation layer parameters of the portrait generation model based on the first loss value and the second loss value.

[0105] Figure 8 The model training device described above can perform Figure 5 The implementation principle and technical effects of the model training method described in the illustrated embodiments will not be repeated here. The specific methods by which each module and unit of the model training device in the above embodiments performs its operations have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0106] Figure 9 This is a schematic diagram of the structure of one embodiment of a computing device provided in this application. Figure 9 As shown, in practice, the computing device may include a storage component 901 and a processing component 902.

[0107] Storage component 901 is used to store computer programs and can be configured to store various other data to support operation on a computing device. Examples of this data include instructions for any application or method used to operate on the computing device, data structures, contact data, phone book data, messages, pictures, videos, etc.

[0108] Processing component 902, coupled to storage component 901, is used to execute computer programs in storage component 901 for implementing, etc. Figure 2 The method for generating the promotion strategy shown, or the implementation of such a strategy. Figure 5 The model training method shown.

[0109] Furthermore, such as Figure 9As shown, the computing device may also include other components such as a communication component 903, a display component 904, a power supply component 905, and an audio component 906. Figure 9 The diagram only shows some components and does not mean that the device includes only these components. Figure 9 The components shown. Additionally... Figure 9 The components within the dashed box are optional, not mandatory, and their specific requirements depend on the product form of the computing device. The computing device in this embodiment can be a terminal device such as a desktop computer, laptop computer, smartphone, or IoT (Internet of Things) device, or a server-side device such as a conventional server, cloud server, or server array. If the computing device in this embodiment is implemented as a terminal device such as a desktop computer, laptop computer, or smartphone, it may include... Figure 9 The components within the dashed box; if the computing device in this embodiment is implemented as a conventional server, cloud server, or server array, etc., then it may not include... Figure 9 The component within the dashed box.

[0110] The processing component described above includes one or more processors to execute computer instructions to complete all or part of the steps in the method described above. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the method described above.

[0111] The aforementioned storage components can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0112] The aforementioned communication component is configured to facilitate wired or wireless communication between the device housing the communication component and other devices. The device housing the communication component can access wireless networks based on communication standards, such as mobile communication networks, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel.

[0113] The aforementioned display components may include a screen, which may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0114] The aforementioned power supply components provide power to various components within the device in which they reside. These power supply components may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device in which they reside.

[0115] The aforementioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals can be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0116] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the above-described method embodiments. The computer-readable storage medium includes volatile or non-volatile components, or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium. Accordingly, this application also provides a computer program product, which includes a computer program or instructions that, when executed by a processor, cause the processor to implement the steps in the above method embodiments. It should be understood that each step or combination of steps in the above method flow can be implemented by the computer program or instructions. Furthermore, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, enabling the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to function as an apparatus for implementing the corresponding functions in the above method embodiments.

[0117] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0118] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0119] Finally, it should be noted that the above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for generating a promotion strategy, characterized in that, Applied to the server side, the method includes: In response to the strategy generation event, obtain multi-dimensional target data corresponding to the target account; Obtain the rule profile label and rule confidence level corresponding to the first profile rule matching the multi-dimensional target data; Using a profile generation model, the basic features of the multi-dimensional target data are identified, and the correlation strength between the first profile rule and the basic features is determined; combined with the rule confidence, the rule features of the first profile rule are identified; based on the correlation strength, the weighted features of the basic features and the rule features are calculated; and the target profile label is generated by combining the prior features of the rule profile label and the weighted features; the profile generation model is trained based on multi-dimensional sample data of sample accounts and their matched second profile rules and labeled profile labels, as well as the sample confidence and sample profile labels corresponding to the second profile rule. Based on the target profile tags, determine the decision-maker information of the target account and generate profile summary information; Based on the profile summary information and guided by the decision-maker information, the target audience is identified, and a promotion strategy is generated for the target audience; the promotion strategy is used to promote the target audience to the decision-maker's client.

2. The method according to claim 1, characterized in that, Also includes: Obtain feedback behavior information regarding the promotion strategy; The feedback behavior information is used to characterize whether the target account has completed the expected operation behavior after the promotion target is promoted to the decision-maker's client. Based on the feedback behavior information, the first portrait rule and / or the portrait generation model are fine-tuned.

3. The method according to claim 1, characterized in that, The rule profile label and rule confidence level corresponding to the first profile rule for matching the multi-dimensional target data include: For any given profile category, determine at least one account feature corresponding to that profile category; Based on the multi-dimensional target data, determine the sub-feature values ​​corresponding to each of the at least one account feature item; The sub-feature values ​​corresponding to each of the at least one account feature item are weighted and fused to obtain the total feature value of the portrait category; The first portrait rule is selected from at least one portrait rule corresponding to the portrait category that matches at least one sub-feature value and the total feature value. The rule portrait label and rule confidence corresponding to the first portrait rule are then obtained.

