Push model training method and device based on large model, and information push method and device

Abstract

The embodiment of the invention provides a push model training method and device based on a large model and an information push method and device. In order to benefit an industrial push system from knowledge of a large model, seed users can be firstly sampled from total users, and high-quality and reusable push knowledge is extracted from the large model based on the seed users and added into a reasoning pool. When the push model is trained, the push knowledge in the reasoning pool is expanded to the whole user group through vector retrieval, the retrieved push knowledge is used as supplementary information to be fused with the embedded vector of the first user sample, and the fusion representation of the first user sample is obtained; and updating the push model by using the enhanced fusion representation. After being trained, the pushing model can be applied to an industrial scene for information pushing. Historical behavior data of a user can be used in the execution process of the method, the data belong to private data, and privacy protection needs to be carried out in the data processing process.

Description

Technical Field

[0001] One or more embodiments of the present specification relate to the field of computer processing technology, and in particular, to a push model training method, information push method and device based on a large model. Background Art

[0002] In order to improve the user experience, service platforms often use push models to determine the items that users are interested in based on their historical behaviors and push them to users. For example, in the user-item application field, the service platform can push items that users are interested in to users for selection; in the movie viewing field, the service platform can push movies that users are interested in to users. The user's historical behavior data is obtained after the user's authorization, and privacy is protected during the use of the data. The training of the push model is often conducted on a closed-loop user-item interaction dataset, which inevitably has serious exposure bias and popularity bias. This affects the generalization ability of the push model and the effectiveness of determining the push information, which in turn leads to insufficient effectiveness and accuracy when using the push model to push information to users.

[0003] Therefore, it is hoped that there will be an improved solution that can improve the effectiveness and accuracy of user push information in the push field. Summary of the invention

[0004] One or more embodiments of this specification describe a push model training method, an information push method and a device based on a large model to improve the effectiveness and accuracy of user push information in the push field. The specific technical solution is as follows.

[0005] In a first aspect, the embodiment provides a push model training method based on a large model, including:

[0006] Based on the matching results between the first user sample and several seed users in the inference pool, determine the first pushed knowledge corresponding to the first user sample; wherein the inference pool contains several seed users and corresponding pushed knowledge, and the pushed knowledge is information obtained by the large model through inference calculation of historical behavior data of the input seed users;

[0007] Fusing the embedding vector of the first user sample with the representation vector corresponding to the first pushed knowledge to obtain a fused representation of the first user sample;

[0008] The push model is updated based on the fused representation of the first user sample.

[0009] In one implementation, the seed user is determined in the following manner:

[0010] Clustering all users based on their historical behavior data to obtain multiple clusters, and sampling a number of representative users from the multiple clusters;

[0011] Based on the number of interactions corresponding to the historical behavior data of the entire set of users, a number of important users are sampled from the entire set of users;

[0012] A number of seed users are determined based on the representative users and the important users.

[0013] In one implementation, the push knowledge of any seed user is inferred in the following manner:

[0014] Prompt data is constructed based on the historical behavior data of the seed user, the prompt data is input into the big model, and the push knowledge of the seed user is determined by the big model; the prompt data includes a first indication text and a second indication text; the first indication text is used to indicate the summary of the behavior preference of the seed user, and the second indication text is used to indicate the determination of the push knowledge based on the summarized behavior preference.

[0015] In one implementation, the step of fusing the embedding vector of the first user sample with the representation vector corresponding to the first pushed knowledge includes:

[0016] Determine a first weight of the embedding vector of the first user sample and a second weight of the representation vector corresponding to the first pushed knowledge based on the first parameter;

[0017] The first weight and the second weight are used to perform weighted fusion on the embedding vector of the first user sample and the representation vector corresponding to the first pushed knowledge.

[0018] In one embodiment, the method further includes: updating the first parameter based on the fused representation of the first user sample.

[0019] In one implementation, the push knowledge includes categories and items of push objects.

[0020] In a second aspect, the embodiment provides an information push method based on a large model, including:

[0021] Based on the matching results between the second user and several seed users in the inference pool, determine the second pushed knowledge corresponding to the second user; wherein the inference pool contains several seed users and corresponding pushed knowledge, and the pushed knowledge is information obtained by the large model through inference calculation of historical behavior data of the input seed users;

[0022] Fusing the embedding vector of the second user with the representation vector corresponding to the second pushed knowledge to obtain a fused representation of the second user;

[0023] The fused representation of the second user is input into the trained push model, and push information for the second user is determined by the push model.

[0024] In one implementation, the step of fusing the embedding vector of the second user with the representation vector corresponding to the second pushed knowledge includes:

[0025] Determine a first weight of the embedding vector of the second user and a second weight of the representation vector corresponding to the second pushed knowledge based on the first parameter obtained through training;

[0026] The first weight and the second weight are used to perform weighted fusion on the embedding vector of the second user and the representation vector corresponding to the second pushed knowledge.

