Training method and apparatus for information prediction model, and device, computer-readable storage medium and computer program product

Abstract

Provided in the present disclosure are a training method and apparatus for an information prediction model, and a device, a computer-readable storage medium and a computer program product. The method comprises: acquiring a positive feedback interaction sequence and a negative feedback interaction sequence of a sample object; inputting both the positive feedback interaction sequence and the negative feedback interaction sequence into an information prediction model for information prediction, so as to obtain positive feedback interaction information representations and negative feedback interaction information representations; on the basis of the difference between a first positive sample predicted value calculated from each positive feedback interaction information representation and a first negative sample predicted value calculated from a corresponding other feedback interaction information representation, determining a first loss value; on the basis of the difference between a second negative sample predicted value calculated from each negative feedback interaction information representation and a second positive sample predicted value calculated from target positive feedback interaction information representation, determining a second loss value; and jointly using the first loss values and the second loss values to train the information prediction model, so as to obtain a trained information prediction model.

Description

Training methods, apparatus, equipment, computer-readable storage media, and computer program products for information prediction models

[0001] Cross-references to related applications

[0002] This application is based on and claims priority to Chinese Patent Application No. 2025100662397, filed on January 15, 2025, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This application relates to the field of natural language processing technology, and in particular to a training method, apparatus, device, and computer storage medium for an information prediction model. Background Technology

[0004] Recommendation systems, as a crucial component of the modern internet, provide personalized information to target users by analyzing their preferences. The main technology employed in recommendation systems employs a layered architecture, first encoding the items to be recommended, then modeling the target user's interests based on the encoding results, and finally implementing specific push scenarios according to these modeled interests. However, this architecture has limitations in understanding the target user's interests, leading not only to low recommendation accuracy but also to excessive redundant computation due to the need to handle large amounts of invalid encoding and one-sided interest modeling, increasing the system's processing burden and reducing the response efficiency of information recommendation. Summary of the Invention

[0005] This disclosure provides a training method for an information prediction model, as well as related methods, apparatus, devices, and media, which can reduce redundant calculations and improve the resource utilization of electronic devices.

[0006] According to one aspect of the present disclosure, a training method for an information prediction model is provided. The method includes: acquiring a positive feedback interaction sequence and a negative feedback interaction sequence corresponding to a sample object, wherein the positive feedback interaction sequence contains at least two positive feedback interaction pieces, and the negative feedback interaction sequence contains at least two negative feedback interaction pieces; inputting the positive feedback interaction sequence and the negative feedback interaction sequence into an information prediction model for information prediction, thereby obtaining a positive feedback interaction information representation corresponding to each positive feedback interaction piece in the positive feedback interaction sequence, and a negative feedback interaction information representation corresponding to each negative feedback interaction piece in the negative feedback interaction sequence; acquiring other feedback interaction information corresponding to other sample objects for each positive feedback interaction piece, and based on each positive feedback interaction... The difference between the calculated first positive sample prediction value and the first negative sample prediction value calculated from the corresponding other feedback interaction information is used to determine a first loss value. The target time information corresponding to each negative feedback interaction information representation is obtained, and the target positive feedback interaction information representation corresponding to each negative feedback interaction information representation is determined based on the target time information. The difference between the calculated second negative sample prediction value and the calculated second positive sample prediction value from the corresponding target positive feedback interaction information representation is used to determine a second loss value. The information prediction model is trained by combining the first loss value and the second loss value to obtain a trained information prediction model.

[0007] According to another aspect of the embodiments of this disclosure, a method for selecting push information is provided. The method includes: acquiring a target positive feedback interaction sequence and a target negative feedback interaction sequence corresponding to a target object; inputting the target positive feedback interaction sequence and the target negative feedback interaction sequence into a trained information prediction model for object representation processing to obtain a positive feedback interaction object representation and a negative feedback interaction object representation corresponding to the target object; based on the trained information prediction model, performing information representation processing on each candidate push information at the current time to obtain a candidate push information representation corresponding to each candidate push information; calculating the similarity between each candidate push information representation and the positive feedback interaction object representation to obtain a first similarity for each candidate push information, and calculating the similarity between each candidate push information representation and the negative feedback interaction object representation to obtain a second similarity for each candidate push information; obtaining a target similarity score for each candidate push information based on the difference between the first and second similarities; and determining the candidate push information corresponding to the largest target similarity score as the target push information; wherein the trained information prediction model is trained according to the training method described above.

[0008] According to another aspect of the present disclosure, a training apparatus for an information prediction model is provided. The apparatus includes: a sample sequence acquisition module configured to acquire a positive feedback interaction sequence and a negative feedback interaction sequence corresponding to a sample object, wherein the positive feedback interaction sequence contains at least two positive feedback interaction information and the negative feedback interaction sequence contains at least two negative feedback interaction information.

[0009] The information prediction module is configured to input the positive feedback interaction sequence and the negative feedback interaction sequence into the information prediction model to perform information prediction, thereby obtaining the positive feedback interaction information representation corresponding to each positive feedback interaction information in the positive feedback interaction sequence and the negative feedback interaction information representation corresponding to each negative feedback interaction information in the negative feedback interaction sequence.

[0010] The first loss determination module is configured to obtain other feedback interaction information of other sample objects corresponding to each positive feedback interaction information, and determine the first loss value based on the difference between the first positive sample prediction value calculated based on the representation of each positive feedback interaction information and the first negative sample prediction value calculated based on the representation of other feedback interaction information of the corresponding other feedback interaction information.

[0011] The target representation acquisition module is configured to acquire the current time information corresponding to each negative feedback interaction information representation, and determine the target positive feedback interaction information representation of the sample object corresponding to each negative feedback interaction information representation based on the current time information;

[0012] The second loss determination module is configured to determine a second loss value based on the difference between the second negative sample prediction value calculated from each negative feedback interaction information representation and the second positive sample prediction value calculated from the corresponding target positive feedback interaction information representation.

[0013] The model joint training module is configured to train the information prediction model by combining the first loss value and the second loss value, thereby obtaining the trained information prediction model.

[0014] According to another aspect of the present disclosure, a push notification selection device is provided, the device comprising:

[0015] The target sequence acquisition module is configured to acquire the target positive feedback interaction sequence and the target negative feedback interaction sequence corresponding to the target object.

[0016] The object representation generation module is configured to input the target positive feedback interaction sequence and the target negative feedback interaction sequence into the trained information prediction model for object representation processing, so as to obtain the positive feedback interaction object representation and the negative feedback interaction object representation corresponding to the target object.

[0017] The information representation generation module is configured to perform information representation processing on each candidate push information at the current time based on the trained information prediction model, and obtain the candidate push information representation corresponding to each candidate push information.

[0018] The similarity calculation module is configured to perform similarity calculation between each candidate push information representation and the positive feedback interaction object representation to obtain a first similarity corresponding to each candidate push information, and to perform similarity calculation between each candidate push information representation and the negative feedback interaction object representation to obtain a second similarity corresponding to each candidate push information.

[0019] The similarity difference calculation module is configured to obtain the similarity score for each candidate push information based on the difference between the first similarity and the second similarity corresponding to each candidate push information.

[0020] The push information determination module is configured to determine the candidate push information corresponding to the highest similarity score as the target push information;

[0021] The trained information prediction model is obtained by training according to the training method of the information prediction model described above.

[0022] According to another aspect of the present disclosure, a computer-readable storage medium is provided that stores a computer program, wherein the computer program, when executed by a processor, performs the above-described training method and related methods for the information prediction model.

[0023] According to another aspect of the present disclosure, a computer device is provided, the computer device including a processor and a memory, the memory storing a computer program, which, when called by the processor, executes the above-described training method and related methods for the information prediction model.

[0024] According to another aspect of the present disclosure, a computer program product is provided, which includes a computer program stored in a storage medium; a processor of a computer device reads the computer program from the storage medium and executes the computer program, causing the computer device to perform the training method and related methods of the information prediction model described above.

[0025] This embodiment not only acquires the positive feedback interaction sequence corresponding to the sample object, but also acquires the negative feedback interaction sequence as training data for the information prediction model. This avoids the sample bias caused by a single feedback sequence, and can accurately construct positive and negative sample pairs by leveraging cross-sample feedback information and temporal features. This eliminates the need for the information prediction model to process a large amount of irrelevant or repetitive sample data, thereby effectively reducing redundant computation and improving the resource utilization of electronic devices. Simultaneously, the joint optimization of the dual loss values ​​enables the information prediction model to learn the interaction preferences and positive and negative feedback boundaries of the sample object more efficiently. This not only improves the model's accuracy in information prediction, its ability to distinguish between positive and negative feedback, and its ability to capture temporal correlation interests, but also enhances the model's generalization performance.

[0026] Other features and advantages of embodiments of this disclosure will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of this disclosure. The objects and other advantages of embodiments of this disclosure may be realized and obtained by means of the structures particularly pointed out in the description, claims, and drawings. Attached Figure Description

[0027] Figure 1 shows a schematic diagram of a system architecture provided in an embodiment of this disclosure;

[0028] Figure 2 shows a schematic diagram of a system deployment provided in an embodiment of this disclosure;

[0029] Figure 3 shows a schematic diagram of an application scenario for an article recommendation scenario provided by an embodiment of this disclosure;

[0030] Figure 4 shows a schematic diagram of an e-commerce recommendation scenario provided by an embodiment of this disclosure;

[0031] Figure 5 shows a flowchart illustrating a training method for an information prediction model provided in an embodiment of this disclosure;

[0032] Figure 6 shows a schematic diagram of a training set provided in an embodiment of this disclosure;

[0033] Figure 7 illustrates a schematic diagram of a historical positive feedback interaction record provided in an embodiment of this disclosure;

[0034] Figure 8 illustrates a schematic diagram of content information generation according to an embodiment of this disclosure;

[0035] Figure 9 shows a schematic diagram of a raw positive feedback interaction sequence provided in an embodiment of this disclosure;

[0036] Figure 10 shows a model architecture diagram of an information prediction model provided in an embodiment of this disclosure;

[0037] Figure 11 shows a schematic diagram of a feature extraction module provided in an embodiment of this disclosure;

[0038] Figure 12 shows a flowchart of a push information selection method provided in an embodiment of this disclosure;

[0039] Figure 13 is a block diagram of a training device for an information prediction model provided in an embodiment of this disclosure;

[0040] Figure 14 is a block diagram of a question-answering device provided in an embodiment of this disclosure;

[0041] Figure 15 is a block diagram of a computer device provided in an embodiment of this disclosure;

[0042] Figure 16 is a block diagram of a computer-readable storage medium provided in an embodiment of this disclosure. Detailed Implementation

[0043] To enable those skilled in the art to better understand the solutions disclosed herein, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments. The implementation methods described with reference to the accompanying drawings are exemplary and are only used to explain this disclosure, and should not be construed as limiting this disclosure. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0044] In some processes described in the specification, claims, and the foregoing drawings, multiple steps are included in a specific order. However, it should be clearly understood that these steps may not be performed in the order they appear herein, or may be performed in parallel. The step numbers are merely used to distinguish different steps and do not themselves represent any execution order. Furthermore, descriptions such as "first," "second," or "objective" in this document are used to distinguish similar objects and are not necessarily used to describe a specific order, sequence, or quantity.

[0045] It is worth noting that, in the implementation of the embodiments of this disclosure, related data such as positive feedback interaction sequences and negative feedback interaction sequences corresponding to the sample objects are involved. When the above embodiments of this disclosure are applied to specific products or technologies, permission or consent from the user (hereinafter collectively referred to as the object) is required, and the collection, use, and processing of related data must comply with relevant laws, regulations, and standards. For example, when this disclosure requires obtaining a positive feedback interaction sequence, separate permission or consent for the positive feedback interaction sequence can be obtained through pop-up windows or redirection to a confirmation page. After obtaining separate permission or consent, the necessary related data for enabling the embodiments of this disclosure to operate normally can then be obtained.

[0046] Red dot recommendations are a common strategy in recommendation systems, primarily used to alert target users to new information or content updates. Sometimes referred to as content push notifications, their goal is to attract attention and guide user actions, thereby increasing user activity and engagement. Red dot recommendations manifest differently across various products. For example, the main chat interface of messaging apps displays a social media activity bar, which collects notifications from multiple social media accounts followed by the target user. When a social media account sends a new message to the messaging app, the activity bar icon displays a small red dot, along with the title of the new message to be pushed to the target user.

[0047] Typically, red dot recommendations not only emphasize the timeliness of content but also require recommendation algorithms to calculate the candidate items that the target audience is most likely to be interested in. The implementation of this recommendation method relies on the quality of behavioral sequence modeling. Currently, relevant technologies mainly focus on the positive feedback behaviors of the target audience, such as liking or saving articles. However, this approach based on positive feedback behavior modeling cannot effectively capture the target audience's interests, resulting in low accuracy in recommending items to the target audience. To address the above problems, embodiments of this disclosure propose a training method and related methods for an information prediction model. The system implementation environment and related application scenarios of the methods involved in the embodiments of this disclosure will be described below.

[0048] Please refer to Figure 1, which illustrates a system architecture diagram provided by an embodiment of this disclosure. The training method and related methods of the information prediction model provided by this disclosure can be applied in system 100. The sample acquisition device 1110 is configured to acquire a large amount of training data, which may include positive feedback interaction sequences and negative feedback interaction sequences corresponding to sample objects. Optionally, the sample acquisition device 1110 can acquire historical positive feedback interaction records and historical negative feedback interaction records corresponding to sample objects, and generate corresponding positive feedback interaction sequences and negative feedback interaction sequences based on these records. The sample acquisition device 1110 can store the training data in database 1120 for model training. The training device 1130 can train a preset model based on the training data in database 1120 until the preset model meets the training termination condition, thus obtaining the trained information prediction model, i.e., the target model 101.

