Information processing method, electronic device, storage medium and computer program product

Abstract

The invention provides an information processing method, electronic equipment, a storage medium and a computer program product. The method comprises the steps of performing a first prediction operation on an interaction behavior of a first object and first information based on a first feature of the first object and a second feature of the first information to obtain a first interaction probability of the first object and the first information; determining a first confidence coefficient of the first prediction operation based on the first interaction probability; under the condition that the first confidence coefficient is smaller than a first preset threshold value, determining second information based on the first feature and the second feature; and determining recommendation information from the first information based on the third feature, the first feature and the second feature of the second information. According to the invention, the accuracy of information recommendation can be improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to an information processing method, electronic equipment, storage medium, and computer program product. Background Art

[0002] Information recommendation systems primarily rely on technologies such as information retrieval, data mining, machine learning, and natural language processing to analyze user behavior data and deliver personalized information recommendations. However, when faced with insufficient data or high uncertainty, the accuracy of information recommendation models in related technologies decreases, leading to low user satisfaction. Summary of the Invention

[0003] Embodiments of the present application provide an information processing method, electronic device, storage medium, and computer program product, which can improve the accuracy of information recommendation.

[0004] The technical solution of the embodiment of the present application is implemented as follows:

[0005] An embodiment of the present application provides an information processing method, comprising: performing a first prediction operation on an interaction behavior between a first object and the first information based on a first feature of the first object and a second feature of the first information to obtain a first interaction probability between the first object and the first information; determining a first confidence level of the first prediction operation based on the first interaction probability; determining second information based on the first feature and the second feature when the first confidence level is less than a first preset threshold; and determining recommended information from the first information based on a third feature of the second information, the first feature, and the second feature.

[0006] An embodiment of the present application provides an information processing device, including: a prediction module, used to perform a first prediction operation on the interaction behavior of the first object and the first information based on a first feature of the first object and a second feature of the first information, to obtain a first interaction probability between the first object and the first information; a confidence determination module, used to determine a first confidence of the first prediction operation based on the first interaction probability; a retrieval module, used to determine second information based on the first feature and the second feature when the first confidence is less than a first preset threshold; and a recommendation information determination module, used to determine recommended information from the first information based on a third feature of the second information, the first feature, and the second feature.

[0007] An embodiment of the present application provides an electronic device, comprising:

[0008] a memory for storing computer-executable instructions or computer programs;

[0009] The processor is used to implement the information processing method provided in the embodiment of the present application when executing the computer-executable instructions or computer programs stored in the memory.

[0010] An embodiment of the present application provides a computer-readable storage medium storing a computer program or computer-executable instructions for implementing the information processing method provided in the embodiment of the present application when executed by a processor.

[0011] An embodiment of the present application provides a computer program product, including a computer program or computer-executable instructions. When the computer program or computer-executable instructions are executed by a processor, the information processing method provided in the embodiment of the present application is implemented.

[0012] The embodiments of the present application have the following beneficial effects:

[0013] Based on the first feature of the first object and the second feature of the first information, a first prediction operation is performed on the interaction behavior of the first object with respect to the first information to obtain a first interaction probability, and a first confidence level is calculated based on the first interaction probability so that the reliability of the information recommendation can be more accurately evaluated based on the first confidence level. When the first confidence level is less than a first preset threshold, the second information is determined based on the first feature and the second feature, and the recommended information is re-determined from the first information based on the third feature, the first feature, and the second feature of the second information. This allows the second information to be re-determined when the information recommendation result is inaccurate, and the third feature of the second information is used as an additional feature to re-determine the information recommendation result, thereby ensuring the reliability of the information recommendation result and thereby improving the accuracy of the information recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a schematic diagram of the architecture of the information processing system provided by an embodiment of the present application;

[0015] Figure 2 is a structural diagram of an electronic device provided in an embodiment of the present application;

[0016] Figure 3 This is a flow diagram of the information processing method provided in the embodiment of the present application. Figure 1 ;

[0017] Figure 4 This is a flow diagram of the information processing method provided in the embodiment of the present application. Figure 2 ;

[0018] Figure 5 This is a flow diagram of the information processing method provided in the embodiment of the present application. Figure 3 ;

[0019] Figure 6 This is a flow diagram of the information processing method provided in the embodiment of the present application. Figure 4;

[0020] Figure 7 This is a flow diagram of the information processing method provided in the embodiment of the present application. Figure 5 ;

[0021] Figure 8 This is a flow diagram of the information processing method provided in the embodiment of the present application. Figure 6 ;

[0022] Figure 9 This is another optional flowchart of the information processing method provided in the embodiment of the present application;

[0023] Figure 10 This is a flowchart of the information processing method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0025] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0026] In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0027] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0028] Unless otherwise defined, all technical and scientific terms used in the embodiments of the present application have the same meanings as those commonly understood by those skilled in the art. The terms used in the embodiments of the present application are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0029] The relevant data collection and processing in the embodiments of this application should be strictly in accordance with the requirements of relevant laws and regulations when applied in examples, and the informed consent or separate consent of the personal information subject should be obtained. Subsequent data use and processing should be carried out within the scope of authorization of laws and regulations and the personal information subject.

[0030] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0031] 1) Information Retrieval: Information retrieval technology is the foundation of ad recommendations. It is used to match users with ad content and can be applied to scenarios such as text matching and keyword extraction. Information retrieval technology helps the system find relevant ads based on the user's query. For example, if a user searches for "sports shoes," the system will retrieve and recommend relevant ads.

[0032] 2) Information recommendation: refers to the process of automatically recommending content or information that best suits users' needs and interests by analyzing their behavior, preferences, and other relevant information.

[0033] 3) First Information: In an information recommendation system, first information refers to information that may be relevant or of interest to the user. This information may be in various forms, such as text, images, videos, and audio. For example, on a short video platform, the first information may be a short video; on a shopping platform, the first information may be product information or product images; or it may be an advertisement placed by an internet advertising platform.

[0034] 4) First object: refers to a specific user or user group. The information recommendation system should provide recommended information for the first object.

[0035] 5) First Feature: A set of comprehensive features describing the user, including the sixth, seventh, and eighth features. The sixth feature characterizes the user's basic attributes, such as age, gender, occupation, and interests. The seventh feature characterizes the user's surroundings, such as time, location, and device information. The eighth feature characterizes the user's historical interactions with historical recommendations, such as clicking on ad A → viewing ad B → clicking on ad C → clicking on ad D → viewing ad E.

[0036] 6) Secondary features: These are used to describe the attributes or characteristics of the first information, including but not limited to the content type, keywords, and release date of the first information. For example, in an ad recommendation scenario, information features include but are not limited to the content of the ad, ad delivery time, ad click-through rate, ad description, and target audience.

[0037] 7) Interactive behavior: The interactive behavior between users and information, such as clicking, browsing, collecting and commenting.

[0038] 8) First interaction probability: refers to the probability of a user interacting with the first information as predicted by the information recommendation model. For example, in an advertising recommendation scenario, the first interaction probability refers to the probability of a user clicking on an advertisement as predicted by the information recommendation model.

[0039] 9) Confidence: It is a quantitative assessment of the credibility of the first interaction probability, usually a value between 0 and 1, indicating the reliability of the first interaction probability. The first confidence of the first information refers to the certainty or confidence of the information recommendation model in the predicted first interaction probability of the first information. For example, in the advertising recommendation scenario, confidence is mainly used to measure the accuracy of the prediction of the user's interest in or click on an advertisement. A confidence level greater than a preset threshold indicates that the advertising recommendation is very confident that the user will be interested in the recommended advertisement; conversely, if the confidence level is less than the preset threshold, it means that the advertising recommendation model is uncertain about the effectiveness of the recommendation.

[0040] 10) Uncertainty: refers to the degree of uncertainty of the information recommendation model when predicting user interests or behaviors. The uncertainty of the information recommendation model can be measured by indicators such as confidence, variance or entropy.

[0041] 11) Vector Space Model: It is used to represent text as vectors through methods such as term frequency-inverse document frequency (TF-IDF), thereby calculating the similarity between documents.

[0042] 12) Boolean search model: used to perform simple keyword matching using Boolean logic (such as AND, OR, NOT).

[0043] 13) Machine Learning: Machine learning algorithms are the core of information recommendation systems, predicting user behavior and optimizing information selection in a data-driven manner. Common machine learning techniques include: supervised learning, logistic regression, support vector machines, unsupervised learning, clustering algorithms, dimensionality reduction techniques, and reinforcement learning. Supervised learning refers to the use of labeled data to train models, such as classifiers and regression models, to predict the probability of user clicks or conversions. Logistic regression is used for binary classification problems and predicting ad click-through rates (CTR). Support vector machines are used for classification tasks in high-dimensional feature spaces. Unsupervised learning is used for data clustering or dimensionality reduction, such as user group analysis. Clustering algorithms: such as K-means, are used to cluster user interests. Dimensionality reduction techniques: such as principal component analysis (PCA), are used for dimensionality reduction processing of high-dimensional data. Reinforcement learning: Used to dynamically optimize advertising delivery strategies, and the system obtains the best strategy through continuous trial and error.

[0044] 14) Data Mining: This is used to extract useful information from big data such as user behavior and advertising effectiveness. For example, ad recommendation systems can use this information to optimize advertising strategies and improve effectiveness. Data mining can include association rule mining and sequential pattern mining. Association rule mining, such as the Apriori algorithm, is used to discover user behavior patterns (such as purchase combinations). Sequential pattern mining is used to discover temporal patterns in user behavior.

[0045] 15) Natural Language Processing: Used to analyze and understand ad text and user comments, and improve the matching accuracy of ad recommendations.

[0046] 16) Text analysis: such as sentiment analysis, used to understand the emotional tendencies of user comments.

[0047] 17) Information extraction: such as entity recognition, which is used to extract key entities (such as product names) from advertisement descriptions.

[0048] 18) Recommendation algorithm: The core of the advertising recommendation system lies in the recommendation algorithm, which mainly includes the following categories: content-based recommendation: recommending advertisements based on the similarity between user historical behavior and advertisement content; feature engineering: extracting and matching features between users and advertisements; collaborative filtering: making recommendations based on the similarity between users or advertisements; user-based collaborative filtering: recommending advertisements based on the behavior of similar users; item-based collaborative filtering: recommending advertisements based on the historical performance of similar advertisements; hybrid recommendation system: combining content and collaborative filtering methods to improve recommendation accuracy.

