CN-122001844-A - Message pushing method, device, equipment and medium

Abstract

The embodiment of the application provides a message pushing method, a device, equipment and a medium, which are characterized in that a template selection model is called by responding to a message pushing request, a target message template related to a target message scene is selected from a verification template set and an experiment template set based on user characteristic data of a target user, wherein a message document of the target message template is generated in advance by a generation type model, meanwhile, a time prediction model is called, the target pushing time is predicted based on historical behavior data of the target user, further, the target message template is pushed to the target user based on the target pushing time, on-line index data corresponding to the target message template is obtained after the message is pushed, the on-line index data comprises a cumulative click rate, a conversion rate and a negative feedback rate for the target message template, and on the basis, a document prompt word under the target message scene is updated according to the on-line index data so that the generation type model can generate a subsequent message document.

Inventors

• ZHANG LELE
• BU YIFAN

Assignees

• 支付宝(杭州)数字服务技术有限公司

Dates

Publication Date

20260508

Application Date

20260401

Claims (18)

1. 1. A message pushing method, comprising: responding to a message pushing request, wherein the message pushing request indicates that a target user in a target message scene is pushed with a message template; invoking a template selection model, selecting a target message template associated with the target message scene from a verification template set and an experimental template set based on user characteristic data of the target user, wherein message templates preconfigured for different message scenes are stored in the verification template set and the experimental template set, and message texts of all message templates are pre-generated by a generation model; Invoking a time prediction model to predict target pushing time based on historical behavior data of the target user; pushing the target message template to the target user based on the target pushing time; Acquiring on-line index data corresponding to the target message template, wherein the on-line index data comprises a click rate, a conversion rate and a negative feedback rate which are accumulated for the target message template; And updating the text prompt words in the target message scene according to the online index data so as to enable the generating model to generate a subsequent message text.
2. 2. The method of claim 1, further comprising, prior to responding to the message push request: Responding to a periodically issued generation instruction, wherein the generation instruction is used for indicating to generate a message file for the target message scene; Calling the generating model, and generating various message documents based on document prompt words pre-configured for the target message scene; Based on a preset auditing rule, performing rule auditing on the plurality of message documents; and configuring the message file passing the verification as a message template and distributing the message template to the experiment template set.
3. 3. The method of claim 2, wherein configuring and assigning audited message documents as message templates to the set of experiment templates comprises: under the condition that the target message scene is determined to need manual auditing, manually auditing the message document passing through the rule auditing based on a manual auditing rule preset for the target message scene; and configuring the message file passing the manual audit as a message template and distributing the message template to the experiment template set.
4. 4. A method according to claim 2 or 3, wherein assigning audited message templates to the set of experiment templates comprises: Defining distribution configuration parameters for the message templates passing the auditing, wherein the distribution configuration parameters comprise initial selection weight, target pushing quantity, and preset upgrading conditions and preset degrading conditions, and the preset upgrading conditions and the preset degrading conditions are determined based on at least one of a click rate threshold, a conversion rate threshold and a negative feedback rate threshold; and distributing the message templates defining the distribution configuration parameters to the experiment template set.
5. 5. The method as recited in claim 4, further comprising: And updating the definition rule of the distribution configuration parameters according to the on-line index data accumulated for each message scene so as to optimize the distribution strategy of the message template.
6. 6. The method as recited in claim 4, further comprising: acquiring on-line index data of each historical message template in the target message scene; according to the online index data, determining a historical message template which is positioned in the experiment template set and meets preset upgrading conditions; Migrating a respective history message template from the set of experimental templates to the set of verification templates.
7. 7. The method as recited in claim 6, further comprising: According to the online index data, determining a historical message template which is positioned in the verification template set and meets a preset degradation condition; migrating a corresponding historical message template from the set of verification templates to the set of experiment templates.
8. 8. The method as recited in claim 6, further comprising: According to the online index data, determining a historical message template which is positioned in the experiment template set and meets a preset degradation condition; and executing the offline operation on the corresponding historical message template.