[0049] The training termination conditions may include: the total loss value being less than a preset value, the total loss value being within a threshold range, or the training iterations reaching a preset number. The target model 101 may be a deep neural network composed of one or more neural networks. For example, the information prediction model of the target model 101 may be a deep neural network with a Large Language Model (LLM) as its backbone. The information prediction model may include a pre-trained LLM, embedding layers for embedding different types of information (such as time, text, etc.), etc., without limitation.

[0050] The execution device 1140 can deploy the target model 101 trained by the training device 1130. During the computation and other related processing performed by the processing module 141 of the execution device 1140, the processing module 141 can call computer programs, data, etc. in the data storage system 142 for corresponding computational processing, and store the processing results and instructions obtained from the computational processing into the data storage system 142. Data transmission between the execution device 1140 and the user device 1150 can be performed through an input / output (I / O) interface.

[0051] For example, when the execution device 1140 receives multiple candidate push information, target positive feedback interaction sequence and target negative feedback interaction sequence corresponding to the target object sent by the user device 1150, it can call the processing module 141 to use the target model 101, that is, the trained information prediction model, to determine the target push information from the multiple candidate push information based on the target positive feedback interaction sequence and target negative feedback interaction sequence corresponding to the target object, and return the target push information to the user device 1150 through the I / O port.

[0052] The aforementioned training device 1130, execution device 1140, and user device 1150 can be computer devices such as servers or terminals. The server can be a standalone 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 (CDNs), blockchain, and big data and artificial intelligence platforms. The terminal can be a laptop, desktop computer, or smartwatch, etc.

[0053] It should be noted that Figure 1 is merely a schematic diagram of a system architecture provided by an embodiment of this disclosure. The system architecture and application scenarios described in this disclosure are for the purpose of more clearly illustrating the technical solutions of this disclosure and do not constitute a limitation on the technical solutions provided by this disclosure. For example, the execution device 1140 in Figure 1 can be configured to deploy the target model 101 trained by the training device 1130. In other cases, the user device 1150 can also directly deploy the target model 101 trained by the training device 1130 and directly select push information.

[0054] Please refer to Figure 2, which illustrates a system deployment diagram provided by an embodiment of this disclosure. Exemplarily, the training method and related methods for the information prediction model provided by this disclosure can be deployed in a client / server (C / S) architecture system as shown in Figure 2. This system may include servers, gateways, the internet, and computer equipment, etc.

[0055] A server is a computer system that provides certain services to the service provider. Compared to ordinary computer equipment, servers have higher requirements in terms of stability, security, and performance. As shown in Figure 2, a server 2110 can be a single high-performance computer in a network platform, a cluster of multiple high-performance computers, a portion of a single high-performance computer (e.g., a virtual machine), or a combination of portions of multiple high-performance computers (e.g., virtual machines).

[0056] A gateway, also known as an internetwork connector or protocol converter, is a computer system or device that acts as a translator between two systems using different communication protocols, data formats, languages, or even completely different architectures. As shown in Figure 2, a gateway 2220 enables network interconnection at the transport layer via the Internet 2330. It serves as a translator. Gateways can also provide filtering and security functions.

[0057] Computer devices can send requests to servers, requesting access to or use resources, data, or services on the server. As shown in Figure 2, computer devices can include various forms such as tablet computers 240a, desktop computers 240b, and laptop computers 240c. Furthermore, it can be a single computer device or a collection of multiple computer devices.

[0058] Computer devices can communicate with the Internet and exchange data via wired or wireless means. As shown in Figure 2, tablet computer 240a can communicate with the Internet 2330 through wireless router 2550. Similarly, messages sent by desktop computer 240b to server 2110 can be transmitted to the corresponding server 2110 through gateway 2220. Messages sent by server 2110 to computer devices can also be transmitted to the corresponding desktop computer 240b through gateway 2220.

[0059] As one implementation method, the training method and related methods of the information prediction model provided in this disclosure can be implemented partly on a server and partly on a computer device. The server can obtain the positive feedback interaction sequence and negative feedback interaction sequence corresponding to the sample object, and use these sequences as training data to train the information prediction model, thus obtaining the trained information prediction model. The computer device can deploy the trained information prediction model and use it to perform object representation processing on the target positive feedback interaction sequence and target negative feedback interaction sequence corresponding to the target object, respectively, to obtain the positive feedback interaction object representation and negative feedback interaction object representation corresponding to the target object. Then, based on the positive feedback interaction object representation and negative feedback interaction object representation, the computer device determines the target push information from multiple candidate push information.

[0060] It should be noted that Figure 2 is merely a schematic diagram of a system deployment provided by an embodiment of this disclosure. The system deployment scheme described in this disclosure is only for the purpose of more clearly illustrating the technical solutions of this disclosure and does not constitute a limitation on the technical solutions provided by this disclosure. For example, "computer device" can refer to one of multiple devices, and this embodiment only uses the three computer devices shown in Figure 2 (tablet computer 240a, desktop computer 240b, and laptop computer 240c) as examples. Those skilled in the art will understand that as system deployment schemes evolve, the technical solutions provided by this disclosure are also applicable to similar technical problems.

[0061] The training method and related methods of the information prediction model provided in this disclosure can be applied in various scenarios. See Figures 3 and 4. Figure 3 shows a schematic diagram of an application scenario of an article recommendation scenario provided in this disclosure, and Figure 4 shows a schematic diagram of an application scenario of an e-commerce recommendation scenario provided in this disclosure. The article recommendation scenario shown in Figure 3 and the e-commerce recommendation scenario shown in Figure 4 are examples.

[0062] (1) Article recommendation scenario.

[0063] An article recommendation system can accurately recommend articles to a target audience. An article recommendation system can be developed based on the training method and related methods of the information prediction model according to embodiments of this disclosure. As shown in Figure 3, the article recommendation system may include a server 311 deployed on a cloud server 3110 and a client 331 deployed on a tablet 3330. The server 311 is mainly responsible for performing complex calculations or logical processing. The client 331 is mainly responsible for providing an interactive interface, sending interactive information obtained based on the target audience's operations on the interactive interface to the server 311, and displaying push information sent by the server 311 on the interactive interface. In practical applications:

[0064] Client 331 can obtain the historical interaction records of the target object with the XX Communication news box, and send the historical records to server 311 on cloud server 3110 via tablet 3330. Server 311 can generate a target positive feedback interaction sequence and a target negative feedback interaction sequence corresponding to the target object based on the received historical records. Then, the target positive feedback interaction sequence and the target negative feedback interaction sequence are respectively input into the trained information prediction model for object representation processing to obtain the target object's positive feedback interaction object representation and negative feedback interaction object representation. After obtaining multiple articles to be sent within a certain time period, server 311 can use the trained information prediction model to generate an article representation for each article to be sent. Then, by calculating the similarity between the article representation and the positive feedback interaction object representation and the negative feedback interaction object representation, the target tweet is determined from the multiple articles to be sent, and the target tweet is sent to client 331. Client 331 then displays a "red dot" and the title of the target tweet on the news box.

[0065] (2) E-commerce recommendation scenario.

[0066] E-commerce recommendation platforms enable accurate product recommendations to target audiences. The training method and related methods of the information prediction model according to embodiments of this disclosure can be applied to online shopping stores. As shown in Figure 4, the shopping interface 4110 of the online shopping store displays a notification message bar 411, which is used to display promotional advertisements for products in the online shopping store. The online shopping store can send the target audience's daily viewing history of promotional advertisements to the e-commerce recommendation platform. The e-commerce recommendation platform then generates positive feedback interaction object representations and negative feedback interaction object representations corresponding to the target audience based on the historical records. These representations contain the target audience's changing interest patterns in which types of promotional advertisements they regularly view. Furthermore, by calculating the similarity between the representation of the product to be pushed and the positive and negative feedback interaction object representations, the target product is determined from multiple products to be pushed and sent to the online shopping store. The shopping interface 4110 then displays the relevant information of the target product.

[0067] It should be noted that Figures 3 and 4 are merely schematic diagrams of application scenarios provided by the embodiments of this disclosure. The application scenarios described in the embodiments of this disclosure are only for more clearly illustrating the technical solutions of the embodiments of this disclosure and do not constitute a limitation on the technical solutions provided by the embodiments of this disclosure. The training method and related methods of the information prediction model of the embodiments of this disclosure can also be used in other scenarios requiring recommendations. As those skilled in the art will understand, with the evolution of application scenarios, the technical solutions of the embodiments of this disclosure are also applicable to similar technical problems.

[0068] According to one embodiment of this disclosure, a method for training an information prediction model is provided. This method can be executed by a computer device (e.g., a desktop computer, laptop computer, or other terminal device with certain processing capabilities), which at least has storage, computing, and communication functions. Referring to Figure 5, Figure 5 shows a flowchart illustrating a method for training an information prediction model according to an embodiment of this disclosure. This method may include the following steps.

[0069] Step 110: Obtain the positive feedback interaction sequence and negative feedback interaction sequence corresponding to the sample object.

[0070] Typically, objects and items are the two core components of a recommendation system. Objects refer to individuals using the recommendation system, while items refer to the entities being recommended, such as products, articles, and videos. Since the core task of a recommendation system is to understand an object's preferences and push items that the object might be interested in, related technologies primarily focus on the positive feedback interactions between objects and items in the recommendation scenario to obtain object preferences. These positive feedback interactions refer to behaviors that demonstrate a clear interest or approval of an item, such as clicking, saving, liking, sharing, or purchasing. These behaviors indicate that the object has a positive evaluation or active attention to the relevant items.

[0071] However, focusing solely on positive feedback interactions of an object can lead to a limited understanding of its preferences. This is because positive feedback interactions only reflect an object's interest in items—that is, what items the object likes. But in real-world applications, an object's interests are dynamic; an object may show dislike for certain items or lose interest in items it was previously interested in at a certain time—that is, negative feedback interactions. Therefore, simply knowing what items an object likes cannot reveal the full picture of its interests, resulting in lower recommendation accuracy. To address this, embodiments of this disclosure propose a method for training an information prediction model for recommendation scenarios based on the object's positive and negative feedback interaction sequences. This information prediction model comprehensively characterizes the evolution of an object's interests, thereby improving recommendation accuracy.

[0072] In this embodiment of the disclosure, the training data used to train the information prediction model includes positive feedback interaction sequences and negative feedback interaction sequences corresponding to the sample objects. Each positive feedback interaction sequence contains at least two positive feedback interaction messages, and each negative feedback interaction sequence contains at least two negative feedback interaction messages. In some embodiments, the positive feedback interaction sequence is a sequence composed of multiple positive feedback interaction messages arranged in chronological order, and the negative feedback interaction sequence is a sequence composed of multiple negative feedback interaction messages arranged in chronological order. Positive feedback interaction information refers to the attribute information of items that have had positive feedback interactions with the object, and negative feedback interaction information refers to the attribute information of items that have had negative feedback interactions with the object. Negative feedback interaction is a negative interaction behavior generated by the sample object in response to the push notification, indicating disinterest. Both positive and negative feedback interaction information include time information, identification information, and content information.

[0073] Among these, time information refers to the interaction time when an object and an item engage in positive or negative feedback interaction. Time information can include multiple time dimensions, such as month, day, and hour. Identification information refers to the source identifier corresponding to the item that engages in positive or negative feedback interaction with the object; for example, in a news push scenario, it's the news publisher's account. Content information refers to the theme and modality corresponding to the item that engages in positive or negative feedback interaction with the object. Modality describes the modality of the item's content, such as text, voice, or video. The theme describes the central idea or core content of the item, such as an article title or product description.

[0074] For example, please refer to FIG6, which shows a schematic diagram of a training set provided in an embodiment of the present disclosure. As shown in FIG6, the training set... n>0 and n∈N * Where n is the number of training samples in the training set S, and the i-th training sample in the training set S includes the positive feedback interaction sequence corresponding to the i-th sample object. and negative feedback interaction sequence With positive feedback interaction sequence For example, positive feedback interaction sequence m>0 and m∈N * m is the positive feedback interaction sequence The number of positive feedback interaction messages, N * It is the set of positive natural numbers, that is, the set of all integers greater than 0. In a positive feedback interaction sequence... In the middle, the j-th positive feedback interaction information It includes time information, identifier information, and content information. The time information includes five dimensions: [month, day, hour, minute, day and hour]. The identifier information includes the source identifier. The content information includes the token sequence after word segmentation based on topic and modality.

[0075] Since positive feedback interaction sequences are sequences of positive feedback interaction information arranged chronologically, and negative feedback interaction sequences are sequences of negative feedback interaction information arranged chronologically, they can jointly reflect the changing trends of object preferences over time. That is, what an object likes at some times and dislikes at others. Therefore, by training an information prediction model using positive and negative feedback interaction sequences, the model can learn to capture the evolutionary path of an object's interests from interaction sequences in two different interaction scenarios, thus providing a comprehensive understanding of object interests. To this end, positive feedback interaction information can be collected from the historical interaction records of different objects to construct positive feedback interaction sequences, and negative feedback interaction information can be collected to construct negative feedback interaction sequences.