[0027] In one implementation, the push model is trained using any one of the methods described in the first aspect.

[0028] In a third aspect, the embodiment provides a push model training device based on a large model, including:

[0029] A first matching module is configured to determine first pushed knowledge corresponding to the first user sample based on a matching result between the first user sample and a plurality of seed users in an inference pool; wherein the inference pool includes a plurality of seed users and corresponding pushed knowledge, and the pushed knowledge is information obtained by the large model through inference calculation of historical behavior data of the input seed users;

[0030] A first fusion module, configured to fuse the embedding vector of the first user sample with the representation vector corresponding to the first pushed knowledge to obtain a fused representation of the first user sample;

[0031] A model updating module is configured to update the push model based on the fusion representation of the first user sample.

[0032] In a fourth aspect, the embodiment provides an information push device based on a large model, including:

[0033] A second matching module is configured to determine second pushed knowledge corresponding to the second user based on the matching results between the second user and a plurality of seed users in the inference pool; wherein the inference pool includes a plurality of seed users and corresponding pushed knowledge, and the pushed knowledge is information obtained by the large model through inference calculation of historical behavior data of the input seed users;

[0034] A second fusion module is configured to fuse the embedding vector of the second user with the representation vector corresponding to the second pushed knowledge to obtain a fused representation of the second user;

[0035] The push determination module is configured to input the fused representation of the second user into the trained push model, and determine the push information for the second user through the push model.

[0036] In a fifth aspect, an embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute any one of the methods in the first aspect and the second aspect.

[0037] In a sixth aspect, an embodiment provides a computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method described in any one of the first aspect and the second aspect is implemented.

[0038] In the method and device provided in the embodiments of this specification, a large model is used to perform generative push for seed users to obtain push knowledge, and an inference pool containing seed users and push knowledge is constructed; then, the large model push knowledge that is reliably associated with user samples is retrieved through vector retrieval, and the push knowledge is fused into the user's embedding vector, thereby enhancing the amount of information in the user representation, and using the enhanced user representation, i.e., the fused representation, to train the push model can improve the generalization ability of the model and the effectiveness in determining push information, thereby improving the effectiveness and accuracy of the determined user push information. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0040] Figure 1 A schematic diagram of an implementation scenario of an embodiment disclosed in this specification;

[0041] Figure 2 A schematic diagram of an implementation architecture of an embodiment disclosed in this specification;

[0042] Figure 3 A flowchart of a push model training method based on a large model provided in an embodiment;

[0043] Figure 4 A schematic diagram of a flow chart of a large model-based information push method provided in an embodiment;

[0044] Figure 5 A schematic block diagram of a push model training device based on a large model provided in an embodiment;

[0045] Figure 6 A schematic block diagram of an information push device based on a large model is provided for an embodiment. DETAILED DESCRIPTION

[0046] The solution provided in this specification is described below in conjunction with the accompanying drawings.

[0047] The push model is used to determine the items that the user is interested in from multiple candidate items (i.e., the push results) based on the input user embedding vector, and the determined items that the user is interested in are used to push to the user. The user embedding vector is a representation determined based on data related to the user. The candidate items are items that are used to push to the user. The current task is to determine the items that the user is interested in from the candidate items and push them to the user.

[0048] The selected items or projects are the specific contents contained in the push object. In this specification, the push object can be understood as an abstract concept, and the project can be understood as the specific object contained in the abstract concept. For example, the push object is a mobile phone (mobile phone is an abstract concept), and the selected items can include multiple mobile phones of different brands and models. Push objects include items (or commodities), videos, music, movies, articles and other categories. Push objects are not limited to the above. As long as the user needs it, it can be used as a push object.

[0049] As mentioned above, existing push models are often trained on a closed-loop user-candidate item interaction dataset, which inevitably has serious exposure bias and popularity bias. Exposure bias usually occurs when the data distribution of training data and model service is inconsistent. Taking items as an example, during the model training phase, the model can only observe the feedback of items that are exposed (i.e., displayed to users), but cannot observe the potential feedback of unexposed items. This leads to the data distribution learned by the model during training being different from the data distribution during actual service, because the push model needs to push all items in actual applications, not just those that are exposed in the training data. Popularity bias refers to the tendency of the model to push items that are already popular, resulting in an excessively high proportion of popular items in the push results. This bias is not conducive to the diversity and personalization of the push system, because it may ignore items that are not so popular but may be of interest to users. Therefore, in addition to the narrow information present in the original dataset, it is crucial to combine open-world knowledge to promote a more comprehensive and generalized understanding of user historical behavior.

