A method and related device for personalized recommendation of a car based on a large language model

Abstract

The application relates to the technical field of personalized recommendation, and provides a car personalized recommendation method based on a large language model and related equipment, which comprises the following steps: calculating an attribute satisfaction score vector, an attribute regret tendency score vector, an attribute competitiveness score vector and an attribute motivation score vector according to user comment data of a car type by using a large language model; calculating a comment recommendation index based on the attribute satisfaction score vector, the attribute regret tendency score vector, the attribute competitiveness score vector and the attribute motivation score vector; calculating a video emotion score vector, a video attribute competitiveness score vector and a video motivation score vector according to video comment data; calculating a video recommendation index based on the video emotion score vector, the video attribute competitiveness score vector and the video motivation score vector; and performing personalized car type recommendation on a user to be recommended according to the comment recommendation index and the video recommendation index. The method can improve the accuracy of car personalized recommendation for users.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of personalized recommendation, in particular to a car personalized recommendation method based on a large language model and related equipment. BACKGROUND

[0002] Under the background of the continuous growth of today's social economy and the rapid development of the new energy automobile industry, the number of automobile drivers is showing a steady upward trend. As a high-value durable consumer product, the user comments of the car have a very key reference significance for the car purchase decision of potential consumers. With the increasingly fierce competition in the automobile market, consumers often consider many factors such as performance, configuration, price and the like of multiple car models when purchasing a car. Therefore, in-depth analysis of the emotional tendency contained in the car user comment text and video content, especially the research on the comparison comments between different car models, has extremely important practical significance and academic value for comprehensively and accurately grasping the attitude and preference of users to car products and providing a scientific basis for the decision of potential consumers. However, the current analysis ability of the attitude and preference of users is poor, which leads to low accuracy of car personalized recommendation for users. SUMMARY

[0003] The embodiments of the present application provide a car personalized recommendation method based on a large language model and related equipment, which can solve the problem of low accuracy of car personalized recommendation for users.

[0004] In a first aspect, the embodiments of the present application provide a car personalized recommendation method based on a large language model, which comprises:

[0005] Obtaining multiple user comment data and multiple video comment data of multiple car models; the user comment data comprises comment text, scores of multiple car attributes and user feature data, and the video comment data comprises video text content and video heat data;

[0006] Using a large language model to calculate attribute satisfaction score vectors, attribute regret tendency score vectors, attribute competitiveness score vectors and attribute motivation score vectors of each car model according to all user comment data of each car model;

[0007] Calculating a comment recommendation index of each car model based on the attribute satisfaction score vectors, the attribute regret tendency score vectors, the attribute competitiveness score vectors and the attribute motivation score vectors of each car model;

[0008] Using a large language model to calculate video emotion score vectors, video attribute competitiveness score vectors and video motivation score vectors of each car model according to all video comment data of each car model;

[0009] Based on the video sentiment score vector, the video attribute competitiveness score vector, and the video motivation score vector of each vehicle model, a video recommendation index of each vehicle model is calculated.

[0010] According to the comment recommendation index and the video recommendation index of all vehicle models, personalized vehicle model recommendation is performed on a user to be recommended.

[0011] In a second aspect, the embodiments of the present application provide a terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the above-mentioned automobile personalized recommendation method based on a large language model when executing the computer program.

[0012] The above-mentioned scheme of the present application has the following beneficial effects:

[0013] In the embodiments of the present application, by analyzing the information of the vehicle model based on the user comments and video comments of the vehicle model, the diversity and comprehensiveness of the vehicle model information can be improved, multiple scoring vectors of the vehicle model in the comments are calculated, user evaluations of the satisfaction, regret tendency, purchase motivation, and competitiveness of the vehicle model are considered, the accuracy of the comment recommendation index and the video recommendation index of the vehicle model is improved, and the accuracy of the automobile personalized recommendation for the user is further improved.

[0014] Other beneficial effects of the present application will be described in detail in the subsequent specific implementation manner part. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0016] Figure 1 The flowchart of the automobile personalized recommendation method based on a large language model provided by an embodiment of the present application;

[0017] Figure 2 The structural schematic diagram of the terminal device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0018] In the following description, specific details such as specific system structures, techniques, etc. are presented for the purpose of explanation, not for the purpose of limitation, so that the embodiments of the present application can be thoroughly understood. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details.

[0019] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0020] To address the issue of low accuracy in existing personalized car recommendations for users, this application provides a car personalization recommendation method based on a large language model. This method analyzes car model information based on user reviews and video reviews, improving the diversity and comprehensiveness of car model information. It calculates multiple rating vectors for car models in the reviews, considering user evaluations of car models from multiple aspects such as satisfaction, regret tendency, purchase motivation, and competitiveness, thereby improving the accuracy of car model review recommendation index and video recommendation index, and ultimately enhancing the accuracy of personalized recommendations for users.

[0021] The following section provides an illustrative example of the car personalization recommendation method based on a large language model provided in this application.

[0022] like Figure 1 As shown, the personalized car recommendation based on a large language model provided in this application includes the following steps:

[0023] Step 11: Obtain multiple user review data and multiple video review data for multiple car models.