9. 9. The method of claim 1, wherein invoking a template selection model to select a target message template associated with the target message scene from a set of verification templates and a set of experimental templates based on user characteristic data of the target user comprises: acquiring an online history message template from a verification template set and acquiring a message template to be verified from an experiment template set to form a candidate set; converting each message template in the candidate set into a text-embedded vector by a text-coding model, the text-coding model comprising a pre-trained language model or a fine-tuned semantic encoder; Acquiring user characteristic data of the target user, and generating a preference embedding vector corresponding to the user characteristic data through a pre-trained user characterization model; Recall M candidate message templates from the candidate set based on the similarity of the text-embedding vector and the preference-embedding vector; click rate prediction is carried out on the M candidate message templates based on the ranking model, and corresponding ranking scores are obtained; And carrying out weighting processing and attenuation processing on the sequencing scores to determine a target message template.
10. 10. The method of claim 9, further comprising, prior to obtaining an online historical message template from the set of verification templates and obtaining a message template to be verified from the set of experimental templates: acquiring a history message template receiving record of the target user in the target message scene, wherein the history message template record comprises receiving times and receiving time aiming at each history message template; determining the fatigue degree of the target user on each historical message template based on the historical message template receiving records; and according to a preset fatigue threshold, eliminating the historical message templates with the fatigue degree larger than the preset fatigue degree threshold from the verification template set or the experiment template set.
11. 11. The method of claim 9, wherein weighting the ranking score comprises: For each candidate message template recalled from the experimental template set, dynamically adjusting the selection weight of the corresponding candidate message template in the rearrangement stage according to the preset target pushing amount and the accumulated historical pushing amount; For each candidate message template recalled from the verification template set, maintaining an original selection weight or dynamically adjusting the selection weight of the corresponding candidate message template based on accumulated online indicator data.
12. 12. The method of claim 9, wherein attenuating the ranking score comprises: Obtaining a pre-configured user preference rule, wherein the user preference rule is used for representing negative preference of the target user on a specific type of message template; And applying an attenuation factor to the ranking scores of the corresponding message templates to attenuate the ranking scores under the condition that the message templates belonging to the specific type exist in the M candidate message templates.
13. 13. The method of claim 1, wherein invoking a temporal prediction model to predict a target push time based on historical behavioral data of the target user comprises: Acquiring historical behavior data of the target user, wherein the historical behavior data comprises an operation time sequence of a received historical message template; and inputting the historical behavior data into a time prediction model, predicting a plurality of candidate times and predicted click rates thereof through the time prediction model, determining the candidate times which are not lower than a preset click rate threshold value from the candidate times as target pushing times, and outputting the target pushing times.
14. 14. The method of claim 1, wherein updating the text prompt in the target message scene based on the online metrics data comprises: According to the on-line index data, positive feedback samples not lower than a preset positive threshold value are determined, and negative feedback samples lower than a preset negative threshold value are determined; Identifying the text characteristics corresponding to the positive feedback sample and the negative feedback sample; based on the text characteristics, updating the text prompt words in the target message scene.
15. 15. The method as recited in claim 1, further comprising: according to the on-line index data accumulated for each message scene, a text training sample is constructed; Optimizing or fine-tuning the template selection model and the temporal prediction model based on the document training samples.
16. 16. A message pushing device, comprising: the receiving module is used for responding to a message pushing request, and the message pushing request indicates to push a message template to a target user in a target message scene; The selection module is used for calling a template selection model, selecting a target message template associated with the target message scene from a verification template set and an experimental template set based on user characteristic data of the target user, wherein message templates preconfigured for different message scenes are stored in the verification template set and the experimental template set, and message texts of all the message templates are pre-generated by a generation type model; The prediction module is used for calling a time prediction model and predicting target pushing time based on the historical behavior data of the target user; The pushing module is used for pushing the target message template to the target user based on the target pushing time; The acquisition module is used for acquiring on-line index data corresponding to the target message template, wherein the on-line index data comprises a click rate, a conversion rate and a negative feedback rate which are accumulated for the target message template; And the updating module is used for updating the text prompt word in the target message scene according to the online index data so as to enable the generating model to generate a subsequent message text.
17. 17. An electronic device, the electronic device comprising: a memory for storing a computer program product; A processor for executing a computer program product stored in said memory, which, when executed, implements the method of any of the preceding claims 1-15.
18. 18. A computer readable storage medium storing a computer program, characterized in that the computer program is configured to implement the method of any one of claims 1-15 when executed by a processor.

