Method and apparatus for generating push information

Abstract

Provided in the embodiments of the present description are a method and apparatus for generating push information. On one hand, personalized push information is generated on the basis of seed information combined with user information, such that the problems of divergence and hallucination of a generation model can be effectively avoided; and on the other hand, users are first clustered during the process of generating seed information, and corresponding seed information is generated on the basis of a user category, such that a huge workload caused by marking a large number of users can be avoided. In summary, the implementations provided in the present specification can improve the effectiveness of personalized information push.

Description

Method and apparatus for generating push notifications

[0001] This application claims priority to Chinese Patent Application No. CN202510125369.3, filed on January 26, 2025, entitled "Method and Apparatus for Generating Push Information", the entire contents of which are incorporated herein by reference. Technical Field

[0002] This specification relates to the field of computer technology in one or more embodiments, and in particular to a method and apparatus for generating push notifications. Background Technology

[0003] Push notifications, also known as information delivery, are a new technology that uses specific technical standards or protocols to proactively deliver information to users on the internet based on triggering conditions (such as regular time intervals or opening relevant pages of an application), thereby reducing information overload. Push technology reduces the time spent searching online by automatically delivering information to users, helping them efficiently discover valuable information. Here, "push notification" refers to candidate information to be pushed to the user or information to be sent to the user. Push notifications can concisely and quickly describe the target information. For example, advertising images in public service announcements or event promotional pages are a type of push notification. Advertising images typically include the image information of the target information and an advertising tagline. With the development of artificial intelligence, to improve user experience, more and more push notification products are focusing on personalization, that is, selecting relevant candidate information for users based on their user profiles.

[0004] Personalized information push can include the following two aspects: selecting one or more pre-determined candidate information to display as personalized push notifications to different users; and generating personalized push notifications for users. In the case of generating personalized push notifications, how to avoid model illusions and overly divergent generated results while providing more personalized and targeted push notifications is a technical problem worthy of research. Summary of the Invention

[0005] This specification describes one or more embodiments of a method and apparatus for generating push notifications, which addresses one or more problems mentioned in the background art.

[0006] According to a first aspect, a method for generating push notifications is provided, the method comprising: obtaining user information of a first user; matching at least one target user category from various user categories for the first user based on the user information of the first user, wherein the user categories are determined by clustering multiple users using user information; obtaining seed information corresponding to each target user category, wherein a single user category corresponds to a single set of seed information; using the seed information, the user information of the first user, and business scenario information as input to a generation model, and determining push notifications to be pushed to the first user based on the output of the generation model.

[0007] In one embodiment, each user category is determined by: clustering the user information of the multiple users in multiple clustering directions to obtain several clusters in a single clustering direction; taking one cluster from each of the multiple clustering directions and combining them to obtain multiple cluster combinations, with each cluster combination corresponding to a single user category.

[0008] In a further embodiment, matching at least one target user category for the first user from various user categories based on the first user's user information includes: determining the cluster whose cluster center is closest to the first user in a single clustering direction as a single target cluster; and determining the user category corresponding to the combination of target clusters in each clustering direction as the target user category.

[0009] In a further embodiment, if multiple cluster centers in a single clustering direction have distances to the first user that are all less than a first threshold, and the differences between these distances are less than a second threshold, then all of the multiple clusters are determined as target clusters in that clustering direction. The step of determining the user category corresponding to the combination of target clusters in each clustering direction as the target user category further includes: combining each target cluster in that clustering direction with target clusters in other directions, and determining each user category corresponding to each target cluster group as the target user category.

[0010] In one embodiment, a single seed information includes at least one of text information, image information, and audio information, and a single set of seed information includes at least one seed information, which is generated by: inputting the data corresponding to the corresponding single user category into the seed generation model, generating multiple candidate seed information through the seed generation model, wherein the data corresponding to the single user category is one of the following: user information described in natural language, or a feature vector composed of the feature values ​​of each cluster center corresponding to the corresponding user category in the corresponding cluster dimension; and selecting at least one from each candidate seed information as a single set of seed information.

[0011] In one embodiment, the generative model is a large language model, and the step of using the seed information and the user information of the first user as input to the generative model includes: converting the user information of the first user into natural language information expressed in natural language; using the seed information, the natural language information, and the business scenario information of the current push business scenario together as prompt information input to the large language model, wherein the business scenario information includes at least one of the following: business domain, business object, target to be pushed, length and format of push information.

