A text-based business recommendation method, apparatus, storage medium, and device

Abstract

This specification discloses a text-based service recommendation method. The method inputs the target user's user characteristic information into a causal inference model. The causal inference model determines the probability gain that the target user will perform the recommended service after sending recommended text under each preset topic. Based on the probability gain, the target topic to which the recommended text to be sent to the target user belongs is determined. Then, the target topic and user characteristic information are input into an LLM (Limited Language Model), which generates recommended text under that target topic and sends it to the target user. This method allows the causal inference model to predict the impact of each preset topic on whether the target user will perform the recommended service, determines the target topic based on this impact, and then uses an LLM to generate recommended text under that target topic, enabling efficient and accurate delivery of recommended text to the target user.

Description

Technical Field

[0001] This specification relates to the field of computer technology, and in particular to a text-based business recommendation method, apparatus, storage medium, and device. Background Technology

[0002] With the development of the times, all kinds of services for users are emerging. For service providers, how to recommend their services to users is a perennial problem.

[0003] Currently, one common method for recommendation services is to send recommendation texts to users via SMS or instant messaging. However, recommendation texts are usually pre-written manually, resulting in a long update cycle. This is not only time-consuming and labor-intensive, but also inefficient. Furthermore, due to the significant differences in preferences among different users, the same recommendation text is unlikely to pique the interest of different users.

[0004] Therefore, how to efficiently and accurately send recommendation texts for recommendation services to different users is an urgent problem to be solved. Summary of the Invention

[0005] This specification provides a text-based business recommendation method, apparatus, storage medium, and electronic device to partially solve the problems existing in the prior art.

[0006] The embodiments in this specification adopt the following technical solutions: This specification provides a text-based business recommendation method, the method comprising: For each target user, the user characteristic information of that target user is input into the causal inference model; Using the causal inference model, for each preset topic corresponding to the service to be recommended, the probability gain of the target user executing the service to be recommended after sending the recommended text under the preset topic to the target user is predicted; Based on the probability gain determined for each preset topic, select the target topic from among the preset topics; The user feature information and the target topic are input into a large language model (LLM). The LLM generates recommended text under the target topic, which is then used as the target text. The target text is sent to the target user.

[0007] This specification provides a text-based business recommendation device, the device comprising: The first input module is used to input the user characteristic information of each target user into the causal inference model; The topic prediction module is used to predict, through the causal inference model, the probability gain of the target user executing the recommended service after sending the recommended text under the preset topic to the target user for each preset topic corresponding to the recommended service; The topic selection module is used to select a target topic from each preset topic based on the probability gain determined for each preset topic. The second input module is used to input the user feature information and the target topic into the Large Language Model (LLM), and generate recommended text under the target topic through the LLM as the target text. The sending module is used to send the target text to the target user.

[0008] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned text-based business recommendation method.

[0009] This specification provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned text-based business recommendation method.

[0010] This specification provides a computer program product, which includes a computer program that, when executed by a processor, implements the aforementioned text-based business recommendation method.

[0011] The above-described at least one technical solution used in the embodiments of this specification can achieve the following beneficial effects: This specification discloses a text-based service recommendation method. The method inputs the target user's user characteristic information into a causal inference model. The causal inference model determines the probability gain that the target user will perform the recommended service after sending recommended text under each preset topic. Based on the probability gain, the target topic to which the recommended text to be sent to the target user belongs is determined. Then, the target topic and user characteristic information are input into an LLM (Limited Language Model), which generates recommended text under that target topic and sends it to the target user. This method allows the causal inference model to predict the impact of each preset topic on whether the target user will perform the recommended service, determines the target topic based on this impact, and then uses an LLM to generate recommended text under that target topic, enabling efficient and accurate delivery of recommended text to the target user. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a text-based business recommendation method provided in the embodiments of this specification; Figure 2 A schematic diagram of a text-based service recommendation device provided in the embodiments of this specification; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this specification. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification 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 specification, and not all of them. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this specification.

