Item pushing method and apparatus, and electronic device and storage medium

Abstract

An item pushing method, which is executed by a computer device. The method comprises: on the basis of respective item features of historical interaction items involved in historical interaction operations executed by a push recipient, obtaining, by means of large language model reasoning, a set of association relationships between the historical interaction items, wherein each association relationship includes: at least two associated historical interaction items and an association reason (S21); in the set of association relationships, respectively replacing a plurality of selected candidate historical interaction items with masks of the candidate historical interaction items, and executing the following operations for each mask: on the basis of item features of other historical interaction items associated with one mask, performing contextual semantic analysis; on the basis of an analysis result, selecting from among the plurality of candidate historical interaction items an estimated historical interaction item corresponding to the mask; and when the estimated historical interaction item does not match a corresponding replaced candidate historical interaction item, deleting the replaced candidate historical interaction item and item features thereof (S22); and on the basis of push item features of an item to be pushed and respective target item features of target historical interaction items that have not been deleted, obtaining, by means of large language model reasoning, a target association relationship between the item to be pushed and at least one target historical interaction item; generating push information on the basis of the item to be pushed and a target association reason included in the target association relationship; and pushing the push information to the push recipient (S23).

Description

Methods, devices, electronic devices and storage media for pushing items

[0001] Related applications

[0002] This application claims priority to Chinese patent application filed on August 30, 2024, application number 2024112146775, entitled "A method, apparatus, electronic device and storage medium for recommending articles", the entire contents of which are incorporated herein by reference. Technical Field

[0003] This application relates to the field of computer technology, and in particular to a method, apparatus, electronic device, and storage medium for pushing items. Background Technology

[0004] With the development of the internet, users can easily access desired items online, such as articles, videos, topics, news, and products. To filter items of interest from this vast amount of information, push notification systems have emerged. When pushing items to users, these systems typically determine items the user might be interested in based on their historical behavioral data, and then select items from this pool to push, thus generating the desired results.

[0005] However, the push notification results output by the push system often lack interpretability. The object cannot understand why the push system pushes certain items, thus lacking trust and satisfaction with the pushed items. This makes it less likely that the object will perform interactive operations on the pushed items. For example, interactive operations include but are not limited to clicking, browsing, purchasing, and sharing, resulting in poor push effects.

[0006] Therefore, how to increase the probability that the pushed items in the push system will be interacted with by objects, thereby improving the push effect, is a problem that needs to be solved. Summary of the Invention

[0007] This application provides an item pushing method, apparatus, electronic device, and storage medium.

[0008] On one hand, embodiments of this application provide an item pushing method, executed by a computer device, the method comprising:

[0009] Based on the item characteristics of each historical interaction item involved in each historical interaction operation performed by the pushed object, the set of association relationships between the historical interaction items is obtained through large language model reasoning; each association relationship includes: at least two related historical interaction items and the reason for association;

[0010] In the set of association relationships, multiple selected candidate historical interaction items are replaced with masks of the candidate historical interaction items, and the following operations are performed for each mask: Contextual semantic analysis is performed based on the item features of other historical interaction items associated with a mask; based on the analysis results, an estimated historical interaction item corresponding to the mask is selected from the multiple candidate historical interaction items; when the estimated historical interaction item does not match the corresponding replaced candidate historical interaction item, the replaced candidate historical interaction item and its item features are deleted; and

[0011] Based on the push item characteristics of the item to be pushed and the target item characteristics of each target historical interaction item that has not been deleted, the target association relationship between the item to be pushed and at least one target historical interaction item is obtained through large language model reasoning. Push information is generated according to the target association reason contained in the item to be pushed and the target association relationship, and the push information is pushed to the target recipient.

[0012] On one hand, an embodiment of this application provides an item pushing device, comprising:

[0013] The reasoning unit is used to obtain a set of association relationships between the historical interactive items based on the item characteristics of each historical interactive item involved in each historical interactive operation performed by the pushed object, and through reasoning by a large language model; each association relationship includes: at least two related historical interactive items and the reason for association;

[0014] A verification unit is configured to replace multiple selected candidate historical interaction items with masks of the candidate historical interaction items in the association set, and perform the following operations for each mask: perform contextual semantic analysis based on the item features of other historical interaction items associated with a mask, and select the estimated historical interaction item corresponding to the mask from the multiple candidate historical interaction items based on the analysis results; when the estimated historical interaction item does not match the corresponding replaced candidate historical interaction item, delete the replaced candidate historical interaction item and its item features; and

[0015] The push unit is used to obtain the target association relationship between the item to be pushed and at least one target historical interaction item based on the push item characteristics of the item to be pushed and the target item characteristics of each target historical interaction item that has not been deleted, through large language model reasoning, and generate push information according to the target association reason contained in the item to be pushed and the target association relationship, and push the push information to the push recipient.

[0016] On one hand, an electronic device provided in this application includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of any of the above-described item pushing methods.

[0017] On one hand, embodiments of this application provide a computer-readable storage medium including a computer program, which, when run on an electronic device, causes the electronic device to perform the steps of any of the above-described item pushing methods.

[0018] On one hand, embodiments of this application provide a computer program product, the computer program product including a computer program stored in a computer-readable storage medium; when the processor of an electronic device reads the computer program from the computer-readable storage medium, the processor executes the computer program, causing the electronic device to perform the steps of any of the above-described item pushing methods.

[0019] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0021] Figure 1 is a schematic diagram of an application scenario of an item push method according to an embodiment of this application;

[0022] Figure 2 is a flowchart of an item push method in an embodiment of this application;

[0023] Figure 3 is a schematic diagram of an association logic diagram in an embodiment of this application;

[0024] Figure 4 is a schematic diagram of an association logic diagram including a mask in an embodiment of this application;

[0025] Figure 5 is a schematic diagram of a process for reasoning about a set of association relationships based on an initial large language model in an embodiment of this application;

[0026] Figure 6 is a schematic diagram of the conversion process of an association set in an embodiment of this application;

[0027] Figure 7 is a schematic diagram of a mask prediction process based on an initial large language model in an embodiment of this application;

[0028] Figure 8 is a schematic diagram of the verification process of estimated historical interaction items and candidate historical interaction items in an embodiment of this application;

[0029] Figure 9 is a schematic diagram of the fine-tuning process of an initial large language model in an embodiment of this application;

[0030] Figure 10 is a schematic diagram of the reasoning process for pushing an item to be pushed in an embodiment of this application;

[0031] Figure 11 is a schematic diagram of an article push interface in an embodiment of this application;

[0032] Figure 12 is a schematic diagram of the prediction process of an interactive probability prediction model in an embodiment of this application;

[0033] Figure 13 is a schematic diagram of the composition structure of an item pushing device in an embodiment of this application;

[0034] Figure 14 is a schematic diagram of the composition structure of an electronic device according to an embodiment of this application;

[0035] Figure 15 is a schematic diagram of the composition structure of another electronic device using an embodiment of this application. Detailed Implementation

[0036] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this application. Obviously, the described embodiments are only some embodiments of the technical solutions of this application, and not all embodiments. Based on the embodiments recorded in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the technical solutions of this application.

[0038] The following describes some of the concepts involved in the embodiments of this application.

[0039] 1. Large Language Model (LLM): LLM is a type of natural language processing technology based on deep learning. Its main purpose is to enable machines to better understand and generate human natural language text, such as articles, dialogues, and searches.

[0040] 2. Embedding: A numerical vector composed of multiple floating-point numbers, describing the various attributes and properties of an item in high-dimensional space. It is generally obtained by training a neural network model. Specifically, the original features of the item can be input into the neural network, undergo multiple layers of nonlinear transformations, and finally output as a fixed-dimensional vector through a fully connected layer, which is the item's embedding.

[0041] 3. Association Logic Graph: Used to describe the association relationships between multiple nodes, with each node connected to one or more other nodes. In this embodiment, each node represents a historical interaction item, and the edge between two nodes can represent the association reason between the two historical interaction items.

[0042] The word “exemplary” as used below means “serving as an example, embodiment, or illustration.” Any embodiment illustrated as an “exemplary” need not be construed as superior to or better than other embodiments.

[0043] The terms "first" and "second" used in this document are for descriptive purposes only and should not be construed as indicating relative importance or implying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0044] The design concept of the embodiments of this application will be briefly introduced below.

[0045] Currently, push notification systems typically determine items an object might be interested in based on its historical behavior data, and then select items from those potentially interesting to push. However, the push notification results often lack interpretability. Objects cannot understand why the system is pushing certain items, leading to a lack of trust and satisfaction with the pushed items. This results in a lower probability that the object will interact with the pushed items, leading to poor push notification performance.

[0046] In related technologies, while some push notification systems can generate reasons for recommending items, these reasons are usually based on the item's own characteristics (such as tags, attributes, keywords, etc.). For example, in movie recommendations, the system can analyze a user's preferred movie genres, directors, actors, etc., and recommend movies with similar characteristics, using these characteristics as the recommendation reason. However, for the user, a recommendation reason related to the item's own characteristics may not be persuasive enough.

[0047] In view of this, in order to generate convincing reasons for pushing items to increase the probability that the pushed items will be interacted with by the target object, thereby improving the push effect, this application provides an item push method, apparatus, electronic device, and storage medium. Based on the item characteristics of each historically interacted item of the target object, a set of association relationships between historically interacted items is obtained through large language model reasoning. Each association relationship involves at least two related historically interacted items and the reasons for their association. Then, by verifying the above association relationship set, unconvincing candidate historically interacted items are deleted to obtain convincing target historically interacted items. Afterwards, in obtaining... After obtaining the item to be pushed, based on the push item characteristics of the item to be pushed and the target item characteristics of each target historical interaction item, the target association relationship between the item to be pushed and at least one target historical interaction item can be accurately inferred. The target association reason contained in the target association relationship is used as the push reason and pushed to the target object along with the item to be pushed. In this way, the obtained push reason is related to the target historical interaction items of the target object and is persuasive. Through this push reason, the target object can understand why the push system pushed the item, thereby increasing the target object's trust and satisfaction with the pushed item, increasing the probability that the object will perform an interaction operation on the pushed item, and thus improving the push effect.

[0048] The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this application. Furthermore, the embodiments and features in the embodiments of this application can be combined with each other without conflict.

[0049] Figure 1 illustrates an application scenario of an embodiment of this application. The application scenario includes a terminal device 110 and a server 120.

[0050] In one alternative implementation, the terminal device 110 and the server 120 can communicate via a communication network; wherein the communication network can be a wired network or a wireless network.

[0051] In this embodiment, the terminal device 110 includes, but is not limited to, mobile phones, tablets, laptops, desktop computers, e-book readers, smart voice interaction devices, smart home appliances, and in-vehicle terminals. The terminal device may have a client installed related to item push notifications. This client can be software (e.g., browsers, video software, shopping software, music software), or a webpage, mini-program, etc. The server 120 can be a backend server corresponding to the software, webpage, or mini-program, or a server specifically used for item push notifications; this application does not impose specific limitations. The server 120 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0052] It should be noted that the item push method in each embodiment of this application can be executed by the terminal device 110 or the server 120, or it can be executed by both the terminal device 110 and the server 120.

[0053] In some embodiments, taking server 120 as an example, a client related to item push is installed in terminal device 110. The push system in server 120 can push items to objects using the client. The pushed object can perform interactive operations on the pushed item in the client. For example, if the pushed item is an article, the interactive operation can be clicking, reading, sharing, etc. Server 120 can obtain all historical interaction items involved in the historical interaction operations performed by the target object, and based on the item characteristics of each historical interaction item, obtain a set of association relationships between the historical interaction items through large language model inference. Each association relationship involves at least two related historical interaction items and the reason for their association. Then, by verifying the above association relationship set, unconvincing candidate historical interaction items are deleted from the association relationship set to obtain convincing target historical interaction items. After obtaining the item to be pushed, based on the push item characteristics of the item to be pushed and the target item characteristics of each target historical interaction item, the target association relationship between the item to be pushed and at least one target historical interaction item can be accurately inferred. Push information is generated according to the target association reason contained in the item to be pushed and the target association relationship, and the push information is pushed to the client of the target object. The client displays the push item and the push reason at the same time, so that the target object can intuitively understand why the push system pushes the item to be pushed.