[0024] The aforementioned user review data includes review text, ratings for multiple car attributes, and user characteristic data. Video review data includes video text content (i.e., the text converted from the user's voice evaluation of the car model in the video) and video popularity data. Multiple car attributes include space, driving experience, range, exterior, interior, value for money, and intelligent features. Review text is categorized as either satisfactory or unsatisfactory (this can be determined based on the average ratings of the corresponding car attributes; if the average rating of the car attributes corresponding to the review text is less than the average threshold, the review text is considered unsatisfactory; if the average rating is greater than or equal to the average threshold, the review text is considered satisfactory). User characteristic data includes the authentication status of the corresponding user account (e.g., whether the user account is real-name authenticated or not), the recommendation level of the corresponding user review data (this can be the recommendation level set by the platform where the review is located, or it can be set based on the corresponding interaction index; the higher the interaction index, the higher the recommendation level), and the interaction index of the corresponding user review data (which can be the sum of user review data views, likes, etc.). Video popularity data includes the number of likes, comments, favorites, and shares of the video.

[0025] In some embodiments of this application, user comment data can be obtained by accessing car sales platforms, etc., and video comment data can be obtained by accessing short video platforms, etc.

[0026] Step 12: Using a large language model, calculate the attribute satisfaction rating vector, attribute regret tendency rating vector, attribute competitiveness rating vector, and attribute motivation rating vector for each vehicle model based on all user review data for each vehicle model.

[0027] The aforementioned attribute satisfaction rating vector is used to describe the degree of user satisfaction with each car attribute of a model on platforms primarily composed of text information, such as sales platforms. The aforementioned attribute regret tendency rating vector is used to describe the degree of user dissatisfaction with each car attribute of a model on platforms primarily composed of text information. The attribute competitiveness rating vector is used to describe the competitiveness of a model with other models on each car attribute on platforms primarily composed of text information. The attribute motivation rating vector is used to describe the priority of users' consideration of each car attribute when purchasing a model on platforms primarily composed of text information (i.e., the higher the priority of consideration of a car attribute, the more likely the user is to consider that car attribute first when purchasing the model).

[0028] In the above embodiments of this application, the steps of using a large language model to calculate the attribute satisfaction score vector, attribute regret tendency score vector, attribute competitiveness score vector, and attribute motivation score vector for each vehicle model based on all user review data for each vehicle model include:

[0029] For each vehicle model, perform the following steps:

[0030] The first step is to calculate the credibility of user review data for each vehicle model based on user characteristic data and ratings of multiple vehicle attributes.

[0031] Specifically, through the formula:

[0032] ;

[0033] Calculate user review data Credibility .

[0034] in, , This represents a set of IDs representing user review data for a particular car model. Indicates the first The weights of each reliable indicator and , Represents user comment data The The value of a reliable indicator:

[0035] ;

[0036] = ;

[0037] ;

[0038] ;

[0039] in, Represents user comment data The value of the first credibility indicator, Represents user comment data The value of the second credibility indicator, Represents user comment data The value of the third credibility indicator, Represents user comment data The value of the fourth credibility indicator, A set of numbers representing vehicle attributes. This refers to the user review data for all car models regarding car attributes. The average rating, Represents user comment data Regarding car attributes The rating, Represents user comment data The interaction index.

[0040] The second step involves identifying the attribute mentions in each satisfactory review text for a vehicle model based on a large language model. Then, based on the attribute mentions in all satisfactory review texts and the corresponding credibility of all satisfactory review texts, the attribute satisfaction rating vector for the vehicle model is calculated.

[0041] The attribute mentions in the satisfaction review text are used to describe the car attributes mentioned in the satisfaction review text, as well as the emotional intensity of the car attributes. The attribute satisfaction rating vector includes the attribute satisfaction rating for each car attribute of the car model.

[0042] The emotional intensity mentioned above describes the degree of emotional impact that users have on the car's attributes as described in their reviews.

[0043] In some embodiments of this application, the large language model can be chatgpt, Tongyi Qianwen Max large model, etc. The steps for calculating the attribute satisfaction score vector of the vehicle model based on the attribute mention results of all satisfactory comment texts and the credibility corresponding to all satisfactory comment texts include:

[0044] Through the formula:

[0045] ;

[0046] ;

[0047] ;

[0048] ;

[0049] Calculate the attribute satisfaction rating vector for vehicle models .

[0050] in, This represents the satisfaction rating of attribute 1 of the car model. Indicates the vehicle attributes of the model. Attribute satisfaction rating A set of numbers representing vehicle attributes. The set of numbers representing the texts of satisfactory comments. This indicates the transpose operation. Indicates the vehicle attributes of the model. Attribute satisfaction rating Text indicating satisfaction Regarding car attributes The rating mentioned Text indicating satisfaction Regarding car attributes The intensity of emotions, express Aggregate weights, Text indicating satisfaction Corresponding credibility Transformation function:

[0051] ;

[0052] For example, prompt words are set for a large language model, and then the text of a positive review is input into the large language model. The large language model infers based on the prompt words to obtain the attribute mention results. An example of prompt words is shown below:

[0053] Scenario: You are a professional car review analyst. Please complete the attribute satisfaction and sentiment score based on the car owner reviews provided by users.