[0012] In one embodiment, the output of the generative model includes multiple candidate information items. The step of determining the push information to be pushed to the first user based on the output of the generative model includes: randomly selecting one candidate information item as the push information to be pushed to the first user; or, using an evaluation model to determine an evaluation score for each candidate information item; and selecting candidate information items in descending order of evaluation scores as the push information to be pushed to the first user.

[0013] In one embodiment, the method further includes: obtaining the current user's operation data on the pushed information, and using the operation data as a label to fine-tune the generation model when update conditions are met, wherein the update conditions include at least one of the following: the time period has been reached, and the number of operation data has reached a predetermined number.

[0014] According to a second aspect, an apparatus for generating push notifications is provided, the apparatus comprising:

[0015] The acquisition unit is configured to acquire the user information of the first user.

[0016] The matching unit is configured to match at least one target user category from various user categories for the first user based on the user information of the first user, wherein the user categories are determined by clustering multiple users using user information; and,

[0017] Obtain the seed information for each target user category, where each user category corresponds to a single set of seed information.

[0018] The generation unit is configured to take the seed information of each group, the user information of the first user, and the business scenario information as input to the generation model, and determine the push information to be pushed to the first user based on the output of the generation model.

[0019] According to a third aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of the first aspect.

[0020] According to a fourth aspect, a computing device is provided, including a memory and a processor, wherein executable code is stored in the memory, and the processor, when executing the executable code, implements the method of the first aspect.

[0021] In information push scenarios, using the apparatus and methods provided in the embodiments of this specification, multiple users can first be clustered according to their user information to obtain multiple user categories (when there is only one clustering direction, a single user category can correspond to a single cluster; when there are multiple clustering directions, a single user category corresponds to a cluster combination consisting of one cluster from each clustering direction), and seed information can be determined for each user category. When pushing information, based on the user information, the corresponding target user category can be matched, the seed information corresponding to the target user category can be obtained, and this seed information, along with the user information and business scenario information, is processed through a generative model (such as a large language model) to obtain push information for pushing information to the current user.

[0022] Thus, personalized information prediction based on seed information provides a certain reference or constraint for the prediction results, avoiding problems such as model illusion caused by overly divergent results generated by the generative model, and reducing reliance on the capabilities of the generative model. Furthermore, in the process of determining seed information, multiple user categories are obtained through user clustering, and a single set of seed information is determined for each user category, which greatly reduces the workload of determining seed information. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 is a schematic diagram of a specific implementation architecture based on the technical concept of this specification;

[0025] Figure 2 illustrates a flowchart of generating push notifications according to one embodiment of this specification;

[0026] Figure 3 illustrates a specific application architecture diagram according to an embodiment of this specification;

[0027] Figure 4 shows a structural block diagram of an apparatus for generating push notifications according to one embodiment of this specification. Detailed Implementation

[0028] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0029] Figure 1 illustrates a specific implementation architecture diagram of this specification. As shown in Figure 1, the implementation architecture provided in this specification includes a business terminal corresponding to the business party, a user terminal corresponding to the user, and a server. Both the business terminal and the user terminal can be various terminal devices capable of human-computer interaction with the user, such as smartphones, laptops, tablets, and smartwatches. The business terminal is the terminal used by the business party. In the implementation architecture shown in Figure 1, the business party can be various business entities that push relevant information of the target to be pushed to the user, such as merchants, enterprises, institutions, event organizers, etc. Correspondingly, the target to be pushed can be, for example, goods, policies and regulations to be promoted, event invitations, etc. The user terminal can be any terminal device corresponding to a user who may receive the push information. The user terminal can install and run various terminal applications, such as information push applications, news applications, shopping applications, food delivery platform applications, Q&A platform applications, etc. The server can provide business support for the terminal applications, such as providing business support for news applications. The server can have a local computing platform or be connected to a computing platform on other devices to determine the business content provided to the user terminal. For example, the server side of push notification applications can generate personalized push messages through a computing platform and push them to user terminals. For instance, in Figure 1, under the same triggering conditions (such as the user opening the application or the business actively broadcasting), message 1 can be pushed to user 1, while message 2 can be pushed to user 2, and so on.

[0030] It is understandable that in the architecture shown in Figure 1, the server and the business terminal can be different terminals or the same terminal. That is, the business party itself is the service provider that provides services to the user terminal, and the roles of the business party and the user can also be interchanged. For example, in one information push service, the business party acts as the service provider, providing the targets to be pushed; in another information push service, the business party can act as the user, receiving push information. In addition, the architecture shown in Figure 1 is only a specific example. In practice, the number of business terminals and the number of user terminals can be any possible number, such as 100,000 or 100 million, etc., and are not limited here.