[0054] Item push notifications have wide applications in many fields. Below are some specific application scenarios for item push notifications and their detailed explanations:

[0055] 1. E-commerce. On e-commerce websites, push systems can recommend products that users might be interested in, thereby increasing sales and user satisfaction. For example, personalized product recommendations suggest related products based on a user's browsing history, purchase history, and click behavior; cross-recommendations recommend other products related to those already purchased; new product recommendations suggest new products, especially those similar to those previously purchased; and products from subscribed stores can also be recommended.

[0056] 2. Video Streaming Services. Video streaming services recommend video content to users through a push system to increase user engagement and viewing time. For example, they may recommend similar videos based on a user's viewing history; offer trending recommendations, suggesting currently popular video content; generate personalized playlists for users, recommending videos they might be interested in; and recommend subscribed video content.

[0057] 3. News Recommendations. News push systems can recommend personalized news content to users. For example, they can recommend relevant news articles based on a user's reading history and preferences; they can also recommend current trending topics and related news; and they can recommend subscribed news content.

[0058] 4. Music Recommendations. Music recommendation systems can push songs that users might be interested in, enhancing their music experience. For example, they can generate personalized playlists based on a user's listening history and preferences; recommend newly released songs, especially those matching their preferences; recommend relevant concert and live performance information based on their music preferences; and even recommend subscribed music.

[0059] 5. Social Media. Social media platforms recommend content that users might be interested in through push notifications. For example, based on users' interests and historical interaction behavior, they can recommend posts, articles, videos, and content from public accounts that they might be interested in; they can also recommend content that users have subscribed to on social media platforms.

[0060] It should be noted that the application scenarios listed above are just simple examples. Other application scenarios are also applicable to the embodiments of this application, and will not be described in detail here.

[0061] Furthermore, it should be emphasized that in the specific embodiments of this application, the historical interactive operations, historical interactive items, item characteristics, and other data involved in the objects involved need to obtain the object's permission or consent when the embodiments of this application are applied to specific products or technologies, and the collection, use, and processing of related data need to comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0062] The following describes the item push method provided by the exemplary embodiments of this application in conjunction with the application scenarios described above and with reference to the accompanying drawings. It should be noted that the above application scenarios are only shown to facilitate understanding of the spirit and principles of this application, and the embodiments of this application are not limited in any way in this respect.

[0063] Referring to Figure 2, which is a flowchart of an implementation method for pushing items according to an embodiment of this application, taking the server as the executing entity as an example, the specific implementation process of this method includes the following S21-S23:

[0064] S21. Based on the item characteristics of each historical interaction item involved in each historical interaction operation performed by the pushed object, obtain the set of association relationships between each historical interaction item through large language model reasoning; each association relationship includes: at least two related historical interaction items and the reason for association.

[0065] In this application embodiment, historical interaction items include, but are not limited to: videos, information streams (e.g., articles, news), Q&A, topics, encyclopedias, posts, products, music, e-books, etc.

[0066] The server's item push platform can store a large number of historical interaction item sequences of object interactions. For each object, the historical interaction item sequence can include all historical interaction items involved in the object's various historical interaction operations within a set historical time period, such as 100 historical interaction items that the object has interacted with in the last 90 days. Each historical interaction item is identified by a unique item ID (Identifier), and the historical interaction items can be arranged in the order of interaction time.

[0067] The target of the push notification can be any object from a large number of objects, specifically a user. The historical interaction operations performed by the target on historically interacted items include, but are not limited to: browsing, clicking, reading (including long-term reading), playing, liking, commenting, sharing, purchasing, etc.

[0068] Each historical interaction item's characteristics can include: item attributes and historical interaction type. Item attributes can characterize the historical interaction item, including the item ID mentioned above, and can also include other characterizing information. For example, if the historical interaction item is an article, other characterizing information includes, but is not limited to, the article title, author (or account), and article category; if the historical interaction item is a video, other characterizing information includes, but is not limited to, the video title, video source, and video category; if the historical interaction item is a product, other characterizing information includes, but is not limited to, the product name, brand, model, and category. The historical interaction type can be understood as the operation type of the aforementioned historical interaction actions. When the recipient performs multiple historical interaction actions on the same historical interaction item, they can arbitrarily select one historical interaction action from among the multiple actions, or they can select a specific historical interaction action, and the operation type of the selected historical interaction action will be used as the historical interaction type.

[0069] After obtaining the item features of each historically interacted item in the pushed content recipient's history, these features can be arranged according to the order of interaction to form an item feature sequence. Then, association analysis can be performed on this sequence, and a set of association relationships between historically interacted items can be obtained through large language model inference. Specifically, the semantic similarity between item attributes in each item feature is calculated, and a similarity threshold T is set. When the semantic similarity between two item attributes is greater than T, the two item attributes are considered related. Combining the order of interaction and the historical interaction types in each item feature, if two historically interacted items have related item attributes and their interaction order follows a certain logic (e.g., interacting with item A before interacting with item B), and their historical interaction types also have a certain correlation (e.g., clicking on item A triggers reading item B), then these two historically interacted items are considered to be related. By traversing the item feature sequence, multiple association relationships between historically interacted items are inferred, thus obtaining a set of association relationships. The historically interacted items in the association relationship set can be some or all of the historically interacted items in the initially obtained historically interacted item sequence. Where T is a pre-set threshold used to determine whether the semantic similarity of item attributes is related.

[0070] On the one hand, each association in the association set can include two related historical interaction items and an association reason. For example, for related historical interaction items 1 and 2, the interaction of the pushed object with historical interaction item 1 triggers the interaction of the pushed object with historical interaction item 2. At this time, historical interaction item 2 corresponds to an association reason, that is, an interaction reason.

[0071] On the other hand, each association in the association set can include two or more historical interaction items. When an association includes two or more historical interaction items, there are multiple association reasons. Two or more historical interaction items can be associated sequentially to form an association chain. For example, for the associated historical interaction items 1, 2, and 3, the interaction of the pushed object with historical interaction item 1 triggers the interaction of the pushed object with historical interaction item 2. At this time, historical interaction item 2 corresponds to an association reason. The interaction of the pushed object with historical interaction item 2 then triggers the interaction of the pushed object with historical interaction item 3. At this time, historical interaction item 3 also corresponds to an association reason.

[0072] Specifically, the set of relationships can be converted into a relationship logic graph, which includes multiple nodes. Each node represents a historical interaction item, and the node information of each node contains the item characteristics of the corresponding historical interaction item. Each node is connected to one or more other nodes, and the edge between each two nodes has an arrow to point from one node to another. Each edge can correspond to a relationship reason, that is, the node pointed to by the arrow of each edge also contains the relationship reason in its node information.

[0073] For example, as shown in Figure 3, assuming the association logic diagram includes nodes V1, V2, V3, V4, V5, V6, and V7, when each association link is considered as an association relationship, V1->V2->V6 is one association relationship, V1->V3->V4 is another association relationship, and V1->V5->V7 is yet another association relationship. In addition, each association relationship can also include two associated nodes (i.e., historical interaction items), such as V1->V2, V2->V6, V1->V3, V3->V4, V1->V5, and V5->V7.

[0074] One implementation method is to use a pre-stored initial large language model. Based on the item feature sequences of the aforementioned historical interactive items, the large language model infers the set of relationships between these historical interactive items. The initial large language model possesses powerful text semantic understanding capabilities and a large amount of universally applicable knowledge, which can be used to infer the interaction logic of the pushed object with each historical interactive item. The initial large language model can be an existing large language model, such as a Generative Pre-trained Transformer (GPT) model.

[0075] In practical applications, appropriate prompts can be constructed based on the item feature sequences of the aforementioned historical interactive items to guide the initial large language model inference to obtain the set of relationships between the historical interactive items. The construction process of the prompts will be described in detail in subsequent embodiments.

[0076] S22. In the association set, replace the selected candidate historical interaction items with the mask of the candidate historical interaction item, and perform the following operations for each mask: perform contextual semantic analysis based on the item features of other historical interaction items associated with a mask, and select the estimated historical interaction item corresponding to the mask from the multiple candidate historical interaction items based on the analysis results. When the estimated historical interaction item does not match the corresponding replaced candidate historical interaction item, delete the replaced candidate historical interaction item and its item features.

[0077] To test the set of associations obtained through large language model inference, some historical interaction items in the set of associations were masked, and the historical interaction items corresponding to each masked position were estimated.

[0078] Specifically, from the set of historical interaction items included in the association relationship, multiple candidate historical interaction items are randomly selected, and each candidate historical interaction item is replaced with a mask. When a candidate historical interaction item corresponds to an association reason, both the candidate historical interaction item and the corresponding association reason can be replaced with the mask. In some embodiments, when a candidate historical interaction item corresponds to an association reason, only the candidate historical interaction item is replaced with the mask, while the corresponding association reason is retained for subsequent analysis, ensuring the accuracy of the reasoning.

[0079] As an optional implementation, when replacing multiple selected candidate historical interaction items with their respective masks in the association set, the following operations can be performed on each of the multiple historical interaction items included in the association set: Based on a set probability, randomly determine whether to select a historical interaction item; if a historical interaction item is selected, designate it as a candidate historical interaction item and replace it with its corresponding mask. This ensures that each historical interaction item in the association set has the same probability of being selected.

[0080] The set probability refers to the probability value used when selecting candidate historical interaction items from the association set. For multiple historical interaction items included in the association set, the selection of a historical interaction item is randomly determined according to this set probability. If the randomly generated number between 0 and 1 is less than or equal to the set probability, then that historical interaction item is selected as a candidate historical interaction item, ensuring that each historical interaction item in the association set has the same probability of being selected. The set probability can be set as needed, for example, to 0.2, 0.3, etc., without limitation.

[0081] In practice, for a given historical interactive item, a random number generator can be used to generate a random number between 0 and 1. If this random number is less than or equal to a set probability, the historical interactive item is considered selected; otherwise, it is not selected. Specifically, according to a set probability p, a random number generator is used to generate a random number r in the interval [0,1]. If r ≤ p, the historical interactive item is considered selected, treated as a candidate historical interactive item, and replaced with the corresponding mask; if r > p, the historical interactive item is not selected. Here, p is a pre-set probability value, and r is the random number generated by the random number generator.

[0082] For example, the association logic diagram after the association set is transformed is shown in Figure 3. This association logic diagram is only a schematic diagram and does not include all the nodes corresponding to the historical interaction items. The candidate historical interaction items randomly selected in the association set are replaced with masks, such as MASK1 and MASK2 in Figure 4. The candidate historical interaction item corresponding to MASK1 is node V2, and the candidate historical interaction item corresponding to MASK2 is node V7.

[0083] After obtaining the set of associations for each mask, for each mask, contextual semantic analysis is performed based on the item features of other historical interaction items associated with that mask to select the estimated historical interaction item corresponding to that mask from multiple candidate historical interaction items. Specifically, when the above set of associations is obtained through inference using an initial large language model, the initial large language model can be used to predict the estimated historical interaction item corresponding to each mask.

[0084] In practical applications, appropriate prompts can be constructed based on the set of associations for each mask and multiple selected candidate historical interaction items. This guides the initial large language model to obtain the estimated historical interaction item corresponding to each mask through large language model inference. The construction process of the prompts will be described in detail in subsequent embodiments.

[0085] For each mask, if the estimated historical interaction item does not match the corresponding candidate historical interaction item to be replaced, it is considered that the association relationship corresponding to the candidate historical interaction item cannot be accurately inferred, that is, the candidate historical interaction item is not trusted, and the candidate historical interaction item and its item features are deleted.

[0086] After deleting one or more candidate historical interaction items and their features that cannot be accurately inferred from the initially obtained historical interaction items, the remaining historical interaction items are taken as target historical interaction items, thereby obtaining the target historical interaction items and their features that were not deleted, so as to be used for subsequent inference of the relationship between the item to be pushed and the target historical interaction items.

[0087] S23. Based on the push item features of the item to be pushed and the target item features of each target historical interaction item that has not been deleted, the target association relationship between the item to be pushed and at least one target historical interaction item is obtained through large language model reasoning. Push information is generated according to the target association reason contained in the item to be pushed and the target association relationship, and the push information is pushed to the target object.

[0088] Among these, the item push platform can push items to recipients. Specifically, it can push items based on the recipient's subscription information or proactively push items to recipients. Before pushing an item, the platform can identify the reason for the push.