[0050] Large Language Models (LLMs) have shown great potential in solving the data isolation problem faced by push systems and improving the performance of push systems due to their rich open-world knowledge and excellent language generation capabilities. However, despite the performance being comparable to traditional push methods, current LLM-based push methods are too costly for industrial applications. This is because, when dealing with hundreds of millions of users, it is very expensive to perform separate reasoning for each user using any mainstream LLM.

[0051] In order to improve the effectiveness of pushing information while reducing the cost of using LLM for pushing, the embodiments of this specification provide a push model training method and information push method based on a large model. Specifically, the embodiments of this specification use LLM to perform generative push on seed users obtained by sampling a small portion of the full number of users, thereby building an LLM reasoning pool. Then, vector retrieval is used to associate reliable large model knowledge with the full number of user groups, so that the push system can use the powerful reasoning ability of the large model to effectively improve the push performance of all users, while at the same time, there is no need to directly use LLM to push to all users, reducing the cost of using LLM.

[0052] LLM is a type of deep learning model that usually has a large number of parameters (possibly billions or even tens of billions), has strong semantic understanding capabilities, and can understand and generate natural language text. Large models can learn rich data representations and patterns by training on large-scale data sets, and thus perform well on a variety of tasks, including text generation, language understanding, image recognition, etc. Large models can process text data, audio data, images, etc. Audio data can be converted to text data, so in summary, the input data of large models can be text, images, etc.

[0053] Figure 1This is a schematic diagram of an implementation scenario of an embodiment disclosed in this specification. It includes multiple computing devices, such as device 1, device 2 and device 3. Device 1 is used to train a push model and use the trained push model to determine the push items of the user. Device 2 is used to provide inference computing services for a large model. Device 3 is used to push items to users. Specifically, device 1 samples seed users from all users and sends the historical behavior data of the seed users to device 2, thereby calling the large model to determine the push knowledge of the seed users. After determining the push knowledge of the seed users through the large model, device 2 sends the push knowledge to device 1. Then, device 1 builds an LLM reasoning pool based on the seed users and the corresponding push knowledge, and can train the push model based on the LLM reasoning pool. Device 3 can send the user of the item to be pushed to device 2, thereby calling the push model, and device 1 uses the trained push model and the LLM reasoning pool to determine the push item of the user. After receiving the push item sent by device 2, device 3 pushes it to the user. Figure 1 This is just one of many implementation scenarios. In actual applications, device 1 itself can also provide inference computing services for large models; or, the operation of pushing items to users can also be performed by device 1, and so on.

[0054] The following describes the embodiments from the method level. Figure 2 This is a schematic diagram of the implementation architecture of an embodiment disclosed in this specification. It includes the LLM reasoning pool construction phase and the LLM reasoning pool application phase. The reasoning pool construction phase is used to extract the push knowledge of LLM based on the sampled seed users, and the reasoning pool application phase is used to apply the push knowledge of LLM to all users through vector retrieval. The reasoning pool application phase may include the training phase of the push model and the application phase of the push model. The training phase of the push model is used to train the push model based on the reasoning pool, and the application phase of the push model is used to determine the push items using the trained push model.

[0055] Combine the following Figure 2 , the construction phase of the LLM reasoning pool, the training phase of the push model, and the application phase of the push model are explained through multiple embodiments.

[0056] In the construction phase of the LLM reasoning pool, the computing device as the execution subject can build the reasoning pool by executing the following steps 1 and 2.

[0057] Step 1: Sample seed users from all users and add them to the seed user set S.

[0058] Step 2: For any seed user, input the historical behavior data of the seed user into the big model, determine the push knowledge for the seed user based on the big model, and add the seed user and the corresponding push knowledge to the LLM reasoning pool. Perform this operation for each seed user to obtain its push knowledge. The push knowledge can be represented in the form of text data.

[0059] In step 1, seed users can be determined by following steps 11 to 13. Step 11: cluster all users based on their historical behavior data to obtain multiple clusters, sample several representative users from the multiple clusters, and add them to the seed user set S R Step 12: Based on the number of interactions corresponding to the historical behavior data of all users, a number of important users are sampled from all users and added to the seed user set S I Step 13: Based on the representative users (i.e., the seed user set S R ) and important users (i.e., seed user set S I )Identify a number of seed users.

[0060] The sampling strategy implemented in step 11 can be called representativeness-aware sampling. All users refer to all users in the data set. Here, all users are called ordinary users. When users who meet the conditions are screened out from ordinary users, these screened users are called seed users. When clustering, the K-means algorithm can be used to cluster all users into k clusters, and seed users are sampled from each cluster. The number of seed users sampled in each cluster is proportional to the cluster size, so as to obtain a representative seed user set S. R The cluster size refers to the number of users included in the cluster. Other clustering algorithms may also be used during clustering, which is not limited in this embodiment.