[0076] In some embodiments, the computer device can acquire historical positive feedback interaction records and historical negative feedback interaction records corresponding to the sample object, construct a sequence based on the historical positive feedback interaction records to obtain the original positive feedback interaction sequence corresponding to the sample object, construct a sequence based on the historical negative feedback interaction records to obtain the original negative feedback interaction sequence corresponding to the sample object, segment the original positive feedback interaction sequence to obtain a positive feedback interaction sequence of a preset length, and segment the original negative feedback interaction sequence to obtain a negative feedback interaction sequence of a preset length.

[0077] In practice, the historical positive feedback interaction records of the sample object within a preset time range can be cleaned to obtain clean historical positive feedback interaction records. Data cleaning includes removing duplicate records, erroneous operation records, short-term exit records, and other abnormal interaction records. Based on the clean historical positive feedback interaction records, they are arranged according to preset rules (such as descending or ascending order of interaction time) to form an ordered data set, i.e., a positive feedback interaction sequence. Similarly, the historical negative feedback interaction records of the sample object within a preset time range can be cleaned to obtain clean historical negative feedback interaction records. Data cleaning includes removing duplicate records, erroneous operation records, short-term exit records, and other abnormal interaction records. Based on the clean historical negative feedback interaction records, they are arranged according to preset rules (such as descending or ascending order of interaction time) to form an ordered data set, i.e., a negative feedback interaction sequence.

[0078] The historical positive feedback interaction records include descriptions, source identifiers, and interaction times corresponding to multiple positive feedback interactions in positive feedback interaction scenarios. For example, the sample object's history of collecting, liking, or commenting on articles in the Red Dot Recommendation's Red Dot consumption scenario (i.e., a positive feedback interaction scenario) includes the title and modality (i.e., descriptions), publisher account (i.e., source identifier), and like time (i.e., interaction time) for each article. The historical negative feedback interaction records include descriptions, source identifiers, and interaction times corresponding to multiple negative feedback interactions in negative feedback interaction scenarios. For example, negative feedback interaction information could be the sample object's history of clicking red dots and quickly exiting for some articles in the Red Dot Recommendation's Red Dot removal scenario (i.e., a negative feedback interaction scenario).

[0079] For example, Figure 7 illustrates a schematic diagram of a historical positive feedback interaction record provided in an embodiment of this disclosure. As shown in Figure 7, the main interface of XX Communications displays an entry point for viewing tweets called "Information Box." When XX Communications receives a new tweet, a "red dot" can be displayed on the icon of the Information Box. When the Information Box is clicked by an object, XX Communications can display the tweet interface corresponding to the tweet viewed by the object. If the object likes the tweet in the tweet interface, then the like operation is a positive feedback interaction, and the historical record corresponding to the tweet is the historical positive feedback interaction record. The historical positive feedback interaction record includes: providing a source identifier (i.e., the ID of the tweet publisher "XX Network"), interaction time (i.e., the time when the object likes the tweet), topic (i.e., the tweet title), and modality (i.e., the modality of the tweet content: image and text).

[0080] As one implementation method, by preprocessing the description content, interaction time, and source identifier corresponding to each interaction information in the historical interaction records, the original positive feedback interaction sequence corresponding to the historical positive feedback interaction records and the original negative feedback interaction sequence corresponding to the historical negative feedback interaction records can be constructed based on the preprocessing results. The construction steps are described below using the construction of the original positive feedback interaction sequence as an example.

[0081] (1) Segment the description content corresponding to each historical positive feedback interaction information to obtain the content information corresponding to each historical positive feedback interaction information.

[0082] Considering that the technology learns object preferences from interaction sequences, it usually uses independent encoders to process different modal information of the content of the recommended items, resulting in a lack of cross-modal understanding mechanism. However, information from different modalities may be highly correlated. For example, the clothing description and clothing image in a fashion tweet may both point to a certain clothing design style. If the clothing description and clothing image are encoded independently, it means that the clothing design style associated between the clothing description and clothing image cannot be learned, that is, there is a lack of cross-modal understanding.

[0083] However, in recommendation scenarios, the characteristics of object interactions often reflect the object's preference for combinations of different modal features. For example, an object might enjoy reading medical knowledge articles from a certain subscription account not only because the textual explanations of medical knowledge are professional, but also because the accompanying images are very simple and easy to understand. Therefore, the characteristics of object interactions may not be fully characterized by a single modal feature, resulting in the loss of semantic or relevance information between modalities, thereby reducing the accuracy of the final recommendation. To address this, embodiments of this disclosure propose explicitly labeling the topics and modalities in the description content (as shown in Figure 7) in the form of prompts, enabling the model to achieve cross-modal understanding during training.

[0084] It should be noted that word segmentation of descriptive content refers to a standardized data processing operation that uses pre-defined language rules or word segmentation tools to break down descriptive content (such as text titles, summaries, etc.) used to describe interactive information topics and modalities into the smallest units (such as words, sub-words, or characters) with independent semantics. Word segmentation of descriptive content transforms unstructured text descriptions into structured semantic sequences that the model can recognize and encode. This allows the model to capture the core semantics within the descriptive content through these semantic units, thereby accurately understanding user preferences for different topics and modalities.

[0085] The descriptive content describes the information theme and information modality of the interactive information. The information theme refers to the core content direction around which the interactive information revolves (such as technology, finance, entertainment, etc.), which is reflected through keywords in the descriptive content. It is a core dimension constituting user interests and helps the model identify user preferences for specific content directions. The information modality refers to the presentation format of the interactive information (such as text, video, audio, etc.), which can be determined through explicit identifiers or implicit features in the descriptive content. It is an important factor influencing user interaction behavior and provides a basis for the model to adapt to multimodal content recommendation.

[0086] It is understandable that content information refers to the smallest set of units with independent semantics obtained after segmenting the descriptive content, such as words and sub-word sequences. It is a structured representation of the information theme and information modality of interactive information, providing basic data for models to capture core semantic features and understand user preferences.

[0087] For example, a computer device can generate a prompt based on the topic and modality in the description content corresponding to historical positive feedback interaction information. The computer device can input the prompt into a word segmenter for word segmentation to obtain a segmented word sequence, and use the segmented word sequence as the content information corresponding to the historical positive feedback interaction information. In this way, the content information corresponding to each historical positive feedback interaction information is obtained.

[0088] Referring to Figure 8, which illustrates a schematic diagram of content information generation according to an embodiment of this disclosure, a prompt template is selected from a prompt template library, and the topic and modality of the description content are filled into the prompt template to obtain a prompt. In some embodiments, the prompt is input to a tokenizer for word segmentation. The tokenizer can convert the text in the prompt into a word segmentation sequence and add an indicator marker [ITEM] at the end of the word segmentation sequence. This indicator marker is used to instruct the model to encode, thereby determining the word segmentation sequence as content information. Optionally, a length threshold (e.g., 128) can be set for the word segmentation sequence. When the length of the generated word segmentation sequence does not reach the length threshold, the length of the sequence can be reached by adding a padding marker [PAD] to the word segmentation sequence, so that the content information corresponding to each historical positive feedback interaction information is a word segmentation sequence of the same length.

[0089] (2) Structure the interaction time corresponding to each historical positive feedback interaction information to obtain the time information corresponding to each historical positive feedback interaction information. The time information includes at least the day-time dimension.

[0090] In the process of learning object preferences from interaction sequences, existing technologies typically employ positional encoding or simple timestamp encoding to model object interaction preferences over time. However, these modeling methods struggle to capture the timeliness of object interaction preferences, thus failing to fully utilize object interaction patterns over time, namely positive and negative feedback interactions. Therefore, embodiments of this disclosure propose a structured temporal representation to enhance the model's perception of the timeliness of object interaction preferences, thereby improving the accuracy of the final recommendation.

[0091] Among them, time information refers to the multi-dimensional numerical feature set obtained after structuring the original interaction time, which includes at least the day-hour dimension. It is the core data characterizing the temporal attributes of interaction behavior, providing a basis for the time ordering of sequence construction and supporting the model in perceiving the temporal distribution pattern of user interests. The day-hour dimension refers to the combined dimension of "date" and "hour" in the time information (such as "5th day 19:00"). It is the core component of time information and can effectively reflect the differences in user interaction preferences at different dates and time periods, providing key features for the model to accurately capture temporal interests.

[0092] It should be noted that structuring refers to the standardized organization of raw interaction time, such as time records in different formats, extracting key time elements according to preset unified rules and formatting them to transform them into standardized time information.

[0093] For example, a computer device can extract time features from the interaction time in four time dimensions: month, day, hour, and minute. Based on the time features of the "day" and "hour" time dimensions, it can construct a "day-hour" time feature. For instance, the four time dimensions of month, day, hour, and minute have 12, 31, 24, and 60 possible values, respectively. Correspondingly, the "day-hour" time dimension has 31*24 possible values. As shown in Figure 7, the interaction time "202411012107" can be structured to obtain the time information "11 / 01 / 21 / 07 / 1121". By adding this cross-time dimension of "day-hour", the distinguishability of object interaction patterns can be improved, enabling the model to accurately learn the timeliness of object interaction preferences.

[0094] (3) Use the source identifier corresponding to each historical positive feedback interaction information as the corresponding identifier information.

[0095] It should be noted that providing a source identifier refers to a unique symbol or code that identifies the entity publishing interactive information, such as a content creator or institutional account. This identifier is used to clearly distinguish the source channels of different information and is a key basis for the model to identify user preferences for specific publishing entities. Identification information refers to the feature data that represents the source of information in a structured sequence, supporting the model's accurate modeling of user source preferences.

[0096] For example, as shown in Figure 7, the account of the tweet publisher "XX.net" is the source identifier "00101", which is used as the identification information.

[0097] (4) Construct a sequence of time information, identification information and content information of each historical positive feedback interaction according to the time dimension to obtain the original positive feedback interaction sequence corresponding to the sample object.

[0098] For example, after obtaining the time information, identification information, and content information of each historical positive feedback interaction, the computer device can refer to the time information to assemble each historical positive feedback interaction into an original positive feedback interaction sequence according to the time dimension. Thus, the original positive feedback interaction sequence includes a time sequence composed of the time information of each historical positive feedback interaction, an identification sequence composed of the identification information of each historical positive feedback interaction, and a content sequence composed of the content information of each historical positive feedback interaction.

[0099] For example, referring to Figure 9, which shows a schematic diagram of an original positive feedback interaction sequence provided in an embodiment of this disclosure, the time information of historical positive feedback interaction information A in Figure 9 is [11, 01, 13, 24, 1113], and the time information of historical positive feedback interaction information B is [11, 01, 21, 07, 1121]. According to the chronological order, historical positive feedback interaction information B is arranged after historical positive feedback interaction information A. The same method is used to arrange other historical positive feedback interaction information in sequence to obtain the original positive feedback interaction sequence. The original positive feedback interaction sequence includes a time sequence, an identifier sequence, and a content sequence.

[0100] Thus, word segmentation of descriptive content transforms unstructured text into content information that the model can recognize, enabling the model to directly capture information themes and modal features, thereby understanding user content preferences. Structuring interaction time while retaining the day-time dimension transforms raw time into computable temporal features, allowing the model to perceive user interaction patterns at different times and dates. Utilizing source identifiers as identification information clearly conveys the source attributes of information, helping the model identify user preferences for specific publishing entities. Furthermore, integrating multi-dimensional features along the time dimension to construct sequences fully presents the temporal evolution trajectory of user interests, allowing the model to simultaneously learn the relationships between content, time sequence, and source. Ultimately, through precise interest characterization, high-quality input is provided for subsequent processing and model training, thereby enhancing the model's personalization and accuracy.

[0101] After constructing the original positive feedback interaction sequence and the original negative feedback interaction sequence, in order to build the training set, the original positive feedback interaction sequence can be segmented to obtain a positive feedback interaction sequence of a preset length, and the original negative feedback interaction sequence can be segmented to obtain a negative feedback interaction sequence of a preset length. The preset length is set according to the training requirements.

[0102] It should be noted that sequence segmentation refers to the standardized truncation or padding of original positive feedback interaction sequences and original negative feedback interaction sequences of variable length, in order to adapt to the input format of the information prediction model and improve training efficiency and stability.

[0103] For example, if the maximum sequence length of the original positive feedback interaction sequence and the original negative feedback interaction sequence is 1000, then the preset length can be set to 50, 51, or 64, etc., without limitation. The positive and negative feedback interaction sequences corresponding to a single sample object are used as a training sample, thus obtaining a training set composed of multiple training samples. The information prediction model is then trained based on this training set, enabling the trained model to effectively capture the cross-scene evolution patterns of object interests.

[0104] Step 120: Input the positive feedback interaction sequence and the negative feedback interaction sequence into the information prediction model to predict the information, and obtain the positive feedback interaction information representation corresponding to each positive feedback interaction information in the positive feedback interaction sequence, and the negative feedback interaction information representation corresponding to each negative feedback interaction information in the negative feedback interaction sequence.

[0105] To enable the information prediction model to learn the changes in an object's interest in different scenarios, positive and negative feedback interaction sequences need to be input into the model. The model then predicts the positive and negative feedback interaction information from these sequences, obtaining corresponding representations. Based on these representations, a loss value is calculated to train the model.

