Information recommendation method and device based on artificial intelligence, equipment and storage medium

Abstract

The application provides an information recommendation method and device based on artificial intelligence, an electronic device and a computer readable storage medium; the method comprises: performing feature cross processing on a first part of attribute information in a plurality of attribute information of to-be-recommended information to obtain cross features of the to-be-recommended information; performing encoding processing on a second part of attribute information in the plurality of attribute information of the to-be-recommended information to obtain encoding features of the to-be-recommended information; performing attention processing based on the cross features and the encoding features to obtain attention features of the to-be-recommended information; performing tendency prediction processing based on the attention features of the to-be-recommended information to obtain a recommended scene to which the to-be-recommended information is adapted; and performing a recommendation operation based on the recommended scene to which the to-be-recommended information is adapted. Through the application, information can be recommended in a targeted manner, and the accuracy of information recommendation is improved.

Description

Technical Field

[0001] This application relates to artificial intelligence technology, and more particularly to an information recommendation method, apparatus, electronic device, and computer-readable storage medium based on artificial intelligence. Background Technology

[0002] Artificial Intelligence (AI) is a comprehensive technology within computer science that studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities. AI technology is a multidisciplinary field, encompassing a wide range of areas, including natural language processing and machine learning / deep learning. With technological advancements, AI will be applied in more fields and play an increasingly important role.

[0003] In recommender systems, both information cold start and user cold start are crucial research areas. Related technologies employ numerous cold start attempts to supplement user interaction data and the interaction data of the information to be recommended, thereby achieving both information cold start and user cold start.

[0004] However, this cold start attempt is achieved through random probing, which severely impacts the accuracy of information recommendation. Summary of the Invention

[0005] This application provides an information recommendation method, apparatus, electronic device, and computer-readable storage medium based on artificial intelligence, which can recommend information in a targeted manner and improve the accuracy of information recommendation.

[0006] The technical solution of this application embodiment is implemented as follows:

[0007] This application provides an information recommendation method based on artificial intelligence, including:

[0008] The first part of the attribute information among the multiple attribute information of the information to be recommended is subjected to feature cross processing to obtain the cross features of the information to be recommended.

[0009] Encode the second part of the attribute information among the multiple attribute information of the information to be recommended to obtain the encoded features of the information to be recommended.

[0010] Attention processing is performed based on the cross features and the encoding features to obtain the attention features of the information to be recommended.

[0011] Based on the attention features of the information to be recommended, a bias prediction process is performed to obtain the recommendation scenario that the information to be recommended is adapted to.

[0012] The recommendation operation is performed based on the recommendation scenario adapted to the information to be recommended.

[0013] In the above technical solution, the step of performing feature cross-processing on the first part of the attribute information among the multiple attribute information of the information to be recommended to obtain the cross-features of the information to be recommended includes:

[0014] Encode the first part of the attribute information among the multiple attribute information of the information to be recommended to obtain the encoded features of the first part of the attribute information;

[0015] The encoded features of the first part of attribute information are subjected to second-order feature cross processing to obtain the second-order cross features of the first part of attribute information.

[0016] The second-order cross features of the first part of the attribute information are mapped to obtain the cross features of the information to be recommended.

[0017] In the above technical solution, the first part of the attribute information includes fixed attribute information and self-media attribute information;

[0018] The step of encoding the first part of the attribute information among the multiple attribute information of the information to be recommended to obtain the encoded features of the first part of the attribute information includes:

[0019] The fixed attribute information of the information to be recommended is discretized to obtain the discrete features of the fixed attribute information;

[0020] The discrete features of the fixed attribute information are vectorized to obtain the encoded features of the fixed attribute information.

[0021] The self-media attribute information of the information to be recommended is vectorized to obtain the encoded features of the self-media attribute information.

[0022] In the above technical solution, the step of performing attention processing based on the cross features and the encoding features to obtain the attention features of the information to be recommended includes:

[0023] The cross features and the encoded features are concatenated to obtain the concatenated features of the information to be recommended.

[0024] Self-attention processing is performed on the spliced ​​features of the information to be recommended to obtain the attention features of the information to be recommended.

[0025] In the above technical solution, the step of performing bias prediction processing based on the attention features of the information to be recommended to obtain the recommendation scenario adapted to the information to be recommended includes:

[0026] The attention features of the information to be recommended are mapped to obtain the probability distribution of the candidate recommendation scenarios;

[0027] The candidate recommendation scenario corresponding to the highest probability in the probability distribution is determined as the recommendation scenario that the information to be recommended is adapted to.

[0028] In the above technical solution, the step of performing a recommendation operation based on the recommendation scenario adapted to the information to be recommended includes:

[0029] Based on the recommendation scenario adapted to the information to be recommended, the information to be recommended is stored in the database corresponding to the recommendation scenario;

[0030] When it is determined that the target user is suitable for the recommendation scenario based on the target user's profile information, at least one piece of information to be recommended from the database corresponding to the recommendation scenario is recommended to the target user. This application embodiment provides an artificial intelligence-based information recommendation method, including:

[0031] Information recommendation tasks are achieved by invoking a bias model;

[0032] The training process of the propensity model includes the following steps:

[0033] Retrieve information samples from the database corresponding to the recommended scenario samples;

[0034] The first part of the attribute information in the multiple attribute information of the information sample is subjected to feature cross processing to obtain the cross features of the information sample;

[0035] Based on the encoding features and the cross features of the information sample, the attention features of the information sample are determined;

[0036] Based on the attention features of the information sample, a biased prediction process is performed to obtain the predicted recommendation scenario of the information sample.

[0037] Based on the predicted recommendation scenario of the information sample and the recommendation scenario sample, the loss function of the bias model is constructed.

[0038] Update the parameters of the bias model, and use the updated parameters of the bias model when the loss function converges as the parameters of the trained bias model.

[0039] The method in the above technical solution further includes:

[0040] Determine the attribute information of the user samples corresponding to the recommendation scenario samples;

[0041] Obtain the user set formed by users possessing the aforementioned attribute information;

[0042] Obtain a set of recommendation information formed by historical recommendation information that has interacted with users in the user set;

[0043] Obtain the interaction data between each user in the user set and each piece of historical recommendation information in the recommendation information set;

[0044] Based on the interaction data, historical recommendation information that matches the recommendation scenario sample is selected from the recommendation information set as information samples to construct a database corresponding to the recommendation scenario sample.

[0045] In the above technical solution, the step of selecting recommendation information that matches the recommendation scenario sample from the recommendation information set based on the interaction data as an information sample includes:

[0046] Based on the interaction data between any user and each historical recommendation in the recommendation information set, the historical recommendation information in the recommendation information set is sorted in descending order, and the portion of the historical recommendation information that appears first in the descending order is used as an information sample that is adapted to the recommendation scenario sample.

[0047] In the above technical solution, the step of selecting recommendation information that matches the recommendation scenario sample from the set of promotion information based on the interaction data includes:

[0048] The average interaction value of the user sample with respect to the historical recommendation information is obtained by aggregating and averaging the interaction data between each user in the user set and any historical recommendation information.

[0049] The average interaction value of all the historical recommendation information based on the user sample is averaged to obtain the total average interaction value of the user sample.

[0050] When the average interaction value of the user sample with respect to the historical recommendation information is greater than the total average interaction value, the historical recommendation information is used as an information sample that is adapted to the recommendation scenario sample.

[0051] This application provides an information recommendation device based on artificial intelligence, including:

[0052] The feature cross module is used to perform feature cross processing on the first part of the attribute information among the multiple attribute information of the information to be recommended, so as to obtain the cross features of the information to be recommended.

[0053] An encoding module is used to encode the second part of the attribute information among the multiple attribute information of the information to be recommended, so as to obtain the encoded features of the information to be recommended.

[0054] The attention module is used to perform attention processing based on the cross features and the encoded features to obtain the attention features of the information to be recommended.

[0055] The prediction module is used to perform bias prediction processing based on the attention features of the information to be recommended, so as to obtain the recommendation scenario that the information to be recommended is suitable for;

[0056] The recommendation module is used to perform recommendation operations based on the recommendation scenario adapted to the information to be recommended.

[0057] In the above technical solution, the feature cross module is further used to encode the first part of the attribute information in the multiple attribute information of the information to be recommended, so as to obtain the encoded features of the first part of the attribute information.

[0058] The encoded features of the first part of attribute information are subjected to second-order feature cross processing to obtain the second-order cross features of the first part of attribute information.

[0059] The second-order cross features of the first part of the attribute information are mapped to obtain the cross features of the information to be recommended.

[0060] In the above technical solution, the first part of the attribute information includes fixed attribute information and self-media attribute information; the feature cross module is also used to discretize the fixed attribute information of the information to be recommended to obtain the discrete features of the fixed attribute information.

[0061] The discrete features of the fixed attribute information are vectorized to obtain the encoded features of the fixed attribute information.

[0062] The self-media attribute information of the information to be recommended is vectorized to obtain the encoded features of the self-media attribute information.

[0063] In the above technical solution, the first part of the attribute information includes text attribute information; the feature cross module is also used to perform word segmentation on the text attribute information of the information to be recommended, to obtain multiple words of the text attribute information;

[0064] The text attribute information is vectorized to obtain word vectors corresponding to each word.

[0065] The word vectors corresponding to the multiple words are averaged to obtain the encoding features of the text attribute information.

[0066] In the above technical solution, the feature crossing module is further used to perform second-order feature crossing processing on the encoding features corresponding to any two attribute information in the first part of attribute information, to obtain the second-order crossing features of the two attribute information.

[0067] The second-order cross features of any two attribute information and the encoding features of the first part of attribute information are fused to obtain the second-order cross features of the first part of attribute information.

[0068] In the above technical solution, the feature cross module is further used to perform matrix decomposition processing on the second-order cross features of the first part of attribute information to obtain the decomposed features of the first part of attribute information.

[0069] The decomposed features of the first part of the attribute information are subjected to nonlinear mapping processing to obtain the cross features of the information to be recommended.

[0070] In the above technical solution, the second part of the attribute information includes text attribute information and image attribute information; the encoding module is also used to perform text encoding processing on multiple text attribute information of the information to be recommended to obtain the text encoding features of the information to be recommended;

[0071] The image attribute information of the information to be recommended is processed by image encoding to obtain the image encoding features of the information to be recommended.