[0054] Skill: Skill 1: Read Text. Reads the text input by {input}, which is the comment from the owner of {car}.

[0055] Skill 2. Dimensional Analysis: For comments, sentiment scoring is performed based on the following 7 dimensions:

[0056] Space, Driving Experience, Range, Exterior, Interior, Cost Performance, Intelligence.

[0057] Skill 3: If the user mentions content related to this dimension, output (1, "score").

[0058] If the user does not mention this dimension, output (0, "score").

[0059] Scoring rules: Satisfaction rating range (i.e. emotional intensity): 0~10.

[0060] JSON Output: Output Requirements: 1. Strictly use JSON format. 2. Output missing dimensions as (0, 0). 3. Do not return comments / explanatory text.

[0061] Example of information extraction: Input: "The space design of this car is quite reasonable. The front and rear seats don't feel cramped, and the trunk can hold quite a lot of things. Overall, it's not bad, but not particularly large. The driving experience is really good. The steering wheel is very light, the chassis is very stable, and I feel confident when cornering and at high speeds. It's very easy to drive and the experience is great; I'm very satisfied. The interior is my favorite part. The quality is excellent, the workmanship is very fine, and it feels very classy as soon as you enter the car. The layout is also simple and practical, and sitting inside improves my mood. In terms of value for money, I think it's quite worthwhile. It has rich features, and the price isn't too high. Compared with other models in the same class, it has a significant advantage, so I'm willing to recommend it to my friends."

[0062] Output results: {"records": [{"fields": {"Space": "(1, 6)", "Driving_Experience": "(1, 8)", "Range": "(0, 0)", "Exterior": "(0, 0)", "Interior": "(1, 8)", "Cost_Performance": "(1, 6)", "Intelligence": "(0, 0)"}}]}.

[0063] The third step involves identifying the attribute mentions of each unsatisfactory review text for a vehicle model based on a large language model. Then, based on the attribute mentions of all unsatisfactory review texts for a vehicle model and the credibility of all unsatisfactory review texts, the attribute regret tendency score vector for the vehicle model is calculated.

[0064] The attribute mentions in the unsatisfactory review text are used to describe the car attributes mentioned in the unsatisfactory review text, as well as the emotional intensity of the car attributes. The attribute regret tendency score vector includes the attribute regret tendency score for each car attribute of the car model.

[0065] It should be noted that the formula for calculating the attribute regret tendency score vector for vehicle models is the same as the formula for calculating the attribute satisfaction score vector mentioned above. Substituting the attribute mentions from the dissatisfied review text into the formula yields the attribute regret tendency score vector for the vehicle model, i.e.:

[0066] Through the formula:

[0067] ;

[0068] ;

[0069] ;

[0070] ;

[0071] Calculate the regret tendency score vector of vehicle type attributes .

[0072] in, This indicates the regret tendency score for the vehicle's attribute 1. Indicates the vehicle attributes of the model. The attribute of regret tendency score, Indicates the vehicle attributes of the model. The attribute of regret tendency score, A set of indexed comments expressing dissatisfaction. express Aggregate weights, Expressing dissatisfaction with the text Regarding car attributes The intensity of emotions, Expressing dissatisfaction with the text Regarding car attributes The rating mentioned Expressing dissatisfaction with the text Credibility.

[0073] For example, prompt words are set for a large language model, and then the text of dissatisfied comments is input into the large language model. The large language model infers based on the prompt words and obtains the comparison attribute mention results. An example of a prompt word is:

[0074] Scenario: You are a professional car review analyst. Please complete the negative sentiment rating based on the car owner reviews provided by users.

[0075] Skill: Skill 1: Read Text

[0076] Read the text input by {input}, which is the comment from the owner of {car}.

[0077] Pay special attention to whether they express dissatisfaction, disappointment, or regret regarding certain aspects of the vehicle.

[0078] Skill 2: Dimensional Analysis

[0079] Please rate your negative emotions based on the following 7 dimensions:

[0080] Space, Driving Experience, Range, Exterior, Interior, Cost Performance, Intelligence.

[0081] Skill 3: Scoring Rules:

[0082] If a user expresses dissatisfaction, disappointment, complaint, or explicitly regrets regarding this dimension:

[0083] Output format: (1, "score"), where the score ranges from -10 to 0, with lower scores for more negative results.

[0084] If the user did not mention this dimension or did not show obvious negative emotions:

[0085] Output format: (0, "0").

[0086] JSON Output: Output Requirements: 1. Strictly use JSON format. 2. Output missing dimensions as (0, 0). 3. Do not return comments / explanatory text.

[0087] Example of information extraction: Input: "I really regret buying this car. First, the space is too small; it's very cramped for three people in the back, making long journeys very uncomfortable. Second, the range is seriously overstated; the official claim is 500 kilometers, but in reality, with the air conditioning on in winter, it only goes about 300 kilometers, requiring frequent charging. The most frustrating thing is the car's infotainment system; it's incredibly laggy, the navigation system crashes frequently, and the voice recognition is practically unusable. It feels like the 'intelligent' features are just a marketing gimmick."