[0089] The characteristics of the item to be pushed can include the item attributes and the set interaction type. The item attributes are similar to those of the historically interacted items. The set interaction type can be a specified interaction type or any one of several interaction types. For example, if the item to be pushed is an article, the set interaction type can be a specified interaction type, such as click, or any one of click, read, share, like, etc.

[0090] Each target historical interaction item has its own target item features, which are arranged according to the order of interaction of each target historical interaction item to obtain the target item feature sequence. Based on the target item feature sequence and the push item features, the target association relationship between the push item and at least one target historical interaction item is inferred. This inference process is similar to the inference process of the item feature sequence in S21 above.

[0091] Optionally, the initial large language model described above can be used to infer the target association relationship between the target item and at least one target historical interaction item based on the target item feature sequence and the push item features. Furthermore, the initial large language model can be fine-tuned. The fine-tuned sample set can be obtained from the aforementioned association set. For example, some associations in the association set can be used as sample associations. These partial associations can be the associations to which each candidate historical interaction item belongs, as determined in S22 above, and can be accurately inferred. Then, the fine-tuned target large language model is used to infer the aforementioned target association relationship.

[0092] When the target association between the item to be pushed and a target historical interaction item is obtained through large language model inference, the target association reason in the target association relationship is used as the push reason; when the target association between the item to be pushed and multiple target historical interaction items is obtained through large language model inference, the item to be pushed and multiple target historical interaction items form an association link. This target association relationship contains multiple target association reasons, that is, there is a target association reason between every two items (the item to be pushed or the target historical interaction item). At this time, the target association reason corresponding to the item to be pushed is used as the push reason, and the push reason and the item to be pushed are pushed to the target recipient.

[0093] It should be noted that when multiple items to be pushed are obtained corresponding to the pushed object, the above S23 can be executed for each item to be pushed, or the above S23 can be executed for multiple items to be pushed at the same time. That is, based on the push item features of each of the multiple items to be pushed and the above target item feature sequence, the target association relationship between each item to be pushed and at least one target historical interaction item is inferred.

[0094] In this embodiment, based on the item characteristics of each historically interacted item of the target object, a set of association relationships between historically interacted items is obtained through large language model reasoning. Each association relationship involves at least two related historically interacted items and the reason for their association. Then, by verifying the above association relationship set, unconvincing candidate historically interacted items are deleted to obtain convincing target historically interacted items. After obtaining the item to be pushed, based on the push item characteristics of the item to be pushed and the target item characteristics of each target historically interacted item, the target association relationship between the item to be pushed and at least one target historically interacted item can be accurately inferred. The target association reason contained in the target association relationship is used as the push reason and pushed to the target object along with the item to be pushed. In this way, the obtained push reason is related to the target historically interacted items of the target object and is persuasive. Through this push reason, the target object can understand why the push system pushes the item, thereby increasing the target object's trust and satisfaction with the item, increasing the probability that the object will perform an interaction operation on the item, and thus improving the push effect.

[0095] In some embodiments, in S21 above, a pre-stored initial large language model can be used to obtain a set of association relationships between historical interactive items based on the item characteristics of each historical interactive item involved in each historical interactive operation performed by the pushed object, through large language model reasoning. Specifically, this includes the following steps A1-A2:

[0096] A1. Construct the first prompt information based on the individual item characteristics of each historical interactive item.

[0097] Based on the above embodiments of this application, the item characteristics of each historical interaction item may include: the item attributes of the corresponding historical interaction item and the historical interaction type of the pushed object to the corresponding historical interaction item; and, the historical interaction items are arranged in the order of interaction, and correspondingly, the item characteristics of each historical interaction item are also arranged in the order of interaction, forming an item characteristic sequence.

[0098] To enable the initial large language model to infer the set of relationships between historical interactive items, the above item feature sequence can be textualized according to the preset first prompt word template to construct the first prompt information. This first prompt information is used to describe the item features of each historical interactive item in the item feature sequence, as well as the analysis method of the item feature sequence and the setting format of the output results.

[0099] The first prompt message is constructed based on the item characteristics of each historical interactive item involved in the various historical interactions performed by the pushed object. Following a preset first prompt word template, it transcribes the item characteristic sequence into text, describing the item characteristics of each historical interactive item in the sequence, the analysis method of the item characteristic sequence, and the format of the output results. This guides the initial large language model to analyze the behavioral logic of the pushed object's interactions with each historical interactive item, thereby outputting a set of relationships between these historical interactive items.

[0100] The aforementioned first prompt can guide the initial large language model to analyze the behavioral logic of the pushed object interacting with each historical interactive item, thereby outputting a set of association relationships between each historical interactive item.

[0101] A2. Using a pre-stored initial large language model, based on the item attributes and historical interaction types of each historical interactive item contained in the first prompt information, as well as the interaction sequence of each historical interactive item, analyze the relationships between each historical interactive item and the reasons for each relationship, and obtain the set of relationships between each historical interactive item.

[0102] After the initial prompt information is input into the initial large language model, the model performs semantic analysis on it. Then, based on the item attributes and historical interaction types of each historical interactive item contained in the prompt information, as well as the interaction order of these items, the following analysis method is used: For each historical interactive item, starting from it, based on the semantic similarity of its item attributes and the association of its historical interaction types, potentially related items are searched among subsequent historical interactive items. If the semantic similarity of the item attributes of two historical interactive items is greater than a set threshold T1, and there is a causal relationship in their historical interaction types (e.g., clicking triggers reading), then these two historical interactive items are considered to be related. The reason for the association can be determined based on the specific circumstances of the item attributes and historical interaction types (e.g., similar interest categories, specific user behavior habits, etc.). Finally, according to the output format indicated by the first prompt information, the set of association relationships between each historical interactive item is output. Here, T1 is a pre-set threshold used to determine whether the semantic similarity of item attributes is sufficient to determine an association relationship.

[0103] For example, as shown in Figure 5, the item feature sequence of each historical interaction item of the pushed object is S[V1, V2, V3, ..., V...]. n Assuming the historical interaction items are articles, each item's characteristics include: item ID, title, account author, article category, and historical interaction type. For example, the first prompt information constructed based on the item characteristic sequence is as follows:

[0104] The following input information describes a sequence of item characteristics for multiple historical interaction items of the pushed object, with the items arranged in chronological order. Each historical interaction item's characteristics include five categories of information: unique ID, title, account author, article category, and interaction type, recorded within parentheses "()".

[0105] (1, "Under the algorithm, real problems are ignored", "xxAPP", "Internet", "click"), ..., (998, ...), ..., (10000, ...).

[0106] Please analyze the reasons for each historical interaction of the pushed items, trying to extract these reasons from a global perspective of the item feature sequence. Uncertain historical interaction items can be ignored. Output the analysis results in the following format:

[0107] {{Interaction Type 1}} Item {{ID1}} triggered {{Interaction Type 2}} Item {{ID2}} because it was the user's {{Behavior Reason 1}}.

[0108] The first prompt message can use special identifiers, such as "{{}}", to represent input information or information that needs to be output according to a set format.

[0109] After inputting the first prompt information into the initial large language model, the initial large language model analyzes the set of relationships between historical interactive items based on the first prompt information. Each relationship can be understood as a causal relationship between at least two historical interactive items. The order of each historical interactive item in the historical interactive item sequence is reorganized, and the reasons for the recipient's interaction with each historical interactive item are analyzed. For example, the output result is:

[0110] "Clicking item {10000} triggered reading time for item {998} because they belong to the user's category of shared interests; sharing item {81} triggered clicks for item {82} because they are viewed by the user's accounts..."

[0111] The reason for interacting with each historical interactive item serves as the associated reason for that historical interactive item.

[0112] The initial output of the large language model is a set of association relationships between historically interacted items, as shown in Figure 6. After converting this set of association relationships into an association logic graph, the nodes and adjacency relationships in the association logic graph can be formatted and stored to obtain an adjacency list. For example: (1, Click, "") -> (2, High-duration reading, "Same interest category"), (2, Share, "") -> (6, Click, "Accounts viewed"), that is, each node in the adjacency list can correspond to the item ID, historical interaction type, and association reason (if any). "" indicates empty, that is, no association reason.

[0113] In this embodiment of the application, since the initial large language model has a powerful text semantic understanding capability and has accumulated a large amount of universal knowledge, it is able to analyze the relationship between historical interactive items based on the first prompt information and the item feature sequence of each historical interactive item, and generate or summarize convincing reasons for the relationship.

[0114] In one possible implementation, considering the vast textual expression space, the association reasons in each association relationship in the initial large language model inference may be highly divergent and lack refinement. To address this issue, this embodiment of the application defines a preset set of association reasons in the aforementioned first prompt information, instructing the initial large language model to select association reasons from this set for each association relationship. The set of association reasons can be set empirically.

[0115] The set of association reasons is a pre-defined set of association reasons, established based on experience. Introducing this set when constructing the first prompt information instructs the initial large language model to select association reasons from each association relationship from the set of association reasons, thus addressing the problem of high divergence and insufficient refinement of association reasons in the initial large language model's reasoning.

[0116] At this point, in step A1 above, the first prompt information can be constructed based on the item characteristics of each historical interactive item and the preset set of associated reasons.

[0117] Furthermore, using the initial large language model, based on the item attributes and historical interaction types of each historical interactive item contained in the first prompt information, as well as the interaction sequence of each historical interactive item, we analyze the relationships between each historical interactive item. For each relationship obtained from the analysis, we select the corresponding relationship reason from the set of relationship reasons to obtain the set of relationship relationships between each historical interactive item.

[0118] In this embodiment of the application, by introducing a preset set of association reasons in the first prompt information, the initial large language model selects appropriate association reasons from the set of association reasons when reasoning about each association relationship, thus ensuring that the association reasons reasoned by the initial large language model are relatively concise.

[0119] In some embodiments, in S22 above, an initial large language model can be used to predict the estimated historical interaction items corresponding to each mask. Specifically, a second prompt information can be constructed based on the association set containing each mask and multiple candidate historical interaction items. Using the initial large language model, the following operations are performed for each mask: based on the item features of other historical interaction items associated with a mask contained in the second prompt information, and the semantic similarity between these features and the item features of each of the multiple candidate historical interaction items, the candidate historical interaction items whose semantic similarity meets the set conditions are selected from the multiple candidate historical interaction items as the estimated historical interaction item corresponding to a mask.

[0120] In some embodiments, using an initial large language model, the following operations are performed for each mask: For the item features of other historically interacted items associated with a mask contained in the second prompt information, their item attributes and historical interaction types are vectorized respectively to obtain vector V1; For the item features of multiple candidate historically interacted items, their item attributes and historical interaction types are similarly vectorized respectively to obtain vector V2. The cosine similarity S between V1 and V2 is calculated using the formula: Where · represents the vector dot product, and |V1| and |V2| represent the moduli of vectors V1 and V2, respectively. From multiple candidate historical interaction items, the candidate historical interaction items whose cosine similarity S satisfies the set conditions (such as S being the largest) are selected as the predicted historical interaction items corresponding to a mask.

[0121] The initial large language model possesses powerful semantic understanding capabilities. It can analyze the semantic information of the item features of other historical interaction items and the semantic information of the item features of multiple candidate historical interaction items, thereby determining the semantic similarity between the item features of other historical interaction items and the item features of multiple candidate historical interaction items. The specific condition can be that the semantic similarity is maximized.

[0122] Specifically, the set of relationships containing each mask can be represented as a logical relationship diagram containing each mask. Based on this logical relationship diagram and multiple candidate historical interaction items, combined with a preset second prompt word template, a second prompt message can be constructed.

[0123] The second prompt information is constructed based on the association set containing each mask (represented as an association logic graph containing each mask) and multiple candidate historical interaction items, combined with a preset second prompt word template. Its function is to guide the initial large language model, for each mask, to select candidate historical interaction items whose semantic similarity meets the set conditions based on the item features of other historical interaction items associated with that mask and the individual item features of multiple candidate historical interaction items. These candidate historical interaction items are then used as the predicted historical interaction items corresponding to that mask.

[0124] For example, as shown in Figure 7, the set of relationships containing each mask is represented as a relationship logic graph. This relationship graph is only exemplary and does not show all historical interaction items. Based on the preset first prompt word template, the relationship logic graph, and multiple candidate historical interaction items, the following second prompt information is constructed:

[0125] "The given input information consists of some of the historical interaction items and interaction reasons of the target object. Select the most logical candidate historical interaction item from the candidate historical interaction item set and fill it into [MASK] to complete the set of association relationships."