[0072] In the above technical solution, the encoding module is further used to perform feature extraction processing on multiple text attribute information of the information to be recommended, so as to obtain the text features of the text attribute information;

[0073] The text features of multiple text attribute information are mapped to obtain the text encoding features of the information to be recommended.

[0074] In the above technical solution, the encoding module is further used to perform convolution processing on multiple text attribute information of the information to be recommended, so as to obtain the convolution features of the text attribute information;

[0075] The convolutional features of the text attribute information are pooled to obtain the text features of the text attribute information.

[0076] In the above technical solution, the encoding module is further used to perform feature extraction processing on the image attribute information of the information to be recommended, so as to obtain the image features of the image attribute information;

[0077] The image features of the image attribute information are mapped to obtain the image encoding features of the information to be recommended.

[0078] In the above technical solution, the encoding module is further used to perform residual connection processing on the image attribute information of the information to be recommended, so as to obtain the residual features of the image attribute information;

[0079] The residual features of the image attribute information are subjected to nonlinear mapping processing to obtain the image features of the image attribute information.

[0080] In the above technical solution, the encoding module is also used to perform dimensionality reduction processing on the image attribute information of the information to be recommended, so as to obtain the dimensionality-reduced image features;

[0081] The dimensionality-reduced image features are subjected to convolution processing to obtain the convolutional features of the image attribute information;

[0082] The convolutional features of the image attribute information are subjected to dimensionality upscaling to obtain the dimensionality upscaled image features;

[0083] The image attribute information of the information to be recommended is summed with the image features after dimensionality increase to obtain the residual features of the image attribute information.

[0084] In the above technical solution, the attention module is further used to concatenate the cross features and the encoded features to obtain the concatenated features of the information to be recommended;

[0085] Self-attention processing is performed on the spliced ​​features of the information to be recommended to obtain the attention features of the information to be recommended.

[0086] In the above technical solution, the prediction module is further used to map the attention features of the information to be recommended to obtain the probability distribution of the candidate recommendation scenario;

[0087] The candidate recommendation scenario corresponding to the highest probability in the probability distribution is determined as the recommendation scenario that the information to be recommended is adapted to.

[0088] In the above technical solution, the recommendation module is further used to store the information to be recommended into the database corresponding to the recommendation scenario based on the recommendation scenario adapted to the information to be recommended;

[0089] When it is determined that the target user is suitable for the recommendation scenario based on the target user's profile information, at least one of the recommended information from the database corresponding to the recommendation scenario is recommended to the target user.

[0090] This application provides an information recommendation device based on artificial intelligence, including:

[0091] The training module is used to obtain information samples from the database corresponding to the recommended scenario samples;

[0092] The first part of the attribute information in the multiple attribute information of the information sample is subjected to feature cross processing to obtain the cross features of the information sample;

[0093] Based on the encoding features and the cross features of the information sample, the attention features of the information sample are determined;

[0094] Based on the attention features of the information sample, a biased prediction process is performed to obtain the predicted recommendation scenario of the information sample.

[0095] Based on the predicted recommendation scenario of the information sample and the recommendation scenario sample, the loss function of the bias model is constructed.

[0096] Update the parameters of the bias model, and use the updated parameters of the bias model when the loss function converges as the parameters of the trained bias model.

[0097] In the above technical solution, the device further includes:

[0098] The module is used to determine the attribute information of user samples corresponding to the recommendation scenario samples;

[0099] Obtain the user set formed by users possessing the aforementioned attribute information;

[0100] Obtain a set of recommendation information formed by historical recommendation information that has interacted with users in the user set;

[0101] Obtain the interaction data between each user in the user set and each piece of historical recommendation information in the recommendation information set;

[0102] Based on the interaction data, historical recommendation information that matches the recommendation scenario sample is selected from the recommendation information set as information samples to construct a database corresponding to the recommendation scenario sample.

[0103] In the above technical solution, the construction module is further configured to sort the historical recommendation information in the recommendation information set in descending order based on the interaction data between any user and each historical recommendation information in the recommendation information set, and take the portion of historical recommendation information that appears first in the descending order as an information sample that is adapted to the recommendation scenario sample.

[0104] In the above technical solution, the construction module is further configured to perform aggregation and averaging processing on the interaction data between each user in the user set and any historical recommendation information to obtain the average interaction value of the user sample with respect to the historical recommendation information;

[0105] The average interaction value of all the historical recommendation information based on the user sample is averaged to obtain the total average interaction value of the user sample.

[0106] When the average interaction value of the user sample with respect to the historical recommendation information is greater than the total average interaction value, the historical recommendation information is used as an information sample that is adapted to the recommendation scenario sample.

[0107] This application provides an electronic device for information recommendation, the electronic device comprising:

[0108] Memory, used to store executable instructions;

[0109] When the processor executes the executable instructions stored in the memory, it implements the artificial intelligence-based information recommendation method provided in the embodiments of this application.

[0110] This application provides a computer-readable storage medium storing executable instructions for inducing a processor to execute and implement the artificial intelligence-based information recommendation method provided in this application.

[0111] The embodiments of this application have the following beneficial effects:

[0112] By performing feature cross processing on the attribute information of the information to be recommended, the resulting cross features can retain the boundary information of the attribute information and fully learn the attribute information of the information to be recommended. By performing attention processing on multiple features of the information to be recommended (including cross features and encoded features obtained by encoding the second part of the attributes), the contextual dependencies of the information to be recommended can be fully learned. Based on the attention features of the information to be recommended, bias prediction can be performed to make targeted recommendations based on the recommendation scenario adapted to the information to be recommended, thereby improving the accuracy and efficiency of information recommendation in the case of cold start. Attached Figure Description

[0113] Figure 1 This is a schematic diagram illustrating an application scenario of the recommendation system provided in the embodiments of this application;

[0114] Figure 2 This is a schematic diagram of the structure of an electronic device for information recommendation provided in an embodiment of this application;

[0115] Figures 3A-3C This is a flowchart illustrating the information recommendation method based on artificial intelligence provided in an embodiment of this application;

[0116] Figure 4 This is a schematic diagram of the structure of a product-based neural network provided in an embodiment of this application;

[0117] Figure 5 This is a schematic diagram of text encoding provided in an embodiment of this application;

[0118] Figure 6 This is a schematic diagram of the residual network structure provided in the embodiments of this application;

[0119] Figure 7 This is a flowchart illustrating the information recommendation method based on artificial intelligence provided in an embodiment of this application;

[0120] Figure 8 This is a schematic diagram of element-level feature intersections provided by related technologies;

[0121] Figure 9 This is a schematic diagram of vector-level feature crossing provided in an embodiment of this application;

[0122] Figure 10 This is a schematic diagram of the structure of the tendency model provided in the embodiments of this application;

[0123] Figure 11 This is a schematic diagram of the system architecture provided in the embodiments of this application. Detailed Implementation

[0124] To make the objectives, 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 limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0125] In the following description, the terms "first" and "second" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first" and "second" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0126] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0127] In the implementation of this application, the collection and processing of relevant data should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0128] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0129] 1) Deep Learning (DL): Originating from research on artificial neural networks, it uses multilayer perceptrons with multiple hidden layers to combine low-level features to form more abstract high-level representations of attribute categories or features, thereby discovering distributed feature representations of data. It can learn the inherent patterns and representation levels of sample data, with the ultimate goal of enabling machines to have analytical and learning capabilities like humans, and to recognize data such as text, images, and sounds.

[0130] 2) Convolutional Neural Networks (CNNs): A class of feedforward neural networks (FNNs) that include convolutional computations and have a deep structure, CNNs are one of the representative algorithms of deep learning. CNNs possess representation learning capabilities, enabling them to perform shift-invariant classification of input images according to their hierarchical structure.

[0131] 3) Content cold start: When content is first published on a content site, it does not yet have enough exposure or historical behavior data for the recommendation system. Such content is defined as new content and needs to be cold-started by exposing it to suitable users to obtain user behavior data to assist in accurate recommendations later.

[0132] 4) User cold start: When a user first registers, there is not enough historical interaction behavior to determine the user's preferences (biases), so it is impossible to use the user's preferences for the recommendation system. Such users are defined as new users. Cold start is required for new users. Recommend information to new users to obtain user interaction behavior data to assist in accurate recommendations later.

[0133] 5) User persona: Also known as user role, it is an effective tool for outlining target users and connecting user needs with design direction. User personas have been widely used in various fields. In practice, they are often used in the simplest and most relatable language to connect user attributes, behaviors, and expectations, serving as virtual representatives of actual users.

[0134] 6) Content: The information recommended by the client to users can be videos, pictures, text, and graphic content. Graphic content can be small vertical or short horizontal graphic content actively edited and published by self-media. Videos can be professionally produced content (PGC) or user-generated content (UGC) provided in the form of feeds.

[0135] 7) Feeds: Also known as news sources, these are continuously updated information streams presented to users. Feeds aggregate content by combining several news sources that users actively subscribe to, helping users continuously obtain the latest content from their subscriptions. These subscription sources are typically news websites and blogs. Feeds can be displayed in various formats, primarily timelines and ranks. A timeline displays content in the order it was updated, such as on Weibo or WeChat Moments. A rank calculates the weight of content based on certain factors, thus determining the order in which the content is displayed.

[0136] 8) Recommendation Scenario: Also known as target preference, it reflects the degree to which a certain group of people like information in a certain application scenario. That is, the target audience and application scenario to which the information is matched. For example, if a video is matched with the recommendation scenario of young people, the video can be effectively exposed if it is recommended to young people, thereby improving the recommendation efficiency.

[0137] 9) Attribute Information: An abstract description of an object. A concrete thing always has many properties and relationships; these properties and relationships are the attributes of that thing. Things and attributes are inseparable; the similarity or difference between one thing and another is the similarity or difference between the attributes of one thing and the attributes of another. Due to the similarity or difference of the attributes of things, many different types of things are formed in the objective world. Things with the same attributes form one type, and things with different attributes form different types.

[0138] This application provides an information recommendation method, apparatus, electronic device, and computer-readable storage medium based on artificial intelligence, which can recommend information in a targeted manner and improve the accuracy of information recommendation.

[0139] The AI-based information recommendation method provided in this application can be implemented by a terminal / server alone; or it can be implemented collaboratively by a terminal and a server. For example, the terminal can independently implement the AI-based information recommendation method described below, or the terminal can send an information recommendation request for a target user to the server, and the server can execute the AI-based information recommendation method according to the received information recommendation request for the target user. Based on the attention features of the information to be recommended, the server can perform bias prediction processing to obtain a recommendation scenario suitable for the information to be recommended, and execute a recommendation operation for the target user based on the recommendation scenario suitable for the information to be recommended, thereby realizing the cold start of the user and the cold start of the information.