[0106] In some embodiments, the computer device can input a positive feedback interaction sequence into an information prediction model, and based on the information prediction model, extract features from the positive feedback interaction sequence to obtain a positive feedback interaction information feature representation corresponding to each positive feedback interaction information in the positive feedback interaction sequence. Then, context fusion is performed on each positive feedback interaction information feature representation to obtain a positive feedback interaction information representation corresponding to each positive feedback interaction information. Simultaneously, the computer device can input a negative feedback interaction sequence into an information prediction model, and based on the information prediction model, extract features from the negative feedback interaction sequence to obtain a negative feedback interaction information feature representation corresponding to each negative feedback interaction information in the negative feedback interaction sequence. Then, context fusion is performed on each negative feedback interaction information feature representation to obtain a negative feedback interaction information representation corresponding to each negative feedback interaction information.

[0107] It should be noted that the feature representation of positive feedback interaction information refers to the high-dimensional feature vector formed by transforming a single positive feedback interaction information after feature extraction of the positive feedback interaction sequence. It is an independent encoding and preliminary integration of the content features, temporal features, and source features of the positive feedback interaction information itself, focusing only on the inherent attributes of a single positive feedback interaction information. The feature representation of negative feedback interaction information refers to the high-dimensional feature vector formed by transforming a single negative feedback interaction information after feature extraction of the negative feedback interaction sequence. It is completely consistent with the feature representation of positive feedback interaction information in terms of dimension and data type. It is also an independent encoding and preliminary integration of the content features, temporal features, and source features of the negative feedback interaction information itself, representing only the inherent attributes of a single interaction information. Positive feedback interaction information representation refers to a fixed-dimensional high-dimensional feature vector formed after contextual fusion of the positive feedback interaction information feature representation. Negative feedback interaction information representation refers to a fixed-dimensional high-dimensional feature vector formed after contextual fusion of the negative feedback interaction information feature representation.

[0108] Please refer to Figure 10, which illustrates the model architecture of an information prediction model provided in this embodiment of the present disclosure. As shown in Figure 10, the information prediction model may include a feature extraction module, a context fusion module, and a representation fusion module. The feature extraction module is used to extract features from the identifier information, content information, and time information corresponding to each positive feedback interaction information / negative feedback interaction information in the output positive feedback interaction sequence / negative feedback interaction sequence, and fuse the extraction results to obtain the corresponding positive feedback interaction information feature representation / negative feedback interaction information feature representation without context information. The context fusion module is used to perform context fusion on the positive feedback interaction information feature representation / negative feedback interaction information feature representation to obtain the corresponding positive feedback interaction information representation / negative feedback interaction information representation with context information. The representation fusion module is used to fuse each positive feedback interaction information representation / negative feedback interaction information representation into a positive sample object representation corresponding to the positive feedback interaction sequence / a negative sample object representation corresponding to the negative feedback interaction sequence.

[0109] As one implementation method, taking a positive feedback interaction sequence as an example, feature extraction is performed on the identifier information of each positive feedback interaction information in the positive feedback interaction sequence to obtain the identifier feature representation corresponding to each positive feedback interaction information; feature extraction is performed on the time information of each positive feedback interaction information in the positive feedback interaction sequence to obtain the time feature representation corresponding to each positive feedback interaction information; feature extraction is performed on the content information of each positive feedback interaction information in the positive feedback interaction sequence to obtain the content feature representation corresponding to each positive feedback interaction information; the identifier feature representation, time feature representation, and content feature representation corresponding to each positive feedback interaction information are fused to obtain the positive feedback interaction information feature representation corresponding to each positive feedback interaction information.

[0110] It should be noted that: Identifier feature representation refers to a low-dimensional feature vector formed by encoding the source identifier of positive feedback interaction information. It is a structured representation of the information source attribute, used to reflect the characteristics of the subject providing the interaction information, and provides basic data for the source dimension in multi-dimensional feature fusion. Content feature representation refers to a low-dimensional feature vector formed by encoding the descriptive content of positive feedback interaction information. It is a structured representation of content attributes such as information theme and modality, used to reflect the core semantic features of the interaction information, and provides basic data for the content dimension in multi-dimensional feature fusion. Temporal feature representation refers to a low-dimensional feature vector formed by encoding the structured temporal information of positive feedback interaction information. It is a structured representation of the interaction temporal sequence attribute, used to reflect the regularity of interaction time, and provides basic data for the temporal dimension in multi-dimensional feature fusion.

[0111] In some embodiments, the identifier feature representation and content feature representation corresponding to each positive feedback interaction information are concatenated to obtain a first concatenated feature representation corresponding to each positive feedback interaction information; each first concatenated feature representation is aligned to obtain an aligned feature representation corresponding to each positive feedback interaction information; the aligned feature representation and time feature representation corresponding to each positive feedback interaction information are concatenated to obtain a second concatenated feature representation corresponding to each positive feedback interaction information; each second concatenated feature representation is aligned to obtain a positive feedback interaction information feature representation corresponding to each positive feedback interaction information.

[0112] Please refer to Figure 11, which illustrates a feature extraction module according to an embodiment of this disclosure. Taking positive feedback interaction information as an example, as shown in Figure 11, the feature extraction module may include an identifier layer, a content layer, a temporal layer, and a fully connected layer. The identifier layer is used to extract identifier feature representations from identifier information, the content layer is used to extract content feature representations from content information, and the temporal layer is used to extract temporal feature representations from temporal information. The identifier layer, content layer, and temporal layer are all deep neural networks; for example, the content layer may be BERT, etc., without limitation. The identifier feature representation, temporal feature representation, and content feature representation are vector-based embeddings.

[0113] The identifier feature representation and content feature representation are concatenated to obtain a first concatenated feature representation. This first concatenated feature representation is then input into a fully connected layer, and the output of the fully connected layer undergoes a nonlinear transformation to obtain an alignment feature representation. This alignment feature representation and the temporal feature representation are then concatenated to obtain a second concatenated feature representation. This second concatenated feature representation is then input into a fully connected layer, and the output of the fully connected layer undergoes a nonlinear transformation to obtain a positive feedback interaction information feature representation. The nonlinear transformation can be implemented using an activation function, such as the ReLU (Rectified Linear Unit) function and its variants, which are not limited here. Similarly, the negative feedback interaction information feature representation corresponding to the negative feedback interaction information can be obtained through a feature extraction module.

[0114] It should be noted that the fully connected layer fuses the discrete information of the two types of features into a coherent high-dimensional feature vector through internal weight matrix operations. At the same time, it standardizes the feature dimension, provides adaptive input for subsequent nonlinear transformations, and strengthens the semantic association between the identifier features and the content features.

[0115] For example, taking a user's positive feedback interaction with a technology article as an example: the identifier feature of this interaction is represented as a 64-dimensional vector, and the content feature is represented as a 128-dimensional vector. The two are concatenated in sequence to obtain a first concatenated feature representation of 64 + 128 = 192 dimensions. The 192-dimensional feature is mapped to 128 dimensions through a fully connected layer and then batch normalized to obtain a 128-dimensional aligned feature representation. This aligned feature representation is then concatenated with a 32-dimensional time feature representation to obtain a second concatenated feature representation of 128 + 32 = 160 dimensions. Finally, the 160-dimensional feature is mapped to 128 dimensions through a fully connected layer with the same structure and normalization process, thus obtaining a positive feedback interaction information feature representation that includes three-dimensional features of source, content, and time.

[0116] To enable information prediction models to better understand the semantic and temporal relationships of object interaction sequences and improve the accuracy of recommendation results, context fusion can be performed on the interaction information feature representations corresponding to each interaction in the interaction sequence. Take a positive feedback interaction sequence as an example:

[0117] As one implementation method, based on the time information corresponding to each positive feedback interaction information, a sequence of positive feedback interaction information feature representations corresponding to each positive feedback interaction information in the positive feedback interaction sequence is constructed to obtain a positive feedback interaction information feature representation sequence. In some embodiments, the positive feedback interaction information feature representation sequence is fused with context to obtain the positive feedback interaction information representation corresponding to each positive feedback interaction information.

[0118] It should be noted that sequence construction refers to the process of arranging and integrating the scattered individual positive feedback interaction feature representations in the positive feedback interaction sequence in chronological order, based on the time information corresponding to each positive feedback interaction, to form a temporally coherent sequence of positive feedback interaction feature representations, used to reconstruct the true temporal logic of user interaction. Context fusion refers to the process of capturing the temporal dependencies and semantic relationships between each feature representation within the sequence through temporal modeling, and integrating the related information into individual feature representations. This allows each feature representation to simultaneously contain its own inherent attributes and the semantics of the sequence context, forming a more comprehensive and accurate feature representation, providing support for characterizing user interests. The positive feedback interaction feature representation sequence is an ordered set of features formed by arranging all individual positive feedback interaction feature representations in the positive feedback interaction sequence in chronological order after sequence construction. The positive feedback interaction feature representation sequence fully preserves the temporal dependencies of each interaction and is the direct object of context fusion.

[0119] As one implementation method, the sequence of positive feedback interaction information feature representations is input into a large language model based on the Transformer architecture; based on the large language model, the sequence of positive feedback interaction information feature representations is fused with context to obtain the positive feedback interaction information representation corresponding to each positive feedback interaction information.

[0120] For example, the context fusion module of the information prediction model can be a large language model. For instance, the backbone network of this large language model might consist of 32 Transformer Blocks, each containing a multi-head attention mechanism and a feedforward network, and layer normalization is performed using RMSnorm. The feature representation sequence of each positive feedback interaction is input into the large language model for context fusion to obtain the positive feedback interaction representation corresponding to each positive feedback interaction. Similarly, the context fusion module can be used to obtain the negative feedback interaction representation corresponding to each negative feedback interaction in the negative feedback interaction sequence.

[0121] By fusing the feature representations of positive and negative feedback interactions within a contextual framework, the information prediction model can capture the complex and deep-seated long-term dependencies between different interaction modes (positive and negative feedback interactions). Simultaneously, it enhances the contextual semantics of each interaction representation, enabling it to not only reflect the item's own characteristics but also, in conjunction with the interaction information of other items in the sequence, reveal the evolution of the object's interest under different interaction modes. This approach provides rich semantic information for personalized recommendations, allowing the recommendation system to more accurately predict the object's potential interests, thereby improving recommendation accuracy. Furthermore, compared to generating item representations solely using item tags, generating interaction representations based on interaction information including temporal, tagging, and content information enhances the model's ability to express the content level of interaction information. For example, the positive / negative interaction representations generated by the information prediction model can characterize whether an object is interested in or uninterested in an item of mode Z published by source Y at time X. It also addresses the problem of not being able to generate interaction representations for new items during cold starts due to the lack of historical interaction records.

[0122] This disclosure proposes a method for training an information prediction model for recommendation scenarios based on positive feedback interaction sequences and negative feedback interaction sequences of objects. This information prediction model comprehensively characterizes the interest evolution patterns of objects under different interaction modes, thereby improving recommendation accuracy. To this end, this disclosure can set positive feedback training tasks corresponding to the positive feedback interaction modes based on the positive feedback interaction sequences, and simultaneously set negative feedback training tasks corresponding to the negative feedback interaction modes based on the negative feedback interaction sequences.

[0123] The purpose of this approach is to enable the information prediction model to learn the preferences of users under positive feedback interaction patterns through positive feedback training tasks. For example, it's about when users prefer to read which types of articles and what the content modality of these articles is. Conversely, it's about enabling the information prediction model to learn the interest boundaries of users under both positive and negative feedback interaction scenarios through negative feedback training tasks. For example, it's about which types of goods users dislike and which types they like at a given time. The training process is as follows.

[0124] Step 130: Obtain the other feedback interaction information of other sample objects corresponding to each positive feedback interaction information, and determine the first loss value based on the difference between the first positive sample prediction value calculated based on the representation of each positive feedback interaction information and the first negative sample prediction value calculated based on the representation of the other feedback interaction information of the corresponding other feedback interaction information.

[0125] To enable the information prediction model to learn the preferences of objects under positive feedback interaction patterns, the next item can be predicted for each item corresponding to a positive feedback interaction in the positive feedback interaction sequence. Therefore, it is necessary to set positive and negative samples based on the representation of each positive feedback interaction in the positive feedback interaction sequence. In some embodiments, the positive feedback interaction representation is used as the first positive sample, and other feedback interaction information of other sample objects corresponding to that positive feedback interaction is obtained; the other feedback interaction information representation of those other feedback interaction information is used as the first negative sample.

[0126] It should be noted that "other sample objects" refers to users or entities other than the target sample object corresponding to the current calculation of the first positive sample prediction value. These users provide the feedback interaction information needed to construct the negative sample, ensuring that the negative sample and positive sample have consistent scenarios. "Other feedback interaction information" refers to the feedback interaction information generated by other sample objects, and this information forms a negative contrast with the positive feedback interaction information of the target sample object in core attributes, such as interaction intent and associated content. This information is the core data for constructing the first negative sample, used to form an effective comparison with the positive sample to optimize the model.

[0127] As one implementation method, obtaining other feedback interaction information may include: obtaining other positive feedback interaction sequences and other negative feedback interaction sequences corresponding to other sample objects; randomly sampling the other positive feedback interaction sequences and other negative feedback interaction sequences to obtain a preset number of feedback interaction information as other feedback interaction information. The preset number of information is the same as the number of positive feedback interaction information contained in the positive feedback interaction sequence, and each piece of other feedback interaction information is different from any positive feedback interaction information in the positive feedback interaction sequence.

[0128] In one implementation, the other feedback interaction information corresponding to each positive feedback interaction information is input into the information prediction model for information prediction, thereby obtaining the other feedback interaction information representation corresponding to each other feedback interaction information. A representation mapping is performed on each positive feedback interaction information representation to obtain the first positive sample prediction value corresponding to each positive feedback interaction information representation. A representation mapping is also performed on each other feedback interaction information representation to obtain the first negative sample prediction value corresponding to each other feedback interaction information representation. Then, the difference between each first positive sample prediction value and the corresponding first negative sample prediction value is statistically analyzed to determine the first loss value.