[0031] This specification provides a method for generating personalized push notifications consisting of at least one of text, image, and audio in an information push scenario.

[0032] In the process of generating personalized push notifications for users, the push notifications generated by the generative model, which combines user information and business scenarios, may have a divergent nature. This is especially true when using a large language model as the generative model, where the training corpus is relatively rich, which may lead to model illusion.

[0033] In view of this, this specification provides a technical concept that, in the process of generating personalized push information for users, not only user information and scene information are used as input data for the generation model, but seed information is also provided to the generation model as a reference or constraint, so that the model can be improved based on the scene information and user information, thereby avoiding the model from being too divergent or model illusion, and making the generated push information more effective.

[0034] The technical concept of this specification will now be described in detail with reference to the embodiments shown in the accompanying drawings.

[0035] Figure 2 illustrates a flowchart of generating push notifications according to one embodiment. The entity executing this process can be a computer, device, or server with certain computing power, more specifically, such as the computing platform shown in Figure 1. When the push notification conditions are met, personalized push notifications are generated and sent to the user through the process shown in Figure 2. These push notification conditions can be, for example, the arrival of a predetermined push time (e.g., a fixed time each day when a sports organization pushes daily step count data from a friend's social circle), the user opening a predetermined page (e.g., the user opening a product browsing page), or the activity initiator initiating a business push notification request (e.g., a public welfare organization initiating a public welfare activity push notification request), and so on.

[0036] As shown in Figure 2, the process for generating push information provided in this embodiment may include: Step 201, obtaining user information of a first user; Step 202, matching at least one target user category from various user categories for the first user based on the user information of the first user, wherein the user categories are determined by clustering multiple users using user information; Step 203, obtaining seed information corresponding to each group for each target user category, wherein a single user category corresponds to a single group of seed information; Step 204, using the seed information, the user information of the first user, and business scenario information as input to the generation model, and determining the push information to be pushed to the first user based on the output of the generation model.

[0037] First, in step 201, the user information of the first user is obtained.

[0038] Here, the first user can be any network user who meets the information push conditions. The network can be the Internet or a telecommunications operator's network (such as the GSM network corresponding to a SIM card); there are no restrictions here. The first user can correspond to a client in the current information push service scenario, as well as the corresponding client identifier, such as the application's user ID, SIM card (card number or ISMI, etc.), etc.

[0039] User information can be various descriptive information about a user, such as one or more of the following: registration information, behavioral information, and interest preference information on a business platform. Registration information is typically relatively fixed, such as the user's identity information, phone number, and date of birth. User behavioral information can include information about various actions performed by the user through the device running the corresponding client, such as at least one of the following obtained from device operation logs: clicks, browsing, searching, forwarding, supporting, opposing, favoriting, form submissions, location, etc. More specifically, browsing information may include the pages viewed and the duration of page viewing, while form submission information may include shopping information and click information. When the device running the client is a portable mobile device (such as a mobile phone), the user's location information may also correspond to one or more of the following: movement trajectory information, dwell time information, dwell duration information, etc. Interest preference information can be user-defined information or information mined from user behavior; for example, on a navigation business platform, a user can set their own preference information, such as "avoid highways," etc.

[0040] The user information of the first user can be obtained from the first user's client, or from the business server (such as the server shown in Figure 1), or partly from the client and partly from the server; no limitation is made here.

[0041] Next, via step 202, at least one target user category is matched for the first user from various user categories based on the first user's user information.

[0042] User categories can be determined by clustering user information from multiple users on a business platform. Under the technical concept of this specification, seed information needs to be determined according to user categories; therefore, a user category needs to be determined for the first user. The user category to which the first user belongs is denoted as the target user category. There must be at least one target user category. The method for determining the target user category is related to the clustering method of the user categories.

[0043] Based on several possible implementation methods, multiple profile feature dimensions can be manually defined using user information. From each dimension, corresponding feature values ​​are extracted, and clustering is performed using the feature vectors formed by these feature values. For example, multiple profile dimensions might include at least one of the following: occupation, date of birth, consumption preferences, average monthly spending, browsing preferences (types of pages viewed), click preferences (types of hyperlinks clicked), form submission preferences (e.g., shopping type, direct submission, centralized submission), and so on. Each user can correspond to a single feature vector, and clustering between feature vectors can be performed using vector similarity (e.g., measured by cosine similarity, Euclidean distance, etc.).