[0140] The electronic device for information recommendation provided in this application can be various types of terminal devices or servers. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. The terminal can be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

[0141] Taking servers as an example, such as server clusters deployed in the cloud, AI as a Service (AIaaS) is offered to users. The AIaaS platform breaks down several common AI services and provides them as independent or packaged services in the cloud. This service model is similar to an AI-themed marketplace, where all users can access and use one or more AI services provided by the AIaaS platform through application programming interfaces.

[0142] For example, one type of AI cloud service could be an information recommendation service, where a cloud server encapsulates the information recommendation program provided in this application embodiment. Users invoke the information recommendation service in the cloud service through a terminal (running a client, such as a music client, news client, etc.), causing the cloud-deployed server to invoke the encapsulated information recommendation program. Based on the attention characteristics of the information to be recommended, the program performs bias prediction processing to obtain a recommendation scenario suitable for the information. Based on this recommendation scenario, it executes recommendation operations targeting specific users. For example, in a music application, since the music to be recommended is new (lacking sufficient behavioral data), it's impossible to perform related recommendation processing based on the music's behavioral data. Therefore, based on the recommendation scenario suitable for the music, it identifies target users from multiple candidate users who fit the recommendation scenario, recommending the music to these target users. This distributes the music to users who match their interests, quickly obtains user behavior data, accelerates the cold start efficiency of videos, and improves the effectiveness of subsequent music recommendations based on user behavior.

[0143] See Figure 1 , Figure 1 This is a schematic diagram of the application scenario of the recommendation system 10 provided in the embodiments of this application. The terminal (exemplarily shown as terminal 200-1, terminal 200-2 and terminal 200-3) connects to the server 100 through the network 300. The network 300 can be a wide area network or a local area network, or a combination of the two.

[0144] A terminal (running a client, such as a music client or a news client) can be used to obtain information recommendation requests for candidate users. For example, when a candidate user (e.g., a new user or an old customer) opens a news client running on the terminal, the terminal automatically obtains news recommendation requests for the candidate user.

[0145] In some embodiments, after the terminal obtains an information recommendation request for a candidate user, it calls the information recommendation interface of the server 100 (which can be provided as a cloud service, i.e., an information recommendation service). Based on the information recommendation request for the candidate user, the server 100 performs bias prediction processing based on the attention features of the information to be recommended to obtain a recommendation scenario that matches the information to be recommended. Based on the recommendation scenario that matches the information to be recommended, the server 100 determines the target user that matches the recommendation scenario from multiple candidate users and recommends the information to be recommended to the target user in response to the information recommendation request for the candidate user.

[0146] As an application example, for a music application, since the music to be recommended is new music (with insufficient exposure and historical behavior data for recommendation), it is impossible to perform related recommendation processing based on the behavior data of the music to be recommended. The terminal calls the server's information recommendation interface based on the music recommendation request for the candidate user. The server determines the target user that is suitable for the recommendation scenario from multiple candidate users based on the recommendation scenario to recommend the music to the target user, thereby distributing the music to be recommended to users who meet the interest needs, quickly obtaining user behavior data, speeding up the music cold start efficiency, and improving the effect of music recommendation based on user behavior in the later stage.

[0147] In some embodiments, an information recommendation plugin may be embedded in the client running on the terminal to implement an AI-based information recommendation method locally on the client. For example, after the terminal receives a request to recommend information to a target user, it calls the information recommendation plugin to implement an AI-based information recommendation method. Based on the attention features of the information to be recommended, it performs bias prediction processing to obtain a recommendation scenario suitable for the information to be recommended. From multiple recommendation scenarios, it determines the recommendation scenario suitable for the target user and recommends the information to be recommended corresponding to the recommendation scenario suitable for the target user to the target user, thereby responding to the information recommendation request for the target user.

[0148] As an application example, for news applications, since the target user is a new user (with insufficient interaction behavior data), it is impossible to perform relevant recommendation processing based on the target user's interaction behavior data. The terminal determines the recommendation scenarios corresponding to multiple news items to be recommended based on the news recommendation request for the target user. From the multiple recommendation scenarios, it selects the recommendation scenario that is suitable for the target user and recommends the news items corresponding to the recommendation scenario that is suitable for the target user to the target user. In this way, the new user can quickly obtain news that meets their interests, so as to quickly obtain user behavior data, speed up the user's cold start efficiency, and improve the effect of news recommendation based on user behavior in the later stage.

[0149] The structure of the electronic device for information recommendation provided in the embodiments of this application is described below. See also... Figure 2 , Figure 2 This is a schematic diagram of the structure of an electronic device 500 for information recommendation provided in an embodiment of this application. The example given is that the electronic device 500 is a server. Figure 2 The illustrated electronic device 500 for information recommendation includes at least one processor 510, a memory 550, at least one network interface 520, and a user interface 530. The various components in the electronic device 500 are coupled together via a bus system 540. It is understood that the bus system 540 is used to implement communication between these components. In addition to a data bus, the bus system 540 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in… Figure 2 The general labeled all buses as Bus System 540.

[0150] The processor 510 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. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0151] Memory 550 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 550 described in this application embodiment is intended to include any suitable type of memory. Memory 550 may optionally include one or more storage devices physically located away from processor 510.

[0152] In some embodiments, memory 550 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0153] Operating system 551 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks;

[0154] The network communication module 552 is used to reach other computing devices via one or more (wired or wireless) network interfaces 520, exemplary network interfaces 520 including: Bluetooth, WiFi, and Universal Serial Bus (USB), etc.

[0155] In some embodiments, the information recommendation device provided in this application can be implemented in software, for example, it can be an information recommendation service in the server described above, or an information recommendation plugin in the terminal described above. Of course, it is not limited to this; the artificial intelligence-based information recommendation device provided in this application can be provided in various software embodiments, including various forms such as applications, software, software modules, scripts, or code.

[0156] Figure 2 An information recommendation device 555 stored in memory 550 is shown. It can be software in the form of programs and plugins, such as an information recommendation plugin, and includes a series of modules, including a feature interaction module 5551, an encoding module 5552, an attention module 5553, a prediction module 5554, a recommendation module 5555, a training module 5556, and a construction module 5557. The feature interaction module 5551, encoding module 5552, attention module 5553, prediction module 5554, and recommendation module 5555 are used to implement the information recommendation function provided in this embodiment. The training module 5556 and construction module 5557 are used to train a preference model. The training module 5556 and construction module 5557 can also be located in a device different from the information recommendation device 555, or in a device different from the electronic device 500; that is, training the preference model and the information recommendation function are two independent processes.

[0157] As mentioned above, the AI-based information recommendation method provided in this application can be implemented by various types of electronic devices. See also Figure 3A , Figure 3A This is a flowchart illustrating the information recommendation method based on artificial intelligence provided in the embodiments of this application, combined with... Figure 3A The steps shown are explained.

[0158] In the following steps, the information to be recommended can be text, images, graphic text, video, or other data. Attribute information is the prior data of the information to be recommended; it is inherent information of the information to be recommended and is unrelated to the user's or the behavioral data of the information to be recommended, that is, it is unrelated to the posterior data of the information to be recommended or the user's posterior data.

[0159] In step 101, feature cross-processing is performed on the first part of the attribute information among the multiple attribute information of the information to be recommended to obtain the cross features of the information to be recommended.

[0160] As an example of acquiring information to be recommended, the publisher uploads the information via a terminal. Upon receiving the information, the terminal determines multiple attributes, such as content text, title, category, theme, duration, and social media information. It then selects a subset of these attributes for feature cross-validation to obtain the cross-features of the information to be recommended. By performing feature cross-validation on these attributes, different cross-features are captured, enhancing the feature representation capability and avoiding the omission of feature boundaries, thus enabling subsequent prediction processing based on accurate cross-features.

[0161] See Figure 3B , Figure 3B This is an optional flowchart illustrating an artificial intelligence-based information recommendation method provided in an embodiment of this application. Figure 3B Show Figure 3A Step 101 can be implemented through steps 1011-1013: In step 1011, the first part of the attribute information in the multiple attribute information of the information to be recommended is encoded to obtain the encoded features of the first part of the attribute information; in step 1012, the encoded features of the first part of the attribute information are subjected to second-order feature cross processing to obtain the second-order cross features of the first part of the attribute information; in step 1013, the second-order cross features of the first part of the attribute information are mapped to obtain the cross features of the information to be recommended.

[0162] like Figure 4 As shown, the first part of the attribute information includes fixed attribute information (video duration, publication time, etc.), self-media attribute information (self-media identifier, self-media account level, etc.), and text attributes (content tags, self-media tags, etc.). First, the first part of the attribute information is encoded to obtain the encoded features (embedding vectors) of each attribute. Then, second-order feature crossing is performed based on the encoded features of each attribute in the first part of the attribute information to obtain the second-order cross features of the first part of the attribute information. Finally, an activation function is used to map the second-order cross features of the first part of the attribute information to obtain the cross features of the information to be recommended, so that attention processing can be performed subsequently based on the cross features.

[0163] In some embodiments, the first part of the attribute information includes fixed attribute information and self-media attribute information; encoding the first part of the attribute information among the multiple attribute information of the information to be recommended to obtain the encoded features of the first part of the attribute information includes: discretizing the fixed attribute information of the information to be recommended to obtain the discrete features of the fixed attribute information; vectorizing the discrete features of the fixed attribute information to obtain the encoded features of the fixed attribute information; and vectorizing the self-media attribute information of the information to be recommended to obtain the encoded features of the self-media attribute information.

[0164] like Figure 4 As shown, different attribute information requires different encoding processes. When the first attribute information is fixed attribute information (video duration, publication timeliness, etc.) or self-media attribute information (self-media identifier, self-media account level, etc.), targeted encoding processing is required. For example, fixed attribute information such as video duration and publication timeliness (with continuous values) is discretized to obtain discrete features (i.e., discrete values). These discrete features are then vectorized to obtain the encoded features (embedded vectors). Self-media attribute information such as self-media identifier and self-media account level (with discrete values) is directly vectorized to obtain the encoded features (embedded vectors). This targeted encoding process yields accurate encoded features for subsequent feature cross-referencing.