[0088] Output results: {"records": [{"fields": {"Space": "(1, -7)", "Driving_Experience": "(0, 0)", "Range": "(1, -8)", "Exterior": "(0, 0)", "Interior": "(0, 0)", "Cost_Performance": "(0, 0)", "Intelligence": "(1,-9)"}}]}.

[0089] The fourth step involves identifying the comparative attribute mentions for each review text of a car model based on a large language model, and calculating the attribute competitiveness score vector of the car model based on all comparative attribute mentions and all credibility levels.

[0090] The comparative attribute mention results are used to describe the car attributes mentioned in the review text when comparing other car models, as well as the emotional intensity of the mentioned car attributes. The attribute competitiveness score vector includes the attribute competitiveness score for each car attribute of the car model.

[0091] The emotional intensity mentioned above describes the degree of emotional impact that users have on the car's attributes as described in their reviews.

[0092] Specifically, through the formula:

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] Calculate the attribute competitiveness score vector of the vehicle model .

[0099] in, This represents the competitiveness score of attribute 1 of the vehicle model. Indicates the vehicle attributes of the model. Attribute competitiveness score Indicates the vehicle attributes of the model. Attribute competitiveness score express Aggregate weights, Comment text When comparing with other models, mention the car's attributes. The intensity of emotions, Comment text When comparing with other car models, the car's attributes The rating mentioned.

[0100] It should be noted that if multiple other car models are compared in the same review text, the above calculation is performed for each of the other car models, and then the average of the attribute competitiveness scores for each car attribute for all other car models is taken. If a car attribute is not mentioned in the review text, the sentiment intensity corresponding to that car attribute is 0.

[0101] For example, prompt words are set for a large language model, and then the comment text is input into the large language model. The large language model infers based on the prompt words and obtains the comparison attribute mention results. An example of a prompt word is:

[0102] Scenario: You are a professional car model information extraction assistant who needs to strictly process user-provided car reviews according to the following rules.

[0103] Skills: Skill 1: Identify all car models mentioned in the text {input} that are not the currently purchased car model. Skill 2: List the full names of all car models (including brand and model). Input format: The user will provide: Purchased car model name: {car}. The car review text {input} to be analyzed. Skill 3: Processing rules:

[0104] 1. Strictly distinguish between the primary vehicle model and other vehicle models:

[0105] The main vehicle model may appear in abbreviated / full name / variant form (such as "Model A" / "Model A of a certain brand" / "Model A of a certain brand 2024 version").

[0106] Other vehicle models must simultaneously meet the following requirements: complete brand + model combination (e.g., "Brand B"); and be an independent model that is clearly different from the main vehicle purchased.

[0107] 2. Special Case Handling: For series models (e.g., "5 Series"), the context must be considered to determine whether it corresponds to a specific model. Historically owned vehicles (e.g., "had owned two Model B vehicles") must be included. Comparisons with reference models (e.g., "more fuel-efficient than Models B and C") must be included. General references that do not specify a model (e.g., "same class of vehicles") are excluded.

[0108] JSON Output: Output Requirements: Strictly adhere to the following JSON structure:

[0109] Example: {"output": [{"car": "Brand B",},{"car": "Brand BC",}] If no other car models exist, output an empty set.

[0110] Example of information extraction: Input: "The interior is really just average. There isn't much soft-touch material, so it doesn't look very high-quality. Compared to the C-Class of a certain brand and the 3 Series of another brand in the same class, the interior is really a bit outdated. Fortunately, the steering wheel is flat-bottomed, which adds a bit of sportiness."

[0111] Output result: { "output": [ { "car": "a certain brand's C-class"}, { "car": "a certain brand's 3-series"} ]}.

[0112] Another example of a prompt word is:

[0113] Scenario: You are a professional automotive review analyst, skilled at identifying the competitive advantages of car models through comparative analysis. Based on the owner reviews provided by users, accurately identify all the models to be compared and complete a multi-dimensional sentiment score.

[0114] Skills: Skill 1: Vehicle Model Recognition: Extract all competing vehicle models directly compared to {car} from the text (e.g., a certain brand's model B / a certain brand's 5 Series in the example). Skill 2: Dimensional Analysis: For each competing vehicle model, perform comparative sentiment scoring from the following 7 dimensions: Space, Driving Experience, Range, Exterior, Interior, Cost Performance, and Intelligence. If the user mentions a content related to this dimension, output (1, "score"). If the user does not mention a content related to this dimension, output (0, "score"). Skill 3: Scoring Rules:

[0115] Rating range: -10 to 10. Negative value: The competitor is significantly better than {car} in this dimension (e.g., the comment explicitly states the competitor's advantage). Positive value: {car} is significantly better than the competitor in this dimension. Zero value: Neutral.

[0116] JSON output: Output requirements:

[0117] 1. Strictly use JSON format. 2. Generate a separate record for each compared vehicle model. 3. Missing dimensions should be kept at zero. 4. Do not return comments / explanatory text.