[0044] Vector clustering methods include K-means, DBSCAN (Density-Based Spatial Clustering of Applications with Noise), Fuzzy Clustering, Mean Shift Clustering, Hierarchical Clustering, and so on.

[0045] Taking K-means clustering as an example, K cluster centers can be randomly assigned among each user. The similarity between the feature vectors of other users and the feature vectors of the K cluster centers is calculated. Each user is then assigned to the cluster corresponding to the cluster center with the highest similarity to their feature vector. The cluster centers for each current cluster are then redefined, and this clustering process is repeated until the cluster centers converge. Here, cluster center convergence can be understood as the change in distance between cluster centers not exceeding a predetermined distance threshold, or the similarity between the feature vectors of the new cluster centers and the cluster centers from the previous round exceeding a predetermined similarity threshold. In this way, multiple clusters are obtained, each representing a different user category. Each cluster center can serve as the center point of the corresponding user category in the feature space, used as a basis for matching user categories.

[0046] In this case, the user category is determined by clustering users based on feature vectors. The data referenced for clustering can be feature vectors. Therefore, when matching the first user to a corresponding user category, the feature vector of the first user can be extracted from the user information, and the distance between the first user and each cluster center (which has a corresponding feature vector) can be calculated using the feature vector. In one example, the cluster center closest to the first user can be determined as the corresponding user category. In another example, considering that the first user may be close to multiple cluster centers, if the distances between multiple cluster centers and the first user are all less than a first threshold (e.g., 0.1), and the difference between each distance is less than a second threshold (e.g., 0.01), then all user categories corresponding to the multiple clusters can be determined as the target user category.

[0047] According to other possible implementations, user information can be clustered across multiple clustering dimensions (or clustering directions). Multiple clusters can be obtained along a single clustering dimension, and the cluster categories along these dimensions can be arbitrarily combined, with each combination corresponding to a single user category. For example, the various profile dimensions mentioned earlier, such as occupation, date of birth, consumption preferences, average monthly spending, browsing preferences, click preferences, and form submission preferences, can be used as clustering dimensions. This expands the user category space and refines the granularity of user category division. For instance, with three clustering dimensions, resulting in m, n, and k clusters respectively, the user category space size can be m×n×k. In optional embodiments, for user information where the clustering direction is difficult to determine explicitly, semantic embedding can be performed to obtain corresponding embedding vectors. These embedding vectors can then be used as a clustering dimension to obtain multiple cluster centers.

[0048] In this case, the method for matching the target category to the first user can be as follows: the cluster whose cluster center is closest to the first user in a single clustering direction is determined as a single target cluster, and the user category corresponding to the combination of each target cluster can be the target user category. For example, if there are three clustering dimensions, and the target cluster a1 is matched in the first clustering dimension, the target cluster b1 is matched in the second clustering dimension, and the target cluster c1 is matched in the third clustering dimension, then the user category corresponding to the combination of each target cluster (a1, b1, c1) can be used as the target user category.

[0049] In one embodiment, considering that the first user may be close to multiple cluster centers in a single cluster dimension, if multiple cluster centers in a single cluster direction have distances to the first user that are all less than a first threshold and the difference between each distance is less than a second threshold, then the corresponding multiple clusters are determined as target clusters in that cluster dimension. Each target cluster in that cluster direction is then combined with target clusters in other directions to form target cluster groups. Thus, the user category corresponding to each target cluster group can be determined as the target user category.

[0050] Other methods for determining user categories can also be used to determine the target user category corresponding to the first user through appropriate means, which will not be listed here.

[0051] It is worth noting that the distance calculations between any two vectors mentioned above can be performed inversely proportional to vector similarity or directly proportional to vector distance. Vector similarity can be, for example, a factor of cosine similarity or dot product similarity, while vector distance can be, for example, a factor of Euclidean distance, variance, or Hamming distance.

[0052] Then, according to step 203, obtain the seed information for each group corresponding to each target user category.

[0053] It is understandable that a single user category can correspond to a single set of seed information. A single set of seed information contains one or more seed information entries, and each seed information entry includes at least one of the following: text, image, and audio. Seed information can be predetermined. Seed information can be determined manually or generated using a seed generation model (such as a large language model). Taking the generation of seed information using a seed generation model as an example, the feature values ​​of the cluster centers can be used as the corresponding user category features input to the seed generation model, and the seed information can be determined based on the output of the seed generation model. A set of seed information can contain one or more entries.