[0165] In some embodiments, the first part of the attribute information includes text attribute information; encoding the first part of the attribute information among the multiple attribute information of the information to be recommended to obtain the encoding features of the first part of the attribute information includes: performing word segmentation on the text attribute information of the information to be recommended to obtain multiple words of the text attribute information; performing vectorization on the multiple words of the text attribute information to obtain word vectors corresponding to the multiple words respectively; and averaging the word vectors corresponding to the multiple words to obtain the encoding features of the text attribute information.

[0166] like Figure 4 As shown, different attribute information requires different encoding processes. When the first attribute information is text attribute information (content tags, self-media tags, etc.), text-based encoding processing is required. For example, content tags, self-media tags, and other text attribute information are segmented into multiple words. Then, these words are vectorized to obtain word vectors corresponding to each word. The word vectors are then averaged to obtain the encoded features of the text attribute information. Thus, through targeted encoding processing, accurate encoded features are obtained for subsequent feature cross-referencing.

[0167] In some embodiments, the encoding features of the first part of attribute information are subjected to second-order feature cross processing to obtain second-order cross features of the first part of attribute information, including: performing second-order feature cross processing on the encoding features corresponding to any two attribute information in the first part of attribute information to obtain second-order cross features of any two attribute information; and performing fusion processing on the second-order cross features of any two attribute information and the encoding features of the first part of attribute information to obtain second-order cross features of the first part of attribute information.

[0168] like Figure 4 As shown, the first part of the attribute information includes category, content tags, self-media tags, theme, duration, level, etc. The encoding feature of the first part of the attribute information is category. Content Tags ,…,grade Then, perform second-order feature cross processing on the encoded features corresponding to any two attribute information in the first part of the attribute information to obtain the second-order cross features of any two attribute information. Second-order cross features for any two attribute information and the encoding features of the first part of the attribute information The data is then concatenated to obtain the second-order cross features of the first part of the attribute information.

[0169] In some embodiments, mapping the second-order cross features of the first part of the attribute information to obtain the cross features of the information to be recommended includes: performing matrix decomposition on the second-order cross features of the first part of the attribute information to obtain the decomposed features of the first part of the attribute information; and performing nonlinear mapping on the decomposed features of the first part of the attribute information to obtain the cross features of the information to be recommended.

[0170] For example, the process of matrix factorization is as follows: , Here, z represents the encoded feature in the second-order cross-feature of the first part of the attribute information, and p represents the second-order cross-feature of any two attributes in the second-order cross-feature of the first part of the attribute information. Matrix factorization can accelerate the subsequent processing. Finally, the decomposed features of the first part of the attribute information are nonlinearly mapped using an activation function to obtain the cross-feature of the information to be recommended, so that attention processing can be performed based on the cross-feature in the subsequent process.

[0171] In step 102, the second part of the attribute information among the multiple attribute information of the information to be recommended is encoded to obtain the encoded features of the information to be recommended.

[0172] The encoding process differs for different attribute information. The second part of the attribute information undergoes text encoding or image encoding to obtain the encoded features of the information to be recommended. This allows for subsequent attention processing based on these different encoded features, thereby learning different encoded features and improving subsequent prediction capabilities.

[0173] For example, the second part of the attribute information includes text attribute information (such as title, text obtained by OCR technology (OCRText)) and image attribute information (such as cover image); the second part of the attribute information among the multiple attribute information of the information to be recommended is encoded to obtain the encoded features of the information to be recommended, including: text encoding of the multiple text attribute information of the information to be recommended to obtain the text encoding features of the information to be recommended; and image encoding of the image attribute information of the information to be recommended to obtain the image encoding features of the information to be recommended.

[0174] In some embodiments, text encoding processing is performed on multiple text attribute information of the information to be recommended to obtain text encoding features of the information to be recommended, including: performing feature extraction processing on each of the multiple text attribute information of the information to be recommended to obtain text features of the text attribute information; and performing mapping processing on the text features of the multiple text attribute information to obtain text encoding features of the information to be recommended.

[0175] For example, such as Figure 5 As shown, for text attribute information (such as title, OCR text), a convolutional neural network is used to extract features from the text attribute information to obtain text features (text vectors). Then, the text features of multiple text attribute information are concatenated and mapped through an activation function to obtain the text encoding features of the information to be recommended.

[0176] For example, feature extraction processing is performed on multiple text attribute information of the recommendation information to obtain text features of the text attribute information. This includes: performing convolution processing on multiple text attribute information of the recommendation information through convolutional layers in a convolutional neural network to obtain convolutional features of the text attribute information, and performing pooling processing on the convolutional features of the text attribute information through pooling layers in a convolutional neural network to obtain text features of the text attribute information.

[0177] In some embodiments, image encoding processing is performed on the image attribute information of the information to be recommended to obtain image encoding features of the information to be recommended, including: performing feature extraction processing on the image attribute information of the information to be recommended to obtain image features of the image attribute information; and performing mapping processing on the image features of the image attribute information to obtain image encoding features of the information to be recommended.

[0178] like Figure 6As shown, the image attribute information of the information to be recommended is processed by feature extraction through cascaded residual modules in a residual network (e.g., ResNet50) to obtain image features of the image attribute information. Then, the image features of the image attribute information are mapped through an activation function (e.g., ReLU) to obtain the image encoding features of the information to be recommended.

[0179] The feature extraction process includes: performing residual connection processing on the image attribute information of the information to be recommended to obtain residual features of the image attribute information; and performing nonlinear mapping processing on the residual features of the image attribute information to obtain image features of the image attribute information.

[0180] In some embodiments, residual connection processing is performed on the image attribute information of the information to be recommended to obtain residual features of the image attribute information, including: dimensionality reduction processing of the image attribute information of the information to be recommended to obtain dimensionality-reduced image features; convolution processing of the dimensionality-reduced image features to obtain convolutional features of the image attribute information; dimensionality increase processing of the convolutional features of the image attribute information to obtain dimensionality-increased image features; and summing the image attribute information of the information to be recommended with the dimensionality-increased image features to obtain residual features of the image attribute information.

[0181] like Figure 6 As shown, the image attribute information of the information to be recommended is dimensionality reduced by the convolutional layer (1*1 kernel) in the residual module to obtain the dimensionality-reduced image features. Then, the dimensionality-reduced image features are convolved by the convolutional layer (3*3 kernel) in the residual module to obtain the convolutional features of the image attribute information. Next, the convolutional features of the image attribute information are increased in dimensionality by the convolutional layer (1*1 kernel) in the residual module to obtain the increased-dimensional image features. Finally, the image attribute information of the information to be recommended is summed with the increased-dimensional image features to obtain the residual features of the image attribute information, thus realizing residual connection.

[0182] In step 103, attention processing is performed based on cross features and encoding features to obtain attention features of the information to be recommended.

[0183] For example, after obtaining the cross features and encoded features, these features can be concatenated to obtain the concatenated features of the information to be recommended. Then, a self-attention mechanism is used to process these concatenated features to obtain the attention features of the information to be recommended. The self-attention mechanism can learn the dependencies between elements in the concatenated features, thereby uncovering important features in the information to be recommended for subsequent prediction processing, achieving accurate prediction.

[0184] In step 104, bias prediction processing is performed based on the attention features of the information to be recommended to obtain the recommendation scenario that matches the information to be recommended.

[0185] For example, after obtaining the attention features of the information to be recommended, a bias model (a classifier) ​​is used to combine the attention features of the information to be recommended to perform bias prediction processing, thereby obtaining the recommendation scenario that the information to be recommended is suitable for, so that subsequent recommendation operations can be performed on the appropriate users based on the recommendation scenario that the information to be recommended is suitable for.

[0186] In some embodiments, bias prediction processing is performed based on the attention features of the information to be recommended to obtain a recommendation scenario that matches the information to be recommended, including: mapping the attention features of the information to be recommended to obtain a probability distribution of candidate recommendation scenarios; and determining the candidate recommendation scenario corresponding to the highest probability in the probability distribution as the recommendation scenario that matches the information to be recommended.

[0187] For example, by mapping the attention features of the information to be recommended through a classifier, the probability distribution of candidate recommendation scenarios can be obtained, such as scenarios favored by young people (0.1), scenarios favored by the elderly (0.2), scenarios favored by women (0.5), and scenarios with high sharing (0.3). Then, scenarios favored by women are used as the recommendation scenarios to be adapted to the information to be recommended, and the information to be recommended is subsequently recommended to female users.

[0188] In step 105, a recommendation operation is performed based on the recommendation scenario adapted to the information to be recommended.

[0189] For example, after determining the recommendation scenario to which the information to be recommended fits, the target user suitable for that recommendation scenario is identified, and the information to be recommended is recommended to that target user. When the information to be recommended is new, since no behavioral data of the information to be recommended is introduced during the process of determining the recommendation scenario, only the attribute information of the information to be recommended is used, thus achieving a cold start for the new information. When the target user is a new user, since no interactive behavioral data of the new user is introduced during the process of determining the recommendation scenario to which the information to be recommended fits, only the attribute information of the information to be recommended is used, thus achieving a cold start for the new user.

[0190] For example, for a cold start for a new user, the information to be recommended is stored in the database corresponding to the recommendation scenario based on the recommendation scenario to be adapted. When it is determined that the target user is suitable for the recommendation scenario based on the target user's profile information, at least one piece of information to be recommended in the database corresponding to the recommendation scenario is recommended to the target user.

[0191] For example, after obtaining that the recommended scenario to be matched with the information to be recommended is a female-oriented scenario, the information to be recommended is stored in a female-oriented database. When the profile information of the target user (new user) indicates that the target user is a female user, at least one piece of information to be recommended is selected from the female-oriented database and recommended to the target user, thereby achieving targeted information recommendation.

[0192] See Figure 3C , Figure 3C This is an optional flowchart illustrating an artificial intelligence-based information recommendation method provided in this application embodiment. The information recommendation method is implemented by calling a bias model. Before applying this information recommendation method, the bias model needs to be trained. Figure 3C The training process of the bias model is shown as follows: In step 106, information samples are obtained from the database corresponding to the recommended scenario samples; in step 107, feature cross-processing is performed on the first part of the attribute information of multiple attribute information of the information samples to obtain the cross features of the information samples; in step 108, the attention features of the information samples are determined based on the encoding features and cross features of the information samples; in step 109, bias prediction processing is performed based on the attention features of the information samples to obtain the predicted recommended scenario of the information samples; in step 110, the loss function of the bias model is constructed based on the predicted recommended scenario of the information samples and the recommended scenario samples; in step 111, the parameters of the bias model are updated, and the updated parameters of the bias model when the loss function converges are used as the parameters of the trained bias model.

