Content recommendation method, electronic equipment and vehicle

Abstract

The invention relates to the technical field of data processing, and provides a content recommendation method, electronic equipment and a vehicle. The method comprises the steps of obtaining historical behaviors of a target user in application software, determining the similarity between the target user and each piece of published content according to the historical behaviors, and determining a preset number of recommended content from multiple pieces of published content according to the similarity; wherein the application software comprises a plurality of published contents; scoring each recommendation content based on a pre-trained large language model to obtain a score, and determining an initial recommendation sequence according to the score; based on a pre-trained large language model, determining topic tags of the preset number of recommended contents, updating the initial recommendation sequence according to the topic tags to obtain a target recommendation sequence, and displaying the preset number of recommended contents according to the target recommendation sequence.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a content recommendation method, an electronic device and a vehicle. BACKGROUND

[0002] At present, in some content recommendation software, a keyword is determined from user search information, and then corresponding information recommendation is made according to the keyword. However, the content recommendation made according to the keyword has the problems of not conforming to user preferences and being single in type.

[0003] Therefore, how to avoid the problems of not conforming to user preferences and being single in type of recommended content has become a technical problem to be solved urgently. SUMMARY

[0004] Therefore, the present application aims to provide a content recommendation method, an electronic device and a vehicle to solve the problems of not conforming to user preferences and being single in type of recommended content in the prior art.

[0005] To achieve the above purpose, a first aspect of the present application provides a content recommendation method, which comprises:

[0006] obtaining historical behaviors of a target user in an application software, determining a similarity between the target user and each published content according to the historical behaviors, and determining a predetermined number of recommended contents from a plurality of published contents according to the similarity; wherein the application software has a plurality of published contents;

[0007] scoring each recommended content based on a pre-trained large language model to obtain a score value, and determining an initial recommendation order according to the score value;

[0008] determining a topic label of the predetermined number of recommended contents based on the pre-trained large language model, updating the initial recommendation order according to the topic label to obtain a target recommendation order, and displaying the predetermined number of recommended contents in the target recommendation order.

[0009] In some embodiments, the determining of the similarity between the target user and each published content according to the historical behaviors and the determining of the predetermined number of recommended contents from the plurality of published contents according to the similarity comprise:

[0010] analyzing and processing the historical behaviors based on a pre-trained behavior extraction model to obtain user preferences;

[0011] obtaining description information of each published content, and analyzing and processing the description information based on a pre-trained content extraction model to obtain content information;

[0012] determine a similarity between the target user and each of the published contents according to the user preference and the content information;

[0013] sort the plurality of published contents in an order from high to low similarity, and determine a predetermined number of recommended contents with high similarity from the sorted plurality of published contents.

[0014] In some embodiments, the analyzing and processing the historical behaviors by the pre-trained behavior extraction model to obtain the user preference comprises:

[0015] analyzing and processing the historical behaviors to obtain user operations and operation content features;

[0016] combining and processing the user operations and the operation content features to obtain operation data, and converting the operation data into operation data in a preset format;

[0017] analyzing and processing the target user data and the operation data in the preset format by the pre-trained behavior extraction model to obtain the user preference.

[0018] In some embodiments, the content extraction model comprises a large language model and a visual language model.

[0019] The obtaining of the description information of each of the published contents comprises analyzing and processing the description information by a pre-trained content extraction model to obtain content information.

[0020] obtaining text description information and picture description information of each of the published contents;

[0021] extracting and processing the text description information by the pre-trained large language model to obtain content abstracts and content titles;

[0022] extracting and processing the picture description information by the pre-trained visual language model to obtain picture content information;

[0023] taking the content abstracts, the content titles and the picture content information as the content information.

[0024] In some embodiments, the determining of the similarity between the target user and each of the published contents according to the user preference and the content information comprises:

[0025] vectorizing the user preference to obtain a user preference vector;

[0026] extracting features of each of the published contents to obtain brand labels and topic labels, and vectorizing the brand labels, the topic labels and the content information to obtain a content feature vector.

[0027] The similarity between the target user and each published content is determined according to the user preference vector and the content characteristic vector.

[0028] In some embodiments, scoring each recommended content based on a pre-trained large language model to obtain a score, and determining an initial recommendation order based on the score, includes:

[0029] Based on a pre-trained large language model, scoring each recommended content according to the user characteristics of the target user and the content characteristics of each recommended content to obtain a score for each recommended content;

[0030] The predetermined number of recommended contents are sorted from high to low according to the scores to obtain an initial recommendation order.

[0031] In some embodiments, determining the topic tags of the predetermined number of recommended contents based on a pre-trained large language model, and updating the initial recommendation order according to the topic tags to obtain a target recommendation order, includes:

[0032] Dividing the predetermined number of recommended contents into a first recommendation group and other recommendation groups according to the initial recommendation order;

[0033] Determining first topic tags of the recommended content in the first recommendation group based on a pre-trained large language model, and determining whether the first topic tags are consistent;

[0034] In response to determining that the first topic tags are inconsistent, taking the initial recommendation order as a target recommendation order;

[0035] In response to determining that the first topic tag is consistent, target recommended content inconsistent with the first topic tag is determined from the other recommendation groups, and the target recommended content is added from the other recommendation groups to the first recommendation group to obtain an updated target recommendation order.

[0036] In some embodiments, the application software is an application software in a vehicle computer;

[0037] The step of obtaining the target user's historical behavior in the application software, determining the similarity between the target user and each published content based on the historical behavior, and determining a predetermined number of recommended content from the plurality of published content based on the similarity includes:

[0038] Obtaining the target user's historical behavior in the application software in the vehicle computer, and retrieving the target user's driving data from the vehicle computer;

[0039] construct a user portrait of the target user based on the historical behavior and the driving data;

[0040] determine a similarity between the target user and each published content according to the user portrait, and determine a predetermined number of recommended contents from the plurality of published contents according to the similarity.

[0041] Based on the same inventive concept, a second aspect of the present application provides an electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable by the processor, wherein the processor implements the method as described above when executing the computer program.

[0042] Based on the same inventive concept, a third aspect of the present application provides a vehicle, comprising the electronic device of the second aspect.