[0014] The technical solutions provided in the various embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0015] Figure 1 This is a schematic diagram of a text-based business recommendation method provided in the embodiments of this specification, which specifically includes the following steps: S100: For each target user, input the user characteristic information of that target user into the causal inference model.

[0016] In the embodiments described in this specification, the following are employed: Figure 1 The device recommended for performing the business operations as shown can be any electronic device, including but not limited to personal computers, mobile phones, tablets, servers, or server clusters consisting of multiple servers. The following explanation uses a backend server as an example.

[0017] In the embodiments described in this specification, the target user refers to the user to whom business recommendations are needed, while the non-target user refers to the user to whom business recommendations are not needed.

[0018] In real-world applications, for any given service, certain users (target users) require service recommendations. These target users might be interested in certain descriptions within the recommended text, thus prompting them to execute the service. For example, in a lending service, target users need to see the server-pushed recommendation text. Some target users might be interested in the "credit limit" mentioned in the recommendation text and execute the loan, while others might be interested in the "interest rate." Therefore, the server needs to send recommendation text to target users, and the content of the recommendation text must be engaging for them to execute the target service. However, for non-target users, either no recommendation text will pique their interest, or they will execute the service even without any recommendation text. Therefore, expending significant resources to push recommendation text to these non-target users is a huge waste, and this specification aims to avoid pushing recommendation text to them as much as possible.

[0019] In addition, for a business to be recommended, when generating corresponding recommendation text for the business to be recommended, recommendation texts with different themes can also be generated according to the type, attributes and other information of the business to be recommended. For example, the lending business in the above example mainly involves different themes such as credit, credit limit, interest rate and period. For a target user, the attractiveness of recommendation texts with different themes to the target user is not the same. Therefore, in the embodiments of this specification, the server also needs to determine the attractiveness of different themes of the business to be recommended to a target user.

[0020] Specifically, for any target user, the server can input the user's characteristic information into a causal inference model to determine the attractiveness of recommended texts under different topics of the recommended service to the target user.

[0021] The aforementioned user characteristic information includes, but is not limited to, user profile information such as gender, age, and location, as well as user historical behavior information.

[0022] S101: Using the causal inference model, for each preset topic corresponding to the service to be recommended, predict the probability gain of the target user executing the service to be recommended after sending the recommended text under the preset topic to the target user.

[0023] The aforementioned causal inference models may include the Uplift model or other causal inference models, and this specification does not impose any restrictions on them.

[0024] The causal inference model described in this specification does not actually predict whether a target user will perform the recommended service if recommended text on a certain topic is pushed to that target user. Rather, it predicts how much the probability of the target user performing the recommended service will increase if recommended text on a certain topic is pushed to that target user compared to not pushing any recommended text to that target user. In other words, it predicts how much additional improvement the intervention action (i.e., pushing push text on a certain topic) will bring to the target user. This is the probability gain described in this specification.

[0025] The server uses this causal inference model to predict the probability gain of a target user executing the recommended service after sending recommended text under that preset topic to the target user for each preset topic corresponding to the aforementioned recommended service. This distinguishes between target and non-target users. Specifically, if the probability gain of all preset topics for the recommended service for the user is less than a set threshold, then any recommended text for that service has little effect on the user, indicating that the user is not a target user and should not be further recommended. Conversely, if some probability gains of all preset topics for the recommended service are not less than the set threshold, then the recommended text under certain preset topics is attractive to the user, indicating that the user is a target user and should be further recommended. On the other hand, this can also determine which preset topic of the recommended service has the greatest impact on a target user, so that the target text under the preset topic with the greatest impact can be generated in a targeted manner for that target user.

[0026] S102: Select the target topic from each preset topic based on the probability gain determined for each preset topic.

[0027] After obtaining the probability gain of executing the recommended service on the target user for each preset topic using a causal inference model, the server can select the preset topic with the highest probability gain from among the preset topics of the recommended service as the target topic. The recommended text under this target topic is the recommended text that has the greatest impact on the target user.