[0049] 19) Big Data Processing Technology: Because ad recommendation systems require processing large amounts of real-time data, big data processing technology plays a key role in ad recommendation. This includes distributed computing and real-time data stream processing. Distributed computing, such as Hadoop and Spark, is used for large-scale data processing and analysis.

[0050] 20) Online Learning and Personalization: Online learning technology enables the system to adapt to user dynamics in real time. By continuously updating the model, personalized ad recommendations can increase user engagement.

[0051] 21) Online Gradient Descent: used to update model parameters in real time.

[0052] In order to better understand the information processing method provided in the embodiments of the present application, the problems faced by the information processing methods of related technologies are first explained.

[0053] In related technologies, information recommendation systems suffer from reduced accuracy when faced with insufficient data or high uncertainty, leading to low user satisfaction. Related technologies struggle to effectively capture user behavior patterns over long time spans, resulting in information recommendation systems underutilizing information about users' long-term dependencies. Furthermore, processing long sequences of data typically requires significant computing resources, and conventional backpropagation algorithms are computationally inefficient and resource-intensive for long sequences. Many information recommendation systems struggle to balance multiple objectives during training, resulting in insufficient generalization and limiting recommendation results to a single dimension.

[0054] Based on the problems existing in the related technologies, the embodiments of the present application provide an information processing method, device, electronic device, computer-readable storage medium and computer program product, which can ensure the reliability of information recommendation results and thereby improve the accuracy of information recommendation.

[0055] The following describes exemplary applications of electronic devices provided in the embodiments of the present application. The electronic devices provided in the embodiments of the present application can be implemented as various types of terminals such as laptop computers, tablet computers, desktop computers, set-top boxes, smart phones, smart speakers, smart watches, smart TVs, and vehicle-mounted terminals, and can also be implemented as servers. Among them, the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal and the server can be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiments of the present application. Below, exemplary applications when the information processing device is implemented as a terminal or a server will be described.

[0056] See also Figure 1 , Figure 1 : It is a schematic diagram of the architecture of the information processing system provided in the embodiment of the present application. In order to perform information processing operations, an information processing application can be provided. For example, the information processing application can be an application dedicated to information processing in videos, or it can be a functional module in other applications (such as an information processing module in a short video application, etc.). The information processing system 100 in the embodiment of the present application includes at least a terminal 400, a network 300 and a server 200, wherein the server 200 is a server for the information processing application. The server 200 can constitute the information processing device of the embodiment of the present application, that is, the information processing method of the embodiment of the present application is implemented by the server 200. The terminal 400 is connected to the server 200 via the network 300, and the network 300 can be a wide area network or a local area network, or a combination of the two.

[0057] See also Figure 1 , the user can perform interactive operations on the client side of the information processing application through the terminal 400. The interactive operations may be, for example, click-to-search information operations, sliding-to-browse information operations, and the like. After receiving the interactive operation of the user, the client side sends the information processing request to the server 200 through the network 300. After receiving the information processing request, the server 200 responds to the information processing request sent by the terminal and performs a first prediction operation on the interactive behavior of the first object and the first information based on the first feature of the first object and the second feature of the first information to obtain a first interaction probability between the first object and the first information; the server 200 determines a first confidence level of the first prediction operation based on the first interaction probability; when the first confidence level is less than a first preset threshold, the server 200 determines the second information based on the first feature and the second feature; the server 200 determines recommended information from the first information based on the third feature, the first feature, and the second feature of the second information. The server 200 may send the recommended information to the terminal 400. The terminal 400 displays the recommended information on the current interface.

[0058] In some embodiments, the information processing method of the embodiment of the present application can also be executed by the terminal 400. That is, after the user performs an interactive operation on the client of the information processing application through the terminal 400, the terminal 400 performs a first prediction operation on the interactive behavior of the first object and the first information based on the first feature of the first object and the second feature of the first information in response to the interactive operation, and obtains a first interaction probability between the first object and the first information; the terminal 400 determines a first confidence level of the first prediction operation based on the first interaction probability; the terminal 400 determines the second information based on the first feature and the second feature when the first confidence level is less than a first preset threshold; and the terminal 400 determines recommended information from the first information based on the third feature, the first feature, and the second feature of the second information. The terminal 400 displays the recommended information on the current interface.

[0059] In the product recommendation scenario of a shopping platform, the server or terminal can predict the user's click behavior on the product (first information) based on the first feature such as the user's shopping history, browsing history, and the second feature such as the attributes of the product information on the platform (such as category, price, evaluation, etc.), and obtain the probability of the user clicking on the product as the first interaction probability. Based on the first interaction probability, the first confidence level that the product was clicked by the user is determined. When the first confidence level is less than the first preset threshold, the second information is retrieved from other platforms (information retrieval library) based on the first feature and the second feature. Based on the third feature, the first feature and the second feature of the second information, the recommended product is determined from the first information. The recommended products are displayed on the interface of the shopping platform to provide users with more accurate product recommendations.

[0060] On a social media platform, when the platform attempts to push a piece of news to a user, it can predict the user's click behavior on the news (first information) based on first features such as the user's reading habits and areas of interest, as well as second features such as the news's subject and keywords, and obtain the probability of the user clicking on the news as a first interaction probability. Based on the first interaction probability, a first confidence level that the news was clicked by the user is determined. If the first confidence level is less than a first preset threshold, the second information is retrieved from other news libraries based on the first and second features. Based on the third feature of the second information, the first and second features, recommended news is determined from the first information. The social media platform pushes the recommended news to the user.

[0061] See also Figure 2 , Figure 2 is a structural diagram of an electronic device provided in an embodiment of the present application, Figure 2The electronic device shown includes: at least one processor 410, a memory 450, at least one network interface 420 and a user interface 430. The various components in the terminal 400 are coupled together via a bus system 440. It is understood that the bus system 440 is used to achieve connection and communication between these components. In addition to including a data bus, the bus system 440 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 440 is not shown in FIG. Figure 2 Various buses are labeled as bus system 440 .

[0062] The processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0063] The user interface 430 includes one or more output devices 431 that enable presentation of media content, including one or more speakers and one or more visual display screens. The user interface 430 also includes one or more input devices 432, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0064] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard drives, optical drives, etc. The memory 450 may optionally include one or more storage devices that are physically remote from the processor 410.

[0065] The memory 450 includes volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 450 described in the embodiments of the present application is intended to include any suitable type of memory.

[0066] In some embodiments, the memory 450 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplified below.

[0067] Operating system 451, including system programs for processing various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, and driver layer, which are used to implement various basic services and process hardware-based tasks;

[0068] A network communication module 452 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 420. Exemplary network interfaces 420 include Bluetooth, Wi-Fi, and Universal Serial Bus (USB);

[0069] a presentation module 453 for enabling presentation of information via one or more output devices 431 (e.g., a display screen, a speaker, etc.) associated with the user interface 430 (e.g., a user interface for operating peripheral devices and displaying content and information);

[0070] The input processing module 454 is configured to detect one or more user inputs or interactions from one of the one or more input devices 432 and to translate the detected inputs or interactions.

[0071] In some embodiments, the information processing device provided in the embodiments of the present application can be implemented in software. Figure 2 Information processing device 455 stored in memory 450 is shown. This may be software in the form of a program or plug-in, and includes the following software modules: a prediction module 4551, a confidence determination module 4552, a retrieval module 4553, and a recommendation information determination module 4554. These modules are logical and can be arbitrarily combined or further separated based on the functions they implement. The functions of each module will be described below.

[0072] In other embodiments, the information processing device provided in the embodiments of the present application can be implemented in hardware. As an example, the device provided in the embodiments of the present application can be a processor in the form of a hardware decoding processor, which is programmed to execute the information processing method provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor can adopt one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs) or other electronic components.

[0073] The information processing method provided in the embodiment of the present application will be explained in combination with the exemplary application and implementation of the server provided in the embodiment of the present application.

[0074] The following describes the information processing method provided by the embodiment of the present application. As mentioned above, the electronic device that implements the information processing method of the embodiment of the present application can be a terminal, a server, or a combination of the two. Therefore, the execution entity of each step will not be repeated below.

[0075] Figure 3 This is a flow diagram of the information processing method provided in the embodiment of the present application. Figure 1 , will combine Figure 3 The steps shown are explained as Figure 3 As shown, the information processing method is described as an example in which the execution subject is a server. The method includes the following steps 101 to 104:

[0076] In step 101, based on a first feature of a first object and a second feature of first information, a first prediction operation is performed on an interaction behavior between the first object and the first information to obtain a first interaction probability between the first object and the first information.

[0077] Here, the first object refers to the user who needs to recommend information in the information recommendation scenario. The first feature of the first object is a comprehensive feature used to describe the basic attributes of the first object and the user's historical interactive behavior with the first information. The first information is information that may be related to the first object or the first object is interested in. The first information may be content in various forms such as text, image, video, audio, etc. For example, on a short video platform, the first information may be a short video to be recommended; on a shopping platform, the first information may be product information or product pictures to be recommended. The number of first information can be one or more. The second feature is a feature used to describe the attributes or characteristics of the first information. The second feature includes at least one of the content, type, and date of the first information. Interactive behavior is the interactive behavior between the first object and the first information, such as operations such as clicking, browsing, collecting, and commenting.

[0078] The information recommendation model can be used to perform a first prediction operation on the interaction between the first object and the first information to obtain a first interaction probability between the first object and the first information. The first feature of the first object and the second features of the plurality of first information are fused and input into the information recommendation model, and the first interaction probability of each first information is output. For each first information, the first interaction probability of the first information is the probability that the first object will interact with the first information. The embodiment of the present application does not limit the model structure of the information recommendation model, which can be a machine learning model, a deep learning model, etc.

[0079] For example, in an ad recommendation scenario, the first subject is user A, and the first information is an advertisement placed by an internet advertising platform. Secondary features include, but are not limited to, ad content, ad type, ad delivery time, ad click-through rate, ad description, and target audience. An information recommendation model can predict a first interaction probability for user A to interact with each ad based on the user's first feature and the second features of multiple ads. The first interaction probability is the probability that user A will click on the ad. The first interaction probability for ad a indicates whether user A will click on ad a.