[0193] Among them, based on the prediction and recommendation scenarios and the recommendation scenario samples, after determining the value of the loss function of the bias model, it can be determined whether the value of the loss function of the bias model exceeds the preset threshold. When the value of the loss function of the bias model exceeds the preset threshold, the error signal of the bias model is determined based on the loss function of the bias model, the error information is backpropagated in the bias model, and the model parameters of each layer are updated during the propagation process.

[0194] Here, we explain backpropagation. Training sample data is input into the input layer of the neural network model, passes through the hidden layers, and finally reaches the output layer to output the result. This is the forward propagation process of the neural network model. Since there is an error between the output result and the actual result, the error between the output result and the actual value is calculated and propagated back from the output layer to the hidden layers until it reaches the input layer. During the backpropagation process, the values ​​of the model parameters are adjusted according to the error. This process is iterated until convergence. The bias model is a type of neural network model.

[0195] In some embodiments, before model training, it is necessary to construct information samples. The method for constructing information samples includes: determining the attribute information of user samples corresponding to the recommendation scenario samples; obtaining a user set formed by users with attribute information; obtaining a set of recommendation information formed by historical recommendation information that has interactions with users in the user set; obtaining the interaction data between each user in the user set and each historical recommendation information in the recommendation information set; and based on the interaction data, selecting historical recommendation information that matches the recommendation scenario samples from the recommendation information set as information samples to construct a database corresponding to the recommendation scenario samples.

[0196] For example, to construct information samples with female bias, firstly, it is determined that the attribute information of the user sample corresponding to the recommendation scenario sample is female. Then, a set of female users is obtained, and a set of recommendation information is obtained, which consists of historical recommendation information that has interactions with female users in the female user set. Then, the interaction data (such as click-through rate, consumption time, etc.) between each female user in the female user set and each historical recommendation information in the recommendation information set is obtained. Based on these interaction data, historical recommendation information that matches female bias is selected from the recommendation information set as information samples to construct a database of female bias.

[0197] In some embodiments, based on interaction data, selecting recommendation information that matches the recommendation scenario sample from the recommendation information set as information samples includes: sorting the historical recommendation information in the recommendation information set in descending order based on the interaction data between any user and each historical recommendation information in the recommendation information set, and taking the part of historical recommendation information that is first in the descending sort result as information samples that match the recommendation scenario sample.

[0198] Following the example above, to construct information samples with female bias, we determine the interaction data between any female user i and each historical recommendation in the recommendation information set. We sort the historical recommendation information in the recommendation information set in descending order and take the first part of the historical recommendation information in the descending order as information samples that are adapted to female bias. Then, we remove duplicates from the information samples that are adapted to female bias obtained based on all female users.

[0199] In some embodiments, based on interaction data, recommending information that matches the recommendation scenario sample is selected from the promotion information set as information samples, including: performing aggregation and averaging processing on the interaction data between each user in the user set and any historical recommending information to obtain the average interaction value of the user sample with respect to the historical recommending information; averaging processing on the average interaction value of the user sample with respect to all historical recommending information to obtain the total average interaction value of the user sample; when the average interaction value of the user sample with respect to historical recommending information is greater than the total average interaction value, the historical recommending information is used as the information sample that matches the recommendation scenario sample.

[0200] Following the example above, for constructing information samples with a female bias, the interaction data between each user in the user set and any historical recommendation information j is aggregated and averaged to obtain the average interaction value of the user sample with respect to historical recommendation information j (e.g., average click-through rate, average consumption time, etc.). The average interaction value of the user sample with respect to all historical recommendation information is then averaged to obtain the total average interaction value of the user sample (e.g., total average click-through rate, total average consumption time, etc.). When the average interaction value of the user sample with respect to historical recommendation information is greater than the total average interaction value, historical recommendation information j is used as an information sample that is adapted to the female bias.

[0201] The information recommendation method based on artificial intelligence provided in this application embodiment will be implemented collaboratively by the terminal (producer and consumer ends) and the server. See below. Figure 7 , Figure 7 This is a flowchart illustrating the information recommendation method based on artificial intelligence provided in the embodiments of this application, combined with... Figure 7 The steps shown are explained.

[0202] In step 201, the production end uploads the information to be recommended to the server.

[0203] For example, the production end provides local or captured images, videos, or image sets for recommendation through the terminal's application programming interface (API). These are the main sources of information for information distribution.

[0204] In step 202, the server invokes the propensity model to predict the propensity of the information to be recommended, thereby obtaining the recommendation scenario that the information to be recommended is suitable for.

[0205] For example, the server performs feature cross processing on the first part of the attribute information of the multiple attribute information of the information to be recommended to obtain the cross features of the information to be recommended. It also encodes the second part of the attribute information of the multiple attribute information of the information to be recommended to obtain the encoded features of the information to be recommended. Based on the cross features and the encoded features, attention processing is performed to obtain the attention features of the information to be recommended. Based on the attention features of the information to be recommended, bias prediction processing is performed to obtain the recommendation scenario that the information to be recommended is suitable for.

[0206] In step 203, the server sends the recommended information to the consumer based on the recommended scenario adapted to the recommended information.

[0207] For example, when the server determines that the consumer is compatible with the recommendation scenario based on the recommendation information to be recommended, it sends the recommendation information to the consumer. For example, if the recommendation scenario to be recommended is female-oriented and the user on the consumer side is a female user, then the recommendation information to be recommended is sent to the consumer side.

[0208] In step 204, the consumer reports user interaction data based on historical recommendation information.

[0209] For example, the consumer end simultaneously reports user playback behavior data, such as buffering, loading time, playback clicks, comments, likes, reposts, favorites, and shares, to the server during the upload and download process.

[0210] There is no obvious order between steps 204-206 and steps 201-203.

[0211] In step 205, the server constructs an information sample based on the user's interaction data.

[0212] For example, based on the interaction data between the user and each historical recommendation in the recommendation information set, the historical recommendation information is sorted in descending order, and the part of historical recommendation information at the top of the descending sort result is used as information sample to adapt to the recommendation scenario sample.

[0213] In step 206, the server trains a propensity model based on information samples.

[0214] For example, feature cross processing is performed on the first part of the attribute information of multiple attribute information of the information sample to obtain the cross features of the information sample. Based on the encoding features and cross features of the information sample, the attention features of the information sample are determined. Based on the attention features of the information sample, bias prediction processing is performed to obtain the predicted recommendation scenario of the information sample. Based on the predicted recommendation scenario and the recommendation scenario sample of the information sample, the loss function of the bias model is constructed, the parameters of the bias model are updated, and the updated parameters of the bias model when the loss function converges are used as the parameters of the trained bias model.

[0215] The following will describe an exemplary application of the embodiments of this application in a real-world application scenario.

[0216] Self-media differs from information dissemination led by professional media organizations; it is information dissemination activity led by the general public. It also refers to an information dissemination method that provides individuals with the ability to produce, accumulate, share, and disseminate content, offering both privacy and public accessibility. This is well-suited for the rapid reading and consumption of information in the mobile era, and is typically displayed in feeds format for users to quickly refresh. Various self-media creators produce their own text and image content, which users can subscribe to. When new content is released, it is pushed to the user, and users can also actively refresh the feeds to get the latest content.

[0217] With the surge in content uploaded to self-media platforms, both text, images, and videos have experienced rapid growth in the mobile internet era. After consuming content, users can express their preferences through various interactive behaviors such as commenting, liking, forwarding, and saving. Recommendation systems for information flow content distribution primarily address the challenge of recommending content based on user profiles within specific scenarios. Cold start is a critical issue for both content and users in recommendation systems. Effective content distribution involves understanding both the user and the content. Many new externally integrated content pieces lack posterior data within the system, making them difficult to distribute through normal recommendation models. Typically, new content needs to be tested first to identify target preferences. Due to the lack of posterior data, recommendation systems often need to conduct numerous cold start trials to supplement user interests and encourage recommendations of new content. Related technologies primarily employ random trial methods, which suffer from slow convergence speeds, consume significant bandwidth, and severely impact distribution efficiency.

[0218] The applicant discovered the following problems during the implementation of the embodiments of this application: 1) General content understanding methods require a large amount of manually labeled data, and the user targeting is poor. The content understanding results do not contain user information, and recommendations cannot be directly optimized using this information; 2) For new users' cold start, the convergence and collection of user interests is slow through interest exploration; 3) For new content's cold start, the corresponding users are found through trial and error, which is very inefficient. During the content lifecycle, especially for high-quality content, it is easy to lead to insufficient content exposure and distribution.

[0219] To address the aforementioned issues, this application provides an information recommendation method and system based on artificial intelligence. It combines user consumption data and application scenarios, obtaining content samples by statistically analyzing these data (posterior data). Then, based on these content samples, a deep learning-based model (a preference model) is constructed to mine the target preference of the content (reflecting the degree to which a certain type of user likes the content in a specific application scenario, i.e., the target audience and application scenario for which the content is suitable). By using user consumption data, posterior data (consumption data) combined with prior data (content features), it can improve recommendation effectiveness for specific user groups, increase content distribution efficiency and user consumption duration; improve the cold start effect for new users; and improve the cold start effect for new content, allowing high-quality content to gain exposure and effective distribution more quickly. This assists the recommendation system, especially in building a target preference content library. For different recommendation scenarios, such as push delivery (server-directed delivery of content to the user's terminal interface in real time), it supports different sub-targets, improving overall recommendation efficiency.

[0220] To address target bias, the key to product-based neural networks and self-attention mechanisms lies in adding a product layer between the embedding layer and the fully connected layer to achieve targeted feature crossing. The self-attention mechanism is used to learn the dependencies between elements (items) in the sequence, ultimately improving the overall mining performance of the model.

[0221] The following describes in detail the information recommendation method based on artificial intelligence provided in the embodiments of this application:

[0222] This application's embodiments combine user behavior posterior data for content understanding, directly utilizing the consumption data of a specific user group (posterior data) to create a data-driven model for understanding content preferences and mining high-quality content. The target preference for content mentioned here reflects the degree to which a particular user group likes certain content in a specific scenario. The goal of this mining is to improve recommendation performance for specific user groups while optimizing the cold start effect for new users or new content.