[0061] During clustering, the historical behavior data of each user can be input into a lightweight behavior encoder to obtain a user representation, and all users can be clustered based on the user representation. Lightweight behavior encoders include encoders such as LightGCN and SASRec.

[0062] The user's historical behavior data includes data related to search, click, purchase and other behaviors. The historical behavior data here can be text data or key-value pair data. It should be noted that the user-related data mentioned in the embodiments of this specification are obtained after the user's authorization, and privacy protection will be performed during the data processing process to ensure that the user's privacy data is not leaked.

[0063] The sampling strategy performed in step 12 can be called importance-aware sampling. The importance sampling strategy gives priority to users with more interactions. When sampling from all users, users with higher interaction times can be randomly sampled from all users based on the number of interactions as important users. Formally, the seed user set S obtained by this importance sampling is I It can be defined as follows:

[0064]

[0065] Among them, D u represents the interaction set of user u, that is, the historical behavior data set, |D u | represents the number of interactions of user u. I(u) is an indicator function. If the function takes the value of 1, user u is selected as an important user. P is the probability. The larger the probability, the easier it is for user u to be selected. Therefore, the sampling process can be random, but this random sampling is constrained by the probability P. The more times user u interacts, the larger the probability P corresponding to user u, and the easier it is for the user to be selected. The user's historical behavior data includes the user's search, click, and purchase of items. The number of user interactions refers to the number of historical behaviors. A search is a historical behavior, and a click and a purchase can also be counted as a historical behavior respectively.

[0066] Seed user set S R and the seed user set S I The number of seed users in can be equal or unequal. In step 13, the two sets can be merged to obtain the final sampled seed user set S. That is, the constructed seed user set S = {u|u∈S R ∪S I The size of the seed user set S is much smaller than the size of the full user set, that is, the total number of seed users is much smaller than the total number of full users.

[0067] The above implementation adopts a sampling strategy that combines representative sampling and importance sampling, and can sample a high-quality, representative set of seed users. In practical applications, seed users can be sampled from all users using only representative sampling or only importance sampling, or other sampling methods.

[0068] In step 2, for any seed user u, the historical behavior data of the seed user u is input into the big model, the push knowledge for the seed user u is determined by LLM, and the seed user u and the corresponding push knowledge are added to the big model reasoning pool. Here, the historical behavior data of user u can be text data or other data, as long as it is a data form that can be recognized by the big model. Specifically, the historical behavior data of the seed user u input into the big model may include behavior and behavior objects, or may not include behavior but include behavior objects. For example, searching, clicking, and buying are all behaviors, and the objects of searching, clicking, and buying are all behavior objects. The objects here may include multiple categories such as items (or commodities), videos, music, movies, articles, etc. The big model used in this embodiment can directly select a pre-trained big model without fine-tuning the big model in the reasoning task.

[0069] See also Figure 2 In the dotted box on the left, when the historical behavior data of the seed user u is input into the big model, prompt data prompt can be constructed based on the historical behavior data of the seed user u, and the prompt data prompt is input into the big model, and the push knowledge of the seed user u is determined by the big model. Figure 2 The historical behavior library in is used to store the historical behavior data of all users. After the seed user is determined, the historical behavior data of the seed user can be obtained from the historical behavior library. The prompt data prompt can be generated according to the preset prompt template.

[0070] Among them, the prompt data prompt includes a first instruction text and a second instruction text. The first instruction text is used to indicate the summary of the behavior preferences of the seed user, and the second instruction text is used to indicate the determination of the push knowledge based on the summarized behavior preferences. For example, the second instruction text can indicate that the push knowledge includes the category and project of the push object. The prompt data prompt also includes the historical behavior data of the seed user u. The two instruction texts can guide the large model to think step by step and generate a chain of thinking to answer the question, so as to extract a more accurate answer from this chain of thinking. Table 1 below is an example of the prompt data prompt and the push knowledge output by the large model.

[0071] Table 1

[0072]

[0073] The input and output of the large model are listed on the left side of Table 1. The input prompt contains instruction text and historical behavior data. The historical behavior data only lists the behavior objects, namely the movie name and movie category. The first instruction text and the second instruction text are marked with brackets in the instruction text part. The output part of the large model contains the behavior preferences and push knowledge obtained based on the first instruction text and the second instruction text. The push knowledge includes movie categories and movie names.

[0074] When the seed user u and the corresponding push knowledge are added to the LLM reasoning pool, the behavior representation of the seed user (ie, user representation) and the corresponding push knowledge may be combined into the push knowledge of the seed user and added to the push pool.