[0129] It should be noted that representation mapping for each positive feedback interaction means processing the vector-form feature representation corresponding to each positive feedback interaction through preset transformation rules, such as fully connected layer operations and nonlinear transformations, ultimately converting it into a quantifiable numerical value that can be used for subsequent loss value calculation. This numerical value is the first positive sample prediction value corresponding to that positive feedback interaction representation. The first positive sample prediction value is the numerical result obtained after performing representation mapping on the positive feedback interaction representation, such as the probability value or similarity score in the interval [0, 1]. The magnitude of the first positive sample prediction value is used to characterize the model's prediction fit for that positive feedback interaction; for example, the higher the first positive sample prediction value, the higher the confidence level of the model in determining that the interaction conforms to user preferences. Similarly, representation mapping for each negative feedback interaction means processing the vector-form feature representation corresponding to each negative feedback interaction through preset transformation rules, such as fully connected layer operations and nonlinear transformations, ultimately converting it into a quantifiable numerical value that can be used for subsequent loss value calculation. This numerical value is the first negative sample prediction value corresponding to that negative feedback interaction representation. The first negative sample prediction value is the numerical result obtained after performing representation mapping on other feedback interaction information, such as the probability value or similarity score in the interval [0, 1]. The magnitude of the first negative sample prediction value is used to characterize the degree of model's prediction fit to the other feedback interaction information. For example, the lower the first negative sample prediction value, the higher the confidence level of the model in determining that the interaction does not conform to the user's preferences.

[0130] Taking a training sample as an example, the formula for calculating the first loss value can be as follows:

[0131] Among them, L1(X + ,X - ) represents the first loss value, and n represents the positive feedback interaction sequence X. + The number of positive feedback interaction messages. For positive feedback interaction sequence X + The positive feedback interaction information corresponding to the i-th positive feedback interaction information is represented as follows: Representation of positive feedback interaction information This represents other feedback interaction information corresponding to other sample objects. σ(·) is the Sigmoid function, used to further represent the mapping and obtain the sample prediction value. This is the predicted value of the first positive sample, which is the probability that the item corresponding to the i-th positive feedback interaction information will be interacted with by the object. τ is the predicted value of the first negative sample, which is the probability that the item corresponding to the i-th negative feedback interaction information will be interacted with by the object. τ is the temperature parameter in the formula for calculating the first loss value, which is used to adjust the smoothness of the probability distribution after the Sigmoid function output value is exponentially operated on, thereby affecting the distinguishability of positive and negative sample prediction values ​​in loss calculation. For example, the smaller τ is, the more significant the difference in probability between positive and negative sample prediction values, and the larger τ is, the smoother the probability distribution.

[0132] First loss value L1(X) + ,X - ) and the predicted value of the first positive sample The correlation is negative, and the first loss value is L1(X). + ,X - ) and the predicted value of the first negative sample There is a positive correlation. This is because, when optimizing the first loss value, the goal is to make the first positive sample prediction value greater than any first negative sample prediction value. That is, as the first loss value decreases (i.e., as the optimization result improves), the first positive sample prediction value increases (i.e., the probability of correctly predicting the next item corresponding to the positive feedback interaction information is higher), while the first negative sample prediction value decreases (i.e., the probability of incorrectly predicting the next item corresponding to the positive feedback interaction information is lower). In this way, during the iterative training process, the positive feedback interaction information generated by the information prediction model can increasingly accurately represent the interest characteristics of the sample object in the positive feedback interaction scenario. For example, the sample object likes to watch recommended sports short videos between 5 pm and 6 pm on Fridays. τ is a temperature coefficient, a hyperparameter. The value of τ is greater than 0 and less than 1. It is used to adjust the scaling range of the first positive sample prediction value and the first negative sample prediction value, balance the weights of positive and negative samples, and ensure that the model performance and convergence speed are in optimal balance.

[0133] Here's an example of how to determine the representation of negative feedback interaction information: Suppose a given training batch B includes 100 training samples. Each training sample consists of a positive feedback interaction sequence and a negative feedback interaction sequence corresponding to a sample object. For example, the second training sample includes the positive feedback interaction sequence corresponding to sample object 2. and negative feedback interaction sequence In training batch B, the interaction information representation corresponding to the positive feedback interaction sequences or negative feedback interaction sequences of the remaining 98 training samples, excluding the second training sample, is the positive feedback interaction sequence. Negative samples.

[0134] Step 140: Obtain the target time information corresponding to each negative feedback interaction information representation, and determine the target positive feedback interaction information representation of the sample object corresponding to each negative feedback interaction information representation based on the target time information.

[0135] To enable the information prediction model to learn the reasons for the negative feedback interaction patterns of an object, the negative feedback interaction information can be used as a second negative sample, and the corresponding positive feedback interaction information of the target at the same time can be used as a second positive sample. By increasing the difference between the second positive sample and the second negative sample, the information prediction model's ability to distinguish between positive and negative feedback interaction patterns is enhanced, thereby enabling the information prediction model to learn the reasons for the negative feedback interaction patterns of an object.

[0136] It should be noted that target time information refers to structured time data that is uniquely bound to a single negative feedback interaction information representation. It can be a precise timestamp of the occurrence of the negative feedback interaction or a standardized time period code. The format of the target time information is completely consistent with the time information of the positive feedback interaction.

[0137] As one implementation method, based on the negative feedback interaction information corresponding to each negative feedback interaction information representation, the target time information in each negative feedback interaction information is obtained. For each target time information, positive feedback interaction information whose time information matches the target time information is selected from the positive feedback interaction sequence corresponding to the sample object. The positive feedback interaction information whose time information matches the target time information is determined as the target positive feedback interaction information corresponding to the negative feedback interaction representation. The positive feedback interaction information representation corresponding to the target positive feedback interaction information is used as the target positive feedback interaction information representation corresponding to the negative feedback interaction information representation.

[0138] It should be noted that selecting positive feedback interaction information that matches the target time information means filtering out positive feedback interaction information whose time information meets preset timing conditions from the positive feedback interaction sequence corresponding to the same sample object that generated the negative feedback interaction information. The preset timing conditions may include that the time information is completely consistent with the target time information, or that the time information is within the preset time window of the target time information.

[0139] Step 150: Determine the second loss value based on the difference between the second negative sample prediction value calculated from each negative feedback interaction information representation and the second positive sample prediction value calculated from the corresponding target positive feedback interaction information representation.

[0140] As one implementation method, a representation mapping is performed on each negative feedback interaction information representation to obtain the second negative sample prediction value corresponding to each negative feedback interaction information representation. Similarly, a representation mapping is performed on the target positive feedback interaction information representation corresponding to each negative feedback interaction information representation to obtain the second positive sample prediction value corresponding to each target positive feedback interaction information representation. Then, the difference between each second positive sample prediction value and its corresponding second negative sample prediction value is statistically analyzed to determine the second loss value. Taking a training sample as an example, the formula for calculating the second loss value can be as follows:

[0141] Among them, L2(X + ,X - ) represents the second loss value, and n represents the negative feedback interaction sequence X. - The number of negative feedback interaction messages. For negative feedback interaction sequence X - The negative feedback interaction information corresponding to the i-th negative feedback interaction information is represented as follows: Representation of negative feedback interaction information The corresponding target positive feedback interaction information is represented. This is the predicted value of the second positive sample. This is the predicted value for the second negative sample. The second loss value is L2(X). + ,X - The second loss value L2(X) is negatively correlated with the predicted value of the second positive sample. + ,X - ) and the predicted value of the second negative sample They are positively correlated.

[0142] The principle behind calculating the second loss value is as follows: it is assumed that the sample object is more interested in items in positive feedback interaction scenarios (positive samples) than in items in negative feedback interaction scenarios (negative samples). Therefore, when optimizing the second loss value, it is hoped that the predicted value of the second positive sample can be greater than the predicted value of the second negative sample. This will maximize the difference between positive and negative samples, thereby improving the information prediction model's ability to capture the interest boundary of the sample object in items in positive and negative feedback interaction scenarios. In other words, it determines which interaction information interests the sample object and which interaction information makes the sample object feel uninterested.

[0143] Therefore, as the second loss value decreases (i.e., the better the optimization result), the second positive sample prediction value increases (i.e., the probability of predicting items in positive feedback interaction scenarios is higher), while the second negative sample prediction value decreases (i.e., the probability of predicting items in negative feedback interaction scenarios is lower). Thus, in the iterative training process, the information prediction model not only enables the generated positive feedback interaction information representation to increasingly accurately characterize the interest features of the sample object in positive feedback interaction scenarios, but also enables the generated negative feedback interaction information representation to increasingly accurately characterize the non-interest features of the sample object in negative feedback interaction scenarios. For example, the sample object may not be interested in a text-based programming tutorial tweet pushed between 10 AM and 12 PM on Monday.

[0144] Step 160: Combine the first loss value and the second loss value to train the information prediction model and obtain the trained information prediction model.

[0145] Upon obtaining the first and second loss values, the information prediction model can be iteratively trained based on the sum of the first and second loss values ​​until the training termination condition is met, resulting in the trained information prediction model. In each iteration of the training process, the information prediction model can process a batch of training samples in parallel. After forward propagation, it calculates the positive feedback interaction information representation under the corresponding positive feedback interaction mode and the negative feedback interaction information representation under the corresponding negative feedback interaction mode. The first and second loss values ​​are then calculated from these representations. Finally, the model parameters of the information prediction model are updated through backpropagation, simultaneously optimizing the first and second loss values.

[0146] In practice, the first weight corresponding to the first loss value and the second weight corresponding to the second loss value can be obtained. Based on the first and second weights, the first and second loss values ​​are weighted and summed to obtain the total loss value. Based on the total loss value, the information prediction model is iteratively trained until the training termination condition is met, resulting in the trained information prediction model. The first and second weights are preset.

[0147] It should be noted that the total loss value is obtained by weighted summation of the first loss value and the second loss value based on the first weight and the second weight, which can be achieved in the following way: multiply the first weight by the first loss value to obtain the first product; multiply the second weight by the second loss value to obtain the second product; add the first product and the second product to obtain the total loss value.

[0148] By training the information prediction model using a combination of first and second loss values, the model can learn not only the preferences of sample objects but also their interest boundaries. Compared to push scenarios that only focus on positive feedback interactions of sample objects, this embodiment allows the information prediction model to learn the evolution of sample objects' interests under different interaction modes by simultaneously focusing on both positive and negative feedback interaction modes. In other words, the information prediction model can learn both what sample objects like and dislike, resulting in a comprehensive understanding of sample object preferences and enabling it to more accurately capture potential interests, thereby effectively improving the recommendation accuracy of downstream push tasks.

[0149] The training termination condition can be: the sum of the first loss value and the second loss value is less than a preset value or falls within a threshold range, or the number of iterations for iterative training reaches a preset number, etc. Optionally, an optimizer can be used to optimize the first loss value and the second loss value, and the learning rate, batch size, and epoch can be set based on experimental experience. The technical solution of this disclosure embodiment can be implemented based on a deep learning framework, such as PyTorch. The optimizer can use AdamW, etc. At the hardware level, the computer device can use a high-performance graphics processing unit.

[0150] One embodiment of this disclosure provides a method for selecting push notifications. This method can be executed by a computer device (e.g., a desktop computer, laptop computer, or other terminal device with certain processing capabilities) after step 160 of the above embodiment. The computer device must at least have storage, computing, and communication functions. Please refer to Figure 12, which shows a flowchart illustrating a push notification selection method provided by an embodiment of this disclosure. The steps of this push notification selection method will be described below with reference to Figure 12:

[0151] Step 210: Obtain the target positive feedback interaction sequence and the target negative feedback interaction sequence corresponding to the target object.

[0152] In this embodiment of the disclosure, the target positive feedback interaction sequence is a sequence composed of multiple selected positive feedback interaction information in chronological order, and the target negative feedback interaction sequence is a sequence composed of multiple selected negative feedback interaction information in chronological order. Selected positive feedback interaction information refers to the attribute information of items that have had positive feedback interactions with the object, and selected negative feedback interaction information refers to the attribute information of items that have had negative feedback interactions with the object.

[0153] It should be noted that attribute information refers to the key characteristic data of items that have positive or negative feedback interactions with objects. This includes both basic attributes of the items, such as the product's identifier, category, and price, and the article's theme and keywords, as well as interaction scenario attributes, such as the interaction time and electronic device.

[0154] In practical recommendation scenarios, the target positive feedback interaction sequence and the target negative feedback interaction sequence are the historical interaction sequences of the target object. Using these historical interaction sequences, a representation of the target object can be generated first. Then, the similarity between this representation and the representations of each piece of information to be pushed is calculated, and the final target push information is determined based on the similarity calculation results. The information to be pushed refers to the items to be recommended in the push scenario, such as products, articles, and videos.

[0155] As one implementation method, the target positive feedback interaction sequence and target negative feedback interaction sequence corresponding to the target object are obtained according to the push cycle. Here, the push cycle refers to a time period for executing the push scenario, such as a push cycle in units of months, weeks, or days. For example, the target positive feedback interaction sequence and target negative feedback interaction sequence for Monday can be obtained for the push scenario on Tuesday. The method for obtaining the target positive feedback interaction sequence and target negative feedback interaction sequence can refer to step 110 of the above embodiment, and will not be repeated here.