[0028] S103: Input the user feature information and the target topic into the Large Language Model (LLM), and generate recommended text under the target topic through the LLM, which serves as the target text.

[0029] In this embodiment of the specification, after the server selects the target topic, it can input the user feature information of the target user and the target topic determined in step S102 into the LLM. Through the LLM, recommended text under the target topic is generated for the target user as the target text.

[0030] This is because even within the same preset topic for the recommended business, the same recommended text under that topic can have varying impacts on different target users. For example, in the lending business example above, the topic of "interest rates" might have different effects on different target users, even though the core message of the recommended texts under that topic is always "low interest rates." Some target users might prefer well-formatted and rigorously worded texts, while others might prefer humorous and elaborate texts. Therefore, the server needs to use LLM (Local Management Model) to generate recommended text tailored to the target user's specific characteristics within that topic—the target text. Even within the same target topic, the target text generated for different target users will differ, achieving a personalized business recommendation experience.

[0031] S104: Send the target text to the target user.

[0032] After generating the target text under the aforementioned target topic, the server can then send the target text to the target user.

[0033] Using the above method, the causal inference model predicts the probability gain of each preset topic of the recommended service on the target user. Subsequently, the service can be recommended only to the target user. On the one hand, the resources consumed by the service recommendation are not wasted on non-target users. On the other hand, the target topic with the greatest impact on the target user can be identified. Then, LLM is used to generate recommendation text suitable for the target user under the target topic, which can send recommendation text to the target user efficiently and accurately.

[0034] Furthermore, the server in this specification can specifically be a server used to provide the aforementioned recommended services to users. The server then performs the aforementioned actions for each target user. Figure 1 After the recommended service is displayed, after a preset time period (such as one day, one week, or one month), it can be determined for each user whether the user has executed the recommended service. The execution result is either that the recommended service was executed or not. Each user mentioned above includes both target users and non-target users.

[0035] For each preset topic of the service to be recommended, the server can determine, based on whether all users have executed the service to be recommended, the first probability that a user will execute the service to be recommended if a recommended text under the preset topic has been sent to a user, and the second probability that a user will execute the service to be recommended if a recommended text under the preset topic has not been sent to a user.

[0036] Specifically, the server can divide all users into four groups: target users who performed the recommended service, target users who did not perform the recommended service, non-target users who performed the recommended service, and non-target users who did not perform the recommended service, denoted as Group 1 to Group 4, respectively. After the preset time period, the server can count the number of users in each group, and then determine the ratio of the number of users in Group 1 to the number of all target users (i.e., the sum of the number of users in Group 1 and Group 2) as the first probability, and determine the ratio of the number of users in Group 3 to the number of all non-target users (i.e., the sum of the number of users in Group 3 and Group 4) as the second probability.

[0037] After determining the first and second probabilities, the server can determine the difference between the first and second probabilities as the probability gain label corresponding to the preset topic. This probability gain label indicates the actual probability gain of the recommended text under the preset topic for target users who did not receive any recommended text after the target text was sent to all target users last time. In other words, how much better the actual execution result of the recommended service was for target users after the target text was sent to all target users last time because the target text under the preset topic was sent.

[0038] The probability gain label can be used to train the aforementioned causal inference model. That is, after obtaining the probability gain label, the server can treat all users as sample users and train the causal inference model based on the probability gain label corresponding to the preset topic and the user feature information of each sample user. This training process is the outer loop in the embodiments of this specification; that is, while employing… Figure 1 The method shown performs business recommendations, while training a causal inference model based on whether the user executes the recommended business after the recommendation.

[0039] When training the causal inference model, the server can input the user feature information of each sample user and each preset topic into the causal inference model to obtain the probability gain predicted by the causal inference model of the sample user executing the recommended service after sending the recommended text under the preset topic to the sample user. Next, if the sample user is the target user, the model parameters of the causal inference model are adjusted with the training objective of reducing the difference between the probability gain predicted by the causal inference model and the probability gain label corresponding to the preset topic; if the sample user is not the target user, the model parameters are adjusted with the training objective of reducing the difference between the probability gain predicted by the causal inference model and 0. That is, a supervised training method is used, using the actual execution results of whether the user executes the recommended service after the previous service recommendation as training data to train the causal inference model.