[0043] As can be seen from the above, the content recommendation method, electronic device and vehicle provided by the present application are provided. The historical behavior of the target user in the application software is obtained, the similarity between the target user and each published content is determined according to the historical behavior, and a predetermined number of recommended contents are determined from the plurality of published contents according to the similarity. In this way, the similarity determined according to the historical behavior can reflect the preference degree of the target user for each published content, and the recommended content determined according to the similarity can meet the user's preference. The score of each recommended content is scored based on the pre-trained large language model, and the initial recommendation order is determined according to the score. In this way, the comprehensive scoring of each recommended content is realized, so as to reflect the matching degree between each recommended content and the target user, so as to realize the accurate sorting of the predetermined number of recommended contents according to the score to determine the initial recommendation order. The topic label of the predetermined number of recommended contents is determined based on the pre-trained large language model, the initial recommendation order is updated according to the topic label to obtain a target recommendation order, and the predetermined number of recommended contents are displayed according to the target recommendation order. In this way, the initial recommendation order is updated according to the topic label, avoiding that multiple recommended contents adjacent in order are the same type of content, thereby avoiding the situation that the recommended contents displayed are of the same type or single type. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the present application or the related art, the following will briefly introduce the drawings needed to be used in the embodiments or related art descriptions. Obviously, the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0045] Figure 1 Flowchart of the content recommendation method of an embodiment of the present application;

[0046] Figure 2 Flow chart of a content recommendation method according to another embodiment of the present application;

[0047] Figure 3 Structural schematic diagram of a content recommendation device according to an embodiment of the present application;

[0048] Figure 4 Structural schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to specific embodiments and drawings.

[0050] It should be noted that, unless otherwise defined, technical terms or scientific terms used in the embodiments of the present application should be understood as their common meanings to those skilled in the art to which the present application belongs. The terms "first", "second" and similar terms used in the embodiments of the present application do not represent any order, number or importance, but are only used to distinguish different components. The terms "include" or "contain" and similar terms mean that the components or objects before the terms encompass the components or objects listed after the terms and their equivalents, without excluding other components or objects. The terms "connect" or "connected" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The terms "upper", "lower", "left", "right" and the like are only used to represent relative positional relationships, and when the absolute positions of the described objects change, the relative positional relationships can also change accordingly.

[0051] Based on the description of the background, in the traditional recommendation system, the following significant technical bottlenecks exist in the recall, precision ranking and re-ranking links:

[0052] 1. Recall link: Traditional recall algorithms (e.g., collaborative filtering, content-based recall) rely on shallow user behavior data and surface features of items, lacking semantic understanding ability of user demand. When a user searches for "lightweight equipment suitable for camping", the traditional recall algorithm is difficult to analyze the semantic demand of "lightweight" and "camping scene adaptation", resulting in that the recall result deviates from the user's true intention. At the same time, the recall result is single in category, which is difficult to cover the user's multiple demands.

[0053] 2. Precision ranking link: The precision ranking model based on deep learning is limited by fixed feature engineering and limited network structure, and it is difficult to automatically learn the complex patterns of user and item interaction. When dealing with high-dimensional sparse data, the precision ranking model cannot effectively mine the high-order semantic association between features, resulting in insufficient recommendation accuracy. In addition, the precision ranking model tends to recommend popular content, causing excessive concentration of traffic, insufficient exposure of long-tail content, and single category and form of the recommended list, affecting user experience.

[0054] 3. Rearrangement link: current rearrangement strategy depends on manual rules or simple statistical methods, lacking in-depth analysis of user demand, item characteristics and scene information. It cannot dynamically balance the diversity, relevance and business goals of the recommendation results, and is difficult to solve the problems of monotonous form and uneven traffic distribution, and cannot meet the dual needs of user personalization and platform ecological health development.

[0055] As described above, how to avoid the recommended content not meeting the user's preferences and the single type of recommended content has become an important research problem.

[0056] The content recommendation method proposed by the embodiments of the present application is applied to a control system (for example, a car machine control system) as shown in the figure. Figure 1 The method includes steps 101-103.

[0057] Step 101, obtaining the historical behavior of the target user in the application software, determining the similarity between the target user and each published content according to the historical behavior, and determining a predetermined number of recommended contents from a plurality of published contents according to the similarity; wherein the application software has a plurality of published contents.

[0058] In specific implementation, the historical behavior of the target user in the application software is obtained. The historical behavior includes user operation and operation content characteristics. The user operation includes clicking, collecting, liking, forwarding and clicking dislike. The operation content characteristics include operation content brand label, operation content topic label and operation content heat value.

[0059] The application software has a plurality of published contents. For example, when the application software is a video software, the published content can be a video that has been published in the application software; when the application software is a social sharing software, the published content can be a video, note or post that has been published in the application software.

[0060] The user preference is obtained by analyzing and processing the historical behavior, and the user preference vector is obtained by vectorizing the user preference. The content information is obtained by analyzing and processing the description information of each published content, and the brand label and topic label are obtained by feature extraction processing of each published content. The content feature vector is obtained by vectorizing the brand label, topic label and content information. The similarity between the target user and each published content is determined according to the user preference vector and the content feature vector. A predetermined number of recommended contents with high similarity are determined from a plurality of published contents.

[0061] Step 102, scoring each recommended content based on a pre-trained large language model to obtain a score value, and determining an initial recommendation order according to the score value.

[0062] In specific implementation, based on the pre-trained large language model, scores of each recommended content are obtained by scoring each recommended content according to the user features of the target user and the content features of each recommended content, wherein the scores can reflect the matching degree between each recommended content and the target user.

[0063] The training process of the large language model comprises: obtaining the user features and the content features of each recommended content; and training the large language model by using the user features and the content features to obtain the trained large language model.

[0064] The predetermined number of recommended contents are sorted in descending order of the scores to obtain an initial recommendation order. In this way, the recommended contents with high matching degrees to the target user can be preferentially recommended according to the initial recommendation order.

[0065] Step 103, based on the pre-trained large language model, determine the topic labels of the predetermined number of recommended contents, update the initial recommendation order according to the topic labels to obtain a target recommendation order, and display the predetermined number of recommended contents according to the target recommendation order.

[0066] In specific implementation, the predetermined number of recommended contents are divided into a first recommended group and other recommended groups according to the initial recommendation order. The first topic labels of the recommended contents in the first recommended group are determined based on the pre-trained large language model.

[0067] The training process of the large language model comprises: obtaining the topic labels of each recommended content; and training the large language model by using the recommended contents and the corresponding topic labels to obtain the trained large language model.

[0068] When the first topic labels are inconsistent, it indicates that the recommended contents in the first recommended group are different types of contents (i.e., there is no problem of single type of content clustering), and the initial recommendation order corresponding to the first recommended group is taken as the target recommendation order.

[0069] When the first topic labels are consistent, it indicates that the recommended contents in the first recommended group are the same type of content (i.e., there is a problem of single type of content clustering), and the initial recommendation order corresponding to the first recommended group is updated to obtain the target recommendation order.