[0223] The bias and targeting identification of recommended content requires the system to give new high-quality content (such as newly released UGC and PGC content) a certain amount of exposure in order to discover high-quality content. After the recall and ranking models learn about high-quality content, they will distribute it normally; this is the content trial process. However, trial traffic is limited, and there is too much new content. The above problems can be solved by the following methods: scoring new content based on the content delivery target model, with higher-scoring content receiving higher exposure priority, and using different delivery targets for different application scenarios. For example, using share rate for mini programs, video views (VV) for selected video streams, and click-through rate for selected main timeframes, thereby improving the system's ability to discover high-quality hit content.

[0224] Training sample data is constructed using posterior data (click-through rate, share rate, views, duration of consumption, etc.) of content with sufficient exposure and the target preferences of the content (gender preference, age preference, etc.) to predict the delivery preference of insufficiently exposed content. This preference sample data is primarily obtained by mining from the posterior data through user behavior and distribution statistics interfaces using statistical methods. This ensures that the content's click-through rate and duration of consumption are among the highest in the corresponding application scenario and user group. Furthermore, the statistically mined content undergoes manual review before being used as a sample to ensure the quality of the content itself, thus ensuring the quality of the final results obtained by the model. The problem to be solved here is, given content, to predict the trial level (exposure priority) and the corresponding target preferences, with the goal of reducing the gap between the predicted click-through rate and the actual click-through rate.

[0225] Different application scenarios and user groups have different target preferences. For example, in constructing preference samples, if the goal is to discover content with high sharing rates, but actual statistics show that most content has a very low sharing rate, the model is prone to overfitting to low-sharing content, resulting in inaccurate predictions for high-sharing content. The main reason for this problem is insufficient learning of cross-features; for example, the sharing rate of long videos in the sketch comedy category is higher than that of long videos in the sketch comedy category.

[0226] The reasons for the insufficient learning of the aforementioned cross features are as follows: Figure 8As shown, Deep Neural Networks (DNNs) achieve implicit feature crossing by adding element-wise features. The relationships between features are OR, which doesn't effectively represent feature crossing and loses feature boundaries (fields). For example, feature c = feature a*w1 + feature b*w2 + ... DNNs directly achieve feature crossing and combination through multiple fully connected layers, but this approach lacks specificity. First, fully connected layers don't handle crossings between different feature domains; second, the operations of fully connected layers aren't directly designed for feature crossings.

[0227] To address the issue of insufficient learning of cross-features, such as Figure 9 As shown, this embodiment introduces an AND relationship into the model, that is, by multiplying features at the vector level, feature field information is preserved. For example, feature c = feature a * feature + ... Based on the characteristics of the problem, sample data, and fitting target, this embodiment employs a product-based neural network (PNN), the key being the addition of a Product Layer between the embedding layer and the fully connected layer. In practical problems, the importance of feature intersection is self-evident. For example, the intersection of age and gender is a very important grouping feature, containing a large amount of high-value information, and a targeted structure is urgently needed to represent this information. Therefore, this embodiment introduces a Product Layer to achieve targeted feature intersection. During feature vector operations, the inner product operation (IPNN) is used, and its Product operation combines features between different feature domains, capturing different intersection information and enhancing the model's ability to represent different data patterns. Figure 10 As shown, the z part directly translates the embedding layer, preserving the network structure of previously learned implicit high-order features. The p part is the key to PNN, and the feature cross matrix P is a symmetric matrix, which is accelerated by matrix decomposition.

[0228] The calculation process for part p is shown in formulas (1)-(2):

[0229] (1)

[0230] (2)

[0231] Where f represents each feature, N represents the total number of features, and p represents the features resulting from feature crossover. Represents the feature cross matrix, This represents the characteristics of part p.

[0232] The calculation process for the z part is shown in formulas (3)-(4):

[0233] (3)

[0234] (4)

[0235] Where f represents each feature, and N represents the total number of features. This represents the features of the z-part.

[0236] like Figure 10 As shown, the content features used in the bias model include: 1. Text attributes (content title, category, topic, content tags, etc.); 2. Image attributes (content cover image, image obtained through Optical Character Recognition (OCR)); 3. Fixed attributes (content publication timeliness, video duration); 4. Publisher's self-media attributes: a. Self-media identifier (ID) information (ID, self-media account level); b. Publisher's self-media text profile attributes (self-media account name, self-media account category). Specifically, a CNN is introduced to understand the title and the text recognized by OCR technology (OCR Text), a residual network (resnet50) is used to understand the video cover, and a PNN is used to understand some of the above four attributes (e.g., category, video duration, self-media level, etc.). Finally, a Self-Attention mechanism is used to fuse multiple channels to output the target bias. Furthermore, due to the sparsity of tags, the initial tags use vectors trained unsupervised by word2vec.

[0237] like Figure 10 As shown, the concat layer employs a self-attention mechanism. By encoding the sequence using self-attention, the context-dependent embedding representation of each item in the sequence can be obtained. Then, pooling is performed to obtain the embedding representation of the entire sequence and to mine the dependencies between items in the sequence. Through the self-attention mechanism, the overall performance of the bias model mining is ultimately improved.

[0238] like Figure 11 As shown below, the information recommendation method and system based on artificial intelligence provided in the embodiments of this application are described in detail below. The functions of each service module in the system are as follows:

[0239] I. Content Production and Content Consumption

[0240] (1) Content production ends such as PGC, UGC, Multi-Channel Network (MCN), and Professional User Generated Content (PUGC) provide local or photographed text, video or image albums through mobile terminals or backend application programming interfaces (APIs). These are the main sources of content for distribution.

[0241] (2) The content production end obtains the upload server interface address through communication with the upstream and downstream content interface services, and then uploads the local file (to be recommended information). During the shooting process, the local text and image content can choose to match the music, filter templates and text and image beautification functions, etc.

[0242] (3) The content consumer communicates with the upstream and downstream content interface services to obtain the index information of the accessed text or video files (such as the address of the accessed content), and then downloads the corresponding streaming media file (the information to be recommended) and plays it through the local player.

[0243] (4) The content consumer end simultaneously reports user playback behavior data, buffering, loading time, playback clicks, etc. during the upload and download process to the server;

[0244] (5) Interactive information of content consumption on the content consumption end, such as comments, likes, forwards, collections, sharing and other interactive information of content, are reported through the user behavior and distribution statistics reporting interface.

[0245] II. Uplink and Downlink Content Interface Services

[0246] (1) Communicate directly with the content production end to obtain the content submitted by the front end, including the title, publisher, summary, cover image, publication time, and photos and text files (attribute information), and store the files directly into the content database through the server;

[0247] (2) Submit the uploaded file to the dispatch center service for subsequent content processing and transfer.

[0248] III. Content Database

[0249] (1) The core database of content, in which all the metadata of the content published by producers is stored, namely the metadata of the content itself, such as file size, cover image link, bitrate, file format, title, publication time, author, video file size, video format, whether it is original or first publication, etc. It also includes the classification of content during the manual review process (including first, second and third level classification and tag information, for example, an article explaining a certain mobile phone, the first level classification is technology, the second level classification is smartphone, the third level classification is domestic mobile phone, and the tag information is XX);

[0250] (2) When the uplink and downlink content interface service receives the video file, it performs standard transcoding operation on the content. After the transcoding is completed, it asynchronously returns metadata (such as file size, bitrate, specifications, cover image, etc.). This information will be stored in the content database.

[0251] (3) During the manual review process, information in the content database will be read, and the results and status of the manual review will also be sent back to the content database to update the metadata of the content in the content database.

[0252] (4) The content processing of the dispatch center service mainly includes machine processing and manual review processing. Here, machine processing is to call the deduplication service to exclude completely duplicate and similar content. The deduplication result will be written to the content database. Completely duplicate content will not be given to manual secondary processing.

[0253] IV. Dispatch Center Services

[0254] (1) Responsible for the entire scheduling process of content flow, receiving the content into the database through the uplink and downlink content interface service, and then obtaining metadata from the content database;

[0255] (2) Schedule the manual review system and the machine processing system, and control the scheduling order and priority;

[0256] (3) For content, first communicate with the deduplication service to process different bitrates, different resolutions, different sizes, partial black screens, whether there are filters, whether there are logos, inserting some advertising content in similar text and image content, and cropping the beginning and end of the video. This effectively reduces the content that needs to be processed in the manual review process. The deduplication of text and image content uses a similarity algorithm (simhash), while the deduplication of video content uses the construction of video embedding.

[0257] (4) Schedule content that needs to be manually reviewed into the manual review system and filter out duplicate content;

[0258] (5) Finally, the content of the dispatch center service is enabled to be displayed directly to the terminal content consumers through the content distribution outlet service (such as recommendation engine, search engine, operation, etc.), that is, the recommendation operation is performed based on the recommendation scenario of content adaptation;

[0259] (6) Communicate with the bias mining service to schedule the mining and construction of bias content libraries for different groups.

[0260] V. Manual Review System

[0261] (1) It is necessary to read the original information of the text and image content itself in the content database. For example, in a complex system developed based on a web database, manual filtering is used to determine whether the text and image content involves pornography or gambling.

[0262] (2) On the basis of the initial review, the content is reviewed a second time, mainly to classify, label or confirm the content. Since the text and image content itself is not yet fully mature through machine learning (such as deep learning), it is necessary to conduct a second manual review on the basis of machine processing. Through human-machine collaboration, the accuracy and efficiency of the text and image labeling can be improved.

[0263] VI. Deduplication Service

[0264] (1) Provide deduplication services for text, images and videos. The main function is to vectorize text, images, images and videos, then create an index of the vectors, and determine the degree of similarity by comparing the distance between the vectors.

[0265] (2) For text and image content, the text and images are vectorized. Before all deduplication tasks, short texts such as titles are deduplicated first to reduce the amount of computation in subsequent processing.

[0266] VII. User Behavior and Distribution Statistics Reporting Interface

[0267] (1) Communicate with the content consumer terminal to receive interactive information such as content comments, likes, forwards, collections, and shares;

[0268] (2) Collect user consumption behavior data on the content consumption end, such as the content viewed, viewing time, etc.

[0269] 8. Preferred Sample Database

[0270] (1) Based on the characteristics of the user post-consumption behavior data and application scenarios required by the bias model mentioned above, construct a bias sample database (the database corresponding to the recommended scenario samples) based on the user-reported behavior logs and statistics. Ensure that the sample data is the top high-quality content of the current application scenario and target audience. The click-through rate and consumption time of high-quality content should exceed the average value of the corresponding category content. At the same time, the content quality should be filtered by the content delivery target model, that is, screened based on the interaction data between the target audience and the sample data, in order to construct the bias sample database.