[0054] According to one embodiment: when clustering is performed based on feature vectors during user category determination, each cluster corresponds to a single user category, and the features of the cluster centers are described by feature vectors. Therefore, the feature vectors of the cluster centers can be used as input to the corresponding user category feature generation model. Alternatively, when clustering is performed based on clustering dimensions, each user category corresponds to a combination of features from the cluster centers across each dimension. Therefore, the corresponding feature combinations from the cluster centers of each user category across each dimension can be used as input to the corresponding user category feature generation model, which then generates candidate seed information. For example, the clustering dimensions include three: a, b, and c. The cluster center a1 in dimension a has a single feature value a. 11 The cluster center b1 with dimension b has a single feature value b. 12 The cluster center c1 with dimension c has two feature values ​​c. 11 c 12 Then, the user category features corresponding to cluster centers a1, b1, and c1 can have four dimensions, such as (a 11 b 12 c 11 c 12 ).

[0055] It's understandable that, to integrate seed information with business scenarios, the input to the seed generation model in some possible designs may also include scenario features of the current business scenario. These scenario features can be extracted based on the descriptive information of the current business scenario. The descriptive information of the current business scenario can include at least one of the following: business domain, business object, target audience for push notifications, length and format of the push notification, etc. Business domains include, for example, product push notifications, news push notifications, event push notifications (such as public welfare activities), etc.; business objects include, for example, the target audience for the information push (such as office workers), the expected effect of the business (such as 40% of users accepting the message), etc.; and target audiences for push notifications include, for example, products, news, events, etc.; and the length and format of the push notification include requirements such as word count and whether it includes images or audio.

[0056] Business scenario features can be extracted using various conventional techniques, which will not be elaborated here. In an optional embodiment, the scenario description information can be input into the embedding layer of a large language model, where it is processed to obtain the scenario features.

[0057] Scenario features can be shared by multiple users, and under the same business scenario, no updates are needed depending on the triggering conditions of each user at different times. Therefore, scenario features can be determined in the current process, or they can be pre-determined and applied to each user until they are updated. Scenario features and user category features can each be used as input to the seed generation model, or they can be fused together through weighting, summing, concatenation, etc., and then used as input to the seed generation model; no restrictions are imposed here.

[0058] According to another embodiment, when the seed generation model is a natural language processing model, such as BERT or the Big Oracle model, business scenario information and user information of each cluster center corresponding to a single user category can also be input into the seed generation model in the form of information described in natural language, and the seed generation model outputs candidate seed information, which will not be elaborated here.

[0059] The seed generation model can generate one or more candidate seed information. If there is only one candidate seed, it can be directly used as the seed. If there are multiple candidate seeds, all candidates can be selected, or a predetermined number (e.g., one) can be chosen from them by humans, an evaluation model, or a large model. When using a large model for selection, the model can detect the format, fluency, and scene inclusion of the candidate seed information to select the better one. The evaluation model can be a pre-trained scoring model. The input data to the evaluation model in the training samples can include user features, scene features, and candidate seed information. The output is the evaluation score for each candidate seed. The evaluation scores are supervised by manually labeled scores, which are used to adjust the parameters to obtain the evaluation model. When selecting seed information through an evaluation model, user type features, scene features, and candidate seed information can be used as input data for the evaluation model. The evaluation model outputs the evaluation score of each candidate seed information, and then selects a predetermined number (e.g., 1) of candidate seed information whose evaluation scores are greater than a preset threshold or whose scores are ranked from largest to smallest as seed information.

[0060] As described above, a single user category corresponds to one or more seed information entries, which can be referred to as a set of seed information. Each user category and its corresponding set of seed information are stored accordingly. For one or more target user categories determined by the first user, each target user category can correspond to a single set of seed information, thus allowing the first user to obtain the respective sets of seed information for each target user category.

[0061] Furthermore, in step 204, the seed information of each group, the user information of the first user, and the business scenario information are used as inputs to the generation model, and the push information to be pushed to the first user is determined based on the output of the generation model.

[0062] It is understandable that the generative model here can be the same as or different from the seed generation model mentioned earlier. The generative model can be implemented using various models capable of achieving the generation goal, such as the BERT model for generating text, Generative Adversarial Networks (GANs) for generating images, and large language models. The generative model here needs to be improved based on the seed information to generate push notifications suitable for the first user. Its input data uses the first user's user information, the current business scenario's scenario information, and the seed information of the first user's user category.