[0070] By the above scheme, the historical behavior of the target user in the application software is obtained, the similarity between the target user and each published content is determined according to the historical behavior, and the predetermined number of recommended contents are determined from the multiple published contents according to the similarity. In this way, the similarity determined according to the historical behavior can reflect the preference degree of the target user for each published content, and the recommended content determined according to the similarity can meet the user's preference. The score of each recommended content is determined based on the pre-trained large language model, and the initial recommendation order is determined according to the score. In this way, the comprehensive scoring of each recommended content is realized, so as to reflect the matching degree between each recommended content and the target user, so as to realize the accurate sorting of the predetermined number of recommended contents according to the score to determine the initial recommendation order. The topic label of the predetermined number of recommended contents is determined based on the pre-trained large language model, the initial recommendation order is updated according to the topic label to obtain the target recommendation order, and the predetermined number of recommended contents are displayed according to the target recommendation order. In this way, the initial recommendation order is updated according to the topic label, avoiding that multiple recommended contents adjacent in order are the same type of content, thereby avoiding the situation that the recommended contents displayed are of the same type or single type.

[0071] In some embodiments, step 101 comprises:

[0072] Step 1011, based on the pre-trained behavior extraction model, the historical behavior is analyzed and processed to obtain user preferences.

[0073] In specific implementation, the historical behavior is analyzed and processed to obtain user operations and operation content features; the user preferences are determined based on the pre-trained behavior extraction model according to the target user data, the user operations and the operation content features.

[0074] The training process of the behavior extraction model comprises: obtaining historical behavior of a user and corresponding user preferences; and training the behavior extraction model using the historical behavior and the corresponding user preferences to obtain a trained behavior extraction model.

[0075] For example, the historical behavior of the target user includes: clicking 20 pieces of first vehicle type content and 10 pieces of camping content, liking 5 pieces of first vehicle type content and 3 pieces of camping content, no collection content, and clicking 2 pieces of second vehicle type content. The historical behavior is analyzed and processed to obtain user operations and operation content features. The user operations include: clicking, liking and clicking dislike, and the operation content features include: first vehicle type content, camping content and second vehicle type content. The user preferences are determined based on the pre-trained behavior extraction model according to the target user data, the user operations and the operation content features, and the user preferences are first vehicle type content and camping, i.e. the target user may like off-road, outdoor and freedom, and the target user does not like second vehicle type content.

[0076] At step 1012, the description information of each published content is obtained, and the content information is obtained by analyzing and processing the description information based on a pre-trained content extraction model.

[0077] In specific implementation, for a plurality of published contents in the application software, the description information of each published content is obtained, wherein the description information includes text description information and picture description information. The content extraction model includes a large language model and a visual language model. The text content information is obtained by extracting and processing the text description information based on the pre-trained large language model. The picture content information is obtained by extracting and processing the picture description information based on the pre-trained visual language model. The text content information and the picture content information are taken as the content information.

[0078] In specific implementation, the training process of the content extraction model includes obtaining the description information of the published content and the corresponding content information, and training the content extraction model based on the description information and the corresponding content information to obtain the trained content extraction model.

[0079] At step 1013, the similarity between the target user and each published content is determined according to the user preference and the content information.

[0080] In specific implementation, the user preference vector is determined according to the user preference. The content feature vector of each published content is determined according to the content information. The similarity between the target user and each published content is determined according to the user preference vector and the content feature vector.

[0081] At step 1014, the plurality of published contents are sorted in descending order of similarity, and a predetermined number of recommended contents with high similarity are determined from the sorted plurality of published contents.

[0082] In specific implementation, the plurality of published contents can be a plurality of published contents pre-stored in a content library. The plurality of published contents in the content library are sorted in descending order of similarity, and a predetermined number of recommended contents with high similarity are determined from the sorted plurality of published contents.

[0083] For example, 200 published contents are stored in the content library, and the predetermined number is 100, that is, 100 recommended contents matching the preference of the target user are determined from the 200 published contents. The 200 published contents are sorted in descending order of similarity, and the first 100 recommended contents with high similarity are determined from the sorted 200 published contents.

[0084] By the above scheme, the user preference is obtained by analyzing and processing the historical behavior based on the pre-trained behavior extraction model, and the behavior extraction model can accurately identify the user preference of the target user. The description information of each published content is obtained, and the content information is obtained by analyzing and processing the description information based on the pre-trained content extraction model, and the content extraction model can accurately extract the content information of each published content. According to the user preference and the content information, the similarity between the target user and each published content is determined, and the similarity between the target user and each published content can be accurately determined by vector calculation. The plurality of published contents are sorted in order from high to low according to the similarity, and a predetermined number of recommended contents with high similarity are determined from the sorted plurality of published contents, so that the determined recommended contents are recommended contents matching the target user preference.

[0085] In some embodiments, step 1011 includes:

[0086] Step 1011A, analyzing and processing the historical behavior to obtain user operations and operation content features.

[0087] In specific implementation, the historical behavior is analyzed and processed to obtain user operations and operation content features. The user operations include clicking, collecting, liking, forwarding and clicking dislike. The operation content features include operation content brand tags, operation content topic tags and operation content heat values.

[0088] Step 1011B, combining and processing the user operations and the operation content features to obtain operation data, and converting the operation data into operation data in a preset format.

[0089] In specific implementation, the user operations and the operation content features are combined and processed to obtain operation data. For example, the user operations include clicking, liking and clicking dislike, and the operation content features include first vehicle content, camping content and second vehicle content. The user operations and the operation content features are combined and processed to obtain operation data, and the operation data includes that the target user clicks the first vehicle content and the camping content, the target user likes the first vehicle content and the camping content, and the target user clicks dislike the second vehicle.

[0090] The preset format includes a natural language format. The operation data is converted into operation data in a natural language format. For example, the operation data in the natural language format includes clicking 20 first vehicle content and 10 camping content, liking 5 first vehicle content and 3 camping content, no collection content, and clicking dislike 2 second vehicle content.

[0091] Step 1011C, analyzing and processing the target user data and the operation data in the preset format based on the pre-trained behavior extraction model to obtain the user preference.

[0092] In specific implementation, the target user data includes target user gender and target user age. The target user data and the operation data in the preset format are input into the pre-trained behavior extraction model, and user preference is obtained by analyzing and processing the target user data and the operation data in the preset format based on the pre-trained behavior extraction model. For example, the user preference is the first vehicle type content and camping, i.e., the target user may like off-road, outdoor and freedom, and the target user does not like the second vehicle type content.

[0093] The target user data and the operation data in the preset format are input into the pre-trained behavior extraction model, and the target user data and the operation data in the preset format are converted into a prompt word Prompt and input into the pre-trained behavior extraction model. The prompt word Prompt is used to prompt the context of the input information to the behavior extraction model, and the prompt word Prompt can be a question, a text description, a dialogue or any form of text input.

[0094] Through the above scheme, the user operation and the operation content feature are obtained by analyzing and processing the historical behavior, the operation data is obtained by combining and processing the user operation and the operation content feature, the operation data is converted into the operation data in the preset format, so that the behavior extraction model can accurately recognize and extract the operation data in the preset format. The user preference is obtained by analyzing and processing the target user data and the operation data in the preset format based on the pre-trained behavior extraction model, the behavior extraction model can accurately identify the user preference of the target user, and the predetermined number of recommended contents are accurately determined according to the user preference.