[0075] The knowledge in the inference pool determined by the above method is high-quality, reusable knowledge, which can be used in the training stage of the push model and the application stage of the push model. Figure 3 The training phase of the push model is described.

[0076] Figure 3 A flowchart of a push model training method based on a large model is provided in an embodiment. The push model is used to determine a push item for an input user, and the determined push item is used to push to the user. The data input to the push model can specifically be a user embedding vector. The method can be executed by a computing device, and the method includes the following steps.

[0077] Step S310, based on the matching results of the first user sample u1 and several seed users in the inference pool, determine the first push knowledge I1 corresponding to the first user sample u1. This step is used to expand the push knowledge retrieved from the inference pool to the entire user group. The retrieved push knowledge will serve as supplementary information to enhance the traditional push system. This is based on the assumption that users with similar behaviors can share the same LLM push knowledge.

[0078] The inference pool contains several seed users and corresponding push knowledge. The push knowledge of any seed user is the information obtained by the large model through inference calculation of the historical behavior data of the input seed user.

[0079] The first user sample u1 may be any common user among all users, and the common user is used as a user sample for training the push model. The first user sample u1 may have corresponding annotated data. The annotated data may be, for example, items that the user actually searches, clicks, or purchases.

[0080] When this step is executed, it specifically includes matching the user representation of the first user sample u1 with the user representations of several seed users in the inference pool to obtain a matching result. The matching result includes the seed users who are successfully matched. The user representation of the first user sample u1 can be determined by the historical behavior data of the first user sample u1. For example, the historical behavior data can be input into a lightweight behavior encoder to obtain a user representation. The above matching operation can be performed using a matching algorithm between vectors, for example, algorithms such as Euclidean distance and pre-similarity can be used to select a seed user with a higher matching degree as the matching result. After obtaining the matching result, the push knowledge corresponding to the seed user in the matching result can be used as the first push knowledge I1.

[0081] For the first pushed knowledge I1 of the first user sample u1, it can be obtained by the following formula:

[0082]

[0083] Among them, r u It is the first push knowledge, s It is the push knowledge of the matched seed user. u is the user representation of the first user sample, m s is the user representation of the seed user. s belongs to the seed user set S and is determined by the constraint condition argmax, that is, when the user representation of the first user sample is most similar to the user representation of the seed user, the seed user is taken as the matched seed user and its pushed knowledge is taken as the first pushed knowledge. The retrieval operation can associate at least one reliable LLM pushed knowledge for each ordinary user.

[0084] Step S320: fuse the embedding vector of the first user sample u1 with the representation vector corresponding to the first pushed knowledge I1 to obtain a fused representation of the first user sample u1.

[0085] Step S330: updating the push model based on the fusion representation of the first user sample u1.

[0086] When step S320 is specifically performed, a first weight of the embedding vector of the first user sample u1 can be determined based on the first parameter θ, and a second weight of the representation vector corresponding to the first pushed knowledge I1 can be determined based on the first parameter θ, and the embedding vector of the first user sample u1 and the representation vector corresponding to the first pushed knowledge I1 are weightedly fused using the first weight and the second weight. Weighted fusion can be understood as weighted summation.

[0087] The representation vector of the first pushed knowledge I1 can be determined by an existing encoder. The encoder can be, for example, but not limited to, a TextEncoder encoder, which is a BERT encoder. BERT is a pre-trained language model based on the Transformer architecture. The encoder here can also use other types of encoders. The text data of the first pushed knowledge I1 is input into the encoder to obtain its representation vector, that is, to convert the text into a vector.

[0088] The embedding vector of the first user sample u1 can be determined based on the feature data of the first user sample u1, and the feature data can include the user's inherent attributes, historical behavior data, and related environmental data. The method for determining the embedding vector of the first user sample u1 can be determined using the existing method in the original training system of the push model. Both the embedding vector and the user representation are feature vectors of the user. Since the two are determined using different methods, their specific values ​​may be different. Generally speaking, the embedding vector is determined based on richer user-related data, and the user representation can be determined using relatively less user-related data, such as user behavior data, to reduce the cost of using large models.

[0089] The first parameter θ can be set as a learnable parameter, and a more accurate value can be gradually determined through iterative updates. The initial value of the first parameter θ can be determined randomly or empirically. When updating, the first parameter θ can be updated based on the fusion representation of the first user sample u1.

[0090] The first parameter θ can be updated jointly with the push model. The operation corresponding to step S320 can also be implemented by adding a fusion layer to the push model. In this way, when the push model is updated, the update of the first parameter θ is included. The fusion layer (FusionLayer) can be implemented in the following ways:

[0091]

[0092] Among them, TextEncoder(r u ) represents converting the first pushed knowledge into a representation vector, and the function Linear is used to convert the representation vector to the same dimension as the embedding vector of the first user sample u1. and are the first weight and the second weight respectively, e′ u is a fusion representation. Softmax is an activation function used to transform data.