[0063] The user information of the first user, the scenario information of the current business scenario, and the seed information of the user category to which the first user belongs can be input into the generative model in the form of embedding vectors, or in the form of natural language (if the generative model is a large language model). When inputting into the generative model in the form of embedding vectors, the embedding vectors can be merged through concatenation, weighting, etc., before being input into the generative model. When inputting into the generative model in the form of natural language, the user information of the first user, the scenario information, and the seed information can all be converted into natural language descriptions. The text is usually in natural language form; if the seed information includes images and audio, the images and audio can be converted into natural language. The user information and scenario information of the first user may contain data formats such as data tables and key-value pairs, which can be converted into natural language form manually or by a large model.

[0064] The candidate information generated by the generative model can include at least one of the following: text, image, and audio, such as text information, graphic promotional copy, video, and animation. Since the large language model has good natural language processing capabilities, it can be used as a text and image generation model. That is, the large language model can be used as a generation module to process the user information of the first user, seed information, and scenario information of the current business scenario, thereby generating candidate information for push notifications targeting the first user.

[0065] The generative model can output one or multiple candidate messages. When only one candidate message is output, it can be directly used as the push notification to the first user. When multiple candidate messages are output, they can be generated all at once or multiple times. In this case, a predetermined number (e.g., one) of the candidate messages can be selected as the push notification to the first user. The selection method can be: random selection, thus providing more possibilities for collecting user preferences; optimal selection, such as determining evaluation scores through an evaluation model and selecting candidate messages with scores exceeding a predetermined threshold, or those ranked highest in descending order; manual designation; etc.

[0066] Based on some possible designs, before determining the candidate information as the push notification to the aforementioned first user, a quality check can be performed to filter out candidate information of poor quality. The quality check may include at least one of the following: whether the format meets the requirements, the fluency of the text, and the degree of inclusion of scene features. The degree of inclusion of scene features can be understood as the degree of relevance to the current business scenario, which can be determined by the degree of matching between the candidate information and the description information of the current business scenario. For example, as an example, the candidate information and the description information of the current business scenario can be semantically embedded to obtain semantic vectors, and the similarity of the semantic vectors can be calculated. The higher the similarity of the semantic vectors, the higher the degree of inclusion of scene features.

[0067] According to one possible design, during subsequent information push processes, user actions (such as clicking, browsing, submitting forms, deleting, etc.) can be used as labels to indicate the validity of the generated push information, and the generative model can be supervisedly fine-tuned (SFT) based on these labels. When using an evaluation model, user action data can also be converted into evaluation scores to jointly fine-tune both the generative and evaluation models.

[0068] To further clarify the technical concept of this specification, based on the above description, refer to a specific application example in Figure 3. The steps within the dashed boxes in Figure 3 are auxiliary operations outside the information push process, which can be performed in advance (such as preprocessing steps) or after multiple information push processes according to preset conditions (such as the model SFT step).

[0069] Large Language Models (LLMs), also known as large-scale language models or large models, typically have a large number of parameters, such as billions. A large language model is a deep learning-based natural language processing model that can learn the syntax and semantics of natural language, thereby generating human-readable text. Due to their massive corpora, large language models can be used as pre-trained models in various language processing scenarios, such as question-answering, information push, and text or image generation. These scenarios cover various application domains, such as product or service marketing, news broadcasting, and event promotion. It is understood that the scenarios listed here are roughly categorized by purpose; in practice, these scenarios may overlap. For example, text or image generation can occur independently, or in question-answering or information push scenarios. The large language model mentioned in this specification can be the pre-trained large language model itself, or a large language model fine-tuned with a small number of samples based on a pre-trained large language model; no limitation is made here.

[0070] In the example in Figure 3, for a specific business scenario, data preprocessing is performed first. The preprocessing process includes user clustering and seed information labeling. The user clustering step groups multiple users into multiple user categories. The seed information labeling step combines the descriptive information of the current business scenario with seed information to determine the corresponding sets of seed information for each user category. For example, user categories can be stored in a one-to-one correspondence with seed information.

[0071] Referring to Figure 3, when information needs to be pushed to a specific user, user information can be obtained first. Category matching is then performed based on this user information, matching the current user to the appropriate user category within each category. Seed information corresponding to the matched user category is then obtained during preprocessing. Alternatively, natural language processing (NLP) can be applied to the user information, converting it into a natural language description. Then, the seed information, business scenario description, and the NLP results of the seed information and user information are used as input to the large language model of the generative model. This large language model outputs multiple candidate messages. Next, information selection is performed, involving quality control by humans or the large model, or scoring by an evaluation model. This process filters the candidate messages to select the most suitable push information for the current user and sends it to them.