[0118] Information extraction example: Input: The interior is really just average. There isn't much soft-touch material, so it doesn't look very high-quality. Compared to other brands in the same class (D and C), the interior is really outdated. Fortunately, the steering wheel is flat-bottomed, which adds a bit of sportiness.

[0119] Output result:

[0120] {"records": [{"fields": {"Car_Name": "Brand D", "Space": "(0, 0)", "Driving_Experience": "(0, 0)", "Range": "(0, 0)", "Exterior": "(0, 0)", "Interior": "(1, -10)", "Cost_Performance": "(0, 0)", "Intelligence": "(0, 0)"}}, {"fields": {"Car_Name": "Brand C", "Space": "(0, 0)", "Driving_Experience": "(0, 0)", "Range": "(0, 0)", "Exterior": "(0, 0)", "Interior": "(1, -10)", "Cost_Performance": "(0, 0)", "Intelligence": "(0, 0)"}}]}.

[0121] The fifth step involves identifying the purchase motivation attributes of each review text for a car model based on a large language model, and calculating the attribute motivation score vector for the car model based on all purchase motivation attributes.

[0122] The purchase motivation attribute is one of multiple car attributes, and the attribute motivation score vector includes the attribute motivation score for each car attribute of the car model.

[0123] Specifically, through the formula:

[0124] ;

[0125] ;

[0126] ;

[0127] Calculate the attribute motivation rating vector of vehicle model .

[0128] in, This indicates the motivation score for attribute 1 of the car model. Indicates the vehicle attributes of the model. Attribute motivation rating Indicates the vehicle attributes of the model. Attribute motivation rating, Comment text With car attributes The motivational relationship, if the car attributes For comment text Purchase motivation attributes, If the car attributes Not for comment text Purchase motivation attributes, .

[0129] For example, since user review data includes review text, the review text shares the same ID as the user review data it belongs to. Cue words are set for the large language model, and then the review text is input into the large language model. The large language model infers based on the cue words to obtain the purchase motivation attribute. An example of a cue word is:

[0130] Scenario: You are a professional car comparison and analysis assistant, skilled at accurately judging car owners' purchasing motives from their reviews of cars.

[0131] Skills: Skill 1. Carefully analyze the given car review "{input}". Skill 2. From the seven factors of [space, driving experience, range, exterior, interior, value for money, and intelligent features], determine the car owner's most likely purchase motivation based on explicit or implicit clues in the text. Skill 3. Output the words themselves directly; do not add explanations or modifiers.

[0132] JSON Output: Output Requirements: Determine the purchase motivation solely based on car reviews, excluding irrelevant topics. You must strictly choose one of the eight given factors for output. Output must be concise, containing only the selected words.

[0133] Information extraction example: Input: Regarding this car, what I like most is its interior. Upon entering, the sense of luxury is immediately apparent, exuding opulence and grandeur. Every detail inside showcases meticulous design. The layout of the center console is extremely scientific, with everyday essentials within easy reach, greatly facilitating driving operations. The seats are also exceptional, offering extremely high comfort; even on long journeys, there is no fatigue.

[0134] Output: "Interior".

[0135] Step 13: Calculate the review recommendation index for each vehicle model based on the attribute satisfaction rating vector, attribute regret tendency rating vector, attribute competitiveness rating vector, and attribute motivation rating vector.

[0136] The review recommendation index describes the level of user interest in a car model on text-based platforms such as sales platforms. A higher review recommendation index indicates a greater level of user interest in the car model on text-based platforms.

[0137] Specifically, through the formula:

[0138] ;

[0139] ;

[0140] ;

[0141] ;

[0142] ;

[0143] Calculate the review and recommendation index of a car model .

[0144] in, Indicates the competitiveness weight. Indicates the vehicle attributes of the model. The attribute of regret tendency score, , This represents the attribute regret tendency rating vector. Represents group preference value, Represents the normalized attribute motivation rating vector , This indicates the initial calculation of the unweighted recommendation index after user personalization. This represents the user's personal preference value. Indicates car attributes Personalized weights, This indicates the strength of the user's personalized preference group attributes.

[0145] Step 14: Using a large language model, calculate the video sentiment score vector, video attribute competitiveness score vector, and video motivation score vector for each vehicle model based on all video comment data for each vehicle model.

[0146] The aforementioned video sentiment score vector is used to describe the degree of user satisfaction with each car attribute of a model on video platforms and other video-based information platforms. The video attribute competitiveness score vector is used to describe the competitiveness of a model with other models on each car attribute on video-based information platforms. The video motivation score vector is used to describe the priority of users' consideration of each car attribute when purchasing a model on video-based information platforms.

[0147] In some embodiments of this application, the steps described above for using a large language model to calculate the video sentiment score vector, video attribute competitiveness score vector, and video motivation score vector for each vehicle model based on all video comment data for each vehicle model include:

[0148] For each vehicle model, perform the following steps:

[0149] The first step is to calculate the video credibility of each video comment data point for each car model based on the video popularity data.