[0093] When the push model is updated in step S330, it can be specifically implemented in the following manner. The fusion representation of the first user sample u1 is input into the push model, and the prediction values ​​for multiple candidate items are determined by the push model. The prediction loss is determined based on the difference between the prediction value and the labeled data of the first user sample u1, and the parameters of the push model are updated in the direction of reducing the prediction loss, including updating the first parameter θ. For example, the push model can be optimized by the following loss function:

[0094]

[0095]

[0096] in, is the predicted value for candidate item i, e i is the embedding vector of candidate item i, e′ u is a fusion representation, u belongs to the full user set U, i belongs to the candidate item set E, y ui It is the labeled data.

[0097] See also Figure 2 The data input to the push model include the representation vector of the pushed knowledge, the embedding vector of the common user, and the embedding vector of the candidate item. The push model outputs the push result based on the input data. The training process of the push model can be implemented by multiple iterations using existing methods until convergence. The specific process will not be described in detail.

[0098] Looking at step S320 again, when fusing the embedding vector of the first user sample u1 with the representation vector corresponding to the first pushed knowledge I1, it is not limited to the above-mentioned weighted fusion method. For example, a direct addition fusion method, or a weighted fusion using a default weight value, or a concatenation of two vectors are all feasible. Specifically, a more reasonable fusion method can be selected based on actual operation results.

[0099] The above is an explanation of the training phase of the push model. Figure 4 The application stage of the push model is explained. The application stage of the push model is also the stage of applying LLM push knowledge in industry.

[0100] Figure 4 A flowchart of a method for pushing information based on a large model is provided in an embodiment. The method is executed by a computing device and includes the following steps.

[0101] Step S410, based on the matching results between the second user u2 and several seed users in the inference pool, determine the second push knowledge I2 corresponding to the second user u2. This step is used to expand the push knowledge retrieved from the inference pool to users in industrial applications. The retrieved push knowledge will be used as supplementary information to represent the characteristics of these users, thereby improving the accuracy of user prediction. Among them, the second user u2 can be any user to be pushed.

[0102] When this step is executed, it specifically includes matching the user representation of the second user u2 with the user representations of several seed users in the inference pool to obtain a matching result. The matching result includes the seed users who are successfully matched. The user representation of the second user u2 can be determined by the historical behavior data of the second user u2. The historical behavior data includes the historical records of the user's search, click or purchase operation behaviors. For example, the historical behavior data can be input into a lightweight behavior encoder to obtain a user representation. The above matching operation can be performed using a matching algorithm between vectors. For example, algorithms such as Euclidean distance and pre-similarity can be used to select a seed user with a higher matching degree as the matching result. After obtaining the matching result, the push knowledge corresponding to the seed user in the matching result can be used as the second push knowledge I2.

[0103] Step S420: fuse the embedding vector of the second user u2 with the representation vector corresponding to the second pushed knowledge I2 to obtain a fused representation of the second user u2.

[0104] When this step is specifically performed, the first weight of the embedding vector of the second user u2 and the second weight of the representation vector corresponding to the second pushed knowledge I2 can be determined based on the first parameter θ obtained through training, and the embedding vector of the second user u2 and the representation vector corresponding to the second pushed knowledge I2 are weightedly fused using the first weight and the second weight. The weighted fusion here can be understood as weighted summation.

[0105] The representation vector of the second pushed knowledge I2 can be determined by an existing encoder, such as but not limited to a TextEncoder encoder. The encoder here can also use other types of encoders. The text data of the second pushed knowledge I2 is input into the encoder to obtain its representation vector, that is, the text can be converted into a vector. The embedding vector of the second user u2 can be determined based on the feature data of the second user u2, which can include the user's inherent attributes, historical behavior data, and related environmental data. The first parameter θ is obtained by Figure 3 The parameters obtained by training in the embodiment can obtain a more accurate fusion representation by weighted fusion of two vectors through the first parameter θ.

[0106] Step S430: Input the fusion representation of the second user u2 into the trained push model, and determine the push items for the second user u2 through the push model. Figure 3 The method of the embodiment is trained.

[0107] This embodiment integrates the reasoning knowledge in the LLM reasoning pool with the user's original embedding vector and applies the knowledge in the LLM to the push system, thereby achieving low-cost use of large models to solve the closed-loop problem of traditional push methods and effectively improving the accuracy of push information.