[0095] In some embodiments, the content extraction model includes a large language model and a visual language model, and step 1012 includes:

[0096] In step 1012A, text description information and picture description information of each published content are obtained.

[0097] In specific implementation, the published content includes at least one of the following: published video, published note and published post. No matter what type of published content is, there are text and pictures in the published content. In order to more accurately extract the content information of the published content, the text description information and the picture description information of each published content are obtained, and the text content information and the picture content information of the published content are extracted respectively.

[0098] The text description information is description information of the text type of the published content, for example, the text description information can be a long text description of 520 words of the published content. The picture description information is description information of the picture type of the published content, for example, the picture description information can be an image in a published video or a picture published in a published note.

[0099] Step 1012B, extracting and processing the text description information based on the pre-trained large language model to obtain a content summary and a content title.

[0100] In specific implementation, the text description information is extracted and processed to obtain an abstract prompt word Prompt and a title prompt word Prompt. The abstract prompt word Prompt is a prompt word that meets a first preset format and can reflect the key information of the text description information, and the title prompt word Prompt is a prompt word that meets a second preset format and can reflect the main content of the text description information.

[0101] The training process of the large language model includes: obtaining the text description information of the published content and the corresponding content summary and content title; and training the large language model using the text description information and the corresponding content summary and content title to obtain the trained large language model.

[0102] For example, the first preset format is no more than 200 words, and the second preset format is no more than 25 words. The text description information of the published content is a long text of 520 words. The text description information is extracted and processed to obtain an abstract prompt word Prompt of no more than 200 words, which can reflect the key information of the text description information. The text description information is extracted and processed to obtain a title prompt word Prompt of no more than 25 words, which can reflect the main content of the text description information.

[0103] Specifically, the abstract prompt word Prompt and the title prompt word Prompt are input into the pre-trained large language model (Large Language Models, LLM for short), the abstract prompt word Prompt is extracted and processed based on the pre-trained large language model to obtain a content summary, and the title prompt word Prompt is extracted and processed based on the pre-trained large language model to obtain a content title. The content summary is a text that can reflect the key information of the text description information, and the content title is a text that can reflect the main content of the text description information.

[0104] Step 1012C, extracting and processing the picture description information based on the pre-trained visual language model to obtain picture content information.

[0105] In specific implementation, the picture description information is extracted and processed to obtain a picture prompt word Prompt. The picture prompt word Prompt is a prompt word generated according to the objects and scenes in the picture.

[0106] The training process of the visual language model includes: obtaining picture description information and corresponding picture content information of the published content; and training the visual language model by using the picture description information and the corresponding picture content information to obtain a trained visual language model.

[0107] For example, the picture of the published content contains a first vehicle model and a camping scene. The picture description information is extracted to obtain picture prompt words Prompt of the first vehicle model and the camping scene.

[0108] Specifically, the picture prompt words Prompt are input into a pre-trained visual language model (VLM), and the picture prompt words Prompt are extracted based on the pre-trained visual language model to obtain picture content information.

[0109] Step 1012D, the content summary, the content title and the picture content information are taken as the content information.

[0110] In specific implementation, the content summary and the content title are taken as text content information, and the text content information and the picture content information are taken as content information corresponding to the published content, so as to determine the similarity between the target user and the published content according to the user preference and the content information.

[0111] Through the above scheme, the content summary and the content title are extracted from the text description information based on the pre-trained large language model, and the large language model can accurately extract the content summary and the content title. The picture content information is extracted from the picture description information based on the pre-trained visual language model, and the visual language model can accurately extract the picture content information. The content summary, the content title and the picture content information are taken as the content information, so as to determine the similarity between the target user and the published content according to the user preference and the content information.

[0112] In some embodiments, step 1013 includes:

[0113] Step 1013A, the user preference is vectorized to obtain a user preference vector.

[0114] In specific implementation, the user preference is vectorized to obtain a user preference vector. The user preference vector is a vector that can reflect the user preference.

[0115] Step 1013B, the brand label and the topic label are extracted from each published content, and the brand label, the topic label and the content information are vectorized to obtain a content feature vector.

[0116] In specific implementation, the determination process of the content feature vector can be: in step 1013B, brand labels and topic labels are extracted from each published content; and the brand labels, the topic labels and content information are vectorized to obtain the content feature vector.

[0117] In addition, the determination process of the content feature vector can also be: in step 1013b, the content feature vector of each published content is obtained from a content library (or a content vector library); and the content library stores a plurality of published contents and the content feature vector corresponding to each published content. Before step 1013b, the brand labels and the topic labels are extracted from each published content; and the brand labels, the topic labels and the content information are vectorized to obtain the content feature vector, and the content feature vector is stored in the content library.

[0118] In step 1013C, the similarity between the target user and each published content is determined according to the user preference vector and the content feature vector.

[0119] In specific implementation, the similarity between the target user and each published content is calculated according to the user preference vector and the content feature vector.

[0120] According to the above scheme, the user preference is vectorized to obtain the user preference vector, so that the user preference vector can not only reflect the user preference, but also facilitate similarity calculation. The brand labels and the topic labels are extracted from each published content, and the brand labels, the topic labels and the content information are vectorized to obtain the content feature vector, so that the content feature vector can comprehensively reflect the brand features, the topic features and the content features of each published content, and facilitate similarity calculation. The similarity between the target user and each published content is determined according to the user preference vector and the content feature vector, and the similarity between the target user and each published content can be accurately determined through vector calculation.

[0121] In some embodiments, step 102 includes:

[0122] In step 1021, each recommendation content is scored to obtain a score of each recommendation content according to the user features of the target user and the content features of each recommendation content based on a pre-trained large language model.

[0123] In specific implementation, the user features are extracted from the target user, and the user features include user basic features, historical behavior features and user preference features. The content features are extracted from each recommendation content, and the content features include brand labels, topic labels, content abstracts, content titles and picture content information.

[0124] inputting the user feature of the target user and the content feature of each recommended content into the pre-trained large language model; and scoring each recommended content according to the user feature of the target user and the content feature of each recommended content based on the pre-trained large language model to obtain a score of each recommended content; wherein the score can reflect the matching degree between each recommended content and the target user.

[0125] In step 1022, the predetermined number of recommended contents are sorted according to the scores from high to low to obtain an initial recommendation order.

[0126] In particular implementation, the predetermined number of recommended contents are sorted according to the scores from high to low, and the sorted order from high to low is taken as the initial recommendation order.