[0080] In some embodiments, the first feature includes the sixth feature, the seventh feature, and the eighth feature, see Figure 4 , step 101 can be implemented by following steps 1011 to 1013, which are described in detail below.

[0081] In step 1011 , the sixth feature and the second feature are combined to obtain a ninth feature, and the sixth feature and the seventh feature are combined to obtain a tenth feature.

[0082] Here, the sixth feature is used to characterize the basic attributes of the first object, including but not limited to the identification, age, gender, occupation, interests, etc. of the first object. The seventh feature is used to characterize the environmental information of the first object, including but not limited to the time, location, and device information used by the first object. The eighth feature is used to characterize the user's historical interaction behavior in the historical recommendation information. For example, the first object is user A, and the sixth feature may include "user identification: A, age: 30, gender: male". The seventh feature may include "time: Monday morning, location: city X, device information: mobile terminal Y". The eighth feature is click on ad A → browse product B → click on ad C → click on ad D → browse product E.

[0083] The ninth feature is a new feature obtained by combining at least one feature from the sixth feature with at least one feature from the second feature. For example, the user ID from the sixth feature and the information type from the second feature can be combined to obtain the ninth feature. When processing the ninth feature, the information recommendation model can determine the number of historical clicks by user A on each ad category, and further determine the number of historical clicks by user A on ad category 1 as the interest weight of user A for ad category 1. Therefore, the ninth feature can reflect the first subject's preference for the first information. The tenth feature is a new feature obtained by combining at least one feature from the sixth feature with at least one feature from the seventh feature. For example, the user ID from the sixth feature and the time from the seventh feature can be combined to obtain the tenth feature. When processing the tenth feature, the information recommendation model can determine the number of historical clicks by user A on ads at different times, thereby analyzing user A's activity level during different time periods. For example, the tenth feature can be used to determine the number of clicks by user A on Monday mornings, i.e., the number of ad clicks by user A on Monday mornings.

[0084] In step 1012, the ninth feature, the tenth feature, the sixth feature, the second feature, and the eighth feature are fused to obtain the eleventh feature.

[0085] Here, fusing the ninth feature, the tenth feature, the sixth feature, the second feature, and the eighth feature means splicing or merging the ninth feature, the tenth feature, the sixth feature, the second feature, and the eighth feature into a feature set as the eleventh feature.

[0086] In some embodiments, the eighth feature can also be segmented to obtain multiple sub-features. The eighth feature is used to characterize the user's historical interactive behavior in historical recommendation information, that is, the eighth feature is a behavior sequence. Segmenting the eighth feature to obtain multiple sub-features means segmenting the behavior sequence to obtain multiple behavior sub-sequences. The embodiment of the present application does not limit the segmentation method for segmenting the eighth feature into multiple sub-features. For example, the eighth feature can be segmented based on a time window to obtain multiple sub-features. By setting a time window of fixed length and then sliding on the original eighth feature according to this time window, a series of sub-features can be obtained. The length of the time window can be determined according to the actual application, for example, it can be set to one day, one week, etc. Slide on the eighth feature with the defined time window length and cut out the sub-features of this length. When sliding to the end of the eighth feature, if the remaining feature part is less than the size of a complete time window, the remaining feature part can be directly used as the last sub-feature.

[0087] For example, the eighth feature is: click on ad A → browse product B → click on ad C → click on ad D → browse product E. After segmenting the eighth feature, three sub-features are obtained: sub-feature 1, sub-feature 2, and sub-feature 3. Sub-feature 1: click on ad A → browse product B, sub-feature 2: click on ad C → click on ad D, sub-feature 3: browse product E. It should be noted that the present embodiment does not limit the number of sub-features obtained by segmenting the eighth feature and can be set based on actual needs.

[0088] When training the information recommendation model, the input eighth feature must first be segmented into multiple sub-features. These sub-features are then used for training. This allows these sub-features to be processed in each training iteration of the information recommendation model, improving training efficiency. During training, the encoder encodes each sub-feature to produce a vector output, H. For example, sub-feature 1 is input to the encoder, which outputs H1. Sub-feature 2 and H1 are input to the encoder, which outputs H2. Sub-feature 3 and H2 are input to the encoder, which outputs H3. The training process backpropagates only the loss for the current sub-feature, limiting gradient updates to the length of the current sub-feature to improve training efficiency.

[0089] After obtaining the multiple sub-features, the ninth feature, the tenth feature, the sixth feature, the second feature, and the multiple sub-features are spliced ​​or merged into a feature set as the eleventh feature.

[0090] In step 1013, based on the eleventh feature, a first prediction operation is performed on the interaction behavior between the first object and the first information to obtain a first interaction probability.

[0091] Here, the information recommendation model is used to perform prediction processing on the eleventh feature to obtain a first interaction probability of the first object interacting with each first information.

[0092] The present embodiment combines the sixth, seventh, and eighth features in different ways to enrich the input features of the information recommendation model, thereby improving the accuracy of information recommendation. Furthermore, the eighth feature is segmented into multiple sub-features, allowing the information recommendation model to process shorter sub-features, thereby improving model processing efficiency.

[0093] In some embodiments, see Figure 5 , step 1013 can be implemented by following steps 10131 to 10133, which are described in detail below.

[0094] In step 10131, based on the eleventh feature, a third prediction operation is performed on the interaction behavior between the first object and the first information to obtain a third interaction probability between the first object and the first information.

[0095] Here, the third prediction operation can be to predict the interactive behavior of the first object to each first information based on the eleventh feature through the first model, and obtain the third interaction probability of each. The first model is an information recommendation model, which is used to predict the possible interactive behavior of the first object to the first information based on the historical behavior, preference or other relevant characteristics (eleventh feature) of the first object, and output the predicted third interaction probability. The embodiment of the present application does not specifically limit the model structure of the first model, for example, it can be a machine learning model or a deep learning model. Exemplarily, the first model can use a memory-enhanced transformer model (Transformer model), the encoder and decoder in the first model share weights, and the first model also has a memory structure. After the encoder encodes the eleventh feature, the encoding result can be stored in the memory, and the decoder then obtains the encoding result from the memory for prediction. For each first information, the third interaction probability of the first information is the probability of the first object interacting with the first information determined by the first model.

[0096] Here, multiple first models can be used to predict the eleventh feature, respectively, to obtain a third interaction probability of the first object interacting with each piece of first information, output by each first model. It should be noted that the third interaction probability is a probability value. That is, for a first model, the sum of the third interaction probabilities of multiple pieces of first information output by the first model is 1.

[0097] In some embodiments, in step 10131, a third prediction operation is performed on the interaction behavior of the first object and the first information based on the eleventh feature to obtain a third interaction probability between the first object and the first information. This can be achieved in the following manner: first, the eleventh feature is encoded to obtain a first feature vector; then, attention processing is performed on the first feature vector to obtain a twelfth feature; then, gate processing is performed on the twelfth feature to obtain a thirteenth feature; finally, the thirteenth feature is decoded to obtain a third interaction probability.

[0098] Here, the eleventh feature can be encoded by the encoder in the first model to obtain the first feature vector. Since the eleventh feature is obtained by fusing the ninth feature, the tenth feature, the sixth feature, the second feature and multiple eighth features, the encoder can encode the ninth feature, the tenth feature, the sixth feature, the second feature and each eighth feature in the eleventh feature respectively to obtain multiple feature vectors, and concatenate the multiple feature vectors to obtain the first feature vector. The first feature vector H can be expressed as {h1, h2, ..., h T}, T is the length of the first feature vector, and T is an integer greater than 1. The decoder in the first model includes a gated cross-attention layer. The process of performing attention processing on the first feature vector is as follows: the t-th hidden state of the decoder is received through the gated cross-attention layer, t = 1, 2, ..., T, the similarity between the t-th hidden state and the first feature vector is calculated using a normalized exponential function (softmax function), and the product of the similarity and the first feature vector is determined as the t-th twelfth feature. The t-th twelfth feature is determined as the t+1-th hidden state of the decoder. When t = 1, the first hidden state of the decoder is the first feature vector. Gating the twelfth feature refers to processing the twelfth feature based on an activation function (sigmoid function) to obtain the thirteenth feature. The sigmoid function amplifies large values ​​in the twelfth feature to larger values ​​and reduces small values ​​to smaller values, resulting in the thirteenth feature. Decoding the thirteenth feature refers to normalizing the thirteenth feature based on the softmax function to obtain a third interaction probability of the first object interacting with the first information.

[0099] The embodiment of the present application performs gating processing on the twelfth feature, thereby enhancing the weight of the feature that the first object is more interested in among the twelfth features, reducing the processing of irrelevant information, and thereby improving the efficiency and accuracy of information recommendation.

[0100] In step 10132, a weight of the third interaction probability is determined.

[0101] Here, after determining the third interaction probability of the first object to the first information based on multiple first models, the weight of each third interaction probability can be obtained. Each third interaction probability has a corresponding weight, which is used to reflect the importance of the third interaction probability. The weight of each third interaction probability can be dynamically adjusted based on the first feature of the first object and the second features of the multiple first information. In an embodiment of the present application, the weight of each third interaction probability can be determined by minimizing the sum of the loss functions of the multiple first models.

[0102] In some embodiments, determining the weight of the third interaction probability in step 10132 can be achieved in the following manner: first, based on the first weight of the first model, the third interaction probability of the first information is fused to obtain the fourth interaction probability of the first information; second, the loss function of the first model is updated to the conjugate function corresponding to the first model; then, based on the fourth interaction probability, the label of the first object's interaction behavior with respect to the first information, and the conjugate function, the conjugate function value is determined; then, based on the conjugate function value, the first weight is updated to obtain the second weight of the first model; finally, the second weight is determined as the weight of the third interaction probability.

[0103] Here, each first model has a corresponding first weight, which is used to reflect the importance of the first model. The first weight of the first model can be any initial value, which is set based on actual experience. It should be noted that the sum of the first weights of multiple first models is 1. Based on the first weights of the first models, the third interaction probability of the first information is fused, that is, the third interaction probability output by each first model is weighted and summed based on the first weights of the first model to obtain the fourth interaction probability. The first weight of each first model can be multiplied by the third interaction probability output by the corresponding first model to obtain a first product. The first products of multiple first models are summed to obtain the fourth interaction probability of the first information.