[0040] In addition to training the causal inference model through the outer loop mentioned above, the server in this specification can also fine-tune the LLM training, which is the inner loop in this specification.

[0041] Specifically, the server can also use... Figure 1 After a preset time period following the service recommendation process, the method shown determines whether all target users who sent the target text under the target topic have executed the recommended service. This execution result is also either that the recommended service was executed or not. It should be noted that, unlike the outer loop, the inner loop is only for target users, as it is used to train the LLM for generating the recommended text, and non-target users do not need to send recommended text.

[0042] After determining whether the target user who received the target text under the target topic has executed the recommended service, the server can select seed texts from the target texts under the target topic that have been sent, and obtain more derivative texts based on the execution results of the target users who received these target texts.

[0043] Specifically, for a target user who receives target text under the target topic, if the target user performs the recommended service, it indicates that the target text sent to the target user is of high quality; if the target user does not perform the recommended service, it indicates that the target text sent to the target user is of low quality. Therefore, when selecting seed texts from the target texts under the target topic that have been sent, a specified number of high-quality target texts and a specified number of low-quality target texts can be selected.

[0044] Specifically, it can identify the first target user who executed the pending service after sending the target text under the target topic, and the second target user who did not execute the pending service after sending the target text under the target topic. A first seed text is selected from the target texts sent to each first target user, and a second seed text is selected from the target texts sent to each second target user. Specifically, the target texts sent to each first target user can be input into an LLM (List of Target Text Managers), which performs a quality evaluation on the target texts sent to each first target user and selects a specified number of target texts with the highest quality evaluation as the first seed texts. Similarly, the target texts sent to each second target user can be input into an LLM, which performs a quality evaluation on the target texts sent to each second target user and selects a specified number of target texts with the lowest quality evaluation as the second seed texts.

[0045] In addition to the methods described above, when selecting the first seed text, for each type of target text sent to each first target user, the number of first target users who executed the recommended service after sending that target text can be determined. The target texts sent to each first target user are then sorted in descending order of the number of first target users who executed the recommended service, and the top n target texts in the sorted order are selected as the first seed texts, where n is a preset first positive integer. Similarly, when selecting the second seed text, for each type of target text sent to each second target user, the number of second target users who did not execute the recommended service after sending that target text can be determined. The target texts sent to each second target user are then sorted in descending order of the number of second target users who did not execute the recommended service, and the top m target texts in the sorted order are selected as the second seed texts, where m is a preset second positive integer. This eliminates the need for LLM to evaluate the quality of each target text sent under the target topic.

[0046] In summary, for a given target topic, it is sufficient to select the high-quality and low-quality target texts generated during the previous business recommendation process for subsequent training.

[0047] After selecting seed texts, the server can input each selected seed text into the LLM (Limited Learning Model). The LLM generates two or more derivative texts corresponding to each seed text to expand the samples for subsequent training of the LLM. Then, all target users who sent target texts under the target topic are treated as sample users, and each generated derivative text is used as a sample text. Based on the execution results of all target users who received various sub-texts, the quality representation value corresponding to each sample text is determined. Finally, the user feature information of each sample user and the target topic are input into the LLM, and a reinforcement learning training method is used to fine-tune the model parameters of the LLM by improving the quality representation value of the recommended texts under the target topic generated by the LLM.

[0048] Specifically, the Direct Preference Optimization (DPO) reinforcement learning method can be used to fine-tune the model parameters of this LLM. When using DPO to fine-tune the LLM, it is only necessary to input the user feature information of the sample users and the target topic into the LLM, and then use the DPO reinforcement learning method to improve the similarity between the LLM output recommendation text and the first seed text or its corresponding derivative texts in the sample text, and reduce the similarity between the LLM output recommendation text and the second seed text or its corresponding derivative texts in the sample text. This is because DPO reinforcement learning no longer needs to determine a specific score (i.e., quality representation value) for each sample text, but instead shows the LLM a series of "good and bad comparison" examples, that is, the first seed text or its corresponding derivative text in the sample text and the second seed text or its corresponding derivative text. The DPO algorithm will directly adjust the LLM model parameters so that when the model sees the same user feature information and target topic, the probability of generating text similar to the first seed text is much greater than the probability of generating text similar to the second seed text.