[0127] For example, the predetermined number of recommended contents include: content 1, content 2, content 3, content 4 and content 5. Based on the pre-trained large language model, the multiple recommended contents are scored, and the scoring results are: the score of content 1 is 6, the score of content 2 is 9, the score of content 3 is 8, the score of content 4 is 5, and the score of content 5 is 7. Thus, the determined initial recommendation order is: content 2, content 3, content 5, content 1, content 4.

[0128] Through the above scheme, based on the pre-trained large language model, each recommended content is scored according to the user feature of the target user and the content feature of each recommended content to obtain the score of each recommended content, and the matching degree between each recommended content and the target user can be accurately judged according to the score of each recommended content. The predetermined number of recommended contents are sorted according to the scores from high to low to obtain the initial recommendation order, so that the recommended contents with high matching degree to the target user can be preferentially recommended according to the initial recommendation order.

[0129] In some embodiments, step 103 includes:

[0130] In step 1031, the predetermined number of recommended contents are divided into a first recommended group and other recommended groups according to the initial recommendation order.

[0131] In particular implementation, the predetermined number of recommended contents are sorted according to the initial recommendation order, and the sorted recommended contents are divided into multiple recommended groups. The multiple recommended groups include: the first recommended group and the other recommended groups.

[0132] For example, the predetermined number of recommended contents is 100, and the recommended contents in each recommended group are 10. The predetermined number of recommended contents are divided into 10 recommended groups according to the initial recommended order. The predetermined number of recommended contents are sorted according to the initial recommended order, and the first 1-10 recommended contents are taken as the first recommended group, the 11-20 recommended contents are taken as the second recommended group, and so on, and the 91-100 recommended contents are taken as the tenth recommended group. Among them, the second recommended group to the tenth recommended group are other recommended groups.

[0133] In step 1032, the first topic label of the recommended content in the first recommended group is determined based on the pre-trained large language model, and whether the first topic label is consistent is determined.

[0134] In implementation, the recommended content in the first recommended group is extracted to obtain a rearrangement prompt Prompt, and the rearrangement prompt Prompt is input into the pre-trained large language model. The first topic label of the recommended content in the first recommended group is determined based on the pre-trained large language model. Whether the recommended content in the first recommended group has the problem of single type is determined by judging whether the first topic label is consistent.

[0135] In step 1033, in response to determining that the first topic label is inconsistent, the initial recommended order is taken as the target recommended order.

[0136] In implementation, when the first topic label is inconsistent, it indicates that the recommended content in the first recommended group is content of different types (i.e., there is no problem of content type clustering single type), and the initial recommended order corresponding to the first recommended group is taken as the target recommended order. After step 1033, whether the second topic label of the recommended content in the second recommended group is consistent is determined, and the target recommended order corresponding to the second recommended group is determined, until the target recommended order corresponding to the tenth recommended group is determined.

[0137] For example, whether the first topic label in the first recommended group (the first 1-10 recommended contents) is consistent is determined. When the first topic label contains the first vehicle type and the second vehicle type (i.e., the first topic label is inconsistent), the initial recommended order of the recommended content in the first recommended group is taken as the target recommended order.

[0138] In step 1034, in response to determining that the first topic label is consistent, target recommended content inconsistent with the first topic label is determined from the other recommended groups, and the target recommended content is added from the other recommended groups to the first recommended group to obtain an updated target recommended order.

[0139] In a specific implementation, when the first topic labels are consistent, it indicates that the recommended contents in the first recommended group are contents of the same type (i.e., there is a problem of content type clustering and singleness), then the topic labels of the recommended contents in the other recommended groups are sequentially judged according to the initial recommended order, the target recommended contents inconsistent with the first topic labels are determined from the other recommended groups, and the target recommended contents are moved from the other recommended groups to the first recommended group to obtain an updated target recommended order. After step 1034, it is continuously judged whether the second topic labels of the recommended contents in the second recommended group are consistent, and the target recommended order corresponding to the second recommended group is determined, until the target recommended order corresponding to the tenth recommended group is determined.

[0140] For example, it is judged whether the first topic labels of the first recommended group (the first to tenth recommended contents) are consistent. When the first topic labels are all of the first vehicle type (i.e., the first topic labels are consistent), and the topic label of the twelfth recommended content is of the second vehicle type (inconsistent with the first topic labels), the twelfth recommended content is moved to the tenth recommended order to obtain an updated target recommended order.

[0141] Based on the above example, the first recommended group in the initial recommended order is in turn: content 1, content 2, content 3, content 4, content 5, content 6, content 7, content 8, content 9, and content 10. The first recommended group in the updated target recommended order is in turn: content 1, content 2, content 3, content 4, content 5, content 6, content 7, content 8, content 9, and content 12. In this way, the content 12 in the first recommended group in the updated target recommended order is inconsistent with the types of other contents, so that the problem of content type clustering and singleness caused by all recommended contents in the first recommended group being of the same type can be avoided.

[0142] Through the above scheme, a predetermined number of recommended contents are divided into a first recommended group and other recommended groups according to an initial recommended order. The first topic labels of the recommended contents in the first recommended group are determined based on a large language model trained in advance, and it is judged whether the first topic labels are consistent. When the first topic labels are inconsistent, the initial recommended order is taken as a target recommended order, and since the recommended contents in the first recommended group are contents of different types, the problem of content type clustering and singleness does not occur. When the first topic labels are consistent, target recommended contents inconsistent with the first topic labels are determined from the other recommended groups, and the target recommended contents are added from the other recommended groups to the first recommended group to obtain an updated target recommended order. In this way, since the topic labels of the target recommended contents are inconsistent with the first topic labels, the problem of content type clustering and singleness does not occur when the recommended contents are displayed according to the updated target recommended order.

[0143] In some embodiments, the application software is application software in a car machine; step 101 includes:

[0144] Step 101A, obtaining the historical behavior of the target user in the application software in the car machine, and calling the driving data of the target user from the car machine.

[0145] In specific implementation, the application software is arranged in the car machine, and the historical behavior of the target user in the application software in the car machine is recorded in the process that the target user uses the application software. When the content is recommended to the target user, the historical behavior of the target user in the application software in the car machine is obtained, so as to recommend the recommended content interested by the target user according to the historical behavior.

[0146] In addition, the driving data of the target user is recorded in the process that the target user uses the vehicle. Since the driving data of the target user can also reflect the user character of the target user, when the content is recommended to the target user, the driving data of the target user is called from the car machine, so as to recommend the recommended content interested by the target user according to the driving data.

[0147] Step 101B, constructing the user portrait of the target user based on the historical behavior and the driving data.

[0148] In specific implementation, in order to accurately and comprehensively construct the style image of the target user, the user portrait of the target user is constructed based on the historical behavior and the driving data of the target user. In this way, in the process of constructing the user portrait, since the historical behavior of the target user in the application software and the driving data are comprehensively considered, the content interested by the target user and the style of the target user can be comprehensively considered, so as to accurately construct the user portrait of the target user.