[0104] The loss function of the first model can be a single-task loss function or a multi-task loss function. Taking the advertising recommendation scenario as an example, the single-task loss function is the loss function of the click-through rate, which is calculated by the predicted click probability and the actual click label (1 represents click, 0 represents no click). The multi-task loss function is the sum of multiple single-task loss functions. For example, the multi-task loss function is the sum of the loss function of the click-through rate and the loss function of the advertisement ranking. After determining the loss function, updating the loss function of the first model to the conjugate function corresponding to the first model can be achieved in the following way: using Legendre transform to transform the loss function of the first model to obtain the conjugate function corresponding to the first model. The second weight of each first model can be determined by minimizing the sum of the loss functions of multiple first models. After the loss function is transformed into a conjugate function by Legendre transform, the second weight of each first model can be determined by maximizing the conjugate function.

[0105] The conjugate function value is still determined based on the error between the fourth interaction probability of the first object with respect to the first information and the label of the first object's interaction behavior with respect to the first information. Determining the conjugate function value based on the fourth interaction probability, the label of the first object's interaction behavior with respect to the first information, and the conjugate function can be achieved by inputting the fourth interaction probability and the label of the first object's interaction behavior with respect to the first information into the conjugate function to calculate the conjugate function value. For example, the label of the first object's interaction behavior with respect to the first information includes the labels "0" and "1." The label "0" indicates that the first object has not interacted with the first information, and the label "1" indicates that the first object has interacted with the first information. The fourth interaction probability of the first information can have a probability value of 0.8. The conjugate function value can be calculated based on the probability value of 0.8 and the labels 0 or 1. The first weight is updated by maximizing the conjugate function value to obtain the second weight of the first model. The weight determination process described above can be performed using the first information from the first object's historical interaction behavior. After determining the second weight of each first model, the second weight of each first model is determined as the weight of the third interaction probability output by the corresponding first model.

[0106] By transforming the loss function into a conjugate function, the embodiment of the present application can obtain the second weight of each first model by maximizing the conjugate function value of the conjugate function, and then obtain the weight of the third interaction probability, so as to continuously adjust and optimize the weight of the third interaction probability according to the actual recommendation effect to adapt to changes in user preferences and the first information content.

[0107] In step 10133, the third interaction probabilities are fused based on the weights of the third interaction probabilities to obtain the first interaction probability.

[0108] Here, based on the weight of each third interaction probability, a weighted sum is performed on the plurality of third interaction probabilities to obtain the first interaction probability.

[0109] The embodiment of the present application determines a multi-model prediction result through the third interaction probabilities output by multiple first models and the corresponding weights, which can improve the diversity and generalization ability of information recommendation results.

[0110] Continue to see Figure 3 , continue with step 101 above for explanation.

[0111] In step 102 , a first confidence level of a first prediction operation is determined based on the first interaction probability.

[0112] Here, the first confidence level is used to characterize the reliability of the first interaction probability of the first object for each first information obtained by the first model performing the first prediction operation based on the first feature and the second feature, that is, it can be used to characterize the uncertainty of the first interaction probability of each first information. The entropy or variance can be determined based on the first interaction probability of each first information to determine the uncertainty index. The uncertainty index can be inversely quantized to obtain the first confidence level. Exemplarily, the inverse quantization process can be as follows: a threshold is set, and the difference between the threshold and the uncertainty index can be determined as the first confidence level. It should be noted that the embodiment of the present application does not limit the specific method of inversely quantizing the uncertainty index to obtain the first confidence level. It is only necessary to ensure that the larger the uncertainty index, the smaller the first confidence level.

[0113] In some embodiments, the number of first information is multiple, see Figure 6 , Figure 6 The illustrated step 102 can be implemented by following steps 1021 to 1023 , which are described in detail below.

[0114] In step 1021 , for each piece of first information, logarithmic processing is performed on the first interaction probability to obtain a first value.

[0115] Here, for each piece of first information, the first model can predict the first object's interaction behavior with respect to the first information to obtain a first interaction probability for the first information. For each piece of first information, the first interaction probability is logarithmically processed to obtain a first value corresponding to the first information.

[0116] In step 1022, the product of the first value and the first interaction probability is determined as a second value.

[0117] Here, for each piece of first information, a product obtained by multiplying the first value corresponding to the first information by the first interaction probability of the first information is determined as the second value corresponding to the first information.

[0118] In step 1023 , a first confidence level is determined based on the plurality of second values.

[0119] Here, the second values ​​corresponding to the multiple first information can be summed to obtain a numerical sum, and the entropy or variance can be determined based on the numerical sum, and the first confidence level can be determined based on the entropy or variance. For example, the negative value of the sum of the second values ​​corresponding to the multiple first information can be determined as the entropy, a threshold can be set, and the difference between the threshold and the entropy can be determined as the first confidence level.

[0120] The embodiment of the present application improves the accuracy of the first confidence level by summing up the second numerical values ​​of multiple first information and comprehensively evaluating the prediction results of multiple candidate information. The reliability of the information recommendation can then be determined based on the first confidence level, thereby improving the accuracy and stability of the prediction.

[0121] Continue to see Figure 3 , continue with step 102 above for explanation.

[0122] In step 103, when the first confidence level is less than a first preset threshold, second information is determined based on the first feature and the second feature.

[0123] Here, the first preset threshold is a pre-set numerical standard used to determine the standard value of the uncertainty of the first interaction probability. The first preset threshold is used to assess whether the first confidence level is high enough to determine whether to continue selecting recommended information based on the first interaction probability. If the first confidence level is lower than the first preset threshold, the reliability of the first interaction probability is too low, and the first information with the highest first interaction probability cannot be directly selected as the recommended information. This embodiment of the present application does not limit the value of the first preset threshold and can be set based on actual needs. When the first confidence level is lower than the first preset threshold, the second information can be retrieved from the information retrieval database based on the first and second features. An information retrieval system refers to a system that stores large amounts of information or data and supports efficient query and retrieval, allowing for quick retrieval of information or data. In this embodiment of the present application, an information retrieval database refers to a database or other platform other than the database in which the multiple first information are stored. The information retrieval database may include multiple third information items, which are alternative information items, such as alternative advertisements. The second information item is the third information item that is selected from the multiple third information items in the information retrieval database to better meet the preferences of the first subject.

[0124] Taking the advertising recommendation scenario as an example, user A is currently browsing or interacting on information platform 1 (such as a short video platform, a social platform, etc.), the first preset threshold is 0.3, and there are 10 advertisements in the information platform 1, that is, 10 first information. The advertising recommendation model predicts the click probability of user A clicking on each advertisement, and determines the first confidence level based on the entropy of the click probabilities of multiple advertisements. Assume that the calculated first confidence level is 0.25, which is less than the first preset threshold. At this time, based on the first feature of user A and the second feature of the first information, the second information can be retrieved from multiple advertisements (that is, multiple third information) on another information platform 2 (that is, the information retrieval library).

[0125] When the first confidence level is greater than or equal to a first preset threshold, at least one first information having a first interaction probability greater than or equal to a second preset threshold may be selected as the recommended information. Alternatively, the first information having the highest first interaction probability may be selected as the recommended information.

[0126] In some embodiments, see Figure 7 Determining the second information based on the first feature and the second feature in step 103 can be achieved by following steps 1031 to 1033, which are described in detail below.

[0127] In step 1031 , based on the first interaction probability, a fourth feature is obtained by screening the second features.

[0128] Here, based on the first interaction probability, the fourth information can be determined from the first information, and the fourth feature of the fourth information can be determined from the second feature. The first information with the highest first interaction probability is determined as the fourth information. Alternatively, at least one first information corresponding to a first interaction probability greater than a preset threshold is determined as the fourth information. After obtaining the fourth information, the first feature corresponding to the fourth information is determined as the fourth feature.

[0129] In step 1032 , a query statement is generated based on the fourth feature and the first feature.

[0130] Here, at least one fourth feature can be concatenated or combined with the first feature to generate a query statement. For example, the fourth information is Advertisement B, and the fourth feature of the fourth information is that Advertisement B is a clothing advertisement. The first feature is that User A is 18 years old. The generated query statement can be "18 years old, clothing."

[0131] In step 1033 , the similarity between the query statement and the third information in the information retrieval database is determined, and the second information is determined from the third information based on the similarity.

[0132] Here, the query statement can be encoded to obtain a query statement vector, and each third information can be encoded to obtain an encoding vector. The similarity between the query statement vector and each encoding vector is determined, that is, the similarity between the query statement and each third information. The embodiment of the present application does not limit the method for calculating vector similarity, for example, it can be a cosine similarity calculation method. Based on the similarity between the query statement and the third information, multiple third information are sorted from large to small according to the similarity. The third information with the greatest similarity can be determined as the second information. Alternatively, the first M third information can be selected from the sorted multiple third information as the second information, where M is a positive integer.

[0133] In the embodiment of the present application, when the first interaction probability is unreliable, a supplementary search is performed from the information retrieval library based on the first feature and the prediction result. The second information obtained can reflect the preference or interest of the first object, which facilitates the subsequent input of the information features of the second information as additional features into the information recommendation model for re-prediction, thereby improving the accuracy of information recommendation.

[0134] Continue to see Figure 3 , continue with step 103 above for explanation.

[0135] In step 104 , recommended information is determined from the first information based on the third feature, the first feature, and the second feature of the second information.

[0136] Here, recommended information refers to the information ultimately selected and recommended to the first subject. It is the optimal information determined after multiple rounds of screening and evaluation. Taking the ad recommendation scenario as an example, the recommended information is the advertisement recommended to user A. Based on the third feature, the first feature, and the second feature of the second information, the first subject's interaction behavior with each piece of first information can be re-predicted to obtain the fifth interaction probability between the first information and the first subject. The fifth interaction probability is the probability that the first subject will interact with the first information. The first information with the highest fifth interaction probability is selected as the recommended information.