[0049] If a conventional reinforcement learning method is used, it is necessary to determine the quality representation value corresponding to each sample text, and to adjust the model parameters of the LLM by improving the quality representation value of the recommended texts under the target topic generated by the LLM.

[0050] When determining the quality characterization value for each sample text, for each sample text, if the seed text that generated the sample text is a first seed text, then the quality characterization value of the sample text is determined based on the number of first target users who executed the recommended service after sending the first seed text. The quality characterization value of the sample text is positively correlated with the number of first target users who executed the recommended service, that is, the more first target users who executed the recommended service because of the first seed text, the higher the quality characterization value of the sample text generated based on the first seed text. If the seed text that generated the sample text is a second seed text, then the quality characterization value of the sample text is determined based on the number of second target users who did not execute the recommended service after sending the second seed text. The quality characterization value of the sample text is negatively correlated with the number of second target users who did not execute the recommended service, that is, the more second target users who did not execute the recommended service after sending the second seed text, the lower the quality characterization value of the sample text generated based on the second seed text.

[0051] Through the outer and inner loops described above, each iteration uses the following... Figure 1 Following the business recommendation methods shown, the causal inference model and LLM can be trained based on whether the user has executed the recommended business. This allows the causal inference model to increasingly accurately distinguish between target users and non-target users, and to increasingly accurately determine the probability gain of each preset topic corresponding to the recommended business for the target user to execute the recommended business. In turn, it can accurately determine the target topic that is most attractive to the target user. The LLM can also increasingly accurately generate recommendation text that is more attractive to the target user under the target topic based on the target user's user characteristics and the target topic. This allows the entire business recommendation to enter a virtuous cycle, and the efficiency and accuracy of business recommendations will become higher and higher.

[0052] The above is a text-based business recommendation method provided by the embodiments of this specification. Based on the same idea, this specification also provides corresponding devices, storage media and electronic devices.

[0053] Figure 2 This is a schematic diagram of a text-based business recommendation device provided in an embodiment of this specification. The device includes: The first input module 200 is used to input the user characteristic information of each target user into the causal inference model for each target user. The topic prediction module 201 is used to predict, through the causal inference model, the probability gain of the target user executing the recommended service after sending the recommended text under the preset topic to the target user for each preset topic corresponding to the recommended service; The topic selection module 202 is used to select a target topic from each preset topic based on the probability gain determined for each preset topic; The second input module 203 is used to input the user feature information and the target topic into the Large Language Model (LLM), and generate recommended text under the target topic through the LLM as the target text; The sending module 204 is used to send the target text to the target user.

[0054] Optionally, the device further includes: Training module 205 is used to determine, after a preset time period, whether all users have executed the recommended service; the "all users" include target users and non-target users; for each preset topic, based on the execution results of all users, a first probability is determined that the user executes the recommended service when a recommended text under the preset topic is sent to the user, and a second probability is determined that the user executes the recommended service when a recommended text under the preset topic is not sent to the user; based on the first probability and the second probability, a probability gain label corresponding to the preset topic is determined; all users are used as sample users, and the causal inference model is trained based on the probability gain label corresponding to the preset topic and the user feature information of each sample user.

[0055] Optionally, the training module 205 is specifically used to: determine the difference between the first probability and the second probability as the probability gain label corresponding to the preset topic; for each sample user and each preset topic, input the user feature information of the sample user into the causal inference model to obtain the probability gain predicted by the causal inference model when the sample user performs the recommended service after the recommended text under the preset topic is sent to the sample user; when the sample user is a target user, adjust the model parameters of the causal inference model with the training objective of reducing the difference between the probability gain predicted by the causal inference model and the probability gain label corresponding to the preset topic; when the sample user is not a target user, adjust the model parameters of the causal inference model with the training objective of reducing the difference between the probability gain predicted by the causal inference model and 0.