[0271] 9. Preferred Content Library

[0272] (1) Including different content databases of the tendency mining service, that is, calling the tendency model to perform information recommendation method on the content database, performing feature cross-processing on the first part of the attribute information of multiple attributes of the content to obtain the cross-feature of the content, performing encoding processing on the second part of the attribute information of multiple attributes of the content to obtain the encoding feature of the content, performing attention processing based on the cross-feature and the encoding feature to obtain the attention feature of the content, performing tendency prediction processing based on the attention feature to obtain the content-adapted recommendation scenario, and dividing the content into the corresponding tendency content database based on the content-adapted recommendation scenario, such as the female tendency content database for female users of different age groups, the elderly content database, the highly shareable content database, etc.

[0273] (2) Communicate with the content distribution export service to provide the different content libraries with different preferences to the recommendation engine for targeted distribution recommendation optimization and delivery, that is, to perform recommendation operations based on different content libraries with different preferences.

[0274] 10. Tendency Mining Service

[0275] (1) The above-described tendency model is service-oriented, and the mining results are written into the tendency content library;

[0276] (2) Accept the scheduling of the dispatch center service and complete the mining of the content of the tendency.

[0277] XI. Propensity Model

[0278] (1) Read various biased sample data from the biased sample database and train the relevant model according to the biased model described above;

[0279] (2) As the basic model of bias mining service, it provides the basis for service provision.

[0280] In summary, the information recommendation method and system based on artificial intelligence provided in this application have the following beneficial effects:

[0281] 1) Using consumption data of specific user groups in specific scenarios, combining posterior data with prior data can improve the recommendation effect for specific user groups in specific scenarios, increase the efficiency of content distribution and the duration of user consumption, and improve the cold start effect for new users.

[0282] 2) It can improve the cold start effect of new content, allowing high-quality new content to be exposed and effectively distributed faster, which helps the recommendation system, especially the construction of the content library with preferences. It supports different sub-targets for different recommendation scenarios, such as push delivery, and improves the overall recommendation efficiency.

[0283] This concludes the description of the artificial intelligence-based information recommendation method provided in this application, using the exemplary application and implementation of the server provided in the embodiments of this application. This application also provides an information recommendation device. In practical applications, the functional modules in the information recommendation device can be collaboratively implemented using the hardware resources of electronic devices (such as terminal devices, servers, or server clusters), such as computing resources like processors, communication resources (such as those used to support various communication methods like optical fiber and cellular), and memory. Figure 2 An information recommendation device 555 stored in memory 550 is shown. It can be software in the form of programs and plug-ins, such as software modules designed in programming languages ​​such as C / C++ and Java, application software designed in programming languages ​​such as C / C++ and Java, or dedicated software modules, application programming interfaces, plug-ins, cloud services, etc. in large software systems. Examples of different implementation methods are given below.

[0284] Example 1: Information recommendation devices are mobile applications and modules.

[0285] The information recommendation device 555 in this embodiment can provide a software module designed using programming languages ​​such as C / C++ and Java, which is embedded in various mobile applications based on systems such as Android or iOS (stored as executable instructions in the storage medium of the mobile device and executed by the processor of the mobile device), thereby directly using the computing resources of the mobile device itself to complete the relevant information recommendation tasks, and periodically or irregularly transmitting the processing results to a remote server through various network communication methods, or saving them locally on the mobile device.

[0286] Example 2: The information recommendation device is a server application and platform.

[0287] The information recommendation device 555 in this embodiment can be a dedicated software module in an application software or large software system designed using programming languages ​​such as C / C++ and Java. It runs on the server side (stored in the server-side storage medium as executable instructions and executed by the server-side processor). The server uses its own computing resources to complete the relevant information recommendation tasks.

[0288] This application embodiment can also provide an information recommendation platform (for recommendation lists) for use by individuals, groups or organizations by mounting a customized, easy-to-interact web interface or other user interfaces (UI) on a distributed, parallel computing platform composed of multiple servers.

[0289] Example 3: The information recommendation device consists of server-side application programming interfaces (APIs) and plugins.

[0290] The information recommendation device 555 in this embodiment can be provided as a server-side API or plugin for users to call, to execute the AI-based information recommendation method of this embodiment, and to be embedded in various applications.

[0291] Example 4: Information recommendation devices are mobile device client APIs and plugins.

[0292] The information recommendation device 555 in this embodiment can be provided as an API or plugin on a mobile device for users to call in order to execute the artificial intelligence-based information recommendation method of this embodiment.

[0293] Example 5: The information recommendation device is an open service in the cloud.

[0294] The information recommendation device 555 in this embodiment can provide a cloud service for information recommendation developed for users, allowing individuals, groups or organizations to make information recommendations.

[0295] The information recommendation device 555 includes a series of modules, including a feature cross-referencing module 5551, an encoding module 5552, an attention module 5553, a prediction module 5554, a recommendation module 5555, a training module 5556, and a construction module 5557. The following describes the scheme for the cooperation of the various modules in the information recommendation device 555 provided in this embodiment to achieve information recommendation.

[0296] The feature cross-processing module 5551 is used to perform feature cross-processing on the first part of the attribute information among the multiple attribute information of the information to be recommended, to obtain the cross-features of the information to be recommended; the encoding module 5552 is used to encode the second part of the attribute information among the multiple attribute information of the information to be recommended, to obtain the encoded features of the information to be recommended; the attention module 5553 is used to perform attention processing based on the cross-features and the encoded features, to obtain the attention features of the information to be recommended; the prediction module 5554 is used to perform bias prediction processing based on the attention features of the information to be recommended, to obtain the recommendation scenario adapted to the information to be recommended; and the recommendation module 5555 is used to perform a recommendation operation based on the recommendation scenario adapted to the information to be recommended.

[0297] In some embodiments, the feature cross-processing module 5551 is further configured to encode a first part of the attribute information among the multiple attribute information of the information to be recommended, to obtain the encoded features of the first part of the attribute information; to perform second-order feature cross-processing on the encoded features of the first part of the attribute information, to obtain the second-order cross-features of the first part of the attribute information; and to perform mapping processing on the second-order cross-features of the first part of the attribute information, to obtain the cross-features of the information to be recommended.

[0298] In some embodiments, the first part of the attribute information includes fixed attribute information and self-media attribute information; the feature cross module 5551 is further configured to discretize the fixed attribute information of the information to be recommended to obtain discrete features of the fixed attribute information; vectorize the discrete features of the fixed attribute information to obtain the encoded features of the fixed attribute information; and vectorize the self-media attribute information of the information to be recommended to obtain the encoded features of the self-media attribute information.

[0299] In some embodiments, the first part of the attribute information includes text attribute information; the feature cross module 5551 is further configured to perform word segmentation on the text attribute information of the information to be recommended to obtain multiple words of the text attribute information; perform vectorization on the multiple words of the text attribute information to obtain word vectors corresponding to the multiple words respectively; and perform average processing on the word vectors corresponding to the multiple words respectively to obtain the encoding features of the text attribute information.

[0300] In some embodiments, the feature crossing module 5551 is further configured to perform second-order feature crossing processing on the encoding features corresponding to any two attribute information in the first part of attribute information to obtain second-order crossing features of the two attribute information; and to perform fusion processing on the second-order crossing features of the two attribute information and the encoding features of the first part of attribute information to obtain second-order crossing features of the first part of attribute information.

[0301] In some embodiments, the feature cross module 5551 is further configured to perform matrix decomposition processing on the second-order cross features of the first part of the attribute information to obtain the decomposed features of the first part of the attribute information; and to perform nonlinear mapping processing on the decomposed features of the first part of the attribute information to obtain the cross features of the information to be recommended.

[0302] In some embodiments, the second part of the attribute information includes text attribute information and image attribute information; the encoding module 5552 is further configured to perform text encoding processing on the multiple text attribute information of the information to be recommended to obtain the text encoding features of the information to be recommended; and to perform image encoding processing on the image attribute information of the information to be recommended to obtain the image encoding features of the information to be recommended.

[0303] In some embodiments, the encoding module 5552 is further configured to perform feature extraction processing on multiple text attribute information of the information to be recommended, respectively, to obtain text features of the text attribute information; and to perform mapping processing on the text features of the multiple text attribute information to obtain text encoding features of the information to be recommended.

[0304] In some embodiments, the encoding module 5552 is further configured to perform convolution processing on multiple text attribute information of the information to be recommended, respectively, to obtain convolution features of the text attribute information; and to perform pooling processing on the convolution features of the text attribute information to obtain text features of the text attribute information.

[0305] In some embodiments, the encoding module 5552 is further configured to perform feature extraction processing on the image attribute information of the information to be recommended, to obtain the image features of the image attribute information; and to perform mapping processing on the image features of the image attribute information, to obtain the image encoding features of the information to be recommended.

[0306] In some embodiments, the encoding module 5552 is further configured to perform residual connection processing on the image attribute information of the information to be recommended to obtain residual features of the image attribute information; and perform nonlinear mapping processing on the residual features of the image attribute information to obtain image features of the image attribute information.

[0307] In some embodiments, the encoding module 5552 is further configured to perform dimensionality reduction processing on the image attribute information of the information to be recommended to obtain dimensionality-reduced image features; perform convolution processing on the dimensionality-reduced image features to obtain convolutional features of the image attribute information; perform dimensionality increase processing on the convolutional features of the image attribute information to obtain dimensionality-increased image features; and perform summation processing on the image attribute information of the information to be recommended and the dimensionality-increased image features to obtain residual features of the image attribute information.

[0308] In some embodiments, the attention module 5553 is further configured to concatenate the cross features and the encoded features to obtain the concatenated features of the information to be recommended; and to perform self-attention processing on the concatenated features of the information to be recommended to obtain the attention features of the information to be recommended.

[0309] In some embodiments, the prediction module 5554 is further configured to perform mapping processing on the attention features of the information to be recommended to obtain a probability distribution of candidate recommendation scenarios; and determine the candidate recommendation scenario corresponding to the highest probability in the probability distribution as the recommendation scenario that is suitable for the information to be recommended.

[0310] In some embodiments, the recommendation module 5555 is further configured to store the information to be recommended in a database corresponding to the recommendation scenario based on the recommendation scenario to which the information to be recommended is adapted; when it is determined that the target user is adapted to the recommendation scenario based on the target user's profile information, at least one of the information to be recommended in the database corresponding to the recommendation scenario is recommended to the target user.