[0108] In this specification, the word "first" in terms such as the first user sample, the first pushed knowledge, the first indication text, the first weight, etc., and the corresponding "second" (if any) in the text, etc., are merely for the convenience of distinction and description and do not have any limiting meaning.

[0109] In this specification, the computing device can be implemented by any device, equipment, platform, device cluster, etc. with computing and processing capabilities.

[0110] The foregoing describes certain embodiments of the present specification, and other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in an order different from that in the embodiments, and the desired results may still be achieved. In addition, the processes depicted in the accompanying drawings do not necessarily have to be performed in the specific order or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0111] Figure 5 A schematic block diagram of a push model training device based on a large model provided in an embodiment. Figure 3 The apparatus 500 is deployed in a computing device and includes:

[0112] The first matching module 510 is configured to determine the first pushed knowledge corresponding to the first user sample based on the matching results between the first user sample and a plurality of seed users in the inference pool; wherein the inference pool includes a plurality of seed users and corresponding pushed knowledge, and the pushed knowledge is information obtained by the large model through inference calculation of the historical behavior data of the input seed users;

[0113] A first fusion module 520, configured to fuse the embedding vector of the first user sample with the representation vector corresponding to the first pushed knowledge to obtain a fused representation of the first user sample;

[0114] The model updating module 530 is configured to update the push model based on the fusion representation of the first user sample.

[0115] In one embodiment, the device 500 also includes the following modules for determining the seed users: a user clustering module 540, configured to cluster the full set of users based on the historical behavior data of the full set of users to obtain multiple clusters, and sample a number of representative users from the multiple clusters; an importance sampling module 550, configured to sample a number of important users from the full set of users based on the number of interactions corresponding to the historical behavior data of the full set of users; and a user determination module 560, configured to determine a number of seed users based on the representative users and the important users.

[0116] In one embodiment, the device 500 also includes a knowledge reasoning module 570, which is configured to infer the push knowledge of any seed user in the following manner: construct prompt data based on the historical behavior data of the seed user, input the prompt data into the big model, and determine the push knowledge of the seed user through the big model; the prompt data includes a first indication text and a second indication text; the first indication text is used to indicate the summary of the behavior preferences of the seed user, and the second indication text is used to indicate the determination of the push knowledge based on the summarized behavior preferences.

[0117] In one embodiment, the first fusion module 520 includes: a first determination submodule and a first fusion submodule (not shown in the figure). The first determination submodule is configured to determine a first weight of the embedding vector of the first user sample and a second weight of the representation vector corresponding to the first pushed knowledge based on a first parameter. The first fusion submodule is configured to perform weighted fusion on the embedding vector of the first user sample and the representation vector corresponding to the first pushed knowledge using the first weight and the second weight.

[0118] In one implementation, the apparatus 500 further includes: a parameter updating module (not shown in the figure), configured to update the first parameter based on the fusion representation of the first user sample.

[0119] In one implementation, the push knowledge includes categories and items of push objects.

[0120] Figure 6 The present invention provides a schematic block diagram of an information push device based on a large model. Figure 4 The method embodiment shown corresponds to the embodiment shown in the figure. The apparatus 600 is deployed in a computing device, and includes:

[0121] The second matching module 610 is configured to determine the second pushed knowledge corresponding to the second user based on the matching results between the second user and a plurality of seed users in the inference pool; wherein the inference pool includes a plurality of seed users and corresponding pushed knowledge, and the pushed knowledge is information obtained by the large model through inference calculation of the historical behavior data of the input seed users;

[0122] A second fusion module 620 is configured to fuse the embedding vector of the second user with the representation vector corresponding to the second pushed knowledge to obtain a fused representation of the second user;

[0123] The push determination module 630 is configured to input the fused representation of the second user into the trained push model, and determine the push information for the second user through the push model.

[0124] In one embodiment, the second fusion module 620 includes: a second determination submodule and a second fusion submodule (not shown in the figure). The second determination submodule is configured to determine a first weight of the embedding vector of the second user and a second weight of the representation vector corresponding to the second pushed knowledge based on the first parameter obtained by training. The second fusion submodule is configured to use the first weight and the second weight to perform weighted fusion on the embedding vector of the second user and the representation vector corresponding to the second pushed knowledge.

[0125] In one embodiment, the push model adopts Figure 3 The method of the illustrated embodiment is used for training.

[0126] The above-mentioned device embodiments correspond to the method embodiments. For specific descriptions, please refer to the description of the method embodiments, which will not be repeated here. The device embodiments are obtained based on the corresponding method embodiments and have the same technical effects as the corresponding method embodiments. For specific descriptions, please refer to the corresponding method embodiments.

[0127] The present specification also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed in a computer, the computer is caused to execute Figures 1 to 4 Any of the methods described above.