4. The method according to any one of claims 1-3, characterized in that, The image generation model was trained in the following manner: Obtain multi-dimensional sample data of the sample accounts, and the labeled profile tags of the multi-dimensional sample data; Obtain the sample profile labels and sample confidence scores corresponding to the second profile rule matched by the multi-dimensional sample data; The portrait generation model is trained using the multi-dimensional sample data, the second portrait rule and its corresponding sample confidence and sample portrait label as input data, and the labeled portrait label and the sample portrait label as training labels.

5. The method according to claim 4, characterized in that, The step of training the portrait generation model using the multi-dimensional sample data, the second portrait rule and its corresponding sample confidence and sample portrait labels as input data, and the labeled portrait labels and the sample portrait labels as training labels, includes: Using the aforementioned profile generation model, based on the multi-dimensional sample data, the second profile rule and its corresponding sample confidence and sample profile label, the predicted profile label is determined; When the labeled image tag, the predicted image tag, and the sample image tag all include image tags corresponding to multiple image categories, for any image category, a first deviation value between the predicted image tag and the sample image tag corresponding to the image category is determined; Identify at least one deviation profile category whose first deviation value is greater than a deviation threshold, and determine a first loss value based on the first deviation value and / or the number of deviation profile categories corresponding to the at least one deviation profile category. The second loss value is determined based on the second deviation value between the predicted image label and the labeled image label corresponding to different image categories; The image generation model is trained based on the first loss value and the second loss value.

6. The method according to claim 5, characterized in that, The step of training the portrait generation model based on the first loss value and the second loss value includes: Freeze the underlying parameters of the portrait generation model, and adjust the adaptation layer parameters of the portrait generation model based on the first loss value and the second loss value.

7. The method according to any one of claims 1-3, characterized in that, Also includes: Obtain multi-dimensional raw data corresponding to the target account; Based on the multi-dimensional raw data, at least one service characteristic tag is determined; The multi-dimensional raw data and at least one service feature tag are cleaned and / or standardized, and the processed data is associated with the account identifier of the target account and stored in the database. The process of obtaining multi-dimensional target data corresponding to the target account in response to the strategy generation event includes: In response to the policy generation event, determine the account identifier of the target account; The database is used to retrieve data associated with the account identifier as multi-dimensional target data corresponding to the target account.

8. The method according to any one of claims 1-3, characterized in that, Also includes: The promotion strategy is sent to the service personnel's client, so that the service personnel's client can respond to the information input operation for the promotion strategy, determine the promotion information of the target, and send the promotion information to the decision-maker's client.

9. The method according to claim 8, characterized in that, Also includes: The profile summary information and / or the decision-maker information are sent to the service personnel's client.

10. A model training method, characterized in that, include: Obtain multi-dimensional sample data of the sample accounts, and the labeled profile tags of the multi-dimensional sample data; Obtain the sample profile labels and sample confidence scores corresponding to the second profile rule matched by the multi-dimensional sample data; The portrait generation model is trained using the multi-dimensional sample data, the second portrait rule and its corresponding sample confidence and sample portrait label as input data, and the labeled portrait label and the sample portrait label as training labels. The profile generation model is used to identify the basic features of multi-dimensional target data corresponding to the target account, determine the correlation strength between the first profile rule matching the multi-dimensional target data and the basic features, and identify the rule features of the first profile rule by combining the rule confidence of the first profile rule. Based on the correlation strength, calculate the weighted features of the basic features and the rule features; The target profile label is generated by combining the prior features of the rule profile label and the weighted features; the target profile label is used to determine the decision-maker information of the target account and to generate profile summary information; the decision-maker information and the profile summary information are used to determine the target audience and to generate a promotion strategy for the target audience; the promotion strategy is used to promote the target audience to the decision-maker's client.

11. A computing device, characterized in that, This includes processing components and storage components; The storage component stores a computer program; the computer program is invoked and executed by the processing component to implement the method for generating the promotion strategy as described in any one of claims 1-9, or to implement the model training method as described in claim 10.

12. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processing component, implements the method for generating a promotion strategy as described in any one of claims 1-9, or the method for training a model as described in claim 10.

13. A computer program product, characterized in that, It includes a computer program or instructions that, when executed by a processing component, implement the method for generating a promotion strategy as described in any one of claims 1-9, or the model training method as described in claim 10.

Images