[0137] The embodiment of the present application performs a first prediction operation on the interaction behavior of the first object with respect to the first information based on the first feature of the first object and the second feature of the first information, obtains a first interaction probability, and calculates a first confidence level based on the first interaction probability, so that the reliability of the information recommendation can be more accurately evaluated based on the first confidence level. When the first confidence level is less than a first preset threshold, the second information is determined from the information retrieval library based on the first feature and the second feature, and the recommended information is re-determined from the first information based on the third feature, the first feature, and the second feature of the second information, so as to realize the re-determination of the second information when the information recommendation result is inaccurate, and use the third feature of the second information as an additional feature to re-determine the information recommendation result, thereby ensuring the reliability of the information recommendation result and improving the accuracy of the information recommendation.

[0138] In some embodiments, see Figure 8 , step 104 can be implemented by following steps 1041 to 1044, which are described in detail below.

[0139] In step 1041 , the third feature, the first feature, and the second feature are fused to obtain a fifth feature.

[0140] Here, fusing the third feature, the first feature, and the second feature refers to splicing or merging the third feature, the first feature, and the second feature into a feature set as the fifth feature.

[0141] In step 1042 , based on the fifth feature, a second prediction operation is performed on the interaction behavior between the first object and the first information to obtain a second interaction probability between the first object and the first information.

[0142] Here, the second prediction operation may be inputting the fifth feature into the first model, the first model processing the fifth feature, and outputting a second interaction probability of the first object for each first information.

[0143] In step 1043 , a second confidence level of the second prediction operation is determined based on the second interaction probability.

[0144] Here, the process of determining the second confidence level of the second prediction operation based on the second interaction probability may refer to the above step 102 and will not be repeated.

[0145] In step 1044 , when the second confidence level is greater than or equal to the first preset threshold, recommended information is determined from the first information based on the second interaction probability.

[0146] Here, when the second confidence level is greater than or equal to the first preset threshold, the second interaction probability representing each piece of first information output by the first model is highly reliable, and the first information with the highest second interaction probability can be selected as the recommended information. When the second confidence level is less than the first preset threshold, step 103 and step 104 are repeated until the second confidence level is greater than or equal to the first preset threshold.

[0147] In the embodiment of the present application, by re-inputting the third feature of the retrieved second information as an additional feature into the information recommendation model, the prediction results of the information recommendation model can be made more consistent with the user's interests or preferences, thereby improving the accuracy of information recommendation.

[0148] Figure 9 This is another optional flow chart of the information processing method provided in the embodiment of the present application, such as Figure 9 As shown, the method includes the following steps 201 to 209:

[0149] Step 201: The terminal receives an interactive operation from a user.

[0150] Here, the interactive operation can be a click to search for information operation, a slide to browse information operation, etc.

[0151] Step 202: The terminal generates an information processing request in response to the interactive operation.

[0152] Step 203: The terminal sends an information processing request to the server.

[0153] Step 204 : In response to the information processing request, the server performs a first prediction operation on the interaction behavior between the first object and the first information to obtain a first interaction probability between the first object and the first information.

[0154] Here, based on the first feature of the first object and the second feature of the first information, a first prediction operation is performed on the interaction behavior of the first object and the first information. The specific process of obtaining the first interaction probability of the first object and the first information can refer to step 101 in the above embodiment and will not be repeated.

[0155] Step 205: The server determines a first confidence level of the first prediction operation based on the first interaction probability.

[0156] Here, the specific process of determining the first confidence of the first prediction operation based on the first interaction probability can refer to step 102 in the above embodiment, and will not be repeated here.

[0157] Step 206: When the first confidence level is less than a first preset threshold, the server determines second information based on the first feature and the second feature.

[0158] Here, the specific process of determining the second information based on the first feature and the second feature can refer to step 103 in the above embodiment and will not be repeated.

[0159] In step 207 , the server determines recommended information from the first information based on the third feature, the first feature, and the second feature of the second information.

[0160] Here, the specific process of determining the recommended information from the first information based on the third feature, the first feature, and the second feature of the second information can refer to step 104 in the above embodiment and will not be repeated.

[0161] Step 208: The server sends the recommendation information to the terminal.

[0162] Step 209: The terminal displays the recommended information on the current interface.

[0163] In the embodiment of the present application, the server re-retrieves the second information when the confidence level is low, and uses the third feature of the second information as an additional input feature to re-determine the recommended information, thereby improving the accuracy of information recommendation.

[0164] The following describes an exemplary application of the embodiments of the present application in a practical application scenario.

[0165] It should be noted that the embodiments of the present application are described using the scenario of advertising recommendation as an example, but those skilled in the art can apply the information processing method provided in the embodiments of the present application to any scenario of information recommendation based on their understanding of the following text. For example, an e-commerce platform recommends products that may be of interest to users by analyzing data such as users' shopping history and browsing behavior; a social media platform recommends posts, articles or videos that may be of interest to users by analyzing data such as users' social behavior and interest preferences; and a travel service platform recommends tourist attractions that may be of interest to users by analyzing users' travel history and interest preferences.

[0166] The information processing method proposed in the embodiment of the present application is an advertising recommendation method based on memory enhancement and efficient decoding, wherein the advertising recommendation model (corresponding to the first model in the above embodiment) is a Transformer model based on memory enhancement, wherein the encoder and decoder share weights, the weight parameters are small, and the training efficiency of the advertising recommendation model is high. The encoder can encapsulate the user's long behavior sequence (corresponding to the eighth feature in the above embodiment, such as ad clicks, product browsing, etc.) into the memory, and the decoder uses the long behavior sequence in the memory to predict the future ad click probability (corresponding to the first interaction probability in the above embodiment). The embodiment of the present application also introduces time truncated back propagation (TBPTT), which divides the user's long behavior sequence into shorter subsequences (corresponding to the sub-features in the above embodiment), and only calculates and propagates the gradient within the current subsequence each time to maintain computational efficiency and capture information in the user's long behavior sequence. In addition, to improve training and reasoning efficiency, a gated cross-attention mechanism is added to each decoder layer. During the decoding process, a decoding time algorithm is used to obtain a linear combination of the prediction results of multiple advertising recommendation models based on the weights of different advertising recommendation models. An efficient decoding strategy is derived using the Legendre transform. Based on the decoding strategy, the linear combination is solved to obtain the recommended content (corresponding to the recommended information in the above embodiment). The loss function uses multi-task learning loss and is optimized in combination with multiple objectives such as click-through rate (CTR). In addition, when the advertising recommendation model is uncertain about the generated prediction results, the knowledge retrieval mechanism is activated to minimize uncertainty by re-ordering the retrieved relevant information, thereby improving the accuracy and personalization of the recommendation.

[0167] The embodiments of the present application can solve the following technical problems: 1) The problem of long sequence dependency: related technologies are difficult to effectively capture the behavior patterns of users over a long time span, resulting in insufficient utilization of long-range dependency information of users by the advertising recommendation model. The embodiments of the present application use a memory-enhanced Transformer model and time-truncated backpropagation (TBPTT) to process the long behavior sequences of users, divide the long behavior sequences into shorter subsequences, and use the encoder-decoder architecture with shared weights to encapsulate the subsequence fragments into the memory, which can effectively capture and utilize information such as the user's historical behavior, preferences, and long-term interest changes, while reducing computational complexity. 2) Computational efficiency and resource consumption: The processing of long behavior sequences usually requires a lot of computing resources, and conventional backpropagation algorithms have low computational efficiency and high resource consumption on long behavior sequences. The embodiments of the present application limit the sequence length of backpropagation through time-truncated backpropagation (TBPTT), reduce the computational burden, and use the gated cross-attention mechanism to further optimize the computational efficiency of the advertising recommendation model when processing large-scale user data, reducing the processing of irrelevant information. 3) Generalization of the ad recommendation system: Ad recommendation systems struggle to balance multiple objectives during training, resulting in insufficient generalization and recommendations limited to a single dimension. This embodiment of the present application combines a decoding time algorithm with the weights of different ad recommendation models to determine a linear combination of the prediction results of multiple ad recommendation models, improving the diversity and generalization of recommendation results. Furthermore, the decoding strategy is optimized using the Legendre transform, achieving more efficient and accurate ad recommendations. 4) Accuracy in long text processing: Long text processing often faces issues of information redundancy and inaccuracy, making it difficult for ad recommendation models to effectively extract valuable information. This embodiment of the present application introduces a gated cross-attention mechanism to enhance the model's processing capabilities for long texts, improving the accuracy of ad content generation and recommendations. The gated mechanism effectively filters out redundant information, focusing on relevant content and enhancing the user experience. 5) Uncertainty in generated results: When faced with insufficient data or high uncertainty, the quality of recommendations for ad recommendation models can easily decline, leading to low user satisfaction. When the ad recommendation model is highly uncertain about the generated prediction results, this embodiment of the present application activates a knowledge retrieval mechanism to reorder the retrieved information, reducing the uncertainty of the generated results and ensuring the relevance and reliability of the recommendations.

[0168] Figure 10 This is a flow chart of the information processing method provided by the embodiment of the present application. Figure 10, step 301, data preparation and data preprocessing: collect user historical data advertising data and clean it. Step 302, feature encoding: encode user and advertising features and combined features. Step 303, model selection modeling and training: use the memory-enhanced Transformer model as the advertising recommendation model, in which the encoder and decoder share weights, and use the time truncated back propagation (TBPTT) algorithm to improve the training process of the advertising recommendation model. Step 304, loss function design: introduce a multi-task loss function, which is obtained by combining the CTR prediction loss and the advertising ranking loss. Step 305, model reasoning and preference knowledge retrieval: when the advertising recommendation model has a high degree of uncertainty in the generated prediction results, activate the retrieval mechanism to minimize the uncertainty. The details are explained below.

[0169] In step 301, data collection is first performed. User data, advertising data (corresponding to the second feature in the above embodiment) and context data (corresponding to the seventh feature in the above embodiment) can be obtained. Among them, user data includes user attribute data and user behavior data, wherein user attribute data (corresponding to the sixth feature in the above embodiment) includes but is not limited to the age, gender, interests, etc. of the user (corresponding to the first object in the above embodiment), and user behavior data (corresponding to the eighth feature in the above embodiment) includes but is not limited to the user's historical click, browsing, purchase data, social media activities, and geographic location, etc. For example, Table 1 is the user behavior data of the user whose identity document (ID) is 12345. The user attribute data can be: user ID: 12345, age: 30, gender: male, geographic location: city X.