[0126] The following is the input information:

[0127] (1, "Under the algorithm, real problems are ignored", click, "")->[MASK1]->(6, "Information accessibility", share, ""),(1, "Under the algorithm, real problems are ignored", click, "")->(5, "The lack of human factors in data-driven decision-making", share, "")->[MASK2],…

[0128] The following is a collection of candidate historical interaction items:

[0129] (2, "Limitations of the Algorithm", "Same Interest Categories")

[0130] (7, "The real summer's top influencers", "Accounts I've watched")

[0131] …

[0132] Output the analysis results in the following format:

[0133] {{MASK1}}:{{Candidate Historical Interaction Item 1}},{{MASK2}}:{{Candidate Historical Interaction Item 2}},…”

[0134] After inputting the second prompt information into the initial large language model, the output of the initial large language model is the estimated historical interaction item corresponding to each mask, that is, a candidate historical interaction item selected from the set of candidate historical interaction items. For example, the output of the initial large language model is as follows:

[0135] MASK1:(7, "The real summer's top influencer", "Accounts I've watched")

[0136] MASK2:(2, "Limitations of the algorithm", "Same interest category")

[0137] "

[0138] ...

[0139] Compared to the adjacency list storage mentioned above (including item ID + historical interaction type + interaction reason), the node information of each node in the masked association logic graph also includes item attributes (such as title text), which makes it easier for the initial large language model to understand the unmasked historical interaction items and candidate historical interaction items at the semantic level.

[0140] As shown in Figure 8, after obtaining the output of the initial large language model, the candidate historical interaction items replaced by each mask can be obtained according to the aforementioned stored adjacency list. For each mask, the corresponding estimated historical interaction item is checked to see if it matches the replaced candidate historical interaction item. For example, if the item ID of the candidate historical interaction item replaced by MASK1 is 2, while the item ID of the estimated historical interaction item corresponding to MASK1 is 7, they do not match. Similarly, if the item ID of the candidate historical interaction item replaced by MASK2 is 7, while the item ID of the estimated historical interaction item corresponding to MASK2 is 2, they also do not match, and so on, performing the matching process for all masks.

[0141] In this embodiment, for the set of association relationships between historical interaction items obtained by the initial large language model through inference, multiple candidate historical interaction items are masked, and the initial large language model infers the estimated historical interaction items corresponding to each mask. For each mask, the corresponding estimated historical interaction item is matched with the corresponding replaced candidate historical interaction item, and the mismatched candidate historical interaction items are deleted, so that the initial large language model has the ability to self-correct the inferred association relationship set.

[0142] In some embodiments, considering that the initial large language model is a general model rather than a task-specific model, even if some of the associations inferred in the above embodiments are accurate, it cannot be guaranteed that the output associations will be accurate and consistent for the same multiple historical interaction items. Therefore, the initial large language model is fine-tuned to better suit the task of inferring associations. To enable the initial large language model to more accurately infer the associations between the item to be pushed and each target historical interaction item, supervised fine-tuning of the initial large language model can be performed.

[0143] First, obtain the sample set for supervised fine-tuning.

[0144] When using the initial large language model to predict each mask in the association set, for each accurately predicted mask—that is, when the predicted historical interaction item corresponding to the mask matches the corresponding replaced candidate historical interaction item—the initial large language model is considered to have made relatively accurate inferences about the association to which the candidate historical interaction item belongs. Further accuracy verification of the association to which the candidate historical interaction item belongs can be performed. The specific verification method is as follows: recalculate the semantic similarity between the item attributes of each historical interaction item in the association. If the similarity is greater than a set threshold T2, and the association of historical interaction types conforms to preset logical rules (e.g., the probability of triggering reading after clicking is greater than a certain value), then the association is considered to have passed the verification and is used as a sample association. Here, T2 is a pre-set threshold used to determine whether the semantic similarity of item attributes in the association is sufficient to determine the accuracy of the association.

[0145] For example, assuming the set of association relationships containing each mask is converted into the association logic diagram shown in Figure 4 above, when the initial large language model accurately predicts the mask MASK1, the association relationship V2->V6 corresponding to V1->MASK1 can be verified for accuracy, and after the verification is passed, the association relationship is used as a sample association relationship.

[0146] Furthermore, for each accurately predicted mask corresponding to the association relationship, based on the item characteristics of each historical interactive item contained in the association relationship, the association relationship between these historical interactive items can be inferred again through the initial large language model. The inference process is similar to the inference process in S21 of the above embodiment. The accuracy of the re-output association relationship is verified, and after the verification is passed, the re-output association relationship is used as a sample association relationship.

[0147] Based on the above process, multiple sample associations can be obtained, which can serve as a sample set for supervised fine-tuning of the initial large language model.

[0148] Then, to reduce the complexity of fine-tuning the initial large language model, the trained parameters in the initial large language model's defined network layers can be replaced with pre-defined parameters to be trained, resulting in the large language model to be trained. The pre-trained parameters are those used to replace the trained parameters in the defined network layers during supervised fine-tuning of the initial large language model. These are typically low-rank matrices; using them to replace trained parameters reduces the number of parameters and computational complexity while maintaining model performance as much as possible. During training, the loss value is determined based on the difference between the predicted association and the sample association, and these pre-trained parameters are adjusted according to the loss value.

[0149] The network layers in this context can be key layers in the initial large language model, i.e., layers that significantly impact model performance, such as attention networks or feedforward neural network layers. The parameters to be trained can specifically be low-rank matrices, which reduce the number of parameters and computational complexity while preserving model performance as much as possible. Low-rank matrices are used to replace the trained parameters in the network layers during supervised fine-tuning of the initial large language model. They are constructed by performing matrix factorization (e.g., singular value decomposition) on the weight matrix of the linear transformation in the network layers (such as attention networks), retaining the k largest singular values ​​and their corresponding singular vectors. Using low-rank matrices reduces the number of parameters and computational complexity while preserving model performance as much as possible.

[0150] If the network layer is set as an attention network layer, when constructing the low-rank matrix, the weight matrix of the linear transformation in the attention network layer can be decomposed. Singular value decomposition (i.e., eigenvalue decomposition) or other appropriate decomposition methods can be used, for example, expressed as the following equation (1): W=U∑V T (1)

[0151] Where W represents the weight matrix, U and V represent orthogonal matrices, and ∑ represents the singular matrix. Orthogonal matrices are those that appear in singular value decomposition. A characteristic of orthogonal matrices is that their transpose is equal to their inverse, i.e., U = ∑V. T U=I and VT V = I (where I is the identity matrix) plays a crucial role in matrix factorization and the construction of low-rank matrices. The singular matrix, a matrix in singular value decomposition, contains the singular values ​​of the weight matrix W. When constructing a low-rank matrix, the k largest singular values ​​and their corresponding singular vectors are retained for building the low-rank matrix W. k .

[0152] Then, retaining the top k largest singular values ​​and their corresponding singular vectors in ∑, a low-rank matrix is ​​constructed, which can be expressed as follows (2): W k =U k ∑ k V k T (2)

[0153] Among them, U k and V k It contains the first k columns of U and V respectively, ∑ k Let represent the k*k singular values ​​in ∑.

[0154] Using a low-rank matrix W k Replace the original weight matrix W to obtain the large language model to be trained.

[0155] Finally, based on the obtained association relationships among multiple samples, the large language model to be trained undergoes multiple rounds of iterative training to obtain the target large language model; each round of training includes:

[0156] The sample item features of at least two historical interaction items included in a sample association relationship are input into a large language model to obtain the predicted association relationship. Based on the difference between the predicted association relationship and the sample association relationship, a loss value is determined, and the training parameters in a set network layer are adjusted based on this loss value. The loss value refers to the value determined based on the difference between the predicted association relationship and the sample association relationship during the training of the large language model. The loss value is obtained by calculating the Euclidean distance between the vectors of association relationships in the predicted and sample association relationships. Then, based on this loss value, the gradient is calculated using the backpropagation algorithm, and the training parameters in a set network layer are adjusted to minimize the loss value.

[0157] Historical interaction items are the historical interaction items contained in a sample association relationship. During supervised fine-tuning of the initial large language model, the sample item features of at least two historical interaction items contained in a sample association relationship are input into the large language model to obtain the predicted association relationship. The model is trained by comparing the difference between the predicted association relationship and the sample association relationship. Sample item features are the characteristics possessed by historical interaction items, identical to the item features of historical interaction items, including item attributes and historical interaction types. During the training of the large language model, sample item features are input into the model to obtain the predicted association relationship, which is then used to adjust the model parameters. The predicted association relationship refers to the association relationship output by the model after inputting the sample item features of at least two historical interaction items contained in a sample association relationship into the large language model during training. By comparing the difference between the predicted association relationship and the sample association relationship, the loss value is determined, and the model parameters are then adjusted.

[0158] Specifically, for the predicted association and the sample association, the reasons for the association are vectorized to obtain vector V. p and V s Calculate the Euclidean distance D between them, using the formula: Where n is the dimension of the vector, and They are vectors V p and V s The i-th element. A loss function L = D is defined, with the Euclidean distance D as the loss value. The gradient is calculated using the backpropagation algorithm based on this loss value, and then the training parameters in the defined network layers are adjusted. The backpropagation algorithm is used to adjust model parameters based on the loss value when training a large language model. After determining the loss value based on the difference between the predicted association and the sample association, the gradient is calculated using the backpropagation algorithm, and then the training parameters in the defined network layers are adjusted to minimize the loss value, thereby optimizing model performance.

[0159] The sample item characteristics of each historical interaction item are the same as the item characteristics of the historical interaction items in the above embodiments of this application.

[0160] Specifically, the loss value between the predicted association and the sample association is calculated based on a set loss function, such as the cross-entropy loss function. Then, the gradient is calculated using the backpropagation algorithm based on the loss value, and the trainable parameters in the set network layers are adjusted to minimize the loss. The initial stopping condition for the language model iteration can be that the number of iterations reaches a set threshold, or the model loss value is less than a set loss threshold. The cross-entropy loss function is a commonly used loss function that can be used as the set loss function when training a large language model. It measures the difference between the predicted association and the sample association. Based on this loss value, the gradient is calculated using the backpropagation algorithm, and the trainable parameters in the set network layers are adjusted to make the model's prediction result closer to the sample association.

[0161] For example, Figure 9 illustrates the fine-tuning process of the large language model to be trained. Assuming the features of each sample item in a sample association relationship are represented as V1, V2, and V6, V1, V2, and V6 are input into the large language model to be trained (with a low-rank matrix introduced in the network layers). The output is the predicted association relationship, i.e., the predicted association relationship between the sample items corresponding to V1, V2, and V6. Based on the difference between the predicted association relationship and the actual sample association relationship, a loss value is calculated, and the parameters of the low-rank matrix are adjusted based on this loss value. Finally, after the large language model to be trained is fine-tuned, the target large language model is obtained.

[0162] In this embodiment of the application, by supervising the fine-tuning of the initial large language model, the fine-tuned target large language model is made more suitable for the reasoning task of the above-mentioned set of associations. Subsequently, after obtaining the item to be pushed, the association between the item to be pushed and each target historical interaction item can be more accurately inferred, thereby accurately mining the reason for pushing the item to be pushed.

[0163] In some embodiments, when reasoning about the target association between the item to be pushed and at least one target historical interaction item in S23 above, the above-mentioned fine-tuned target large language model can be used.

[0164] The push item features include the push item attributes and set interaction type of the push item, and the target item features include the item attributes and historical interaction types of the corresponding target historical interaction items. Specifically, based on the push item features and the features of each target item, a third prompt message is constructed. Using a target large language model, based on the item attributes and historical interaction types of each target historical interaction item contained in the third prompt message, the push item attributes and set interaction types, and the interaction sequence of each target historical interaction item, a correlation analysis is performed on the push item and each target historical interaction item to obtain the target association relationship between the push item and at least one target historical interaction item.

[0165] The process of constructing the third prompt information based on the characteristics of the pushed items and the characteristics of each target item is similar to the process of constructing the first prompt information based on the item feature sequence in the above embodiment. That is, the above-mentioned first prompt word template can be used to construct the third prompt information.