[0056] Optionally, the device further includes: The training module 205 is used to determine, after a preset time period, whether all target users who sent target text under the target topic have executed the recommended service; based on the execution results of all target users who sent target text under the target topic, select seed text from the target texts under the target topic that have been sent; input the seed text into the LLM, and generate corresponding derived texts through the LLM; use all target users who sent target text under the target topic as sample users, and use each derived text as sample text, and determine the quality representation value corresponding to each sample text based on the execution results of all target users who sent the seed text; input the user feature information of the sample users and the target topic into the LLM, and use a reinforcement learning method to adjust the model parameters of the LLM by improving the quality representation value of the recommended text under the target topic generated by the LLM as the training objective.

[0057] Optionally, the training module 205 is specifically used to: determine a first target user who executed the pending service after sending the target text under the target topic, and a second target user who did not execute the pending service after sending the target text under the target topic; select a first seed text from the target text sent to each first target user, and select a second seed text from the target text sent to each second target user.

[0058] Optionally, the training module 205 is specifically used to: determine the number of first target users who executed the recommendation service after sending each type of target text to each first target user; sort each type of target text sent to each first target user in descending order of the number of first target users who executed the recommendation service; and select the top n sorted target texts as first seed texts; where n is a preset first positive integer. For each type of target text sent to each second target user, determine the number of second target users who did not execute the recommended service after sending the target text; sort each type of target text sent to each second target user in descending order of the number of second target users who did not execute the recommended service; select the top m target texts in the sorting as each second seed text; m is a preset second positive integer.

[0059] Optionally, the training module 205 is specifically configured to: when the seed text for generating the sample text is a first seed text, determine the quality characterization value of the sample text based on the number of first target users who executed the service to be recommended after sending the first seed text, wherein the quality characterization value of the sample text is positively correlated with the number of first target users who executed the service to be recommended; and when the seed text for generating the sample text is a second seed text, determine the quality characterization value of the sample text based on the number of second target users who did not execute the service to be recommended after sending the second seed text, wherein the quality characterization value of the sample text is negatively correlated with the number of second target users who did not execute the service to be recommended.

[0060] This specification also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can be used to perform the text-based business recommendation method described above.

[0061] This specification also provides a computer program product comprising a computer program that, when executed by a processor, implements the text-based business recommendation method described above.

[0062] based on Figure 1 The text-based business recommendation method shown in this specification also provides embodiments that... Figure 3 The diagram shows the structure of the electronic device. Figure 3 At the hardware level, the electronic device includes a processor, internal bus, network interface, memory, and non-volatile storage, and may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile storage into memory and then runs it to implement the text-based business recommendation method described above.

[0063] The above description is merely an embodiment of this specification and is not intended to limit this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this specification.

Claims

1. A text-based business recommendation method, the method comprising: For each target user, the user characteristic information of that target user is input into the causal inference model; Using the causal inference model, for each preset topic corresponding to the service to be recommended, the probability gain of the target user executing the service to be recommended after sending the recommended text under the preset topic to the target user is predicted; Based on the probability gain determined for each preset topic, select the target topic from among the preset topics; The user feature information and the target topic are input into a large language model (LLM). The LLM generates recommended text under the target topic, which is then used as the target text. The target text is sent to the target user.

2. The method of claim 1, further comprising: After a preset time period, it is determined whether all users have executed the recommended service; the "all users" include each target user and each non-target user. For each preset topic, based on the execution results of all users, determine a first probability that a user will execute the recommended service when a recommended text under that preset topic is sent to the user, and a second probability that a user will execute the recommended service when a recommended text under that preset topic is not sent to the user; Based on the first probability and the second probability, determine the probability gain label corresponding to the preset topic; All users are used as sample users, and the causal inference model is trained based on the probability gain label corresponding to the preset topic and the user feature information of each sample user.