[0072] According to one possible design, after information is pushed to users, user actions such as clicks, browsing, and form submissions can be collected. Based on user actions, supervised fine-tuning (i.e., model SFT) can be provided for the large language model.

[0073] It is understood that FIG3 is a specific example of the embodiment described in FIG2, and the various descriptions of the technical concept of this specification in FIG2 can be applied to the example in FIG3.

[0074] In summary, the method for generating push notifications provided under the technical concept of this specification, on the one hand, effectively avoids problems such as model divergence and illusions by combining seed information with user information to generate personalized push notifications, thus improving the accuracy of the generated push notifications in conjunction with business scenarios and users. On the other hand, the process of generating seed information involves first clustering users and generating corresponding seed information according to user categories, which avoids the enormous workload caused by labeling a large number of users. In conclusion, the technical concept provided in this specification, while increasing the workload as much as possible, determines the seed information to provide reference and constraints for the user-generated push notifications, ensuring that the generated push notifications accurately match user characteristics and business scenarios, thereby improving the effectiveness of information push.

[0075] According to another embodiment, an apparatus for generating push notifications is also provided. This apparatus can be located in a computer, terminal, or server with a certain computing power, more specifically, as shown in FIG1, on a computing platform. FIG4 illustrates an apparatus 400 for generating push notifications according to one embodiment. As shown in FIG4, apparatus 400 may include:

[0076] The acquisition unit 401 is configured to acquire the user information of the first user.

[0077] The matching unit 402 is configured to match at least one target user category from various user categories for the first user based on the user information of the first user, wherein the user category is determined by clustering multiple users using user information; and to obtain seed information corresponding to each target user category, wherein a single user category corresponds to a single set of seed information.

[0078] The generation unit 403 is configured to take the seed information of each group, the user information of the first user, and the business scenario information as inputs to the generation model, and determine the push information to be pushed to the first user based on the output of the generation model.

[0079] In one embodiment, the apparatus 400 further includes a classification unit (not shown) configured to determine each user category by: clustering user information of multiple users in multiple clustering directions to obtain several clusters in a single clustering direction; and combining one cluster from each of the multiple clustering directions to obtain multiple cluster combinations, wherein each cluster combination corresponds to a single user category.

[0080] In a further embodiment, the matching unit 402 may be further configured to: determine the cluster whose cluster center is closest to the first user in a single clustering direction as a single target cluster; and determine the user category corresponding to the combination of target clusters in each clustering direction as the target user category.

[0081] In a further embodiment, if the distances between multiple cluster centers and the first user in a single clustering direction are all less than a first threshold and the differences between the distances are less than a second threshold, the matching unit 402 may be further configured to: determine multiple clusters as target clusters in the clustering direction; combine each target cluster in the clustering direction with target clusters in other directions, and determine each user category corresponding to each target cluster group as a target user category.

[0082] In one embodiment, a single seed information includes at least one type of information such as text, image, and audio, and a single set of seed information includes at least one seed information. The device 400 further includes a text generation unit (not shown) configured to generate a single set of seed information by: inputting data corresponding to a single user category into a seed generation model, generating multiple candidate seed information through the seed generation model, wherein the data corresponding to a single user category is one of the following: user information described in natural language, or a feature vector composed of the feature values ​​of each cluster center corresponding to the corresponding user category in the corresponding cluster dimension; and selecting at least one of the candidate seed information as a single set of seed information.

[0083] In one embodiment, the generation model is a large language model, and the generation unit 403 is further configured to: convert the user information of the first user into natural language information expressed in natural language; and use seed information, natural language information, and business scenario information of the current push business scenario as prompt information for the large language model and input them into the large language model. The business scenario information includes at least one of the following: business domain, business object, target to be pushed, length and format of push information.

[0084] In one embodiment, the output of the generation model includes multiple candidate information items. The generation unit 403 can also be configured to: randomly determine one candidate information item as the push information to be pushed to the first user; or, use the evaluation model to determine the evaluation score for each candidate information item; and select candidate information items in descending order of evaluation score as the push information to be pushed to the first user.

[0085] In one embodiment, the device 400 further includes a fine-tuning unit configured to: acquire current user operation data on the pushed information, and use the operation data as labels to fine-tune the generated model when update conditions are met. The update conditions may include at least one of the following: a time period has elapsed, the number of operation data points has reached a predetermined quantity, etc.