[0149] Step 101C, determining the similarity between the target user and each published content according to the user portrait, and determining a predetermined number of recommended contents from the multiple published contents according to the similarity.

[0150] In specific implementation, the user portrait vector is determined according to the user portrait. The content feature vector of each published content is determined according to the content information of each published content. The similarity between the target user and each published content is determined according to the user portrait vector and the content feature vector.

[0151] The multiple published contents can be multiple published contents pre-stored in the content library. The multiple published contents in the content library are sorted in descending order of similarity, and a predetermined number of recommended contents with high similarity are determined from the sorted multiple published contents.

[0152] By the above scheme, the historical behavior of the target user in the application software in the car machine is obtained, and the driving data of the target user is called from the car machine. Based on the historical behavior and the driving data, the user portrait of the target user is constructed, which can comprehensively consider the content interested by the target user and the style of the target user, so as to accurately construct the user portrait of the target user. According to the user portrait, the similarity between the target user and each published content is determined, and according to the similarity, a predetermined number of recommended contents are determined from the multiple published contents, so that the determined recommended contents are recommended contents matched with the style of the target user.

[0153] The following is a specific embodiment of an information recommendation method of the present application, as shown in the car machine control system. Figure 2 The process executed in the car machine control system is specifically implemented.

[0154] Step 1, obtaining the user preference of the target user.

[0155] 1) Extracting the historical behavior of the target user, wherein the user behavior includes at least one of the following: clicking, collecting, liking, forwarding and clicking dislike.

[0156] 2) Obtaining the behavior data features of the historical behavior, wherein the behavior data features include at least one of the following: brand label, topic label and heat value.

[0157] 3) Combining the historical behavior and the behavior data features to obtain operation data, and converting the operation data into natural language format.

[0158] 4) Collecting user understanding prompt words prompt, and inputting the prompt words prompt into a large language model to let the large language model output the user preference of the target user.

[0159] Step 2, determining a predetermined number of recommended contents according to the user preference.

[0160] 1) Content understanding based on the large language model.

[0161] ① Obtaining the long text description of the published content (e.g. published post, published note).

[0162] ② Combine the abstract prompt words prompt and the title prompt words prompt, and input the long text description into the large language model to output the content abstract and the content title of the long text respectively.

[0163] ③ Obtaining the picture of the content.

[0164] ④ Combine the picture prompt words prompt, input the picture prompt words prompt into the picture understanding large language model, and output the text understanding of the picture.

[0165] 2) Inputting the content into the content vector library.

[0166] ①Brand label, topic label, content abstract, content title, and picture understanding of the published content are stored in a vector library.

[0167] 3) Based on the preference requirements of the user, similar content is recalled in the content vector library, and the 100 most similar published content is recalled as recommended content.

[0168] Step 3: Score the recommended content based on the large language model LLM to obtain a score, and determine the initial recommendation order of the recommended content according to the score.

[0169] 1) Obtain the user's basic features, preference features and historical behavior features. Among them, the user's basic features include: user gender and user age. The preference features are the user preferences obtained by using the large language model in step 1. The historical behavior features include: the brand sequence recently clicked by the target user, the topic sequence recently clicked by the target user, and the keyword sequence recently clicked by the target user.

[0170] 2) Obtain the brand label, topic label, content abstract, content title, and picture understanding of the published content. Among them, the content basic features include: publication time, brand label and topic label. The text understanding features of the large language model include: content abstract extracted by the large language model, content title extracted by the large language model, and keywords extracted by the large language model. The picture understanding features of the visual language model include: brand label and scene label of the visual language model.

[0171] 3) Input the features of the target user and the published content into the large language model LLM, and use the large language model LLM to score each published content based on the score to determine the initial recommendation order of the recommended content.

[0172] Step 4: Update the initial recommendation order based on the large language model LLM to obtain the target recommendation order, and display a predetermined number of recommended content according to the target recommendation order.

[0173] 1) Combine the rearrangement prompt, input the initial recommendation order into the large language model, and use the large language model to rearrange the predetermined number of recommended content to obtain the target recommendation order, to optimize the problems of clustering, single form, uneven traffic distribution, etc.

[0174] 2) Output the predetermined number of recommended content to the user according to the target recommendation order.

[0175] Through deep semantic understanding and complex relationship modeling of user demand, item characteristics and scene information by the large language model, the problems of traditional recommendation system in semantic ambiguity in recall, misjudgment of user preference, limitation of feature expression in precision ranking and single strategy in rearrangement are solved, the accuracy, diversity and traffic balance of the recommendation result are realized, and the user experience and platform business value are improved.

[0176] The embodiments of the present disclosure have the following technical effects:

[0177] (1) Full-process large language model deep intervention: breaking through the limitation of independent processing of traditional recommendation system modules, the large language model is applied throughout the whole process of recall, precision ranking and rearrangement, end-to-end semantic understanding and relationship modeling of user demand and item characteristics are realized, and the problems of insufficient semantic understanding and limited feature expression are fundamentally solved.

[0178] (2) Semantic understanding and preference capture: the semantic analysis and knowledge reasoning ability of the large language model is used in the recall link to accurately understand the complex demand of users.

[0179] (3) High-order feature automatic learning: the large language model is used to automatically mine high-order semantic features of user-item interaction in the precision ranking link, breaking through the expression bottleneck of traditional deep learning models.

[0180] (4) Multi-target intelligent rearrangement optimization: the large language model is used to generate a dynamic rearrangement strategy by comprehensively considering diversity, relevance and business targets in the rearrangement link, to solve the problems of category clustering, monotonous form and uneven traffic distribution, and to realize multi-dimensional optimization of the recommendation result.

[0181] It should be noted that the method of the embodiments of the present application can be executed by a single device, such as a computer or a server. The method of the embodiments can also be applied in a distributed scenario, and completed by multiple devices cooperating with each other. In this distributed scenario, one of the multiple devices can only execute one or more steps in the method of the embodiments of the present application, and the multiple devices can interact with each other to complete the method.

[0182] It should be noted that some embodiments of the present application have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in a different order than the order described above and still achieve the desired result. In addition, the processes depicted in the accompanying drawings do not necessarily require the particular order shown or sequential order to achieve the desired results. In certain implementations, multitasking and parallel processing can be advantageous or required.

[0183] Based on the same inventive concept, the present application also provides a content recommendation device corresponding to any of the above-mentioned embodiment methods.

[0184] Reference Figure 3 The device comprises:

[0185] The recommendation content determination module 201 is configured to acquire historical behaviors of a target user in an application software, determine a similarity between the target user and each published content according to the historical behaviors, and determine a predetermined number of recommendation contents from a plurality of published contents according to the similarity; wherein the application software has a plurality of published contents.