[0166] The third prompt information is constructed based on the characteristics of the item to be pushed and the characteristics of the target items in the target's historical interactions. The construction process is similar to that of the first prompt information based on the item feature sequence, and a first prompt word template can be used. It instructs the target large language model to perform association analysis on the item to be pushed and each target historical interaction item based on the item attributes, historical interaction types, pushed item attributes, set interaction types, and the interaction sequence of each target historical interaction item, thereby obtaining the target association relationship between the item to be pushed and at least one target historical interaction item.

[0167] In this embodiment of the application, the semantic reasons for pushing items can be mined based on the target large language model. These reasons mainly revolve around the target's interest categories and historical interaction items. For example, the reasons for pushing include "interest category: cars", "accounts I often watch", and "related events of content I have watched".

[0168] The following example illustrates the reasoning process for the push notification of the item to be pushed, with reference to Figure 10.

[0169] For example, as shown in Figure 10, when the recipient logs into a client that provides push functionality on a terminal device, the item push platform on the server can push the item to be pushed to the recipient. At this time, the reason for pushing the item can be analyzed. Specifically, the server can construct a third prompt based on the item feature sequence S'[V1, V2, V3, ..., Vm] of each target historical interaction item and the push item feature [W] of the item to be pushed. This third prompt is used to instruct the target big language model to infer the target association relationship between the item to be pushed and at least one target historical interaction item. After the third prompt is input into the target big language model, the target big language model analyzes and obtains the target association relationship based on the third prompt: the target association relationship between the target historical interaction item corresponding to V3 and the item to be pushed, as well as the target association reason. The target association reason is used as the push reason and pushed to the recipient along with the item to be pushed.

[0170] For example, as shown in Figure 11, which is a schematic diagram of the message interface after subscription, the item to be pushed to the recipient is the article "Who invented this method? The taste is absolutely amazing", and the reason for the push is "a public account I often read", so that the recipient can intuitively understand why the article is being pushed.

[0171] In some embodiments, based on the latest set of association relationships formed by each target historical interactive item obtained in the above embodiments, a corresponding association logic graph is generated; then, graph feature extraction is performed on the association logic graph to obtain association logic graph features; based on the preset item features and association logic graph features of the item to be pushed, the interaction probability of the pushed object with the item to be pushed is predicted through a pre-stored interaction probability prediction model.

[0172] Graph feature extraction for associative logic graphs can be performed using Graph Neural Networks (GNNs). The information of nodes and edges in the associative logic graph is input into the GNN, which updates the feature representations of nodes through a message-passing mechanism. A Graph Neural Network is a neural network used for graph feature extraction from associative logic graphs. It updates the feature representations of nodes through a message-passing mechanism. The information of nodes and edges in the associative logic graph is input into the GNN. For each node in the associative logic graph, its feature representation at the th layer is updated using a specific formula. After multiple layers of updates, the feature representations of all nodes are concatenated or summed to obtain the features of the associative logic graph.

[0173] Specifically, for each node v in the association logic graph, its feature representation at layer l is... Update using the following formula: Where σ is the activation function (such as ReLU), and Let N(v) be a trainable weight matrix, representing the set of neighboring nodes of node v. After multiple updates, the feature representations of all nodes are concatenated or summed to obtain the association logic graph features. This represents the feature representation of node v at layer l. and Let be the trainable weight matrix used to update node features at layer l.

[0174] The association logic graph represents the interaction logic between various historically interacted items of the pushed object, reflecting rich historical interaction information of the pushed object. By extracting features from the association logic graph, it is easy to integrate it into the interaction probability prediction model as extended feature input, which can improve the prediction accuracy of interaction probabilities. Interaction probabilities include, for example, click-through rate, negative feedback rate, positive feedback rate, etc. Preset item features refer to the pre-defined features of the item to be pushed, used to characterize its item attributes, etc., and together with the association logic graph features, are used as input to predict the interaction probability of the pushed object with the item to be pushed through the pre-stored interaction probability prediction model.

[0175] It should be noted that in practical applications, as the number of historically interacted items of the pushed object continues to increase, the set of inference relationships can be updated periodically to obtain the latest relationship logic diagram.

[0176] Specifically, after obtaining multiple items to be pushed to the target audience, the item push platform on the server can predict the interaction probability of each item and sort the items according to the interaction probability. For example, when the interaction probability is the click rate, the items are arranged from high to low click rate; when the interaction probability is the negative feedback rate, the items are arranged from low to high negative feedback rate; and when the interaction probability is the positive feedback rate, the items are arranged from high to low positive feedback rate.

[0177] For example, as shown in Figure 12, the features of the association logic graph (Node embedding, Graph embedding, etc.) and the preset item features are input into the interaction probability prediction model to predict the interaction probability of the pushed object with the pushed item, such as click rate, negative feedback rate, etc.

[0178] As an optional implementation, the node information of each node includes: the target item characteristics of the corresponding target historical interaction item; the above-mentioned feature extraction of the association logic graph to obtain the association logic graph features may specifically include the following steps B1-B3:

[0179] B1. Perform vector transformation on the node information of each node in the association logic graph to obtain the corresponding node representation vector.

[0180] The vector transformation method for node information can be selected based on the specific content of the node information.

[0181] As an optional implementation, the node information of each node also includes: the association reason included in the association relationship to which the corresponding target historical interaction item belongs; the target item characteristics include the item attributes and historical interaction type of the corresponding target historical interaction item.

[0182] When converting the node information of each node in the association logic graph into a corresponding node representation vector, the association reason, item attribute and historical interaction type of each node can be converted into vectors respectively to obtain the first sub-vector, the second sub-vector and the third sub-vector. Then, the first sub-vector, the second sub-vector and the third sub-vector are merged to obtain the corresponding node representation vector.

[0183] Specifically, the following vector transformation process is performed on the node information of each node (including item ID, historical interaction type, and association reason):

[0184] 1) For the reasons for association (which are text), pre-trained word vector models, such as Word2Vec (Word to Vector) and Global Vectors for Word Representation (GloVe), are used to convert the text corresponding to the reasons for association into vector representations. The Word2Vec model is a pre-trained word vector model used to convert words in text into vector representations. When converting the reasons for association (text) corresponding to nodes in the association logic graph into vector representations, this model can be used to map each word in the text corresponding to the reasons for association to its corresponding word vector, and then these word vectors are averaged or summed to obtain the vector representation of the text corresponding to the reasons for association. The GloVe model is a pre-trained word vector model, similar to the Word2Vec model, used to convert words in text into vector representations. When processing the reasons for association (text) corresponding to nodes in the association logic graph, each word in the text corresponding to the reasons for association can be mapped to its corresponding word vector, and then the vector representation of the text corresponding to the reasons for association can be obtained through averaging or summing.

[0185] The specific operation involves mapping each word in the text corresponding to the association reason to its corresponding word vector, and then averaging or summing these word vectors to obtain the vector representation of the text corresponding to the association reason, i.e., the first sub-vector mentioned above. The first sub-vector refers to the vector obtained after vector transformation of the association reason corresponding to the node when converting the node information of the node in the association logic graph into node representation vectors. A pre-trained word vector model, such as the Word2Vec model or the GloVe model, is used to convert the text corresponding to the association reason into a vector representation. If a node does not contain an association reason, the first sub-vector corresponding to that node defaults to 0.

[0186] The specific operation likely involves mapping each word in the text corresponding to the association reason to its corresponding word vector. When the word importance in the association reason text is relatively balanced, these word vectors are averaged, i.e. To obtain the vector representation of the text corresponding to the association reason, i.e., the first sub-vector mentioned above; when certain words in the association reason text are of high importance and it is necessary to highlight the influence of these words, the vectors of these words are summed, i.e. To obtain the vector representation of the text corresponding to the association reason. Wherein, This represents the averaged vector. This represents the summed vector, where n is the number of words in the related reasoning text. Let be the word vector corresponding to the i-th word.

[0187] It should be noted that when a node does not contain an association reason, the first subvector corresponding to that node can be defaulted to 0.

[0188] 2) For the item ID, a trainable feature extraction layer (i.e., the embedding layer) is used to convert it into a vector representation, namely the second sub-vector mentioned above. The second sub-vector refers to the vector obtained after vector transformation of the item ID corresponding to the node when converting the node information of the node in the association logic graph into a node representation vector. The trainable feature extraction layer (i.e., the embedding layer) is used to convert the item ID into a vector representation.

[0189] 3) For historical interaction types (belonging to categorical data), one-hot encoding is used to convert them into vector representations, i.e., the third sub-vector mentioned above. The third sub-vector refers to the vector obtained after vectorizing the historical interaction type corresponding to the node when converting the node information of the node in the association logic graph into a node representation vector. Since historical interaction types belong to categorical data, one-hot encoding is used to convert them into vector representations. One-hot encoding is an encoding method for converting categorical data into vector representations. When converting the historical interaction types (belonging to categorical data) corresponding to the nodes in the association logic graph into vector representations, one-hot encoding is used to convert them into vectors for subsequent vector processing and analysis.

[0190] 4) Vector Aggregation: This involves concatenating or weighted summing the vector representations of the interaction reason, historical interaction type, and item ID to obtain a comprehensive node representation vector. Vector aggregation refers to the process of merging the first, second, and third sub-vectors obtained from the conversion of node information in the association logic graph into node representation vectors, respectively, based on the association reason, item attributes, and historical interaction type. The specific fusion method can be concatenation or weighted summation to obtain a comprehensive node representation vector.

[0191] B2. Extract features from the node representation vectors of each node in the associated logic graph to obtain the node feature vectors of each node.

[0192] One approach is to use Node Embedding to extract features from the node representation vector of each node. Node Embedding is used to learn the node feature vectors of each node in the interconnected logical graph, ensuring that the node feature vectors of different nodes with similar structures and attributes are also similar.

[0193] Specifically, the main idea of ​​the Node Embedding method is to learn a class of mappings, utilizing the topological structure of the interconnected logical graph itself, to map the node information of each node in the graph into a low-dimensional dense vector space, and optimize these mappings so that the learned node feature vector of each node can truly reflect the topological structure of the interconnected logical graph itself. Specifically, for each node in the interconnected logical graph, a node feature vector of a specified dimension is learned, and each node's respective feature vector is used for downstream tasks. For example, the Node Embedding method can use algorithms such as DeepWalk, Node to Vector (Node2Vec), or graph neural networks.

[0194] Specifically, the Node Embedding method generates node sequences on the interconnected logical graph through random walks, such as using the DeepWalk algorithm. It starts with each node in the interconnected logical graph and performs a random walk to generate a series of node sequences. These node sequences are then used as input to train a Skip-Gram model. The goal of the Skip-Gram model is to predict the context nodes based on the center node. By continuously optimizing the parameters of the Skip-Gram model, the learned node feature vectors of each node accurately reflect the topological structure of the interconnected logical graph itself. For example, for each node in the interconnected logical graph, a node feature vector of a specified dimension is learned, and each node's individual feature vectors are used in downstream tasks. The Skip-Gram model is a model used to train node feature vectors in the Node Embedding method. It takes the node sequences generated by the DeepWalk algorithm as input, aims to predict the context nodes based on the center node, and continuously optimizes the model's parameters to ensure that the learned node feature vectors of each node accurately reflect the topological structure of the interconnected logical graph itself.

[0195] B3. Merge the node feature vectors of multiple nodes to obtain the features of the association logic graph.

[0196] The specific fusion method can be to concatenate the feature vectors of multiple nodes, that is, to arrange and connect the feature vectors of all nodes in order into a longer vector; or it can be to perform a weighted summation, assigning a weight to the feature vector of each node, and then multiplying each node's feature vector by its corresponding weight before summing them. When it is necessary to retain the complete information of each node's feature vector, and the relative importance between the feature vectors of each node is difficult to distinguish, the concatenation method is used, that is, connecting the feature vectors of multiple nodes in order into a longer vector. When the feature vectors of different nodes contribute differently to the features of the association logic graph, and the weights of each node can be obtained based on prior knowledge or training, the weighted summation method is used. in For the features of the association logic graph, N is the number of nodes, w i The feature vector of the i-th node The weight, and

[0197] Specifically, the node feature vectors of multiple nodes are concatenated or weighted and summed to obtain the features of the association logic graph.

[0198] In this embodiment, the node information of each node in the association logic graph is converted into a node vector representation using a suitable vector transformation method, which facilitates feature extraction of each node and obtains the node feature vector of each node. Then, the node feature vectors of multiple nodes are fused into the weight features of the association logic graph, so as to be used for downstream interaction probability prediction tasks and improve the prediction accuracy of interaction probability.