[0156] Step 220: Input the target positive feedback interaction sequence and the target negative feedback interaction sequence into the trained information prediction model for object representation processing to obtain the positive feedback interaction object representation and the negative feedback interaction object representation corresponding to the target object.

[0157] Since the target positive feedback interaction sequence contains multiple selected positive feedback interaction information of the target object in continuous time, and the target negative feedback interaction sequence contains multiple selected negative feedback interaction information of the target object in continuous time, the target positive feedback interaction sequence and the target negative feedback interaction sequence can accurately reflect the interest evolution law of the target object under different interaction modes. Therefore, the target object representation, that is, the positive feedback interaction object representation and the negative feedback interaction object representation, can be generated based on the target positive feedback interaction sequence and the target negative feedback interaction sequence.

[0158] As one implementation method, the target positive feedback interaction sequence is input into the trained information prediction model for information prediction, and the selected positive feedback interaction information representation corresponding to each selected positive feedback interaction information in the target positive feedback interaction sequence is obtained. The representation of each selected positive feedback interaction information representation is fused to obtain the positive feedback interaction object representation corresponding to the target object.

[0159] As one implementation method, the target negative feedback interaction sequence is input into the trained information prediction model for information prediction, and the selected negative feedback interaction information representation corresponding to each selected negative feedback interaction information in the target negative feedback interaction sequence is obtained. The representation fusion of each selected negative feedback interaction information representation is performed to obtain the negative feedback interaction object representation corresponding to the target object.

[0160] Taking the positive feedback interaction object representation as an example, the process of generating the selected positive feedback interaction information representation can refer to step 120 of the above embodiment, and will not be repeated here. After obtaining the selected positive feedback interaction information representation, each selected positive feedback interaction information representation can be fused. For example, a summation operation can be performed on each selected positive feedback interaction information representation to obtain the positive feedback interaction object representation corresponding to the target object. Similarly, the negative feedback interaction object representation corresponding to the target object can also be obtained.

[0161] In this way, by processing and fusing the target positive feedback interaction sequence and the target negative feedback interaction sequence separately, redundant calculations and data interference during mixed processing are avoided. At the same time, the condensed positive feedback interaction object representation and negative feedback interaction object representation significantly reduce the data volume in subsequent matching decisions, reduce device memory usage and computational overhead, and effectively improve the resource utilization efficiency of electronic devices.

[0162] Step 230: Based on the trained information prediction model, perform information representation processing on each candidate push information at the current time to obtain the candidate push information representation corresponding to each candidate push information.

[0163] Typically, push notification scenarios may present multiple candidate push notifications at a time, and then select the final target push notification from these candidates. To do this, a representation of each candidate push notification is calculated so that similarity can be subsequently calculated between it and the object representation corresponding to the target object.

[0164] It should be noted that, based on the trained information prediction model, information representation processing is performed on each candidate push information at the current time to obtain the candidate push information representation corresponding to each candidate push information. This can be achieved in the following way: From each candidate push information at the current time, the original identification information and content information of the candidate push information are extracted, and the time information is extracted in combination with the current time; feature extraction is performed on the identification information to obtain the identification feature representation, on the content information to obtain the content feature representation, and on the time information to obtain the time feature representation; the identification feature representation, the time feature representation, and the content feature representation are fused to obtain the fused standardized feature vector, which is the candidate push information representation.

[0165] As one implementation method, the current time is structured to obtain the corresponding target time information, and the target identifier information and target content information corresponding to each candidate push information at the current time are obtained. The target time information, target identifier information, and target content information corresponding to each candidate push information are input into the trained information prediction model for feature extraction to obtain the candidate push information representation corresponding to each candidate push information. The input of the target time information, target identifier information, and target content information corresponding to each candidate push information into the trained information prediction model for feature extraction can refer to step 120 of the above embodiment, and will not be repeated here.

[0166] Optionally, several time templates can be set according to the time requirements of the push scenario. The number of time templates is determined by the push cycle. For example, if the push cycle is based on weeks, 7 time templates can be set. If the push cycle is based on days, 24 time templates can be set. Then, information representation processing is performed on each candidate push message corresponding to the current time for each time template to obtain the candidate push message representation for each candidate push message. The candidate push message representation for each candidate push message is then cached in the candidate push message vector database.

[0167] Step 240: Calculate the similarity between each candidate push information representation and the positive feedback interaction object representation and the negative feedback interaction object representation respectively, to obtain the first similarity and the second similarity corresponding to each candidate push information.

[0168] As one implementation method, a similarity calculation is performed between each candidate push information representation and the positive feedback interaction object representation to obtain a first similarity for each candidate push information. Then, a similarity calculation is performed between each candidate push information representation and the negative feedback interaction object representation to obtain a second similarity for each candidate push information. The similarity calculation can be cosine similarity, etc., and is not limited here.

[0169] Step 250: Based on the difference between the first similarity and the second similarity corresponding to each candidate push information, obtain the similarity score corresponding to each candidate push information.

[0170] As one implementation method, the first similarity score for each candidate push message is subtracted from the second similarity score, resulting in a similarity difference for each candidate push message. This similarity difference is then used as the corresponding similarity score. Alternatively, the first similarity score for each candidate push message can be directly used as the corresponding similarity score.

[0171] For example, the positive feedback interaction object of the user corresponds to positive preference features, and the negative feedback interaction object corresponds to negative rejection features: the first similarity between the candidate push information A and the positive feedback interaction object is 0.9, and the second similarity between A and the negative feedback interaction object is 0.15. The similarity difference is 0.75, and 0.75 is the similarity score.

[0172] Step 260: Determine the candidate push information corresponding to the highest similarity score as the target push information.

[0173] As one implementation method, the similarity scores corresponding to each candidate push information are sorted in descending order to obtain a similarity score ranking. The candidate push information with the highest similarity score is selected from the similarity score ranking, and then the candidate push information corresponding to the highest similarity score is determined as the target push information.

[0174] By applying the above embodiments of this application, the positive feedback interaction sequence and negative feedback interaction sequence of the target object are condensed into a fixed-dimensional positive feedback interaction object representation and a negative feedback interaction object representation, which greatly reduces the volume of the original interaction data. Moreover, the selection of candidate push information can be completed by only two similarity calculations and difference operations, avoiding complex multi-dimensional data processing and redundant calculations, reducing the memory occupation and computing consumption of electronic devices, and thus effectively improving the resource utilization of electronic devices.

[0175] Please refer to Figure 13, which is a block diagram of a training apparatus for an information prediction model provided in an embodiment of this disclosure. The training apparatus 300 for the information prediction model may include:

[0176] The sample sequence acquisition module 310 is configured to acquire the positive feedback interaction sequence and the negative feedback interaction sequence corresponding to the sample object, wherein the positive feedback interaction sequence contains at least two positive feedback interaction information and the negative feedback interaction sequence contains at least two negative feedback interaction information.

[0177] The information prediction module 320 is configured to input the positive feedback interaction sequence and the negative feedback interaction sequence into the information prediction model for information prediction, thereby obtaining the positive feedback interaction information representation corresponding to each positive feedback interaction information in the positive feedback interaction sequence and the negative feedback interaction information representation corresponding to each negative feedback interaction information in the negative feedback interaction sequence.

[0178] The first loss determination module 330 is configured to obtain other feedback interaction information of other sample objects corresponding to each positive feedback interaction information, and determine the first loss value based on the difference between the first positive sample prediction value calculated based on each positive feedback interaction information and the first negative sample prediction value calculated based on the other feedback interaction information of the corresponding other feedback interaction information.

[0179] The target representation acquisition module 340 is configured to acquire the current time information corresponding to each negative feedback interaction information representation, and determine the target positive feedback interaction information representation of the sample object corresponding to each negative feedback interaction information representation based on the current time information;

[0180] The second loss determination module 350 is configured to determine a second loss value based on the difference between the second negative sample prediction value calculated based on each of the negative feedback interaction information representations and the second positive sample prediction value calculated based on the corresponding target positive feedback interaction information representations.

[0181] The model joint training module 360 ​​is configured to train the information prediction model by combining the first loss value and the second loss value, thereby obtaining the trained information prediction model.

[0182] In some embodiments, the first loss determination module 330 may include:

[0183] The information processing submodule is configured to input the other feedback interaction information corresponding to each positive feedback interaction information representation into the information prediction model to perform information prediction, thereby obtaining the other feedback interaction information representation corresponding to each other feedback interaction information.

[0184] The first mapping submodule is configured to perform representation mapping on each of the positive feedback interaction information representations to obtain the first positive sample prediction value corresponding to each of the positive feedback interaction information representations;

[0185] The second mapping submodule is configured to perform representation mapping on each other feedback interaction information representation to obtain the first negative sample prediction value corresponding to each other feedback interaction information representation;

[0186] The loss determination submodule is configured to calculate the difference between each predicted value of the first positive sample and the corresponding predicted value of the first negative sample in order to determine the first loss value.

[0187] Wherein, the first loss value is negatively correlated with the first positive sample prediction value, and the first loss value is positively correlated with the first negative sample prediction value.

[0188] In some embodiments, the first loss determination module 330 may further include:

[0189] Other feedback submodules are configured to obtain other positive feedback interaction sequences and other negative feedback interaction sequences corresponding to other sample objects;

[0190] The information sampling submodule is configured to randomly sample the other positive feedback interaction sequences and the other negative feedback interaction sequences to obtain a preset number of feedback interaction information as other feedback interaction information.

[0191] The number of preset information is the same as the number of positive feedback interaction information contained in the positive feedback interaction sequence, and each other feedback interaction information is different from any positive feedback interaction information in the positive feedback interaction sequence.

[0192] In some embodiments, the second loss determination module 350 is further configured to:

[0193] A representation mapping is performed on each of the negative feedback interaction information representations to obtain the second negative sample prediction value corresponding to each of the negative feedback interaction information representations;

[0194] A representation mapping is performed on the target positive feedback interaction information representation corresponding to each negative feedback interaction information representation to obtain the second positive sample prediction value corresponding to each target positive feedback interaction information representation;

[0195] The difference between each predicted value of the second positive sample and the corresponding predicted value of the second negative sample is calculated to determine the second loss value;

[0196] The second loss value is negatively correlated with the second positive sample prediction value, and the second loss value is positively correlated with the second negative sample prediction value.

[0197] In some embodiments, the target representation acquisition module 340 is further configured to:

[0198] Based on the negative feedback interaction information corresponding to each negative feedback interaction information, obtain the target time information in each negative feedback interaction information;

[0199] For each target time information, positive feedback interaction information that matches the target time information is selected from the positive feedback interaction sequence corresponding to the sample object, and the positive feedback interaction information that matches the target time information is determined as the target positive feedback interaction information corresponding to the negative feedback interaction representation.

[0200] The positive feedback interaction information corresponding to the target positive feedback interaction information is used as the target positive feedback interaction information corresponding to the negative feedback interaction information.

[0201] In some embodiments, the information prediction module 320 may include:

[0202] The first input submodule is configured to input the positive feedback interaction sequence into the information prediction model;

[0203] The first extraction submodule is configured to perform feature extraction on the positive feedback interaction sequence based on the information prediction model, and obtain the positive feedback interaction information feature representation corresponding to each positive feedback interaction information in the positive feedback interaction sequence.

[0204] The first fusion submodule is configured to perform context fusion on the feature representation of each positive feedback interaction information to obtain the positive feedback interaction information representation corresponding to each positive feedback interaction information.

[0205] The second input submodule is configured to input the negative feedback interaction sequence into the information prediction model;

[0206] The second extraction submodule is configured to perform feature extraction on the negative feedback interaction sequence based on the information prediction model, and obtain the negative feedback interaction information feature representation corresponding to each negative feedback interaction information in the negative feedback interaction sequence.

[0207] The second fusion submodule is configured to perform context fusion on the feature representation of each negative feedback interaction information to obtain the negative feedback interaction information representation corresponding to each negative feedback interaction information.

[0208] In some embodiments, the first extraction submodule may include:

[0209] The identifier extraction unit is configured to extract features from the identifier information of each positive feedback interaction information in the positive feedback interaction sequence to obtain the identifier feature representation corresponding to each positive feedback interaction information.

[0210] The time extraction unit is configured to extract features from the time information of each positive feedback interaction information in the positive feedback interaction sequence to obtain the time feature representation corresponding to each positive feedback interaction information.

[0211] The content extraction unit is configured to extract features from the content information of each positive feedback interaction information in the positive feedback interaction sequence to obtain the content feature representation corresponding to each positive feedback interaction information.

[0212] The representation fusion unit is configured to perform representation fusion on the identifier feature representation, time feature representation and content feature representation corresponding to each positive feedback interaction information to obtain the positive feedback interaction information feature representation corresponding to each positive feedback interaction information.

[0213] In some embodiments, the fusion unit is further configured as follows:

[0214] The identifier feature representation and content feature representation corresponding to each positive feedback interaction information are concatenated to obtain the first concatenated feature representation corresponding to each positive feedback interaction information.

[0215] Each first concatenated feature representation is aligned to obtain the aligned feature representation corresponding to each positive feedback interaction information.

[0216] The alignment feature representation and temporal feature representation corresponding to each positive feedback interaction information are concatenated to obtain the second concatenated feature representation corresponding to each positive feedback interaction information;

[0217] Each second concatenated feature representation is aligned to obtain the positive feedback interaction information feature representation corresponding to each positive feedback interaction information.