[0170] Table 1 User behavior data

[0171] User ID: 12345 Time: 2024-06-01 08:00:00, Action: Click, Ad ID: A5678 Time: 2024-06-02 12:00:00, Action: Browsing, Product ID: P4321 Time: 2024-06-03 15:30:00, Action: Click, Ad ID: B9876 Time: 2024-06-04 10:45:00, Action: Click, Ad ID: C1234 Time: 2024-06-05 14:20:00, Action: Browsing, Product ID: P8765

[0172] Ad data includes ad display data and ad content data. Ad display data includes, but is not limited to, ad ID, ad location, ad delivery time, ad click-through rate, and ad response status. Ad content data includes, but is not limited to, ad category, ad content, ad description, and target audience. For example, Table 2 shows ad display data for an ad. Table 3 shows ad content data for an ad.

[0173] Table 2 Advertisement display data

[0174]

[0175] Table 3 Advertisement content data

[0176] Ad ID: A5678, Category: Electronics, Creative: New Mobile Phones, Description: The latest flagship mobile phone Ad ID: B9876, Category: Home Furnishings, Idea: New Sofa, Description: High-quality modern sofa Ad ID: C1234, Category: Home Appliances, Creative: Promotion, Description: Brand home appliances full discount event

[0177] Contextual data includes but is not limited to time information (such as date, time period), device information (such as device type, operating system), page information (such as page category, content), etc. Then, data preprocessing is performed. Data preprocessing includes data cleaning, data conversion and data aggregation. Data cleaning: Check and process missing data. Among them, for data with missing values, interpolation or mean filling and other methods can be used to process missing data. For data with outliers, abnormal data points can be removed or corrected. Data conversion: Convert the format of the cleaned data, such as standardizing the date format, unifying user identification, etc. For example, the timestamp is converted to date, time and relative time features. Data aggregation: Aggregate behavioral data according to user ID and ad ID to form a long behavior sequence. For example, the long behavior sequence of a user is: User ID: 12345, behavior sequence: [click on ad A5678, browse product P4321, click on ad B9876, click on ad C1234, browse product P8765].

[0178] In step 302, features are first extracted from the acquired data. The extracted features include basic features, combined features, and sequence features. Basic features include user features (corresponding to the sixth feature in the above embodiment, such as age, gender, and interests), ad features (corresponding to the second feature in the above embodiment, such as type, length, and content), and context features (corresponding to the seventh feature in the above embodiment, such as time, location, and device). Combined features can be created by creating user-ad combined features (corresponding to the ninth feature in the above embodiment) or user-context combined features (corresponding to the tenth feature in the above embodiment). User-ad combined features: Key features of the ad and user, such as ad category, user age, and user gender, are selected and combined using a concatenation method. For example, a user interest combined feature: The user's historical clicks on ad categories are summed to obtain the user's interest weights for different ad categories. User A's interest weight for ad category 1 = sum(historical click count for ad category 1), and user A's interest weight for ad category 2 = sum(historical click count for ad category 2). Ad feature combination: Combines ad category, content type, and publishing platform to form more fine-grained ad description features. For example, Ad A's combined features = Ad category + Content type + Publishing platform. Creation of user-context combined features: Considers the user's current environment and contextual information, such as time, location, and device type. These seventh features are combined with the user's historical behavior or personal attributes to better understand user behavior patterns under different contextual conditions. For example, a time-context combined feature: Combines the current hour and day of the week with the user's historical click behavior to analyze user activity during different time periods. User A's click count on Monday morning = count (number of ad clicks by user A on Monday morning). Geographic location combined features: Combines the user's current location with their preferred ad type to analyze the impact of location on ad preferences. User A's interest weight for ad category 1 in city X = sum (number of historical clicks by user A on ad category 1 in city X). Sequence features are long user behavior sequences derived from data aggregation. In the embodiment of the present application, basic features and sequence features can be used as input features of the advertising recommendation model, and basic features, combined features and sequence features can be used as input features of the advertising recommendation model (corresponding to the eleventh feature in the above embodiment).

[0179] The extracted features are then encoded. The processing methods for each type of feature are as follows: Numerical features can be standardized or normalized; categorical features can use One-Hot encoding or embedding encoding; text features can use Word2Vec embedding vectors to represent the text content; and sequence features can be encoded by first dividing a long behavioral sequence into multiple short subsequences, each of which is then encoded using a sequence embedding Long Short-Term Memory (LSTM) network. Exemplary feature encoding can include user feature encoding and ad feature encoding. The encoded user feature vector serves as an additional input to the ad recommendation model, helping the model understand the user's basic characteristics and attributes and tailor recommendations accordingly. For example, the user features "User ID: 12345, Age: 30, Gender: Male, Location: City X" are encoded to "Age Code: 30, Gender Code: 1 (Male), Location Code: City X." The encoded ad feature vector serves as one of the inputs to the ad recommendation model, describing the key features and attributes of the ad, helping the model match and recommend ads based on their content and attributes to improve click-through rates or other target metrics. For example, if the ad features are the ad content data in Table 3, the encoding yields the following category codes: electronic products are [1, 0, 0], household items are [0, 1, 0], and home appliances are [0, 0, 1]. The descriptions are TF-IDF encoded: ad A5678 is [0.2, 0.4, 0.1], ad B9876 is [0.3, 0.2, 0.5], and ad C1234 is [0.1, 0.3, 0.4]. Creative encoding: new mobile phone is [0.8], new sofa is [0.6], and special promotion is [0.7]. Combining the vectors encoded from user and ad features with the vectors encoded from user behavior sequences can achieve more accurate personalized matching. Ad recommendation models can use these vectors to calculate the similarity or match between users and ads, thereby determining which ads to recommend to the user.

[0180] In step 303, the advertising recommendation model chooses to use a memory-enhanced Transformer model in which the encoder and decoder share weights. The encoder can encapsulate the feature vector obtained after input feature encoding (corresponding to the first feature vector in the above embodiment) into the memory, and the decoder uses the feature vector in the memory to predict subsequent segments. For example, assuming that the vector of segment 1 in the feature vector represents browsing ad 1 at time t1, and the vector of segment 2 represents clicking ad 2 at time t2, the decoder can predict the user's behavior at time t3. Each decoder layer includes an additional gated cross-attention layer to improve efficiency and content accuracy. The input to the decoder usually includes the decoder hidden state h at the current time step t t, the output of the encoder is expressed as H = {h1, h2, ..., h T}, T is the length of the encoder input sequence. The gated cross attention layer first receives the current hidden state h of the decoder t , and dynamically selects the encoder output sequence H that matches the current decoder state h according to the decoder's hidden state t The most relevant information, which may include the user's historical behavior, ad attributes, contextual information, etc. The gated cross attention layer calculates the decoder hidden state h t The correlation between the encoder output and the encoder output is used to obtain the weight. The gating mechanism uses these weights to perform a weighted summation on the encoder output sequence H to obtain a weighted encoder context vector, which represents the information that the decoder needs to pay attention to at the current time step. The weighted encoder context vector (corresponding to the thirteenth feature in the above embodiment) is used as the input of the decoder at the next time step to ensure that the generated recommendation content matches the user's current needs. The weight calculation process is as follows: The decoder hidden state h is calculated using the softmax function t The similarity between the encoder output h and each encoder output h is converted into weights to obtain a weight matrix. The weights represent the importance of each encoder hidden state in the current decoder state. The weight matrix is ​​then processed through a sigmoid function to make large weights larger and small weights smaller, resulting in the final attention weights. Based on the calculated attention weights, the encoder output sequence H is weighted and summed to obtain the weighted encoder context vector.

[0181] The gating mechanism effectively filters and focuses attention on information relevant to recommendations, preventing irrelevant information from interfering with recommendation results and improving content accuracy and quality. When predicting click-through rate (CTR), the ad recommendation model receives the current user's user features and ad features as input. At each decoder layer, the gated crisscross attention layer dynamically selects and integrates the most relevant user and ad features based on the current decoder hidden state and encoder output. For example, it can prioritize attention to specific ad types based on the user's current context and historical behavior, or adjust attention weights based on the ad's historical CTR and the degree to which the content matches the user's preferences. After processing with the gated crisscross attention mechanism, the decoder layer generates a representation that incorporates weighted user and ad features. This representation is passed to the output layer for predicting the ad's CTR. The output layer uses the softmax function to calculate the probability of an ad being clicked based on the weights and feature representations learned by the model.

[0182] The advertising recommendation model provided in the embodiment of the present application has also been incrementally optimized, that is, the time-truncated back propagation (TBPTT) algorithm is used to improve the training process of the memory-enhanced Transformer model. Time-truncated back propagation (TBPTT): used to process long behavior sequences of users in advertising recommendation. In the advertising recommendation scenario, the user's behavior history (such as clicking on ads, browsing products, etc.) is a long behavior sequence, and the relationship between behaviors is crucial to the decision-making of the recommendation system. The advertising recommendation system needs to predict the future probability of ad clicks or generate personalized recommendations based on the user's long behavior sequence. This long behavior sequence includes: the user's historical click records, browsing behavior, and the time series of interaction with advertisements (the moment of clicking or browsing ads, and the duration of stay, etc.). The time-truncated back propagation algorithm first performs sequence segmentation: divides the long behavior sequence into shorter subsequences so that these subsequences can be processed in each training iteration of the advertising recommendation model, reducing the computational burden of the advertising recommendation model. Example: A long action sequence is: click ad A → browse product B → click ad C → click ad D → browse product E. After segmentation, the following sequence is: Sequence 1: click ad A → browse product B, Sequence 2: click ad C → click ad D, Sequence 3: browse product E. Then, during model training, forward propagation is performed: for each subsequence, an encoder with shared weights is used to process it, generating a contextual representation H. Example: Sequence 1 is input into the encoder, which outputs H1. Sequence 2 and H1 are input into the encoder, which outputs H2. Sequence 3 and H2 are input into the encoder, which outputs H3. Backpropagation: Backpropagation is performed only on the loss of the current subsequence, limiting the gradient update to the length of the current subsequence to control computational overhead. Example: Backpropagation for sequence 2 is performed only between H1 and H2. State update: The hidden state after each processing is passed to the encoder processing the next subsequence to preserve contextual information. Example: H1 is passed to the encoder processing sequence 2, and H2 is passed to the encoder processing sequence 3. That is, the training process is: sequence 1 → encoder output H1 → calculate loss and back propagate, sequence 2 uses H1 → encoder output H2 → calculate loss and back propagate, sequence 3 uses H2 → encoder output H3 → calculate loss and back propagate.