[0150] Specifically, through the formula:

[0151] ;

[0152] ;

[0153] Calculate video comment data Video credibility .

[0154] in, , This represents the set of IDs for video comment data. Indicates the first The weight of each popularity data point Represents video comment data The The score of each popularity data point Represents video comment data The The value of a heat data point, when At that time, the first The popularity metric is the number of likes. At that time, the first The popularity metric is the number of comments, when... At that time, the first The popularity data is the number of collections, when At that time, the first The popularity data is the number of reposts.

[0155] The second step involves identifying the attribute mentions of each video text content based on a large language model, and calculating the video sentiment score vector based on the attribute mentions of all video text content and the credibility of all videos.

[0156] The attribute mentions in the video text content are used to describe the car attributes mentioned in the video text content, as well as the emotional intensity of the car attributes. The video sentiment score vector includes the video sentiment score for each car attribute of the car model.

[0157] Specifically, through the formula:

[0158] ;

[0159] ;

[0160] ;

[0161] ;

[0162] Calculate the video sentiment score vector .

[0163] in, The video sentiment score represents the car attribute 1 of the vehicle model. Indicates the vehicle attributes of the model. Video sentiment rating Indicates the vehicle attributes of the model. Video sentiment rating Indicates video text content Regarding car attributes The rating mentioned Indicates video text content Regarding car attributes The intensity of emotions, express Aggregate weights.

[0164] For example, since video comment data includes video text content, the video text content shares the same ID as its corresponding video comment data. Cue words are set for the large language model, and then the video comment text is input into the large language model. The large language model infers based on the cue words to obtain the attribute mentions of the video text content. An example is as follows:

[0165] Example of information extraction from video sentiment knowledge extraction prompts:

[0166] The user wrote: "This car has great space; both the front and rear seats are spacious, and the trunk is quite large, more than enough for daily family use. However, the range is really bad. In winter, with the air conditioning on, I have to charge it every two days, which is a real hassle. I usually charge it for an hour and then leave; any longer is too much of a hassle. Sometimes after working overtime, I have to wait here to charge, and an hour can get a lot done, so sitting there is quite uncomfortable. Although the charging experience isn't great, the overall cost of ownership is indeed low. Electricity is much cheaper than gasoline, so in the long run, it's quite cost-effective. I think it deserves full marks for value for money! Remember to like and follow!"

[0167] Output result:

[0168] {"records":[{"fields":{"Space":"(1, 8)", "Driving_Experience":"(0,0)", "Range":"(1, -5)", "Exterior":"(0, 0)", "Interior":"(0, 0)", "Cost_Performance":"(1, 10)", "Intelligence":"(0, 0)"}}]}.

[0169] The third step involves identifying video comparison mentions for each video text content based on a large language model, and calculating a video attribute competitiveness score vector based on all video comparison mentions and all video credibility.

[0170] The video comparison mentions results are used to describe the car attributes mentioned in the video text when comparing other car models, as well as the emotional intensity of the mentioned car attributes. The video attribute competitiveness score vector includes the video attribute competitiveness score for each car attribute of the car model.

[0171] Specifically, through the formula:

[0172] ;

[0173] ;

[0174] ;

[0175] ;

[0176] ;

[0177] Calculate the video attribute competitiveness score vector .

[0178] in, This indicates the video attribute competitiveness score of the vehicle model's vehicle attribute 1. Indicates the vehicle attributes of the model. Video attribute competitiveness score, Indicates the vehicle attributes of the model. Video attribute competitiveness score, This indicates the text of the video comment. When comparing with other models, mention the car's attributes. The intensity of emotions, express Aggregate weights, This indicates the text of the video comment. When comparing with other car models, the car's attributes The rating mentioned.

[0179] For example, prompt words are set for the large language model, and then the video comment text is input into the large language model for inference to obtain video comparison mention results, such as:

[0180] Example of extracting prompt words by comparing attributes:

[0181] Input: What about that brand, model B? They're labeling it as brand Y. Haven't you considered brand Y? I have, I have that option. Actually, if I can't decide which car to buy within a year, I'll just blindly go for brand Y.

[0182] Output result: { "output": [ { "car": "Brand Y"} ]}.

[0183] Example of information extraction from vehicle sentiment rating prompts:

[0184] Input: "To be honest, the interior design of this car is indeed a bit outdated. There aren't many soft-touch materials, the overall quality is average, and it looks rather cheap. Compared to other brands in the same class (D and C), the difference is even more obvious. Although the flat-bottomed steering wheel slightly enhances the sporty feel, the overall interior style still appears somewhat outdated, lacking a sense of technology and luxury."

[0185] Output result:

[0186] {"records": [{"fields": {"Car_Name": "Brand D", "Space": "(0, 0)", "Driving_Experience": "(0, 0)", "Range": "(0, 0)", "Exterior": "(0, 0)", "Interior": "(1, -10)", "Cost_Performance": "(0, 0)", "Intelligence": "(0, 0)"}}, {"fields": {"Car_Name": "Brand C", "Space": "(0, 0)", "Driving_Experience": "(0, 0)", "Range": "(0, 0)", "Exterior": "(0, 0)", "Interior": "(1, -10)", "Cost_Performance": "(0, 0)", "Intelligence": "(0, 0)"}}]}.