[0199] Based on the same inventive concept, this application also provides an item pushing device. The principle of this device in solving the problem is similar to the method in the above embodiments. Therefore, the implementation of this device can refer to the implementation of the above method, and the repeated parts will not be described again.

[0200] Figure 13 shows a schematic diagram of the structure of the item pushing device 1300, which may include:

[0201] The reasoning unit 1301 is used to obtain a set of association relationships between historical interactive items based on the item characteristics of each historical interactive item involved in each historical interactive operation performed by the pushed object, and through reasoning by a large language model; each association relationship includes: at least two related historical interactive items and the reason for association.

[0202] The verification unit 1302 is used to replace the selected multiple candidate historical interaction items with the mask of the candidate historical interaction item in the association set, and perform the following operations for each mask: perform contextual semantic analysis based on the item features of other historical interaction items associated with a mask, and select the estimated historical interaction item corresponding to the mask from multiple candidate historical interaction items based on the analysis results; when the estimated historical interaction item does not match the corresponding replaced candidate historical interaction item, delete the replaced candidate historical interaction item and its item features.

[0203] The push unit 1303 is used to obtain the target association relationship between the item to be pushed and at least one target historical interaction item based on the push item characteristics of the item to be pushed and the target item characteristics of each target historical interaction item that has not been deleted, through large language model reasoning, and generate push information according to the target association reason contained in the item to be pushed and the target association relationship, and push the push information to the push recipient.

[0204] Optionally, the item characteristics of each historical interaction item include: the item attributes of the historical interaction item, the historical interaction types of the pushed object with respect to the historical interaction item; and the historical interaction items are arranged in the order of interaction.

[0205] Inference unit 1301 is specifically used for:

[0206] Based on the individual characteristics of each historical interactive item, construct the first prompt message;

[0207] Using a pre-stored initial large language model, based on the item attributes and historical interaction types of each historical interactive item contained in the first prompt information, as well as the interaction sequence of each historical interactive item, we analyze the relationships between each historical interactive item and the reasons for each relationship, and obtain the set of relationships between each historical interactive item.

[0208] Optionally, the inference unit 1301 is specifically used for:

[0209] Based on the individual characteristics of each historical interactive item and the preset set of associated reasons, construct the first prompt message;

[0210] Using an initial large language model, based on the item attributes and historical interaction types of each historical interactive item contained in the first prompt information, as well as the interaction sequence of each historical interactive item, we analyze the relationships between each historical interactive item. For each relationship obtained from the analysis, we select the corresponding relationship reason from the set of relationship reasons to obtain the set of relationship relationships between each historical interactive item.

[0211] Optionally, the verification unit 1302 is specifically used for:

[0212] Based on the set of relationships containing each mask and multiple candidate historical interaction items, a second prompt message is constructed;

[0213] Using the initial large language model, perform the following operations for each mask: based on the item features of other historical interaction items associated with a mask contained in the second prompt information, calculate the semantic similarity between the items and the item features of multiple candidate historical interaction items, and select the candidate historical interaction items whose semantic similarity meets the set conditions from the multiple candidate historical interaction items as the estimated historical interaction items corresponding to a mask.

[0214] Optionally, the device also includes a training unit for:

[0215] Whenever a mask corresponds to an estimated historical interaction item that matches the corresponding candidate historical interaction item to be replaced, the accuracy of the association relationship to which the candidate historical interaction item to be replaced belongs is checked in the association relationship set, and the association relationship that passes the check is taken as a sample association relationship.

[0216] In the initial large language model's defined network layers, the trained parameters are replaced with the pre-defined parameters to be trained, resulting in a large language model to be trained. Based on the obtained association relationships among multiple samples, the large language model to be trained undergoes multiple rounds of iterative training to obtain the target large language model. Each round of training includes:

[0217] The sample item features of at least two historical interaction items included in a sample association relationship are input into a large language model to obtain the predicted association relationship. Based on the difference between the predicted association relationship and the sample association relationship, the loss value is determined, and the training parameters in the set network layer are adjusted based on the loss value.

[0218] Optionally, the characteristics of the pushed item include: the pushed item attributes and the set interaction type of the item to be pushed; the characteristics of the target item include: the item attributes and historical interaction types of the corresponding target historical interaction item;

[0219] Push unit 1303 is specifically used for:

[0220] Based on the characteristics of the pushed items and the characteristics of each target item, a third prompt message is constructed;

[0221] Using a target large language model, based on the item attributes and historical interaction types of each target historical interaction item contained in the third prompt information, the attributes and set interaction types of the pushed item, and the interaction sequence of each target historical interaction item, a correlation analysis is performed on the item to be pushed and each target historical interaction item to obtain the target association relationship between the item to be pushed and at least one target historical interaction item.

[0222] Optionally, when replacing multiple selected candidate historical interaction items with the mask of the candidate historical interaction item in the association set, the verification unit 1302 is specifically used for:

[0223] For each of the multiple historically interacted items contained in the set of relationships, perform the following operations:

[0224] Based on a set probability, randomly determine whether to select a historical interaction item;

[0225] When determining which historical interaction item to select, a historical interaction item is used as a candidate historical interaction item, and the candidate historical interaction item is replaced with the corresponding mask.

[0226] Optionally, the device also includes:

[0227] The generation unit is used to generate a corresponding association logic graph based on the latest set of association relationships formed by each target historical interaction item that has not been deleted. The association logic graph includes multiple nodes, each node corresponds to a target historical interaction item, and each node is connected to at least one other node.

[0228] The extraction unit is used to extract graph features from the association logic graph to obtain the features of the association logic graph.

[0229] The prediction unit is used to predict the interaction probability of the target object with the target object based on the preset item features and associated logic graph features of the item to be pushed, through a pre-stored interaction probability prediction model.

[0230] Optionally, the node information for each node includes: the target item characteristics of the corresponding target historical interaction items;

[0231] The extraction unit is specifically used for:

[0232] The node information of each node in the association logic graph is transformed into a vector to obtain the corresponding node representation vector;

[0233] Feature extraction is performed on the node representation vectors of multiple nodes in the association logic graph to obtain the node feature vectors of each node.

[0234] The node feature vectors of multiple nodes are fused together to obtain the features of the association logic graph.

[0235] Optionally, the node information for each node may also include: the association reason included in the association relationship to which the corresponding target historical interaction item belongs; and the target item characteristics including the item attributes and historical interaction type of the corresponding target historical interaction item.

[0236] When converting the node information of each node in the association logic graph into a corresponding node representation vector, the extraction unit is specifically used for:

[0237] The associated reason, item attribute and historical interaction type of each node are transformed into vectors to obtain the first sub-vector, the second sub-vector and the third sub-vector;

[0238] The first sub-vector, the second sub-vector, and the third sub-vector are merged to obtain the corresponding node representation vector.

[0239] For ease of description, the above sections are divided into modules (or units) according to their functions and described separately. Of course, in implementing this application, the functions of each module (or unit) can be implemented in one or more software or hardware components.

[0240] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0241] Having introduced the item push method and apparatus according to exemplary embodiments of this application, we will now introduce an electronic device according to another exemplary embodiment of this application.

[0242] Based on the same inventive concept as the above-described method embodiments, this application also provides an electronic device. In one embodiment, the electronic device may be a server, such as the server 120 shown in FIG1. ​​In this embodiment, the structure of the electronic device may be as shown in FIG14, including a memory 1401, a communication module 1403, and one or more processors 1402.

[0243] The memory 1401 is used to store computer programs executed by the processor 1402. The memory 1401 may mainly include a program storage area and a data storage area. The program storage area may store the operating system and programs required to run instant messaging functions, etc.; the data storage area may store various instant messaging information and operation instruction sets, etc.

[0244] Memory 1401 may be volatile memory, such as random-access memory (RAM); memory 1401 may also be non-volatile memory, such as read-only memory, flash memory, hard disk drive (HDD), or solid-state drive (SSD); or memory 1401 may be any other medium capable of carrying or storing a desired computer program having the form of instructions or data structures and accessible by a computer, but is not limited thereto. Memory 1401 may be a combination of the above-described memories.

[0245] The processor 1402 may include one or more central processing units (CPUs) or digital processing units, etc. The processor 1402 is used to implement the aforementioned item pushing method when it invokes a computer program stored in the memory 1401.

[0246] The communication module 1403 is used to communicate with terminal devices and other servers.

[0247] This application embodiment does not limit the specific connection medium between the memory 1401, communication module 1403, and processor 1402. In this application embodiment, the memory 1401 and processor 1402 are connected via a bus 1404 in Figure 14. The bus 1404 is depicted as a thick line in Figure 14. The connection methods between other components are only illustrative and are not intended to be limiting. The bus 1404 can be divided into address bus, data bus, control bus, etc. For ease of description, only one thick line is used to describe it in Figure 12, but it does not mean that there is only one bus or one type of bus.

[0248] The memory 1401 stores a computer storage medium containing computer-executable instructions for implementing the item pushing method of this application embodiment. The processor 1402 is used to execute the above-described item pushing method, as shown in FIG2.

[0249] In another embodiment, the electronic device may also be other electronic devices, such as the terminal device 110 shown in FIG1. ​​In this embodiment, the structure of the electronic device may be as shown in FIG15, including: a communication component 1510, a memory 1520, a display unit 1530, a camera 1540, a sensor 1550, an audio circuit 1560, a Bluetooth module 1570, a processor 1580, and other components.

[0250] The communication component 1510 is used to communicate with the server. In some embodiments, it may include a Circuit-Based Wireless Fidelity (WiFi) module, which is a short-range wireless transmission technology. Electronic devices can use the WiFi module to help users send and receive information.

[0251] The memory 1520 can be used to store software programs and data. The processor 1580 executes various functions of the terminal device 110 and performs data processing by running the software programs or data stored in the memory 1520. The memory 1520 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. The memory 1520 stores an operating system that enables the terminal device 110 to run. In this application, the memory 1520 may store the operating system and various applications, and may also store a computer program that executes the item pushing method of the embodiments of this application.

[0252] The display unit 1530 can also be used to display information input by the user or information provided to the user, as well as various menus of the terminal device 110, in a graphical user interface (GUI). Specifically, the display unit 1530 may include a display screen 1532 disposed on the front of the terminal device 110. The display screen 1532 may be configured as a liquid crystal display, a light-emitting diode, or the like. The display unit 1530 can be used to display the display interface for pushing items, as described in the embodiments of this application.

[0253] The display unit 1530 can also be used to receive input digital or character information and generate signal inputs related to user settings and function control of the terminal device 110. Specifically, the display unit 1530 may include a touch screen 1531 disposed on the front of the terminal device 110, which can collect touch operations of the user on or near it, such as clicking buttons, dragging scroll boxes, etc.

[0254] The touchscreen 1531 can be placed on top of the display screen 1532, or the touchscreen 1531 and the display screen 1532 can be integrated to realize the input and output functions of the terminal device 110. After integration, it can be referred to as a touch display screen. In this application, the display unit 1530 can display the application and the corresponding operation steps.

[0255] Camera 1540 can be used to capture still images, which users can then share via an application. There can be one or multiple cameras 1540. An object is projected onto a photosensitive element through a lens, generating an optical image. This photosensitive element can be a charge-coupled device (CCD) or a complementary metal-oxide-semiconductor (CMOS) phototransistor. The photosensitive element converts the light signal into an electrical signal, which is then transmitted to the processor 1580 for conversion into a digital image signal.

[0256] The terminal device may also include at least one sensor 1550, such as an accelerometer 1551, a proximity sensor 1552, a fingerprint sensor 1553, and a temperature sensor 1554. The terminal device may also be equipped with other sensors such as a gyroscope, barometer, hygrometer, thermometer, infrared sensor, light sensor, and motion sensor.

[0257] Audio circuitry 1560, speaker 1561, and microphone 1562 provide an audio interface between the user and terminal device 110. Audio circuitry 1560 converts received audio data into electrical signals, which are then transmitted to speaker 1561, where they are converted into sound signals for output. Terminal device 110 may also be equipped with volume buttons for adjusting the volume of the sound signal. On the other hand, microphone 1562 converts collected sound signals into electrical signals, which are received by audio circuitry 1560, converted into audio data, and then output to communication component 1510 for transmission to, for example, another terminal device 110, or to memory 1520 for further processing.