[0086] It is worth noting that the device 400 shown in FIG4 corresponds to the method described in FIG2, and the corresponding description in the method embodiment shown in FIG2 also applies to the device 400, and will not be repeated here.

[0087] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform the method described in conjunction with FIG2, etc.

[0088] According to another embodiment, a computing device is also provided, including a memory and a processor, wherein executable code is stored in the memory, and when the processor executes the executable code, the method described in conjunction with FIG2, etc., is implemented. Those skilled in the art will recognize that the functions described in the embodiments of this specification in one or more of the above examples can be implemented using hardware, software, firmware, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0089] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the technical concept in this specification. It should be understood that the above description is only a specific embodiment of the technical concept in this specification and is not intended to limit the scope of protection of the technical concept in this specification. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solutions of the embodiments in this specification should be included within the scope of protection of the technical concept in this specification.

Claims

1. A method for generating push notifications, the method comprising: Obtain the user information of the first user; Based on the user information of the first user, at least one target user category is matched from various user categories for the first user, and the user category is determined by clustering multiple users using user information; Obtain the seed information for each target user category, where each user category corresponds to a single set of seed information. The seed information of each group, the user information of the first user, and the business scenario information are used as inputs to the generation model. The push information to be pushed to the first user is determined based on the output of the generation model.

2. The method as described in claim 1, wherein, Each user category is determined in the following way: The user information of the multiple users is clustered in multiple clustering directions, thereby obtaining several clusters in a single clustering direction; By taking one cluster from each of the multiple clustering directions and combining them, multiple cluster combinations are obtained, and each cluster combination corresponds to a single user category.

3. The method as described in claim 2, wherein, The step of matching at least one target user category from various user categories for the first user based on the user information of the first user includes: The cluster whose cluster center is closest to the first user in a single clustering direction is identified as a single target cluster. The user category corresponding to the target cluster combination in each clustering direction is determined as the target user category.

4. The method of claim 3, wherein, If, in a single clustering direction, the distances between multiple cluster centers and the first user are all less than a first threshold, and the differences between these distances are less than a second threshold, then all of these multiple clusters are determined as the target clusters in that clustering direction. The step of determining the user category corresponding to the target cluster combination in each clustering direction as the target user category also includes: Each target cluster in this clustering direction is combined with target clusters in other directions, and each user category corresponding to each target cluster group is determined as the target user category.

5. The method of claim 1, wherein, A single seed message includes at least one of text, image, or audio information. A single set of seed messages includes at least one seed message and is generated in the following manner: The data corresponding to a single user category is input into the seed generation model, which generates multiple candidate seed information. The data corresponding to a single user category is one of the following: user information described in natural language, or a feature vector composed of the feature values ​​of each cluster center in the corresponding cluster dimension. Select at least one from the candidate seed information as a single group of seed information.

6. The method of claim 1, wherein, The generative model is a large language model, and the step of using the seed information and the user information of the first user as input to the generative model includes: Convert the user information of the first user into natural language information expressed in natural language; The seed information, the natural language information, and the business scenario information of the current push business scenario are used together as prompt information for the large language model. The business scenario information includes at least one of the following: business domain, business object, target to be pushed, length and format of push information.

7. The method of claim 1, wherein, The output of the generative model includes multiple candidate information items, and the step of determining the push information to be pushed to the first user based on the output of the generative model includes: Randomly select a candidate message as the push message to be sent to the first user; or The evaluation model is used to determine the evaluation score for each candidate piece of information; Candidate information is selected based on the evaluation scores from highest to lowest, and this information is then sent to the first user.

8. The method of claim 1, wherein, The method further includes: The system obtains the user's operation data on the pushed information, and uses the operation data as a label to fine-tune the generation model when update conditions are met. The update conditions include at least one of the following: the time period has been reached, or the number of operation data has reached a predetermined number.

9. An apparatus for generating push notifications, the apparatus comprising: The acquisition unit is configured to acquire the user information of the first user. The matching unit is configured to match at least one target user category from various user categories for the first user based on the user information of the first user, wherein the user category is determined by clustering multiple users using user information. as well as, Obtain the seed information for each target user category, where each user category corresponds to a single set of seed information. The generation unit is configured to take the seed information of each group, the user information of the first user, and the business scenario information as input to the generation model, and determine the push information to be pushed to the first user based on the output of the generation model.

10. A computing device, comprising a memory and a processor, characterized in that, The memory stores executable code, and when the processor executes the executable code, it implements the method of any one of claims 1-8.

Images