[0186] The initial recommendation order determination module 202 is configured to score each recommendation content based on a pre-trained large language model to obtain a score value, and determine an initial recommendation order according to the score value.

[0187] The target recommendation order determination module 203 is configured to determine a topic label of the predetermined number of recommendation contents based on a pre-trained large language model, update the initial recommendation order according to the topic label to obtain a target recommendation order, and display the predetermined number of recommendation contents in the target recommendation order.

[0188] In some embodiments, the recommendation content determination module 201 comprises:

[0189] The user preference determination unit is configured to analyze and process the historical behaviors based on a pre-trained behavior extraction model to obtain a user preference.

[0190] The content information determination unit is configured to acquire description information of each published content, and analyze and process the description information based on a pre-trained content extraction model to obtain content information.

[0191] The similarity determination unit is configured to determine a similarity between the target user and each published content according to the user preference and the content information.

[0192] The recommendation content determination unit is configured to sort the plurality of published contents in an order from high to low similarity, and determine a predetermined number of recommendation contents with high similarity from the sorted plurality of published contents.

[0193] In some embodiments, the user preference determination unit comprises:

[0194] The analysis and processing subunit is configured to analyze and process the historical behaviors to obtain user operations and operation content features.

[0195] The data conversion subunit is configured to combine and process the user operations and the operation content features to obtain operation data, and convert the operation data into operation data in a preset format.

[0196] The user preference determining sub-unit is configured to analyze and process the target user data and the operation data in the preset format based on a pre-trained behavior extraction model to obtain the user preference.

[0197] In some embodiments, the content extraction model comprises a large language model and a visual language model.

[0198] The content information determining unit comprises:

[0199] The acquisition sub-unit is configured to acquire text description information and picture description information of each published content.

[0200] The text extraction sub-unit is configured to extract and process the text description information based on the pre-trained large language model to obtain a content abstract and a content title.

[0201] The picture extraction sub-unit is configured to extract and process the picture description information based on the pre-trained visual language model to obtain picture content information.

[0202] The content information determining sub-unit is configured to take the content abstract, the content title, and the picture content information as the content information.

[0203] In some embodiments, the similarity determining unit comprises:

[0204] The first vectorization sub-unit is configured to vectorize the user preference to obtain a user preference vector.

[0205] The second vectorization sub-unit is configured to extract features of each published content to obtain a brand label and a topic label, and vectorize the brand label, the topic label, and the content information to obtain a content feature vector.

[0206] The similarity determining sub-unit is configured to determine the similarity between the target user and each published content according to the user preference vector and the content feature vector.

[0207] In some embodiments, the initial recommendation order determining module 202 comprises:

[0208] The score determining unit is configured to score each recommended content based on the user features of the target user and the content features of each recommended content according to a pre-trained large language model to obtain a score of each recommended content.

[0209] The initial recommendation order determining unit is configured to sort the predetermined number of recommended contents in descending order of the scores to obtain an initial recommendation order.

[0210] In some embodiments, the target recommendation sequence determination module 203 comprises:

[0211] a division unit configured to divide the predetermined number of recommendation contents into a first recommendation group and other recommendation groups according to the initial recommendation sequence;

[0212] a judgment unit configured to determine a first topic label of the recommendation contents in the first recommendation group based on a pre-trained large language model, and determine whether the first topic label is consistent;

[0213] a first target recommendation sequence determination unit configured to, in response to determining that the first topic label is inconsistent, take the initial recommendation sequence as a target recommendation sequence;

[0214] a second target recommendation sequence determination unit configured to, in response to determining that the first topic label is consistent, determine target recommendation contents inconsistent with the first topic label from the other recommendation groups, and add the target recommendation contents from the other recommendation groups to the first recommendation group to obtain an updated target recommendation sequence.

[0215] In some embodiments, the application software is application software in a car machine;

[0216] The recommendation content determination module 201 comprises:

[0217] a driving data calling unit configured to obtain historical behaviors of the target user in the application software in the car machine, and call driving data of the target user from the car machine;

[0218] a user portrait construction unit configured to construct a user portrait of the target user based on the historical behaviors and the driving data;

[0219] a recommendation content determination unit configured to determine a similarity between the target user and each published content according to the user portrait, and determine a predetermined number of recommendation contents from a plurality of published contents according to the similarity.

[0220] For the convenience of description, the above device is described as various modules respectively described in function. Of course, in the implementation of the present application, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0221] The device of the above embodiments is used to implement the corresponding content recommendation method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which are not described here again.

[0222] Based on the same inventive concept, the present application also provides an electronic device corresponding to the method of any of the above embodiments, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the content recommendation method of any of the above embodiments.

[0223] Figure 4 A more specific hardware structure of an electronic device is shown in this embodiment. The device can include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are connected to each other through the bus 1050 for internal communication within the device.

[0224] The processor 1010 can be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits, etc., for executing related programs to implement the technical solutions provided by the embodiments of the present application.

[0225] The memory 1020 can be implemented by a ROM (Read Only Memory), a RAM (Random Access Memory), a static storage device, a dynamic storage device, etc. The memory 1020 can store an operating system and other application programs. When the technical solutions provided by the embodiments of the present application are implemented by software or firmware, the related program codes are stored in the memory 1020 and executed by the processor 1010.

[0226] The input / output interface 1030 is used to connect input / output modules to realize information input and output. The input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. The input device can include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device can include a display, a speaker, a vibrator, an indicator, etc.

[0227] The communication interface 1040 is configured to connect a communication module (not shown in the figure) to realize the communication interaction between the device and other devices. The communication module can realize communication through wired mode (for example, USB (Universal Serial Bus, Universal Serial Bus), network cable, etc.) or wireless mode (for example, mobile network, WIFI (Wireless Fidelity, wireless network communication technology), Bluetooth, etc.).

[0228] The bus 1050 includes a path for transmitting information between various components (for example, the processor 1010, the memory 1020, the input / output interface 1030 and the communication interface 1040) of the device.

[0229] It should be noted that although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040 and the bus 1050, in the specific implementation process, the device can also include other components necessary for normal operation. In addition, those skilled in the art can understand that the above device can also only contain the components necessary for the implementation of the embodiments of the present application, and does not have to contain all the components shown in the figure.

[0230] The electronic device of the above embodiment is used to realize the corresponding content recommendation method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which are not described here.

[0231] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present application also provides a non-transitory computer readable storage medium, which stores computer instructions for causing the computer to execute the content recommendation method according to any of the above embodiments.

[0232] The computer readable medium of the present embodiment includes permanent and non-permanent, removable and non-removable media, which can be realized by any method or technology to store information. The information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage device, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0233] The storage medium of the above embodiments stores computer instructions for causing the computer to perform the content recommendation method according to any one of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0234] Based on the same inventive concept, the present application also provides a vehicle comprising the content recommendation device, or the electronic device, or the storage medium in the above embodiments, which implements the content recommendation method according to any one of the above embodiments.