3. The method as described in claim 2, wherein determining the probability gain label corresponding to the preset topic based on the first probability and the second probability specifically includes: The difference between the first probability and the second probability is determined as the probability gain label corresponding to the preset topic; The causal inference model is trained based on the probability gain annotation and the user feature information of each sample user, specifically including: For each sample user and each preset topic, the user feature information of the sample user is input into the causal inference model to obtain the probability gain of the sample user executing the recommended service after the recommended text under the preset topic is sent to the sample user. When the sample user is the target user, the model parameters of the causal inference model are adjusted with the training objective of reducing the difference between the probability gain predicted by the causal inference model and the probability gain label corresponding to the preset topic. When the sample user is not the target user, the model parameters of the causal inference model are adjusted with the training objective of reducing the difference between the probability gain predicted by the causal inference model and 0.

4. The method of claim 1, further comprising: After a preset time period, determine whether all target users who sent the target text under the target topic have executed the execution result of the service to be recommended; Based on the execution results of all target users who have sent target text under the target topic, select seed text from the target texts that have been sent under the target topic; The seed text is input into the LLM, and the LLM generates the corresponding derived texts of the seed text. All target users who sent the target text under the target topic are taken as sample users, and all derived texts are taken as sample texts. Based on the execution results of all target users who sent the seed text, the quality characterization value corresponding to each sample text is determined. The user feature information of the sample users and the target topic are input into the LLM. A reinforcement learning method is used to improve the quality representation value of the recommended text under the target topic generated by the LLM as a training objective, and the model parameters of the LLM are adjusted.

5. The method as described in claim 4, wherein, based on the execution results of all target users who have sent target text under the target topic, seed text is selected from the target texts already sent under the target topic, specifically including: The system identifies a first target user who executed the pending service after sending the target text under the target topic, and a second target user who did not execute the pending service after sending the target text under the target topic. Select the first seed text from the target texts sent to each first target user, and select the second seed text from the target texts sent to each second target user.

6. The method as described in claim 5, wherein selecting a first seed text from the target text sent to each first target user specifically includes: For each type of target text sent to each first target user, determine the number of first target users who executed the recommended service after sending that target text; Sort each type of target text sent to each first target user in descending order of the number of first target users who have executed the recommended service; Select the top n target texts in the sorted order as the first seed texts for each target text. n is a pre-defined first positive integer; Select the second seed text from the target text sent to each second target user, specifically including: For each type of target text sent to each second target user, determine the number of second target users who did not execute the recommended service after sending the target text; Sort each type of target text sent to each second target user in descending order of the number of second target users who have not performed the recommended service; Select the first m target texts sorted and use them as the second seed texts; m is a preset second positive integer.

7. The method as described in claim 6, wherein determining the quality characterization value corresponding to each sample text based on the execution results of all target users who sent the seed text, specifically includes: When the seed text for generating the sample text is the first seed text, the quality characterization value of the sample text is determined based on the number of first target users who executed the service to be recommended after the first seed text was sent. The quality characterization value of the sample text is positively correlated with the number of first target users who executed the service to be recommended. When the seed text for generating the sample text is the second seed text, the quality characterization value of the sample text is determined based on the number of second target users who did not execute the recommended service after the second seed text was sent. The quality characterization value of the sample text is negatively correlated with the number of second target users who did not execute the recommended service.

8. A text-based business recommendation device, the device comprising: The first input module is used to input the user characteristic information of each target user into the causal inference model; The topic prediction module is used to predict, through the causal inference model, the probability gain of the target user executing the recommended service after sending the recommended text under the preset topic to the target user for each preset topic corresponding to the recommended service; The topic selection module is used to select a target topic from each preset topic based on the probability gain determined for each preset topic. The second input module is used to input the user feature information and the target topic into the Large Language Model (LLM), and generate recommended text under the target topic through the LLM as the target text. The sending module is used to send the target text to the target user.

9. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method according to any one of claims 1-7.

11. A computer program product comprising a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images