[0128] The embodiment of the present specification also provides a computing device, including a memory and a processor, wherein the memory stores an executable code, and when the processor executes the executable code, Figures 1 to 4 Any of the methods described above.

[0129] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the storage medium and computing device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0130] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the embodiments of the present invention can be implemented by hardware, software, firmware, or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0131] The specific implementation methods described above further describe the purpose, technical solutions and beneficial effects of the embodiments of the present invention in detail. It should be understood that the above description is only a specific implementation method of the embodiments of the present invention and is not intended to limit the scope of protection of the present invention. Any modification, equivalent replacement, improvement, etc. made on the basis of the technical solution of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A push model training method based on a large model, comprising: Based on the matching results between the first user sample and several seed users in the inference pool, determine the first pushed knowledge corresponding to the first user sample; wherein the inference pool contains several seed users and corresponding pushed knowledge, and the pushed knowledge is information obtained by the large model through inference calculation of historical behavior data of the input seed users; Fusing the embedding vector of the first user sample with the representation vector corresponding to the first pushed knowledge to obtain a fused representation of the first user sample; The push model is updated based on the fused representation of the first user sample.

2. According to the method of claim 1, the seed user is determined in the following manner: Clustering all users based on their historical behavior data to obtain multiple clusters, and sampling a number of representative users from the multiple clusters; Based on the number of interactions corresponding to the historical behavior data of the entire set of users, a number of important users are sampled from the entire set of users; A number of seed users are determined based on the representative users and the important users.

3. According to the method of claim 1, the push knowledge of any seed user is inferred in the following way: Prompt data is constructed based on the historical behavior data of the seed user, the prompt data is input into the big model, and the push knowledge of the seed user is determined by the big model; the prompt data includes a first indication text and a second indication text; the first indication text is used to indicate the summary of the behavior preference of the seed user, and the second indication text is used to indicate the determination of the push knowledge based on the summarized behavior preference.

4. According to the method of claim 1, the step of fusing the embedding vector of the first user sample with the representation vector corresponding to the first pushed knowledge comprises: Determine a first weight of the embedding vector of the first user sample and a second weight of the representation vector corresponding to the first pushed knowledge based on the first parameter; The first weight and the second weight are used to perform weighted fusion on the embedding vector of the first user sample and the representation vector corresponding to the first pushed knowledge.

5. The method according to claim 4, further comprising: The first parameter is updated based on the fused representation of the first user sample. The method according to claim 1 , wherein the push knowledge includes categories and items of push objects.

7. A method for pushing information based on a large model, comprising: Based on the matching results between the second user and several seed users in the inference pool, determine the second pushed knowledge corresponding to the second user; wherein the inference pool contains several seed users and corresponding pushed knowledge, and the pushed knowledge is information obtained by the large model through inference calculation of historical behavior data of the input seed users; Fusing the embedding vector of the second user with the representation vector corresponding to the second pushed knowledge to obtain a fused representation of the second user; The fused representation of the second user is input into the trained push model, and a push item for the second user is determined by the push model.

8. According to the method of claim 7, the step of fusing the embedding vector of the second user with the representation vector corresponding to the second pushed knowledge comprises: Determine a first weight of the embedding vector of the second user and a second weight of the representation vector corresponding to the second pushed knowledge based on the first parameter obtained through training; The first weight and the second weight are used to perform weighted fusion on the embedding vector of the second user and the representation vector corresponding to the second pushed knowledge.

9. According to the method of claim 7, the push model is trained using the method of claim 1.

10. A push model training device based on a large model, comprising: A first matching module is configured to determine first pushed knowledge corresponding to the first user sample based on a matching result between the first user sample and a plurality of seed users in an inference pool; wherein the inference pool includes a plurality of seed users and corresponding pushed knowledge, and the pushed knowledge is information obtained by the large model through inference calculation of historical behavior data of the input seed users; A first fusion module, configured to fuse the embedding vector of the first user sample with the representation vector corresponding to the first pushed knowledge to obtain a fused representation of the first user sample; A model updating module is configured to update the push model based on the fusion representation of the first user sample.

11. An information push device based on a large model, comprising: A second matching module is configured to determine second pushed knowledge corresponding to the second user based on the matching results between the second user and a plurality of seed users in the inference pool; wherein the inference pool includes a plurality of seed users and corresponding pushed knowledge, and the pushed knowledge is information obtained by the large model through inference calculation of historical behavior data of the input seed users; A second fusion module is configured to fuse the embedding vector of the second user with the representation vector corresponding to the second pushed knowledge to obtain a fused representation of the second user; The push determination module is configured to input the fused representation of the second user into the trained push model, and determine the push item for the second user through the push model.

12. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 9.

13. A computing device, comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method according to any one of claims 1 to 9 is implemented.

Images