[0183] The advertising recommendation system provided in the embodiment of the present application also integrates multiple advertising recommendation models, wherein the decoding time algorithm is used to combine the weights of different advertising recommendation models to output the next recommendation result from the linear combination of the prediction results of all advertising recommendation models. The weights of the advertising recommendation models are dynamically transformed, and the weights of the advertising recommendation models are calculated by Legendre transform. In the advertising recommendation system, multiple advertising recommendation models may give different prediction results for the same input features. In order to obtain the final recommendation result, it is necessary to combine these prediction results. The decoding time algorithm obtains a comprehensive prediction result by linearly combining the prediction results of different models, wherein the weight of each advertising recommendation model is calculated using Legendre transform (corresponding to the weight of the third interaction probability in the above embodiment), and the comprehensive prediction result is determined based on the weight of each advertising recommendation model and the prediction result of each model to obtain the final recommendation result.

[0184] Multi-model prediction: Assume there are N ad recommendation models, each ad recommendation model fi generates a prediction result yi for a given input feature x. The prediction result yi can satisfy the following formula (1).

[0185] yi=fi(x),i=1,2,...,N Formula (1);

[0186] Where N is the number of advertising recommendation models, x is the input feature of the advertising recommendation model, fi is the i-th advertising recommendation model, and yi is the prediction result output by the i-th advertising recommendation model (that is, the probability distribution of the user's click behavior on multiple candidate ads predicted by the advertising recommendation model).

[0187] Weight of the ad recommendation model: Assign different weights α to the prediction results of each ad recommendation model i , these weights represent the importance of different advertising recommendation models. The weight α of the advertising recommendation model i The following formula (2) is satisfied.

[0188]

[0189] in, To convert the weight α1 of the i-th advertising recommendation model to the weight α of the N-th advertising recommendation model N The sum of .

[0190] Linear combination formula: The prediction results of all advertising recommendation models are weighted and combined according to their weights. The resulting linear combination is the linear combination result. The following formula (3) is satisfied.

[0191]

[0192] Among them, α i is the weight of the i-th advertising recommendation model, weight α i Set based on historical data or empirical values ​​in actual applications. yi is the prediction result of the i-th advertising recommendation model, α i yi is the product of the weight of the i-th advertising recommendation model and the prediction result of the i-th advertising recommendation model, The linear combination result of the prediction results of all ad recommendation models.

[0193] Legendre transform is used to convert the original problem into a dual problem that is easier to solve, so as to find a closed-form solution. A closed-form solution usually refers to a mathematical problem (especially an equation or a system of equations) whose solution can be expressed by a clear formula. A closed-form solution means that for any variable in the equation, a specific numerical value can be substituted and the corresponding solution can be obtained through this formula. In the embodiment of the present application, the original problem is to minimize the loss function It can be expressed as The transformation process is to use Legendre transformation to transform the loss function Transformed into a dual function L(λ) (corresponding to the conjugate function in the above embodiment), λ is the dual variable, that is, the weight variable of each advertising recommendation model. The purpose of Legendre transformation is to find the optimal weight α of each advertising recommendation model. i The dual function L(λ) satisfies the following formula (4).

[0194]

[0195] in, Based on the linear combination result Determine the loss function, λ is the dual variable in Legendre transformation, L(λ) is the result of Legendre transformation, that is, the loss function The dual function obtained after the transformation (corresponding to the conjugate function in the above embodiment) is (inverted triangle operator) is a vector differential operator used to represent Partial derivatives of variables.

[0196] Solve the dual problem: Find the solution λ* that maximizes the dual function L(λ). The solution λ* satisfies the following formula (5).

[0197]

[0198] Among them, argmax is a mathematical symbol, indicating the maximum value parameter, That is, within the domain of the function L(λ), we need to find a specific value of λ, denoted as λ*, so that L(λ) reaches its maximum value at λ*. In short, λ* is the value of λ that maximizes L(λ).

[0199] Restore the original solution: Determine the original solution through the solution λ* of the dual problem. The original solution is the optimal linear combination result. For some specific types of loss functions, Legendre transform can provide a closed form solution. For example, when the loss function is square error loss, for the square error loss function After the above steps determine the dual function L(λ) and the solution λ* that maximizes the dual function L(λ), the linear combination result in the original loss function can be calculated based on the maximized solution λ*. Linear combination results Among them, y is the true label of the click behavior (1 represents click), The linear combination result of the prediction results of multiple ad recommendation models.

[0200] According to the linear combination results And the closed solution of Legendre transform, define an efficient decoding strategy: Calculate the weighted sum: Calculate the weighted sum of the prediction results of all advertising recommendation models in real time as the comprehensive prediction result. Comprehensive prediction result The following formula (6) is satisfied.

[0201]

[0202] Among them, y i,t is the prediction result of the i-th advertising recommendation model at time t, is the comprehensive prediction result of N advertising recommendation models at time t, α i is the weight of the i-th ad recommendation model.

[0203] Dynamically adjust weights: Based on the current context or the user's historical behavior, dynamically adjust the weights α of each ad recommendation model i , to adapt to different scenarios. Based on the comprehensive prediction results, the next recommended content is determined. Token is any recommended content in the advertising system. It is a comprehensive prediction result obtained by weighting the prediction results (prediction probability) of N advertising recommendation models for the recommended content token. Argmax is a mathematical symbol representing the maximum parameter. Indicates finding a comprehensive prediction result among all possible recommended content tokens The largest token is selected as the next recommended content. In other words, from all candidate ads, the candidate ad with the highest comprehensive predicted probability is selected as the next recommended content.

[0204] Feedback mechanism: Continuously adjust and optimize the weight α based on the actual recommendation effect i , to improve the accuracy of the decoding strategy. Real-time update: In actual advertising recommendation systems, model prediction results and weights are updated regularly to adapt to changes in user preferences and advertising content.

[0205] In step 304, multi-objective optimization can be performed. A multi-task loss function is introduced, and the multi-task loss function can be obtained by combining multiple objectives (clicks, conversions, and views). For example, a multi-task loss function can be obtained by combining the CTR prediction loss and the ad ranking loss. For the click-through rate (CTR) prediction loss, the cross entropy loss function can be used to measure the difference between the predicted click probability and the actual click. For the ad ranking loss, the pairwise ranking loss function can be used to optimize the ad ranking. The click-through rate (CTR) loss function L(CTR) satisfies the following formula (7).

[0206]

[0207] Among them, L(CTR) is the click-through rate loss function of the advertising recommendation model, M is the number of candidate ads, and y j is the prediction result obtained by the ad recommendation model for the jth candidate ad, that is, the click probability, p j is the actual click behavior label of the jth candidate ad, p j =1 means the jth candidate ad is clicked, p j =0 means that the jth candidate ad was not clicked. Assume that there are three candidate ads in the data example: Ad A5678: click probability = 1 (clicked), click behavior label is 1; Ad B9876: click probability = 0 (not clicked), click behavior label is 0; Ad C1234: click probability = 1 (clicked), click behavior label is 1; then

[0208] In step 305, preference knowledge retrieval refers to activating the retrieval mechanism when the prediction result generated by the advertising recommendation model has a high degree of uncertainty, and reordering the retrieved knowledge fragments (corresponding to the second information in the above embodiment) according to the relevance score to minimize the uncertainty. Preference knowledge retrieval is used to retrieve relevant information from an external knowledge base or preference data (corresponding to the information retrieval library in the above embodiment) when the prediction result of the advertising recommendation model has a high degree of uncertainty, so as to improve the accuracy of the prediction result. The whole process includes four main steps: uncertainty detection, retrieval triggering, knowledge integration and sorting. Uncertainty detection: Use the prediction probability output within the advertising recommendation model to measure uncertainty. For example, analyze the probability distribution of the prediction result and calculate indicators such as entropy or variance. Entropy satisfies the following formula (8).

[0209] Entropy=-∑ j yj l ogy j Formula (8);

[0210] Among them, Entropy is the entropy of the advertising recommendation model, as an uncertainty indicator, y j is the predicted probability output by the ad recommendation model for the jth candidate ad, logy j is the logarithm of the predicted probability of the jth candidate ad, y j logy j is the product of the predicted probability and logarithm of the jth candidate advertisement, ∑ j y j logy j The product y of all advertisements is j logy j The sum of .

[0211] When the uncertainty indicator exceeds a preset threshold, preference knowledge retrieval is triggered. For example, when the entropy value exceeds a certain value, it indicates that the ad recommendation model lacks confidence in the results and a retrieval is required. After the retrieval is triggered, the candidate ad with the highest predicted probability is determined based on the probability distribution output by the ad recommendation model. The feature that best reflects the user's preference is selected from the input features of the ad recommendation model (for example, the ad with the highest predicted probability is a watch, and the feature that reflects the user's preference is the discount the user wants for the watch, etc.). A query statement is generated based on the ad features of the candidate ad with the highest predicted probability and the features that best reflect the user's preference. Query statement = f(preference, ad content), where f is the function that generates the query. Retrieval strategy: Use the generated query statement to search an external knowledge base or preference data (e.g., XX shopping platform) to obtain multiple ads, where multiple ads are not within the range predicted by the ad recommendation model. The similarity between the query statement and each ad is calculated, and multiple ads are selected based on the similarity. The content of the external knowledge base includes articles related to the user's interests, ad data, recommendation history, etc., and can be structured as a structured database or text corpus. Then, re-rank the ads according to the relevance score (similarity) and determine the ad with the highest relevance score. Perform knowledge integration, pre-process the ad with the highest relevance score, and extract ad-related features or information. Fuse the ad-related features or information with the input features to generate enhanced input features. Enhanced input features = h (ad-related features or information, input features), where h is a fusion function that can be fused in a splicing manner. Minimize uncertainty: Re-input the enhanced input features into the ad recommendation model, update the prediction results of the ad recommendation model, and estimate the uncertainty index again. If the uncertainty index is significantly reduced, stop the search; otherwise, further search or strategy adjustment may be required. If the uncertainty index (such as entropy) of the prediction result obtained after the retrieval is lower than the threshold, confirm the recommendation result, otherwise repeat the retrieval process.