[0235] The vehicle of the above embodiments is used to implement the content recommendation method according to any one of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0236] Based on the same inventive concept, the present application also provides a computer program product comprising computer program instructions, which, when executed on a computer, cause the computer to perform the content recommendation method according to any one of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0237] It can be understood that before using the technical solutions of the embodiments of the present application, the type, use range, use scenario, etc. of the personal information involved will be informed to the user in a proper manner, and the authorization of the user will be obtained.

[0238] For example, in response to receiving the active request of the user, prompt information is sent to the user to explicitly prompt the user that the operation requested to be performed will require the acquisition and use of personal information of the user. Thus, the user can voluntarily choose whether to provide personal information to the software or hardware such as electronic device, application program, server or storage medium that performs the technical solutions of the present application according to the prompt information.

[0239] As an optional but not limiting implementation manner, in response to accepting the active request of the user, the manner of sending prompt information to the user may, for example, be a pop-up window manner, in which the prompt information can be presented in the form of text. In addition, the pop-up window can also carry selection controls for the user to select "agree" or "disagree" to provide personal information to the electronic device.

[0240] It can be understood that the above notification and user authorization process is only illustrative, and does not limit the implementation manner of the present application, and other manners meeting the relevant laws and regulations can also be applied to the implementation manner of the present application.

[0241] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present application is limited to these examples. In line with the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0242] In addition, for simplicity of description and discussion, and in order not to make the embodiment of the application difficult to understand, the known power supply / ground connection with integrated circuit (IC) chip and other components may or may not be shown in the accompanying drawings provided. In addition, the device can be shown in the form of a block diagram to avoid making the embodiment of the application difficult to understand, and this also takes into account the following fact, that is, the details of the embodiment of these block diagram devices are highly dependent on the platform to be implemented in the embodiment of the application (that is, these details should be fully within the scope of understanding of those skilled in the art). When specific details (for example, circuit) are set forth to describe exemplary embodiments of the application, it will be apparent to those skilled in the art that the embodiment of the application can be implemented without these specific details or when these specific details are changed. Therefore, these descriptions should be considered to be illustrative rather than restrictive.

[0243] Although the present invention has been described in conjunction with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may utilize the embodiments discussed.

[0244] The embodiments of the present application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the present application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present application should be included in the scope of protection of the present application.

Claims

1. A content recommendation method, characterized in that: The method comprises: Obtaining historical behavior of a target user in an application software, determining similarity between the target user and each published content based on the historical behavior, and determining a predetermined number of recommended content from the multiple published content based on the similarity; wherein the application software has multiple published content; Scoring each recommended content based on a pre-trained large language model to obtain a score, and determining an initial recommendation order based on the score; The topic tags of the predetermined number of recommended contents are determined based on a pre-trained large language model, the initial recommendation order is updated according to the topic tags to obtain a target recommendation order, and the predetermined number of recommended contents are displayed according to the target recommendation order.

2. The method according to claim 1, characterized in that Determining the similarity between the target user and each published content based on the historical behavior, and determining a predetermined number of recommended content from the plurality of published content based on the similarity, includes: Analyzing and processing the historical behaviors based on a pre-trained behavior extraction model to obtain user preferences; Obtaining description information of each published content, and analyzing and processing the description information based on a pre-trained content extraction model to obtain content information; Determining the similarity between the target user and each published content based on the user preference and the content information; The plurality of published contents are sorted in descending order of similarity, and a predetermined number of recommended contents with high similarity are determined from the sorted plurality of published contents.

3. The method according to claim 2, characterized in that The analyzing and processing the historical behaviors based on the pre-trained behavior extraction model to obtain user preferences includes: Analyze and process the historical behavior to obtain user operations and operation content characteristics; Combining and processing the user operation and the operation content feature to obtain operation data, and converting the operation data into operation data in a preset format; The target user data and the operation data in the preset format are analyzed and processed based on a pre-trained behavior extraction model to obtain the user preference.

4. The method according to claim 2, characterized in that The content extraction model includes: a large language model and a visual language model; The obtaining of description information of each published content and analyzing and processing the description information based on a pre-trained content extraction model to obtain content information includes: Obtaining text description information and image description information for each published content; Extracting the text description information based on the pre-trained large language model to obtain a content summary and a content title; Extracting the image description information based on the pre-trained visual language model to obtain image content information; The content summary, the content title, and the picture content information are used as the content information.

5. The method according to claim 2, characterized in that The determining, based on the user preference and the content information, the similarity between the target user and each published content includes: performing vectorization processing on the user preference to obtain a user preference vector; Performing feature extraction processing on each published content to obtain a brand tag and a topic tag, and performing vectorization processing on the brand tag, the topic tag, and the content information to obtain a content feature vector; The similarity between the target user and each published content is determined according to the user preference vector and the content characteristic vector.

6. The method according to claim 1, characterized in that Scoring each recommended content based on the pre-trained large language model to obtain a score, and determining an initial recommendation order according to the score, includes: Based on a pre-trained large language model, scoring each recommended content according to the user characteristics of the target user and the content characteristics of each recommended content to obtain a score for each recommended content; The predetermined number of recommended contents are sorted from high to low according to the scores to obtain an initial recommendation order.

7. The method according to claim 1, characterized in that The determining of topic tags of the predetermined number of recommended contents based on the pre-trained large language model, and updating the initial recommendation order according to the topic tags to obtain a target recommendation order, includes: Dividing the predetermined number of recommended contents into a first recommendation group and other recommendation groups according to the initial recommendation order; Determining first topic tags of the recommended content in the first recommendation group based on a pre-trained large language model, and determining whether the first topic tags are consistent; In response to determining that the first topic tags are inconsistent, taking the initial recommendation order as a target recommendation order; In response to determining that the first topic tag is consistent, target recommended content inconsistent with the first topic tag is determined from the other recommendation groups, and the target recommended content is added from the other recommendation groups to the first recommendation group to obtain an updated target recommendation order.

8. The method according to claim 1, characterized in that The application software is the application software in the vehicle computer; The step of obtaining the target user's historical behavior in the application software, determining the similarity between the target user and each published content based on the historical behavior, and determining a predetermined number of recommended content from the plurality of published content based on the similarity includes: Obtaining the target user's historical behavior in the application software in the vehicle computer, and retrieving the target user's driving data from the vehicle computer; Building a user profile of the target user based on the historical behavior and the driving data; The similarity between the target user and each published content is determined according to the user portrait, and a predetermined number of recommended contents are determined from the multiple published contents according to the similarity.

9. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

10. A vehicle, characterized in that: The electronic device comprising claim 9.

Images