[0218] In some embodiments, the first fusion submodule is further configured as follows:

[0219] Based on the time information corresponding to each positive feedback interaction information, a sequence of positive feedback interaction information feature representations corresponding to each positive feedback interaction information in the positive feedback interaction sequence is constructed to obtain a positive feedback interaction information feature representation sequence.

[0220] Context fusion is performed on the sequence of positive feedback interaction information features to obtain the positive feedback interaction information representation corresponding to each positive feedback interaction information.

[0221] In some embodiments, the sample sequence acquisition module 310 may include:

[0222] The record acquisition submodule is configured to acquire historical positive feedback interaction records and historical negative feedback interaction records corresponding to the sample object;

[0223] The first construction submodule is configured to construct a sequence based on the historical positive feedback interaction records to obtain the original positive feedback interaction sequence corresponding to the sample object.

[0224] The second construction submodule is configured to construct a sequence based on the historical negative feedback interaction records to obtain the original negative feedback interaction sequence corresponding to the sample object.

[0225] The first molecular module is configured to perform sequence segmentation on the original positive feedback interaction sequence to obtain a positive feedback interaction sequence of a preset length;

[0226] The second slicing module is configured to perform sequence segmentation on the original negative feedback interaction sequence to obtain a negative feedback interaction sequence of a preset length.

[0227] In some embodiments, the historical positive feedback interaction record includes descriptions of multiple historical positive feedback interaction information in a positive feedback interaction scenario, provides source identifiers and interaction times, and the first construction submodule is further configured to:

[0228] The descriptive content corresponding to each historical positive feedback interaction is segmented into words to obtain the content information corresponding to each historical positive feedback interaction. The descriptive content is used to describe the information theme and information modality of the interaction information.

[0229] The interaction time corresponding to each historical positive feedback interaction is structured to obtain the time information corresponding to each historical positive feedback interaction, and the time information includes at least the day-hour dimension;

[0230] Use the source identifier corresponding to each historical positive feedback interaction as the corresponding identifier information;

[0231] The time information, the identification information, and the content information of each historical positive feedback interaction are sequenced according to the time dimension to obtain the original positive feedback interaction sequence corresponding to the sample object.

[0232] This embodiment can acquire the positive feedback interaction sequence corresponding to the sample object, and also acquire the negative feedback interaction sequence as training data for the information prediction model. The positive feedback interaction sequence contains at least two positive feedback interaction pieces, and the negative feedback interaction sequence contains at least two negative feedback interaction pieces. In some embodiments, to enable the information prediction model to learn the cross-scene interest changes of the sample object in both positive and negative feedback interaction scenarios during training, the positive and negative feedback interaction sequences are input into the information prediction model. The information prediction model then predicts the information from the positive and negative feedback interaction sequences to obtain corresponding positive and negative feedback interaction representations. Based on these representations, a loss value for training the information prediction model can be calculated. In some embodiments, this embodiment acquires other feedback interaction information corresponding to other sample objects for each positive feedback interaction piece, and determines a first loss value by comparing the difference between the first positive sample prediction value calculated based on the representation of each positive feedback interaction piece and the first negative sample prediction value calculated based on the other feedback interaction representations of the corresponding positive feedback interaction pieces. To optimize the first loss value, the information prediction model can learn the interaction preferences of sample objects in positive feedback interaction scenarios. Simultaneously, the target time information corresponding to each negative feedback interaction information representation is obtained, and based on the target time information, the target positive feedback interaction information representation for each negative feedback interaction information representation is determined. Then, the difference between the second negative sample prediction value calculated for each negative feedback interaction information representation and the second positive sample prediction value calculated for the corresponding target positive feedback interaction information representation is used to determine the second loss value. Optimizing the second loss value allows the information prediction model to learn the interest boundaries of sample objects in both positive and negative feedback interaction scenarios; that is, what scenarios will trigger negative feedback interactions, and what scenarios are more likely to trigger positive feedback interactions. Thus, by training the information prediction model using both the first and second loss values, the model can learn not only the preferences of sample objects but also their interest boundaries. Compared to push scenarios that only focus on positive feedback interactions, this implementation allows the information prediction model to learn the evolution of sample objects' interests under different interaction scenarios by simultaneously considering both positive and negative feedback interactions. In other words, the information prediction model can learn both what sample objects like and dislike, making its understanding of sample objects' interaction preferences more comprehensive and enabling it to more accurately capture potential interests, thereby effectively improving the recommendation accuracy of downstream push tasks.

[0233] Please refer to Figure 14, which is a block diagram of a question-answering device provided in an embodiment of this disclosure. The push information selection device 400 may include:

[0234] The target sequence acquisition module 410 is configured to acquire the target positive feedback interaction sequence and the target negative feedback interaction sequence corresponding to the target object;

[0235] The object representation generation module 420 is configured to input the target positive feedback interaction sequence and the target negative feedback interaction sequence into the trained information prediction model for object representation processing, so as to obtain the positive feedback interaction object representation and the negative feedback interaction object representation corresponding to the target object.

[0236] The information representation generation module 430 is configured to perform information representation processing on each candidate push information at the current time based on the trained information prediction model, and obtain the candidate push information representation corresponding to each candidate push information.

[0237] The similarity calculation module 440 is configured to perform similarity calculation between each candidate push information representation and the positive feedback interaction object representation to obtain a first similarity corresponding to each candidate push information, and to perform similarity calculation between each candidate push information representation and the negative feedback interaction object representation to obtain a second similarity corresponding to each candidate push information.

[0238] The similarity difference calculation module 450 is configured to obtain a similarity score for each candidate push information based on the difference between the first similarity and the second similarity corresponding to each candidate push information.

[0239] The push information determination module 460 is configured to determine the candidate push information corresponding to the highest similarity score as the target push information;

[0240] The trained information prediction model is obtained by training the information prediction model according to the training method described in the above embodiments.

[0241] In some embodiments, the object representation generation module 420 is further configured to:

[0242] The target positive feedback interaction sequence is input into the trained information prediction model to predict information, thereby obtaining the selected positive feedback interaction information representation corresponding to each selected positive feedback interaction information in the target positive feedback interaction sequence.

[0243] The representation of each selected positive feedback interaction information is fused to obtain the positive feedback interaction object representation corresponding to the target object;

[0244] The target negative feedback interaction sequence is input into the trained information prediction model to predict information, thereby obtaining the selected negative feedback interaction information representation corresponding to each selected negative feedback interaction information in the target negative feedback interaction sequence.

[0245] The representation of each selected negative feedback interaction information is fused to obtain the negative feedback interaction object representation corresponding to the target object.

[0246] In some embodiments, the information representation generation module 430 is further configured to:

[0247] The current time is structured to obtain the corresponding current time information;

[0248] Obtain the target identifier information and target content information corresponding to each candidate push information at the current time;

[0249] The current time information, target identifier information, and target content information corresponding to each candidate push information are input into the trained information prediction model for feature extraction, so as to obtain the candidate push information representation corresponding to each candidate push information.

[0250] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the working process of the above-described devices, modules, etc., can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0251] In the several embodiments provided in this disclosure, the coupling between modules can be electrical, mechanical, or other forms of coupling.

[0252] Furthermore, the functional modules in the various embodiments of this disclosure can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0253] As shown in Figure 15, which is a block diagram of a computer device according to an embodiment of this disclosure, the computer device 500 can be in various forms such as a desktop computer or a laptop computer. The computer device 500 includes a processor 510, a memory 520, a power supply 530, and an input unit 540. The memory 520 stores a computer program, which, when called by the processor 510, can execute the various method steps provided in the above embodiments. Those skilled in the art will understand that the structure of the computer device shown in the figures does not constitute a limitation on the computer device; it may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0254] Processor 510 may include one or more processing cores. Processor 510 connects to various parts of the entire battery management system using various interfaces and lines. It executes instructions, programs, instruction sets, or program sets stored in memory 520, calls data stored in memory 520, performs various functions and processes data within the battery management system, and performs various functions and processes data within the computer device, thereby providing overall control of the computer device. Optionally, processor 510 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). Processor 510 may integrate one or more of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem handles wireless communication. It is understood that the modem may also not be integrated into processor 510 and may be implemented separately using a communication chip.

[0255] The memory 520 may include random access memory (RAM) or read-only memory (ROM). The memory 520 may be configured to store instructions, programs, instruction sets, or program assemblies. The memory 520 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for implementing at least one function (such as touch functionality, sound playback functionality, image playback functionality, etc.), and instructions for implementing the various method embodiments described above. The data storage area may also store data created during the use of the computer device (such as phonebook and audio / video data). Accordingly, the memory 520 may also include a memory controller to provide the processor 510 with access to the memory 520.

[0256] The power supply 530 can be logically connected to the processor 510 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 530 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0257] Input unit 540 is configured to receive input digital or character information and generate keyboard, mouse, joystick, optical, or trackball signal inputs related to object settings and function control.

[0258] Although not shown, the computer device 500 may also include an output unit, a display unit, etc., which will not be described in detail here. In the embodiments of this disclosure, the processor 510 of the computer device loads the executable files corresponding to the processes of one or more computer programs into the memory 520 according to the following instructions, and the processor 510 runs the data stored in the memory 520, such as telephone books and audio / video data, thereby implementing the various method steps provided in the foregoing embodiments.

[0259] As shown in Figure 16, which is a block diagram of a computer-readable storage medium provided in an embodiment of the present disclosure, the computer-readable storage medium 600 stores a computer program 610, which can be called by the processor of a computer device to execute various method steps provided in the embodiment of the present disclosure.

[0260] Computer-readable storage media can be electronic storage devices such as flash memory, electrically erasable programmable read-only memory (EEPROM), hard disks, or ROM. Optionally, the computer-readable storage medium can also be a non-volatile computer-readable storage medium. The computer-readable storage medium 600 has storage space for a computer program that performs any of the method steps in the above embodiments. These computer programs can be read from or written to one or more computer program products. The computer programs can be compressed in an appropriate form.

[0261] According to one aspect of the present disclosure, a computer program product is provided, comprising a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the various method steps provided in the above embodiments.

[0262] The terms "comprising" and "including," and any variations thereof, in this disclosure and the foregoing drawings are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or apparatus. "At least one" means one or more, and "more than" means two or more. "Greater than," "less than," "exceeding," etc., are understood to exclude the number itself. "Above," "below," "within," etc., are understood to include the number itself. 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 objective, 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) may be used to implement one or more modules or units. Furthermore, each module or unit may be part of a larger module or unit that includes the functionality of that module or unit.

[0263] In the several embodiments provided in this disclosure, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0264] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units.

[0265] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this disclosure, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a program product. This computer program product is stored in a storage medium and includes several computer programs to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the various embodiments of this disclosure. It should also be understood that the various implementation methods provided by the embodiments of this disclosure can be arbitrarily combined to achieve different technical effects.

[0266] The above are merely preferred embodiments of this disclosure and are not intended to limit this disclosure in any way. Although this disclosure has been disclosed above with reference to preferred embodiments, it is not intended to limit this disclosure. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the technical solution of this disclosure. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of this disclosure without departing from the content of the technical solution of this disclosure shall still fall within the scope of the technical solution of this disclosure.

Claims

1. A method for training an information prediction model, the method being executed by a computer device, the method comprising: Obtain the positive feedback interaction sequence and negative feedback interaction sequence corresponding to the sample object, wherein the positive feedback interaction sequence contains at least two positive feedback interaction information and the negative feedback interaction sequence contains at least two negative feedback interaction information; The positive feedback interaction sequence and the negative feedback interaction sequence are respectively input into the information prediction model for information prediction, so as to obtain the positive feedback interaction information representation corresponding to each positive feedback interaction information in the positive feedback interaction sequence, and the negative feedback interaction information representation corresponding to each negative feedback interaction information in the negative feedback interaction sequence. Obtain other feedback interaction information of other sample objects corresponding to each positive feedback interaction information, and determine the first loss value based on the difference between the first positive sample prediction value calculated based on the representation of each positive feedback interaction information and the first negative sample prediction value calculated based on the representation of other feedback interaction information of the corresponding other feedback interaction information; Obtain the target time information corresponding to each negative feedback interaction information representation, and determine the target positive feedback interaction information representation of the sample object corresponding to each negative feedback interaction information representation based on the target time information; The second loss value is determined by the difference between the second negative sample prediction value calculated based on each of the negative feedback interaction information representations and the second positive sample prediction value calculated based on the corresponding target positive feedback interaction information representation. The information prediction model is trained by combining the first loss value and the second loss value to obtain the trained information prediction model.

2. The method according to claim 1, wherein, The step of determining the first loss value by comparing the first positive sample prediction value calculated based on each of the positive feedback interaction information representations with the first negative sample prediction values ​​calculated based on the other feedback interaction information representations of the corresponding other feedback interaction information includes: Each positive feedback interaction information is represented by other feedback interaction information, which is then input into the information prediction model to predict information, thereby obtaining other feedback interaction information representations corresponding to each other feedback interaction information. Perform a representation mapping on each of the positive feedback interaction information representations to obtain the first positive sample prediction value corresponding to each of the positive feedback interaction information representations; A representation mapping is performed on each other feedback interaction information representation to obtain the first negative sample prediction value corresponding to each other feedback interaction information representation; The difference between each predicted value of the first positive sample and the corresponding predicted value of the first negative sample is calculated to determine the first loss value; Wherein, the first loss value is negatively correlated with the first positive sample prediction value, and the first loss value is positively correlated with the first negative sample prediction value.