[0311] In some embodiments, the information recommendation method is implemented by invoking a bias model; the information recommendation device 555 further includes: a training module 5556, configured to obtain information samples from a database corresponding to recommendation scenario samples; perform feature cross processing on a first part of the attribute information among multiple attribute information of the information sample to obtain cross features of the information sample; determine the attention features of the information sample based on the encoding features of the information sample and the cross features; perform bias prediction processing based on the attention features of the information sample to obtain the predicted recommendation scenario of the information sample; construct the loss function of the bias model based on the predicted recommendation scenario of the information sample and the recommendation scenario sample; update the parameters of the bias model, and use the updated parameters of the bias model when the loss function converges as the parameters of the trained bias model.

[0312] In some embodiments, the information recommendation device 555 further includes: a construction module 5557, configured to: determine attribute information of user samples corresponding to the recommendation scenario sample; obtain a user set formed by users having the attribute information; obtain a recommendation information set formed by historical recommendation information that has interactions with users in the user set; obtain interaction data between each user in the user set and each piece of historical recommendation information in the recommendation information set; and, based on the interaction data, select historical recommendation information that matches the recommendation scenario sample from the recommendation information set as information samples to construct a database corresponding to the recommendation scenario sample.

[0313] In some embodiments, the construction module 5557 is further configured to sort the historical recommendation information in the recommendation information set in descending order based on the interaction data between any user and each historical recommendation information in the recommendation information set, and take the portion of the historical recommendation information that appears first in the descending order as an information sample that is adapted to the recommendation scenario sample.

[0314] In some embodiments, the construction module 5557 is further configured to perform aggregate averaging processing on the interaction data between each user in the user set and any historical recommendation information to obtain the average interaction value of the user sample with respect to the historical recommendation information; to perform averaging processing on the average interaction value of the user sample with respect to all the historical recommendation information to obtain the total average interaction value of the user sample; and when the average interaction value of the user sample with respect to the historical recommendation information is greater than the total average interaction value, the historical recommendation information is used as an information sample adapted to the recommendation scenario sample.

[0315] This application provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the artificial intelligence-based information recommendation method described above in this application.

[0316] This application provides a computer-readable storage medium storing executable instructions. When these executable instructions are executed by a processor, they cause the processor to execute the artificial intelligence-based information recommendation method provided in this application. For example... Figures 3A-3C The information recommendation method based on artificial intelligence shown is as follows: Figure 7 The example shown is an information recommendation method based on artificial intelligence.

[0317] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0318] In some embodiments, executable instructions may take 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 standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0319] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0320] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0321] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. An information recommendation method based on artificial intelligence, characterized in that, The method comprises: encoding processing is performed on a first part of attribute information in a plurality of attribute information of the information to be recommended, to obtain an encoding feature of the first part of attribute information; second-order cross feature processing is performed on the encoding features corresponding to any two attribute information in the first part of attribute information, to obtain a second-order cross feature of the any two attribute information; fusion processing is performed on the second-order cross feature of the any two attribute information and the encoding feature of the first part of attribute information, to obtain a second-order cross feature of the first part of attribute information; mapping processing is performed on the second-order cross feature of the first part of attribute information, to obtain a cross feature of the information to be recommended; encoding processing is performed on a second part of attribute information in the plurality of attribute information of the information to be recommended, to obtain an encoding feature of the information to be recommended; attention processing is performed based on the cross feature and the encoding feature, to obtain an attention feature of the information to be recommended; tendency prediction processing is performed based on the attention feature of the information to be recommended, to obtain a recommended scene adapted to the information to be recommended; a recommendation operation is performed based on the recommended scene adapted to the information to be recommended.

2. The method of claim 1, wherein the first part of attribute information comprises text attribute information; the encoding processing performed on the first part of attribute information in the plurality of attribute information of the information to be recommended to obtain the encoding feature of the first part of attribute information comprises: word segmentation processing is performed on the text attribute information of the information to be recommended, to obtain a plurality of words of the text attribute information; vectorization processing is performed on the plurality of words of the text attribute information, to obtain word vectors corresponding to the plurality of words respectively; average processing is performed on the word vectors corresponding to the plurality of words respectively, to obtain the encoding feature of the text attribute information.

3. The method of claim 1, wherein, the mapping processing performed on the second-order cross feature of the first part of attribute information to obtain the cross feature of the information to be recommended comprises: matrix decomposition processing is performed on the second-order cross feature of the first part of attribute information, to obtain a decomposition feature of the first part of attribute information; nonlinear mapping processing is performed on the decomposition feature of the first part of attribute information, to obtain the cross feature of the information to be recommended.

4. The method of claim 1, wherein the second part of attribute information comprises text attribute information and image attribute information; the encoding processing performed on the second part of attribute information in the plurality of attribute information of the information to be recommended to obtain the encoding feature of the information to be recommended comprises: text encoding processing is performed on a plurality of the text attribute information of the information to be recommended, to obtain a text encoding feature of the information to be recommended; image encoding processing is performed on the image attribute information of the information to be recommended, to obtain an image encoding feature of the information to be recommended.

5. The method of claim 4, wherein, the text encoding processing performed on the plurality of the text attribute information of the information to be recommended to obtain the text encoding feature of the information to be recommended comprises: feature extraction processing is performed on the plurality of the text attribute information of the information to be recommended respectively, to obtain text features of the text attribute information; The text features of the plurality of text attribute information of the to-be-recommended information are mapped to obtain text coding features of the to-be-recommended information.

6. The method of claim 5, wherein, The feature extraction processing of the plurality of text attribute information of the to-be-recommended information respectively obtains text features of the text attribute information, including: The convolution processing of the plurality of text attribute information of the to-be-recommended information respectively obtains convolution features of the text attribute information; The pooling processing of the convolution features of the text attribute information obtains text features of the text attribute information.

7. The method of claim 4, wherein, The image coding processing of the image attribute information of the to-be-recommended information obtains image coding features of the to-be-recommended information, including: The feature extraction processing of the image attribute information of the to-be-recommended information obtains image features of the image attribute information; The mapping processing of the image features of the image attribute information obtains image coding features of the to-be-recommended information.

8. The method of claim 7, wherein, The feature extraction processing of the image attribute information of the to-be-recommended information obtains image features of the image attribute information, including: The residual connection processing of the image attribute information of the to-be-recommended information obtains residual features of the image attribute information; The nonlinear mapping processing of the residual features of the image attribute information obtains image features of the image attribute information.

9. The method of claim 8, wherein, The residual connection processing of the image attribute information of the to-be-recommended information obtains residual features of the image attribute information, including: The dimension reduction processing of the image attribute information of the to-be-recommended information obtains reduced image features; The convolution processing of the reduced image features obtains convolution features of the image attribute information; The dimension reduction processing of the image attribute information of the to-be-recommended information obtains reduced image features; The dimension reduction processing of the image attribute information of the to-be-recommended information obtains reduced image features; 10.A method for recommending information based on artificial intelligence, the method comprising: receiving a user input; determining a user interest based on the user input; and recommending information based on the user interest. The method comprises: The information recommendation task is realized by calling a tendency model; The training process of the tendency model comprises the following processing: Obtaining information samples from a database corresponding to a recommendation scene sample; Encoding processing of a first part of attribute information in the plurality of attribute information of the information sample obtains coding features of the first part of attribute information; Second-order feature cross processing of coding features corresponding to any two attribute information in the first part of attribute information respectively obtains second-order cross features of the any two attribute information; Fusion processing of the second-order cross features of the any two attribute information and the coding features of the first part of attribute information obtains second-order cross features of the first part of attribute information; Mapping processing of the second-order cross features of the first part of attribute information obtains cross features of the information sample; Based on the coding features of the information sample and the cross features, attention features of the information sample are determined; Based on the attention features of the information sample, tendency prediction processing is performed to obtain a predicted recommendation scene of the information sample; construct a loss function of the tendency model based on the predicted recommendation scene of the information sample and the recommendation scene sample; update parameters of the tendency model, and take the updated parameters of the tendency model when the loss function converges as the parameters of the trained tendency model.

11. An information recommendation apparatus based on artificial intelligence, characterized by comprising: The device comprises: a feature cross module configured to perform encoding processing on a first part of attribute information in a plurality of attribute information of the information to be recommended to obtain encoded features of the first part of attribute information, perform second-order feature cross processing on encoded features corresponding to any two attribute information in the first part of attribute information to obtain second-order cross features of the any two attribute information, perform fusion processing on the second-order cross features of the any two attribute information and the encoded features of the first part of attribute information to obtain second-order cross features of the first part of attribute information, and perform mapping processing on the second-order cross features of the first part of attribute information to obtain cross features of the information to be recommended; an encoding module configured to perform encoding processing on a second part of attribute information in the plurality of attribute information of the information to be recommended to obtain encoded features of the information to be recommended; an attention module configured to perform attention processing based on the cross features and the encoded features to obtain attention features of the information to be recommended; a prediction module configured to perform tendency prediction processing based on the attention features of the information to be recommended to obtain a recommended scene to which the information to be recommended is adapted; a recommendation module configured to perform a recommendation operation based on the recommended scene to which the information to be recommended is adapted.

12. The apparatus of claim 11, wherein, The first part of attribute information comprises text attribute information. The feature cross module is further configured to perform word segmentation processing on text attribute information of the information to be recommended to obtain a plurality of words of the text attribute information, perform vectorization processing on the plurality of words of the text attribute information to obtain word vectors corresponding to the plurality of words respectively, and perform average processing on the word vectors corresponding to the plurality of words respectively to obtain the encoded features of the text attribute information.

13. The device of claim 11, wherein the feature cross module is further configured to perform matrix decomposition processing on the second-order cross features of the first part of attribute information to obtain decomposition features of the first part of attribute information; and perform nonlinear mapping processing on the decomposition features of the first part of attribute information to obtain the cross features of the information to be recommended. The electronic device comprises:

14. An electronic device, comprising: a memory configured to store executable instructions; a processor configured to execute the executable instructions stored in the memory to implement the information recommendation method based on artificial intelligence according to any one of claims 1 to 10. The executable instructions are stored in the memory and are configured to be executed by the processor to implement the information recommendation method based on artificial intelligence according to any one of claims 1 to 10.

15. A computer-readable storage medium, characterized in that, The computer instructions are executed by the processor to implement the information recommendation method based on artificial intelligence according to any one of claims 1 to 10.

16. A computer program product comprising computer instructions, characterized in that, ​

Images