Description

Message pushing method, device, equipment and medium Technical Field The present invention relates to the field of message pushing technologies, and in particular, to a message pushing method, device, equipment, and medium. Background In the current internet and mobile application scenarios, message access (such as push notification, in-station messaging, etc.) is an important means for improving user activity and conversion. The traditional scheme usually relies on operators to manually write a small quantity of templated texts, so that diversified user interests and scenes are difficult to cover, the sending decision is mostly based on simple rules or subjective judgment, dynamic perception of user states and contexts is lacking, the sending time is generally in a unified strategy (such as pushing in a fixed period), and timely and on-demand touch cannot be realized. In addition, the text optimization mainly depends on manual trial and error, lacks systematic experiments and feedback mechanisms, and has low efficiency and difficult quantification of effects. Although schemes are tried to combine an artificial intelligence generated content (ARTIFICIAL INTELLIGENCE GENERATED content, AIGC) model in recent years to generate a multi-version text and verify effects through an A/B experiment platform, links such as text generation, auditing, barrel separation, sending strategy and effect feedback are still mutually split, a cooperative closed loop cannot be formed, each module independently operates, so that not only is the cost of manual intervention increased, but also the joint optimization capability of the content and the opportunity is limited. Therefore, a message pushing scheme capable of deep fusion generation, auditing, experiment, personalized recommendation and feedback optimization is needed. Disclosure of Invention In order to realize a fine-grained self-evolutionary message pushing mechanism, the embodiment of the application provides a message pushing method, a device, equipment and a medium. In a first aspect, an embodiment of the present application provides a message pushing method, including responding to a message pushing request, where the message pushing request indicates to push a message template to a target user in a target message scene, invoking a template selection model, selecting a target message template associated with the target message scene from a verification template set and an experiment template set based on user feature data of the target user, where the verification template set and the experiment template set store message templates preconfigured for different message scenes, and a message case of each message template is pre-generated by a generating model, invoking a time prediction model, predicting a target pushing time based on historical behavior data of the target user, pushing the target message template to the target user based on the target pushing time, acquiring online index data corresponding to the target message template, where the online index data includes a click rate, a conversion rate and a negative feedback rate accumulated for the target message template, and updating a text prompt word in the target message scene according to the online index data, so as to generate a subsequent message case by the generating model. In an alternative embodiment, before responding to the message pushing request, the method further comprises responding to a periodically issued generation instruction, wherein the generation instruction is used for indicating to generate message cases for the target message scene, calling the generation model, generating various message cases based on the preset case prompt words for the target message scene, conducting rule auditing on the various message cases based on preset auditing rules, configuring the audited message cases as message templates and distributing the message templates to the experiment template set. In an alternative embodiment, the configuration of the message file passing the audit as the message template and the distribution to the experiment template set comprises the steps of carrying out manual audit on the message file passing the rule audit based on a manual audit rule preset on the target message scene under the condition that the target message scene is determined to need manual audit, and configuring the message file passing the manual audit as the message template and the distribution to the experiment template set. In an alternative embodiment, the method for distributing the message templates passing the examination to the experiment template set comprises the steps of defining distribution configuration parameters for the message templates passing the examination, wherein the distribution configuration parameters comprise initial selection weight, target pushing quantity, preset upgrading conditions and preset degrading conditions, the preset upgrading conditions and the preset degrading conditio