[0187] The fourth step involves identifying the video purchase motivation attributes of each video text content based on a large language model, and calculating the video motivation score vector based on all video purchase motivation attributes and all video credibility.

[0188] The video purchase motivation attribute is one of multiple car attributes, and the video motivation score vector includes the video motivation score for each car attribute of the car model.

[0189] Specifically, through the formula:

[0190] ;

[0191] ;

[0192] ;

[0193] Calculate the video motivation rating vector .

[0194] in, The video motivation score represents the vehicle's attribute 1. Indicates the vehicle attributes of the model. Video motivation rating, Indicates the vehicle attributes of the model. Video motivation rating, This indicates the text of the video comment. With car attributes The motivational relationship, if the car attributes For video comment text The video purchase motivation attribute, then If the car attributes Not for video comment text The video purchase motivation attribute, then .

[0195] For example, prompt words are set for the large language model, and the video comment text is input into the large language model for inference to obtain the video purchase motivation attribute of the video text content, such as:

[0196] Example of information extraction from motivational knowledge extraction prompts:

[0197] The user wrote: "What I like most about this car is its value for money. It's almost unbeatable in its price range, with rich features and comprehensive functionality, completely justifying the price. Both the materials and the technology offer a superior experience. Moreover, the daily operating costs are very low; it can travel a long distance on a single charge, and the energy consumption is well controlled. Overall, this car is truly worth buying and offers excellent value for money!"

[0198] Output: "Cost-effectiveness".

[0199] Step 14: Calculate the video recommendation index for each vehicle model based on the video sentiment score vector, video attribute competitiveness score vector, and video motivation score vector.

[0200] Specifically, through the formula:

[0201] ;

[0202] ;

[0203] ;

[0204] ;

[0205] ;

[0206] Calculate the video recommendation index for car models .

[0207] in, Indicates the competitiveness weight. Indicates the vehicle attributes of the model. Video sentiment rating Indicates the vehicle attributes of the model. Video attribute competitiveness score Indicates the vehicle attributes of the model. Video motivation rating, Represents the normalized video motivation rating vector , This indicates the preliminary calculation of the unweighted video user recommendation index after user personalization. This represents the video group's preference value. This indicates the individual's preference value for the video. Indicates car attributes Video personalization weight, This indicates the strength of the user's personalized video group attributes.

[0208] It should be noted that the strength of the user's personalized preference video group attributes mentioned above is used to describe the user's personalized needs. If the user has no personalized needs, then... If users have personalized needs, then the video personalization weight will be set according to those needs. Furthermore, the sum of the video personalization weights for all car attributes equals 1.

[0209] Step 15: Based on the review recommendation index and video recommendation index of all car models, provide personalized car model recommendations to the users to be recommended.

[0210] Specifically, based on the platforms used by the users to be recommended, as well as the review recommendation index and video recommendation index of all car models, personalized car model recommendations are given to the users to be recommended.

[0211] For example, if the user to be recommended primarily obtains car information through text-based platforms (such as car sales platforms), then car models are recommended to the user's account based on the comment recommendation index. The higher the comment recommendation index, the greater the probability that the corresponding car model will be recommended to the user; the lower the comment recommendation index, the lower the probability that the corresponding car model will be recommended to the user. If the user to be recommended primarily obtains car information through video-based platforms (such as Sina Auto), then car models are recommended to the user's account based on the video recommendation index. The higher the video recommendation index, the greater the probability that the corresponding car model will be recommended to the user; the lower the video recommendation index, the lower the probability that the corresponding car model will be recommended to the user.

[0212] The method of this application will be illustrated below with a specific example.

[0213] Analyzing user review data for brand G, with 7 vehicle attributes, an example of how the large language model infers from satisfactory review texts for this vehicle model is as follows:

[0214] Review text: First, the range is reassuring, basically around 80% of the original range, and electric vehicles are indeed economical, costing just over 10 cents per kilometer. Second, the sound system is exceptionally good, comparable to major brand audio systems. Third, the air conditioning is excellent, providing very rapid heating and cooling.

[0215] Attributes: Range: Mentioned (1,6); Value for money: Mentioned (1,8); Driving experience: Not mentioned (0,0); Exterior: Not mentioned (0,0); Interior: Not mentioned (0,0); Space: Not mentioned (0,0); Intelligent features: Not mentioned (0,0).

[0216] In this example, the satisfaction vector obtained by aggregating all comments is:

[0217] ;

[0218] An example of how a large language model infers from negative reviews of a particular car model is as follows:

[0219] Comment text: First, the comfort is poor, especially in the back seats; it's too bumpy. Does anyone have a good way to improve this? Second, the wireless charging for mobile phones is basically unusable; the power is too low. Third, it's rather plain, like a bare-bones structure.