[0212] Suppose the ad recommendation model generates a recommendation with high uncertainty, such as ad C1234. The retrieval mechanism is activated to reorder the retrieved ads to minimize the uncertainty. User feedback or market trend data related to ad C1234 is retrieved to provide more accurate recommendations or further analysis.

[0213] The embodiment of the present application introduces a gated cross-attention mechanism to enhance the advertising recommendation model's ability to process long texts, thereby improving the accuracy of advertising content generation and recommendation. The gating mechanism effectively filters redundant information, focuses on relevant content, and enhances user experience. The embodiment of the present application uses a time-truncated back propagation TBPTT algorithm to limit the sequence length of back propagation and reduce the computational burden; the gated cross-attention mechanism further optimizes the computational efficiency of the advertising recommendation model when processing large-scale user data, and reduces the processing of irrelevant information. The embodiment of the present application uses a decoding time algorithm to combine the weights of different advertising recommendation models, and improves the diversity and generalization of recommendation results through the predicted linear combination of multiple advertising recommendation models. At the same time, the decoding strategy is optimized using Legendre transform to achieve more efficient and accurate advertising recommendations.

[0214] The following further describes an exemplary structure of the information processing device 455 provided in the embodiment of the present application as a software module. In some embodiments, for example, Figure 2 As shown, the software modules stored in the information processing device 455 of the memory 450 may include:

[0215] The prediction module 4551 is used to perform a first prediction operation on the interaction behavior of the first object and the first information based on the first feature of the first object and the second feature of the first information, and obtain a first interaction probability between the first object and the first information; the confidence determination module 4552 is used to determine the first confidence of the first prediction operation based on the first interaction probability; the retrieval module 4553 is used to determine the second information based on the first feature and the second feature when the first confidence is less than a first preset threshold; the recommendation information determination module 4554 is used to determine recommended information from the first information based on the third feature, the first feature and the second feature of the second information.

[0216] In some embodiments, the retrieval module 4553 is also used to filter out a fourth feature from the second feature based on the first interaction probability; generate a query statement based on the fourth feature and the first feature; determine the similarity between the query statement and the third information in the information retrieval library, and determine the second information from the third information based on the similarity.

[0217] In some embodiments, the recommendation information determination module 4554 is also used to fuse the third feature, the first feature, and the second feature to obtain a fifth feature; based on the fifth feature, perform a second prediction operation on the interaction behavior of the first object and the first information to obtain a second interaction probability between the first object and the first information; based on the second interaction probability, determine a second confidence level of the second prediction operation; when the second confidence level is greater than or equal to the first preset threshold, determine the recommended information from the first information based on the second interaction probability.

[0218] In some embodiments, the first feature includes a sixth feature, a seventh feature, and an eighth feature. The prediction module 4551 is further configured to combine the sixth feature and the second feature to obtain a ninth feature, combine the sixth feature and the seventh feature to obtain a tenth feature, fuse the ninth feature, the tenth feature, the sixth feature, the second feature, and the eighth feature to obtain an eleventh feature, and perform a first prediction operation on the interaction between the first object and the first information based on the eleventh feature to obtain a first interaction probability.

[0219] In some embodiments, the prediction module 4551 is also used to perform a third prediction operation on the interaction behavior of the first object and the first information based on the eleventh feature to obtain a third interaction probability between the first object and the first information; determine the weight of the third interaction probability; and fuse the third interaction probability based on the weight of the third interaction probability to obtain the first interaction probability.

[0220] In some embodiments, the prediction module 4551 is further used to encode the eleventh feature to obtain a first feature vector; perform attention processing on the first feature vector to obtain a twelfth feature; perform gating processing on the twelfth feature to obtain a thirteenth feature; and decode the thirteenth feature to obtain a third interaction probability.

[0221] In some embodiments, the prediction module 4551 is also used to fuse the third interaction probability of the first information based on the first weight of the first model to obtain the fourth interaction probability of the first information; update the loss function of the first model to the conjugate function corresponding to the first model; determine the conjugate function value based on the fourth interaction probability, the label of the first object's interaction behavior with respect to the first information, and the conjugate function; update the first weight based on the conjugate function value to obtain the second weight of the first model; and determine the second weight as the weight of the third interaction probability.

[0222] In some embodiments, there are multiple pieces of first information. The confidence determination module 4552 is further configured to perform logarithmic processing on the first interaction probability for each piece of first information to obtain a first value; multiply the first value by the first interaction probability to obtain a second value; and determine the first confidence based on the multiple second values.

[0223] An embodiment of the present application provides a computer program product, which includes a computer program or computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, causing the electronic device to perform the information processing method described in the embodiment of the present application.

[0224] The embodiment of the present application provides a computer-readable storage medium in which computer-executable instructions or computer programs are stored. When the computer-executable instructions or computer programs are executed by a processor, the processor will execute the information processing method provided by the embodiment of the present application, for example, Figure 3 The information processing method shown.

[0225] In some embodiments, the computer-readable storage medium may be a memory such as RAM, ROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or may be various devices including one or any combination of the above memories.

[0226] In some embodiments, computer-executable instructions may be in the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0227] As an example, computer-executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinating files (e.g., files storing one or more modules, subroutines, or code portions).

[0228] By way of example, computer-executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one site, or on multiple electronic devices distributed across multiple sites and interconnected by a communication network.

[0229] In summary, the gated cross-attention mechanism is introduced through the embodiment of the present application, which enhances the model's ability to process long texts, optimizes the computational efficiency of the information recommendation model when processing large-scale user data, and reduces the processing of irrelevant information. At the same time, the embodiment of the present application improves the diversity and generalization ability of the recommendation results, reduces the uncertainty of the generated results, and ensures the relevance and reliability of the recommendations.

[0230] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present application are included in the scope of protection of the present application.

Claims

1. An information processing method, characterized in that: The method comprises: Based on a first feature of a first object and a second feature of first information, performing a first prediction operation on an interaction behavior between the first object and the first information to obtain a first interaction probability between the first object and the first information; determining a first confidence level of the first prediction operation based on the first interaction probability; When the first confidence level is less than a first preset threshold, determining second information based on the first feature and the second feature; Recommended information is determined from the first information based on the third feature of the second information, the first feature, and the second feature.

2. The method according to claim 1, characterized in that The determining the second information based on the first feature and the second feature includes: Based on the first interaction probability, a fourth feature is obtained by filtering the second features; generating a query statement based on the fourth feature and the first feature; The similarity between the query statement and third information in the information retrieval database is determined, and second information is determined from the third information based on the similarity.

3. The method according to claim 1, characterized in that The determining the recommended information from the first information based on the third feature of the second information, the first feature, and the second feature includes: fusing the third feature, the first feature, and the second feature to obtain a fifth feature; Based on the fifth feature, performing a second prediction operation on the interaction behavior between the first object and the first information to obtain a second interaction probability between the first object and the first information; determining a second confidence level of the second prediction operation based on the second interaction probability; When the second confidence level is greater than or equal to the first preset threshold, recommended information is determined from the first information based on the second interaction probability.

4. The method according to claim 1, wherein The first feature includes a sixth feature, a seventh feature, and an eighth feature. The performing a first prediction operation on the interaction behavior between the first object and the first information based on the first feature of the first object and the second feature of the first information to obtain a first interaction probability between the first object and the first information includes: The sixth feature and the second feature are combined to obtain a ninth feature, and the sixth feature and the seventh feature are combined to obtain a tenth feature; fusing the ninth feature, the tenth feature, the sixth feature, the second feature, and the eighth feature to obtain an eleventh feature; Based on the eleventh feature, a first prediction operation is performed on the interaction behavior between the first object and the first information to obtain the first interaction probability.

5. The method according to claim 4, characterized in that The performing a first prediction operation on the interaction behavior between the first object and the first information based on the eleventh feature to obtain the first interaction probability includes: Based on the eleventh feature, performing a third prediction operation on the interaction behavior between the first object and the first information to obtain a third interaction probability between the first object and the first information; determining a weight of the third interaction probability; The third interaction probabilities are fused based on the weights of the third interaction probabilities to obtain the first interaction probability.

6. The method according to claim 5, characterized in that The performing a third prediction operation on the interaction behavior between the first object and the first information based on the eleventh feature to obtain a third interaction probability between the first object and the first information includes: Encoding the eleventh feature to obtain a first feature vector; Performing attention processing on the first feature vector to obtain a twelfth feature; Performing gate processing on the twelfth feature to obtain a thirteenth feature; The thirteenth feature is decoded to obtain the third interaction probability.

7. The method according to claim 5, characterized in that Determining the weight of the third interaction probability includes: fusing the third interaction probability of the first information based on the first weight of the first model to obtain a fourth interaction probability of the first information; Updating the loss function of the first model to a conjugate function corresponding to the first model; determining a conjugate function value based on the fourth interaction probability, a label of the interaction behavior of the first object with respect to the first information, and the conjugate function; updating the first weight based on the conjugate function value to obtain a second weight of the first model; The second weight is determined as a weight of the third interaction probability.

8. The method according to any one of claims 1 to 7, characterized in that The number of the first information is multiple; The determining, based on the first interaction probability, a first confidence level of the first prediction operation includes: For each piece of first information, performing logarithmic processing on the first interaction probability to obtain the first value; multiplying the first value by the first interaction probability to determine a second value; The first confidence level is determined based on a plurality of the second values.

9. An electronic device, characterized in that: The electronic device comprises: a memory for storing computer-executable instructions or computer programs; The processor is configured to implement the information processing method according to any one of claims 1 to 8 when executing the computer-executable instructions or computer programs stored in the memory.

10. A computer-readable storage medium storing computer-executable instructions or a computer program, characterized in that: When the computer-executable instructions or computer program are executed by a processor, the information processing method according to any one of claims 1 to 8 is implemented.

11. A computer program product comprising computer executable instructions or a computer program, characterized in that When the computer-executable instructions or computer program are executed by a processor, the information processing method according to any one of claims 1 to 8 is implemented.

Images