3. The method according to claim 2, wherein, The step of obtaining other feedback interaction information of other sample objects corresponding to each of the positive feedback interaction information includes: Obtain other positive feedback interaction sequences and other negative feedback interaction sequences corresponding to other sample objects; Randomly sample the other positive feedback interaction sequences and the other negative feedback interaction sequences to obtain a preset number of feedback interaction information as other feedback interaction information; The number of preset information is the same as the number of positive feedback interaction information contained in the positive feedback interaction sequence, and each other feedback interaction information is different from any positive feedback interaction information in the positive feedback interaction sequence.

4. The method according to any one of claims 1 to 3, wherein, The step of determining the second loss value by comparing the second negative sample prediction value calculated based on each of the negative feedback interaction information representations with the second positive sample prediction value calculated based on the corresponding target positive feedback interaction information representation includes: A representation mapping is performed on each of the negative feedback interaction information representations to obtain the second negative sample prediction value corresponding to each of the negative feedback interaction information representations; A representation mapping is performed on the target positive feedback interaction information representation corresponding to each negative feedback interaction information representation to obtain the second positive sample prediction value corresponding to each target positive feedback interaction information representation; The difference between each predicted value of the second positive sample and the corresponding predicted value of the second negative sample is calculated to determine the second loss value; The second loss value is negatively correlated with the second positive sample prediction value, and the second loss value is positively correlated with the second negative sample prediction value.

5. The method according to claim 4, wherein, The step of acquiring the target time information corresponding to each negative feedback interaction information representation, and determining the target positive feedback interaction information representation of the sample object corresponding to each negative feedback interaction information representation based on the target time information, includes: Based on the negative feedback interaction information corresponding to each negative feedback interaction information, obtain the target time information in each negative feedback interaction information; For each target time information, positive feedback interaction information that matches the target time information is selected from the positive feedback interaction sequence corresponding to the sample object, and the positive feedback interaction information that matches the target time information is determined as the target positive feedback interaction information corresponding to the negative feedback interaction representation. The positive feedback interaction information corresponding to the target positive feedback interaction information is used as the target positive feedback interaction information corresponding to the negative feedback interaction information.

6. The method according to any one of claims 1 to 5, wherein, The step of inputting the positive feedback interaction sequence and the negative feedback interaction sequence into the information prediction model for information prediction, respectively, to obtain the positive feedback interaction information representation corresponding to each positive feedback interaction information in the positive feedback interaction sequence, and the negative feedback interaction information representation corresponding to each negative feedback interaction information in the negative feedback interaction sequence, includes: The positive feedback interaction sequence is input into the information prediction model; Based on the information prediction model, feature extraction is performed on the positive feedback interaction sequence to obtain the positive feedback interaction information feature representation corresponding to each positive feedback interaction information in the positive feedback interaction sequence; Context fusion is performed on the feature representation of each positive feedback interaction information to obtain the positive feedback interaction information representation corresponding to each positive feedback interaction information. The negative feedback interaction sequence is input into the information prediction model; Based on the information prediction model, feature extraction is performed on the negative feedback interaction sequence to obtain the negative feedback interaction information feature representation corresponding to each negative feedback interaction information in the negative feedback interaction sequence. Context fusion is performed on the feature representation of each negative feedback interaction information to obtain the negative feedback interaction information representation corresponding to each negative feedback interaction information.

7. The method according to claim 6, wherein, The step of extracting features from the positive feedback interaction sequence to obtain the feature representation of the positive feedback interaction information corresponding to each positive feedback interaction information in the positive feedback interaction sequence includes: Feature extraction is performed on the identifier information of each positive feedback interaction information in the positive feedback interaction sequence to obtain the identifier feature representation corresponding to each positive feedback interaction information; The temporal information of each positive feedback interaction in the positive feedback interaction sequence is extracted to obtain the temporal feature representation of each positive feedback interaction. Feature extraction is performed on the content information of each positive feedback interaction in the positive feedback interaction sequence to obtain the content feature representation corresponding to each positive feedback interaction; The identifier feature representation, time feature representation, and content feature representation corresponding to each positive feedback interaction information are fused to obtain the positive feedback interaction information feature representation corresponding to each positive feedback interaction information.

8. The method according to claim 7, wherein, The process of fusing the identifier feature representation, time feature representation, and content feature representation corresponding to each positive feedback interaction information to obtain the positive feedback interaction information feature representation corresponding to each positive feedback interaction information includes: The identifier feature representation and content feature representation corresponding to each positive feedback interaction information are concatenated to obtain the first concatenated feature representation corresponding to each positive feedback interaction information. Each first concatenated feature representation is aligned to obtain the aligned feature representation corresponding to each positive feedback interaction information. The alignment feature representation and temporal feature representation corresponding to each positive feedback interaction information are concatenated to obtain the second concatenated feature representation corresponding to each positive feedback interaction information; Each second concatenated feature representation is aligned to obtain the positive feedback interaction information feature representation corresponding to each positive feedback interaction information.

9. The method according to claim 7 or 8, wherein, The step of performing context fusion on the feature representation of each positive feedback interaction information to obtain the positive feedback interaction information representation corresponding to each positive feedback interaction information includes: Based on the time information corresponding to each positive feedback interaction information, a sequence of positive feedback interaction information feature representations corresponding to each positive feedback interaction information in the positive feedback interaction sequence is constructed to obtain a positive feedback interaction information feature representation sequence. Context fusion is performed on the sequence of positive feedback interaction information features to obtain the positive feedback interaction information representation corresponding to each positive feedback interaction information.

10. The method according to claim 9, wherein, The step of performing context fusion on the sequence of positive feedback interaction information features to obtain the positive feedback interaction information representation corresponding to each positive feedback interaction information includes: The sequence of positive feedback interaction information features is input into a large language model based on the Transformer architecture. Based on the large language model, the positive feedback interaction information feature representation sequence is fused with context to obtain the positive feedback interaction information representation corresponding to each positive feedback interaction information.

11. The method according to any one of claims 1 to 10, wherein, The acquisition of the positive feedback interaction sequence and negative feedback interaction sequence corresponding to the sample object includes: Obtain the historical positive feedback interaction records and historical negative feedback interaction records corresponding to the sample object; Based on the historical positive feedback interaction records, a sequence is constructed to obtain the original positive feedback interaction sequence corresponding to the sample object; Based on the historical negative feedback interaction records, a sequence is constructed to obtain the original negative feedback interaction sequence corresponding to the sample object; The original positive feedback interaction sequence is segmented to obtain a positive feedback interaction sequence of a preset length; The original negative feedback interaction sequence is segmented to obtain a negative feedback interaction sequence of a preset length.

12. The method according to claim 11, wherein, The historical positive feedback interaction records include descriptions, source identifiers, and interaction times corresponding to multiple historical positive feedback interaction information in the positive feedback interaction scenario. The step of constructing a sequence based on the historical positive feedback interaction records to obtain the original positive feedback interaction sequence corresponding to the sample object includes: The descriptive content corresponding to each historical positive feedback interaction is segmented into words to obtain the content information corresponding to each historical positive feedback interaction. The descriptive content is used to describe the information theme and information modality of the interaction information. The interaction time corresponding to each historical positive feedback interaction is structured to obtain the time information corresponding to each historical positive feedback interaction, and the time information includes at least the day-hour dimension; Use the source identifier corresponding to each historical positive feedback interaction as the corresponding identifier information; The time information, the identification information, and the content information of each historical positive feedback interaction are sequenced according to the time dimension to obtain the original positive feedback interaction sequence corresponding to the sample object.

13. The method according to any one of claims 1 to 12, wherein, The information prediction model is trained by combining the first loss value and the second loss value to obtain the trained information prediction model, including: Obtain the first weight corresponding to the first loss value and the second weight corresponding to the second loss value; Based on the first weight and the second weight, the first loss value and the second loss value are weighted and summed to obtain the total loss value; Based on the total loss value, the information prediction model is trained to obtain the trained information prediction model.

14. A method for selecting push notifications, the method being executed by a computer device, the method comprising: Obtain the target positive feedback interaction sequence and the target negative feedback interaction sequence corresponding to the target object; The target positive feedback interaction sequence and the target negative feedback interaction sequence are respectively input into the trained information prediction model for object representation processing to obtain the positive feedback interaction object representation and the negative feedback interaction object representation corresponding to the target object. Based on the trained information prediction model, information representation processing is performed on each candidate push information at the current time to obtain the candidate push information representation corresponding to each candidate push information. The similarity between each candidate push information representation and the positive feedback interaction object representation is calculated to obtain the first similarity corresponding to each candidate push information. The similarity between each candidate push information representation and the negative feedback interaction object representation is calculated to obtain the second similarity corresponding to each candidate push information. Based on the difference between the first similarity and the second similarity corresponding to each candidate push information, the target similarity score corresponding to each candidate push information is obtained; The candidate push information corresponding to the highest target similarity score is determined as the target push information; The trained information prediction model is obtained by training the information prediction model according to any one of claims 1 to 11.

15. The method according to claim 14, wherein, The step of inputting the target positive feedback interaction sequence and the target negative feedback interaction sequence into the trained information prediction model for object representation processing to obtain the positive feedback interaction object representation and the negative feedback interaction object representation corresponding to the target object includes: The target positive feedback interaction sequence is input into the trained information prediction model to predict information, thereby obtaining the selected positive feedback interaction information representation corresponding to each selected positive feedback interaction information in the target positive feedback interaction sequence. The representation of each selected positive feedback interaction information is fused to obtain the positive feedback interaction object representation corresponding to the target object; The target negative feedback interaction sequence is input into the trained information prediction model to predict information, thereby obtaining the selected negative feedback interaction information representation corresponding to each selected negative feedback interaction information in the target negative feedback interaction sequence. The representation of each selected negative feedback interaction information is fused to obtain the negative feedback interaction object representation corresponding to the target object.

16. The method according to claim 14 or 15, wherein, The trained information prediction model performs information representation processing on each candidate push information at the current time to obtain the candidate push information representation corresponding to each candidate push information, including: The current time is structured to obtain the corresponding current time information; Obtain the target identifier information and target content information corresponding to each candidate push information at the current time; The current time information, target identifier information, and target content information corresponding to each candidate push information are input into the trained information prediction model for feature extraction, so as to obtain the candidate push information representation corresponding to each candidate push information.

17. A training device for an information prediction model, the device comprising: The sample sequence acquisition module is configured to acquire the positive feedback interaction sequence and the negative feedback interaction sequence corresponding to the sample object, wherein the positive feedback interaction sequence contains at least two positive feedback interaction information and the negative feedback interaction sequence contains at least two negative feedback interaction information. The information prediction module is configured to input the positive feedback interaction sequence and the negative feedback interaction sequence into the information prediction model to perform information prediction, thereby obtaining the positive feedback interaction information representation corresponding to each positive feedback interaction information in the positive feedback interaction sequence and the negative feedback interaction information representation corresponding to each negative feedback interaction information in the negative feedback interaction sequence. The first loss determination module is configured to obtain other feedback interaction information of other sample objects corresponding to each positive feedback interaction information, and determine the first loss value based on the difference between the first positive sample prediction value calculated based on the representation of each positive feedback interaction information and the first negative sample prediction value calculated based on the representation of other feedback interaction information of the corresponding other feedback interaction information. The target representation acquisition module is configured to acquire the current time information corresponding to each negative feedback interaction information representation, and determine the target positive feedback interaction information representation of the sample object corresponding to each negative feedback interaction information representation based on the current time information; The second loss determination module is configured to determine a second loss value based on the difference between the second negative sample prediction value calculated from each negative feedback interaction information representation and the second positive sample prediction value calculated from the corresponding target positive feedback interaction information representation. The model joint training module is configured to train the information prediction model by combining the first loss value and the second loss value, thereby obtaining the trained information prediction model.

18. A push notification selection device, the device comprising: The target sequence acquisition module is configured to acquire the target positive feedback interaction sequence and the target negative feedback interaction sequence corresponding to the target object. The object representation generation module is configured to input the target positive feedback interaction sequence and the target negative feedback interaction sequence into the trained information prediction model for object representation processing, so as to obtain the positive feedback interaction object representation and the negative feedback interaction object representation corresponding to the target object. The information representation generation module is configured to perform information representation processing on each candidate push information at the current time based on the trained information prediction model, and obtain the candidate push information representation corresponding to each candidate push information. The similarity calculation module is configured to perform similarity calculation between each candidate push information representation and the positive feedback interaction object representation to obtain a first similarity corresponding to each candidate push information, and to perform similarity calculation between each candidate push information representation and the negative feedback interaction object representation to obtain a second similarity corresponding to each candidate push information. The similarity difference calculation module is configured to obtain the similarity score for each candidate push information based on the difference between the first similarity and the second similarity corresponding to each candidate push information. The push information determination module is configured to determine the candidate push information corresponding to the highest similarity score as the target push information; The trained information prediction model is obtained by training the information prediction model according to any one of claims 1 to 13.

19. A computer-readable storage medium storing a computer program that, when executed by a processor, implements a training method for an information prediction model as described in any one of claims 1 to 13, or a push information selection method as described in any one of claims 14 to 16.

20. A computer device, comprising: Memory is used to store executable instructions or computer programs. The processor, wherein the memory stores a computer program, which, when executed by the processor, implements the training method for the information prediction model as described in any one of claims 1 to 13, or the push information selection method as described in any one of claims 14 to 16.

21. A computer program product, comprising a computer program, which, when executed by a processor, implements the training method for an information prediction model as described in any one of claims 1 to 13, or the push information selection method as described in any one of claims 14 to 16.

Images