[0258] The Bluetooth module 1570 is used to interact with other Bluetooth devices that also have a Bluetooth module via the Bluetooth protocol. For example, a terminal device can establish a Bluetooth connection with a wearable electronic device (such as a smartwatch) that also has a Bluetooth module through the Bluetooth module 1570, thereby exchanging data.

[0259] The processor 1580 is the control center of the terminal device, connecting various parts of the terminal through various interfaces and lines. It executes various functions and processes data by running or executing software programs stored in the memory 1520 and calling data stored in the memory 1520. In some embodiments, the processor 1580 may include one or more processing units; the processor 1580 may also integrate an application processor and a baseband processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the baseband processor mainly handles wireless communication. It is understood that the baseband processor may not be integrated into the processor 1580. In this application, the processor 1580 can run the operating system, applications, user interface display and touch response, as well as the item push method of this embodiment. Furthermore, the processor 1580 is coupled to the display unit 1530.

[0260] In some possible implementations, various aspects of the item push method provided in this application can also be implemented in the form of a program product, which includes a computer program. When the program product is run on an electronic device, the computer program is used to cause the electronic device to perform the steps in the item push method according to various exemplary embodiments of this application described above. For example, the electronic device can perform the steps shown in FIG2.

[0261] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0262] The program product of the embodiments of this application may employ a portable compact disc read-only memory (CD-ROM) and include a computer program, and may run on an electronic device. However, the program product of this application is not limited thereto. In this document, the readable storage medium may be any tangible medium that contains or stores a program that may be used by or in conjunction with a command execution system, apparatus, or device.

[0263] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a readable computer program. This propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with a command execution system, apparatus, or device.

[0264] Computer programs contained on readable media may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0265] Computer programs for performing the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The computer program can execute entirely on the user's electronic device, partially on the user's electronic device, as a standalone software package, partially on the user's electronic device and partially on a remote electronic device, or entirely on a remote electronic device or server. In cases involving remote electronic devices, the remote electronic device can be connected to the user's electronic device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external electronic device (e.g., via the Internet using an Internet service provider).

[0266] In summary, this application provides a method, apparatus, electronic device, computer-readable storage medium, and computer program product for pushing items. By considering the item characteristics of each historical interactive item involved in the historical interaction operations performed by the target user, a set of association relationships between these historical interactive items is obtained through large language model reasoning. Each association relationship includes at least two related historical interactive items and the reason for the association. The large language model possesses powerful semantic understanding and knowledge reasoning capabilities, enabling it to extract potential association relationships and reasons for association from the item characteristics of each historical interactive item. This allows the system to deeply analyze the historical interaction behavior patterns of the target user, providing a more accurate basis for subsequent item recommendations. For example, in a news recommendation scenario, it can discover the interaction associations of the target user with different news topics, such as frequently following related business news after following technology news, thus providing a more targeted direction for subsequent recommendations. Subsequently, in the association set, multiple selected candidate historical interaction items are replaced with masks. Contextual semantic analysis is performed on the item features of other historical interaction items associated with a mask, and the estimated historical interaction item corresponding to that mask is selected. When the estimated historical interaction item does not match the corresponding replaced candidate historical interaction item, the replaced candidate historical interaction item and its item features are deleted. This verification mechanism removes inaccurate associations in the association set, improves the quality of the association set, and ensures the accuracy of subsequent inference. Finally, based on the push item features of the item to be pushed and the target item features of each of the remaining target historical interaction items, a large language model inference is used to obtain the target association relationship between the item to be pushed and at least one target historical interaction item. Push information is generated based on the target association reason contained in the item to be pushed and the target association relationship and pushed to the recipient. The push reason obtained in this way is related to the target historical interaction items in the recipient's past interactions, making it highly persuasive. The recipient can understand why the push system pushed the item through the push reason, thereby increasing trust and satisfaction with the pushed item, increasing the probability that the recipient will perform an interaction with the pushed item, improving the push effect, increasing resource utilization, and avoiding ineffective pushes.

[0267] Furthermore, the characteristics of each historically interacted item include its attributes, the types of interactions the recipient made with that item, and the items arranged in chronological order of interaction. Item attributes describe the features of the historically interacted item in detail, while the interaction types reflect the interaction methods between the recipient and the item. The order of interactions reflects the behavioral logic over time. The comprehensive use of this information allows for a more complete and accurate depiction of the recipient's historical interactive behavior. For example, in video recommendation, combining video category attributes, the recipient's viewing or liking interactions, and the viewing order allows for a more precise grasp of user interest trends. This provides a richer and more accurate data foundation for subsequent inference of relationship sets, improving the accuracy of relationship inference and making recommendations more aligned with users' actual needs.

[0268] When performing historical interactions on the target user and considering the item characteristics of each historical interaction item, the large language model is used to infer the set of relationships between these historical interaction items. Based on the item characteristics of each historical interaction item, a first prompt message is constructed. Using a pre-stored initial large language model, and based on the item attributes, historical interaction types, and interaction order of each historical interaction item contained in the first prompt message, the model analyzes the relationships between these historical interaction items and the reasons for each relationship, thus obtaining the set of relationships between them. The first prompt message organizes the item characteristics of each historical interaction item in a specific format to guide the initial large language model's analysis. The powerful capabilities of the initial large language model enable it to extract the reasons for the target user's interaction with each historical interaction item from a global perspective, generating reasonable relationships and reasons for association. For example, in product recommendations, it can analyze the relationship and reasons why a user purchases related accessories after buying a certain product, improving the quality of the relationship set and providing a more reliable basis for subsequent recommendations.

[0269] Furthermore, based on the individual characteristics of each historical interactive item and a pre-defined set of association reasons, the first prompt information is constructed. Due to the diversity of textual expression, the association reasons inferred by the initial large language model may be highly divergent and lack refinement. The pre-defined set of association reasons provides the initial large language model with a limited range of choices, enabling it to select appropriate reasons from this set when inferring association reasons. This ensures the consistency and standardization of association reasons, avoids overly scattered association reasons, and improves the quality and interpretability of the association set. For example, in music recommendation, the set of association reasons can include "same music style" and "songs by the same artist," making the inferred association reasons clearer and more explicit.

[0270] Using a pre-stored initial large language model, based on the item attributes and historical interaction types of each historical interactive item contained in the first prompt information, as well as the interaction sequence of each historical interactive item, the model analyzes the relationships between each historical interactive item and the reasons for each relationship, obtaining a set of relationships between historical interactive items. For each relationship, a corresponding reason is selected from the set of reasons for association, thus obtaining a set of relationships between historical interactive items. The initial large language model analyzes the relationships between historical interactive items based on the semantic similarity of item attributes and the association of historical interaction types, combined with the interaction sequence, and selects appropriate reasons from the set of reasons for association. This approach can improve the efficiency and accuracy of relationship reasoning, making the set of relationships more logical and interpretable. For example, in video recommendation, it can accurately analyze the relationships between different videos watched by users and select appropriate reasons from the set of reasons for association, such as "related plot" or "same actors."

[0271] When performing the following operations for each mask: performing contextual semantic analysis based on the item features of other historical interaction items associated with a mask, and selecting an estimated historical interaction item corresponding to a mask from multiple candidate historical interaction items based on the analysis results, a second prompt is constructed based on the association set containing each mask and multiple candidate historical interaction items. Using the initial large language model, the following operations are performed for each mask: based on the semantic similarity between the item features of other historical interaction items associated with a mask contained in the second prompt and the item features of each of the multiple candidate historical interaction items, a candidate historical interaction item whose semantic similarity meets the set conditions is selected as an estimated historical interaction item corresponding to a mask. The second prompt guides the initial large language model to select the estimated historical interaction item based on semantic similarity. The semantic understanding capability of the initial large language model can accurately determine the semantic similarity between different item features. Selecting the estimated historical interaction item in this way can more accurately verify the accuracy of the association set. If the estimated historical interaction item does not match the replaced candidate historical interaction item, it indicates that the association of the candidate historical interaction item in the association set may be inaccurate; deleting it can improve the reliability of the association set. For example, in news recommendation, semantic similarity can be used to determine whether the association between a news item and other historical interactive news is reasonable, and unreasonable associations can be removed.

[0272] Furthermore, whenever a predicted historical interaction item corresponding to a mask matches a corresponding candidate historical interaction item to be replaced, the accuracy of the association to which the candidate historical interaction item to be replaced belongs in the association set is verified, and the verified association is taken as a sample association. In the initial large language model's set network layer, the trained parameters in the set network layer are replaced with the set parameters to be trained to obtain the large language model to be trained. Based on the obtained multiple sample associations, the large language model to be trained is trained in multiple rounds of iterative training to obtain the target large language model. Each round of training includes inputting the sample item features of at least two sample historical interaction items contained in a sample association into the large language model to obtain the predicted association. Based on the difference between the predicted association and the sample association, the loss value is determined, and the parameters to be trained in the set network layer are adjusted based on the loss value. By supervising and fine-tuning the initial large language model and training with sample associations, the target large language model can be made more suitable for the inference task of the association set. During the training process, the parameters to be trained are continuously adjusted to make the predicted association closer to the sample association, thereby improving the accuracy and stability of the model. For example, in product recommendation, a finely tuned target large language model can more accurately infer the relationships between products, providing users with more precise product recommendations.

[0273] When obtaining the target association between a proposed item and at least one target historical interaction item, the characteristics of the proposed item include its proposed item attributes and set interaction type, while the characteristics of the target item include the item attributes and historical interaction types of the corresponding target historical interaction items. Based on the characteristics of the proposed item and each target item, a third prompt message is constructed. Using a target big language model, based on the item attributes and historical interaction types of each target historical interaction item contained in the third prompt message, the attributes and set interaction types of the proposed item, and the interaction sequence of each target historical interaction item, an association analysis is performed on the proposed item and each target historical interaction item to obtain the target association between the proposed item and at least one target historical interaction item. After fine-tuning, the target big language model can more accurately analyze the association between the proposed item and the target historical interaction items based on the characteristics of the proposed item and the target item, combined with the interaction sequence. This makes the extracted push reasons more targeted and persuasive. For example, in music recommendation, it can accurately analyze the recommendation reasons based on the user's historical listening records and the attributes of the proposed song, such as "similar to the style of songs you often listen to," improving the push effect.

[0274] In a set of association relationships, when replacing multiple candidate historical interaction items with their masks, the following operations are performed for each historical interaction item in the set: Based on a set probability, it is randomly determined whether to select a historical interaction item. If a historical interaction item is selected, it is designated as a candidate historical interaction item, and its mask is replaced with the corresponding mask. This random selection method ensures that each historical interaction item in the set has an equal probability of being selected, avoiding interference from human factors and making the mask replacement process fairer and more objective. By using random masks, a comprehensive verification of the association relationship set can be performed, improving its accuracy and reliability. For example, in news recommendation, randomly selecting news items for mask verification can identify potentially inaccurate associations in the set, improving the quality of news recommendations.

[0275] Furthermore, based on the latest set of relationships formed by the historical interaction items of each target that have not been deleted, a corresponding relationship logic graph is generated. This graph includes multiple nodes, each corresponding to a target historical interaction item, and each node is connected to at least one other node. Graph feature extraction is performed on the relationship logic graph to obtain its features. Based on the preset item features of the item to be pushed and the relationship logic graph features, a pre-stored interaction probability prediction model is used to predict the interaction probability of the target user with the item to be pushed. The relationship logic graph intuitively displays the relationships between the historical interaction items of each target, and graph feature extraction can transform the information in the relationship logic graph into feature vectors suitable for model processing. Combined with the preset item features of the item to be pushed, the interaction probability prediction model can more accurately predict the interaction probability of the target user with the item to be pushed. For example, in video recommendation, by using the relationship logic graph features and the preset features of the video to be pushed, the probability of user clicks, views, and other interactions with the video can be predicted, providing a more scientific basis for video recommendation and improving the accuracy of the push.

[0276] When extracting features from an association logic graph, each node's information includes the target item features of its corresponding historical interaction items. The node information of each node in the association logic graph is vectorized to obtain a corresponding node representation vector. Feature extraction is then performed on the node representation vectors of multiple nodes in the association logic graph to obtain individual node feature vectors. These feature vectors are then fused to obtain the overall association logic graph features. Vector transformation converts node information into vector form, facilitating model processing. Feature extraction from node representation vectors can uncover latent features within node information. Fusing multiple node feature vectors comprehensively considers the information of each node in the association logic graph, improving the quality of the association logic graph features. For example, in product recommendation, this can more accurately represent the relationships between products, providing more accurate feature information for product recommendations.

[0277] When converting the node information of each node in the association logic graph into a corresponding node representation vector, the node information of each node also includes the association reason contained in the association relationship to which the corresponding target historical interaction item belongs. The target item features include the item attributes and historical interaction types of the corresponding target historical interaction items. The association reason, item attributes, and historical interaction types corresponding to each node are each vectorized to obtain a first sub-vector, a second sub-vector, and a third sub-vector. These three sub-vectors are then fused to obtain the corresponding node representation vector. This method fully considers all aspects of node information, performing vector transformation and fusion of different types of information, thus representing node information more comprehensively. The association reason, item attributes, and historical interaction types describe the target historical interaction item from different perspectives. Fusing them can more accurately represent the characteristics of the node, improve the accuracy of the node representation vector, and thereby improve the quality of the association logic graph features and the accuracy of interaction probability prediction. For example, in news recommendation, it can more comprehensively represent the association relationships between news items, improving the accuracy of news recommendations.

[0278] When using an initial large language model to predict the predicted historical interaction items corresponding to each mask, for the item features of other historical interaction items associated with a mask contained in the second prompt information, their item attributes and historical interaction types are vectorized separately to obtain vectors. Similarly, for the item features of multiple candidate historical interaction items, their item attributes and historical interaction types are also vectorized separately to obtain vectors. The cosine similarity between the two is calculated. From multiple candidate historical interaction items, the candidate historical interaction item whose cosine similarity satisfies a set condition (e.g., maximum) is selected as the predicted historical interaction item corresponding to a mask. By vectorizing and calculating cosine similarity, the semantic similarity between different item features can be measured more accurately. Cosine similarity is a commonly used method for measuring vector similarity; it can accurately determine the degree of similarity between item features in high-dimensional space. Selecting candidate historical interaction items whose cosine similarity satisfies the set condition as predicted historical interaction items can improve the accuracy of the prediction, thereby more effectively verifying the accuracy of the association set.

[0279] During supervised fine-tuning of the initial large language model, the trained parameters in the initial network layers are replaced with pre-defined trainable parameters, which can be low-rank matrices. Replacing trained parameters with low-rank matrices reduces the number of parameters and computational complexity while preserving model performance as much as possible. The low-rank matrix is ​​constructed by performing matrix factorization on the weight matrices of the linear transformations in the initial network layers, retaining the top k largest singular values ​​and their corresponding singular vectors. This reduces computational cost and improves training efficiency without sacrificing significant model performance. During multiple iterations of training, the loss value is determined based on the difference between the predicted association and the sample association, and the low-rank matrix is ​​adjusted according to the loss value, continuously optimizing the model and improving the accuracy of inferring associations.

[0280] When predicting the probability of an audience interacting with a recommended item, graph neural networks (GNNs) can be used to extract graph features from the relational logic graph. GNNs update node feature representations through a message-passing mechanism. Information about nodes and edges in the relational logic graph is input into the GNN. For each node in the relational logic graph, its feature representation at the th layer is updated using a specific formula. After multiple layers of updates, the feature representations of all nodes are concatenated or summed to obtain the relational logic graph features. GNNs can fully utilize the topological structure information of the relational logic graph, and the message-passing mechanism allows information to be passed between nodes to update their feature representations. This allows for more accurate feature extraction from the relational logic graph, improving the quality of the features and thus increasing the accuracy of interaction probability prediction. For example, in video recommendation, it can more accurately capture the relationships between videos, improving the accuracy of predicting user interaction probabilities.

[0281] It should be noted that although several units or sub-units of the device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units.

[0282] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0283] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing a computer-usable computer program.

[0284] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0285] These computer program commands may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the commands stored in the computer-readable storage medium produce an article of manufacture including command means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0286] These computer program commands may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the commands executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0287] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0288] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for pushing items, executed by a computer device, the method comprising: Based on the item characteristics of each historical interaction item involved in each historical interaction operation performed by the pushed object, the set of association relationships between the historical interaction items is obtained through large language model reasoning; each association relationship includes: at least two related historical interaction items and the reason for association; In the set of association relationships, multiple selected candidate historical interaction items are replaced with masks of the candidate historical interaction items, and the following operations are performed for each mask: Contextual semantic analysis is performed based on the item features of other historical interaction items associated with a mask; based on the analysis results, an estimated historical interaction item corresponding to the mask is selected from the multiple candidate historical interaction items; when the estimated historical interaction item does not match the corresponding replaced candidate historical interaction item, the replaced candidate historical interaction item and its item features are deleted; and Based on the push item characteristics of the item to be pushed and the target item characteristics of each target historical interaction item that has not been deleted, the target association relationship between the item to be pushed and at least one target historical interaction item is obtained through large language model reasoning. Push information is generated according to the target association reason contained in the item to be pushed and the target association relationship, and the push information is pushed to the target recipient.

2. The method according to claim 1, wherein the item characteristics of each historical interaction item include: The item attributes of the historically interacted item, and the types of historical interactions of the pushed object with respect to the historically interacted item; The historical interactive items are arranged in the order of their interaction.

3. The method according to claim 2, wherein obtaining the set of association relationships between the historical interactive items based on the item characteristics of each historical interactive item involved in the execution of each historical interactive operation by the pushed object, through large language model inference, includes: Based on the individual item characteristics of each historical interactive item, a first prompt message is constructed; Using a pre-stored initial large language model, based on the item attributes and historical interaction types of each historical interactive item contained in the first prompt information, as well as the interaction sequence of each historical interactive item, the relationships between each historical interactive item and the reasons for each relationship are analyzed to obtain the set of relationships between each historical interactive item.

4. The method according to claim 3, wherein constructing the first prompt information based on the item characteristics of each of the historically interacted items includes: The first prompt information is constructed based on the item characteristics of each historical interactive item and a preset set of associated reasons.

5. The method according to claim 4, wherein the step of using a pre-stored initial large language model, based on the item attributes and historical interaction types of each historical interactive item contained in the first prompt information, and the interaction sequence of each historical interactive item, analyzing the relationships between the historical interactive items and the reasons for each relationship, to obtain a set of relationships between the historical interactive items, includes: Using the initial large language model, based on the item attributes and historical interaction types of each historical interactive item contained in the first prompt information, as well as the interaction sequence of each historical interactive item, the relationships between the historical interactive items are analyzed, and for each relationship obtained from the analysis, the corresponding relationship reason is selected from the set of relationship reasons to obtain the set of relationship relationships between the historical interactive items.

6. The method according to any one of claims 2 to 5, wherein the following operations are performed for each mask: performing contextual semantic analysis based on the item features of other historical interaction items associated with a mask, and selecting the estimated historical interaction item corresponding to the mask from the plurality of candidate historical interaction items based on the analysis results, comprising: Based on the set of associations containing each mask and the multiple candidate historical interaction items, a second prompt message is constructed; Using the initial large language model, perform the following operations for each mask: based on the item features of other historical interaction items associated with a mask contained in the second prompt information, calculate the semantic similarity between the items and the item features of the multiple candidate historical interaction items, and select the candidate historical interaction items whose semantic similarity meets the set conditions from the multiple candidate historical interaction items as the estimated historical interaction items corresponding to the mask.

7. The method according to any one of claims 2 to 6, further comprising: Whenever a mask corresponds to an estimated historical interaction item that matches the corresponding candidate historical interaction item to be replaced, the accuracy of the association relationship to which the candidate historical interaction item to be replaced belongs in the association relationship set is verified, and the association relationship that passes the verification is taken as a sample association relationship. In the initial large language model's defined network layers, the trained parameters in these layers are replaced with predefined parameters to be trained, resulting in a large language model to be trained. Based on the obtained association relationships between multiple samples, the large language model to be trained undergoes multiple rounds of iterative training to obtain the target large language model. Each training round includes: The sample item features of at least two historical interaction items included in a sample association relationship are input into the large language model to obtain the predicted association relationship. Based on the difference between the predicted association relationship and the sample association relationship, a loss value is determined, and the training parameters in the set network layer are adjusted based on the loss value.

8. The method according to claim 7, wherein the characteristics of the pushed item include: The push item attributes and set interaction type of the item to be pushed; The characteristics of the target item include: the item attributes and historical interaction types of the corresponding target historical interaction item; The target association relationship between the item to be pushed and at least one target historical interaction item is obtained through the following method: Based on the characteristics of the pushed items and the characteristics of each target item, a third prompt message is constructed; Using the target large language model, based on the item attributes and historical interaction types of each target historical interaction item contained in the third prompt information, the attributes of the pushed item and the set interaction type, as well as the interaction sequence of each target historical interaction item, an association analysis is performed on the item to be pushed and each target historical interaction item to obtain the target association relationship between the item to be pushed and at least one target historical interaction item.

9. The method according to any one of claims 1 to 8, wherein replacing the selected plurality of candidate historical interaction items in the association set with the masks of the candidate historical interaction items respectively comprises: For each of the multiple historically interacted items contained in the aforementioned set of relationships, perform the following operations respectively: Based on a set probability, randomly determine whether to select a historical interaction item; When selecting a historical interaction item, the historical interaction item is used as a candidate historical interaction item, and the candidate historical interaction item is replaced with the corresponding mask.

10. The method according to any one of claims 1 to 9, wherein the method further comprises: Based on the latest set of association relationships formed by the historical interaction items of each target that have not been deleted, a corresponding association logic graph is generated. The association logic diagram includes multiple nodes, each node corresponds to a target historical interaction item, and each node is connected to at least one other node; Graph feature extraction is performed on the association logic graph to obtain association logic graph features; Based on the preset item features of the item to be pushed and the features of the association logic graph, the interaction probability of the pushed object with the item to be pushed is predicted through a pre-stored interaction probability prediction model.

11. The method according to claim 10, wherein the node information of each node includes: The target item characteristics of the corresponding target historical interaction items; The step of extracting features from the association logic graph to obtain association logic graph features includes: The node information of each node in the association logic graph is vectorized to obtain the corresponding node representation vector; Feature extraction is performed on the node representation vectors of each of the multiple nodes in the associated logic graph to obtain the node feature vectors of each of the multiple nodes. The node feature vectors of the multiple nodes are fused to obtain the features of the association logic graph.

12. The method according to claim 11, wherein the node information of each node further includes: The association relationship of the corresponding target historical interaction item includes the association reason; the characteristics of the target item include the item attributes and historical interaction type of the corresponding target historical interaction item; The step of converting the node information of each node in the association logic graph into a corresponding node representation vector includes: The associated reason, item attribute and historical interaction type of each node are transformed into vectors to obtain the first sub-vector, the second sub-vector and the third sub-vector; The first sub-vector, the second sub-vector, and the third sub-vector are fused together to obtain the corresponding node representation vector.

13. An item pushing device, comprising: The reasoning unit is used to obtain a set of association relationships between the historical interactive items based on the item characteristics of each historical interactive item involved in each historical interactive operation performed by the pushed object, and through reasoning by a large language model; each association relationship includes: at least two related historical interactive items and the reason for association; A verification unit is configured to replace multiple selected candidate historical interaction items with masks of the candidate historical interaction items in the association set, and perform the following operations for each mask: perform contextual semantic analysis based on the item features of other historical interaction items associated with a mask, and select the estimated historical interaction item corresponding to the mask from the multiple candidate historical interaction items based on the analysis results; when the estimated historical interaction item does not match the corresponding replaced candidate historical interaction item, delete the replaced candidate historical interaction item and its item features; and The push unit is used to obtain the target association relationship between the item to be pushed and at least one target historical interaction item based on the push item characteristics of the item to be pushed and the target item characteristics of each target historical interaction item that has not been deleted, through large language model reasoning, and generate push information according to the target association reason contained in the item to be pushed and the target association relationship, and push the push information to the push recipient.

14. An electronic device comprising a processor and a memory, wherein, The memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of any one of the methods described in claims 1 to 12.

15. A computer-readable storage medium comprising a computer program that, when executed on an electronic device, causes the electronic device to perform the steps of any one of claims 1 to 12.

16. A computer program product comprising a computer program stored in a computer-readable storage medium; wherein when a processor of an electronic device reads the computer program from the computer-readable storage medium, the processor executes the computer program such that the electronic device performs the steps of any one of claims 1 to 12.

Images