[0220] Attribute identification: Intelligent: mentioned (1,-5); Interior: mentioned (1,-3); Driving experience: not mentioned (0,0); Value for money: not mentioned (0,0); Exterior: not mentioned (0,0); Range: not mentioned (0,0); Space: not mentioned (0,0).

[0221] In this example, the attribute regret tendency score vector obtained by aggregating all comment analysis is:

[0222] ;

[0223] The comparison attribute mentions are shown in Table 1:

[0224] Table 1 Comparison Attribute Mention Results

[0225]

[0226] The attribute competitiveness score vector obtained by aggregating all comments and analysis is as follows:

[0227] ;

[0228] Here is an example of extracting purchase motivation attributes:

[0229] Review text: First, the driving range is reassuring and unlikely to be exaggerated. Electric vehicles are indeed economical, costing just over 10 cents per kilometer. Second, the sound system is exceptionally good, comparable to major brand systems. Third, the air conditioning is excellent, providing both heating and cooling very quickly.

[0230] Motivation identification: cost-effectiveness.

[0231] The attribute motivation rating vector obtained by aggregating all comments and analysis is as follows:

[0232] ;

[0233] The following is an example of using a large language model to infer attribute mentions in video comment text:

[0234] Video comment text: I initially considered a certain brand, a domestic brand, which is said to be one of the better domestic brands. However, things like autonomous driving, and since I often drive on highways, my requirements for autonomous driving are quite high. Another brand with relatively stable autonomous driving technology is almost...

[0235] Attribute identification: Intelligent: mentioned (1,8); Interior: not mentioned (0,0); Driving experience: not mentioned (0,0); Value for money: not mentioned (0,0); Exterior: not mentioned (0,0); Range: not mentioned (0,0); Space: not mentioned (0,0).

[0236] The results of the video comparison are shown in Table 2.

[0237] Table 2. Video Comparison Mention Results

[0238]

[0239] The video attribute competitiveness score vector obtained from the aggregated analysis of all videos is as follows:

[0240] ;

[0241] The following is an example of identifying video purchase motivation attributes:

[0242] Video comment text: I initially considered a certain brand, a domestic brand, which is said to be one of the better domestic brands. However, things like autonomous driving, and since I often drive on highways, my requirements for autonomous driving are quite high. Another brand with relatively stable autonomous driving technology is almost...

[0243] Motivation identification: Intelligentization.

[0244] The video motivation score vector that aggregates all video analyses is:

[0245] ;

[0246] This example uses three car models for aggregation: brand G, brand D, and brand H. Assuming the value is 0.5, in this example... Assuming a value of 0.5, and assuming users prioritize appearance and personalization. , Assume the range is [0.1, 0.1, 0.1, 0.4, 0.1, 0.1, 0.1].T .

[0247] The text data results of the reviews for brand G are shown in Table 3.

[0248] Table 3. Results of Textual Review Data for Brand G

[0249]

[0250] Normalize means normalization.

[0251] The video text data results for brand G are shown in Table 4.

[0252] Table 4. Video Text Data Results for Brand G

[0253]

[0254] The text data results of the reviews for brand D are shown in Table 5.

[0255] Table 5. Results of Text Data on Brand D's Reviews

[0256]

[0257] The video text data results for brand D are shown in Table 6.

[0258] Table 6. Video Text Data Results for Brand D

[0259]

[0260] The text data results of the reviews for brand H are shown in Table 7.

[0261] Table 7. Results of Textual Review Data for Brand H

[0262]

[0263] The video text data results for brand H are shown in Table 8.

[0264] Table 8. Video Text Data Results for Brand H

[0265]

[0266] The final recommendation index example is shown in Table 9.

[0267] Table 9 Final Recommendation Index

[0268]

[0269] in, Indicates brand G, Brand D is represented. It represents brand H.

[0270] As can be seen, the method in this application can fully analyze user preferences for different car models and is committed to conducting a comprehensive and systematic in-depth analysis of user review texts.

[0271] like Figure 2 As shown, an embodiment of this application provides a terminal device, wherein the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 2 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 executes the computer program D102 to implement the steps in any of the above method embodiments.

[0272] Specifically, when the processor D100 executes the computer program D102, it analyzes the information of the vehicle models based on user reviews and video reviews, which can improve the diversity and comprehensiveness of the vehicle model information. It calculates multiple rating vectors for the vehicle models in the reviews, taking into account user evaluations of multiple aspects such as satisfaction, regret tendency, purchase motivation, and competitiveness of the vehicle models, thereby improving the accuracy of the review recommendation index and video recommendation index of the vehicle models, and thus improving the accuracy of personalized recommendations for users.

[0273] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0274] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0275] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention.

Claims

1. A personalized car recommendation method based on a large language model, characterized in that, include: Obtain multiple user review data and multiple video review data for multiple car models; User review data includes review text, ratings of multiple car attributes, and user characteristic data; video review data includes video text content and video popularity data. Using a large language model, the attribute satisfaction score vector, attribute regret tendency score vector, attribute competitiveness score vector, and attribute motivation score vector are calculated for each vehicle model based on all user review data. The review recommendation index for each vehicle model is calculated based on the attribute satisfaction rating vector, attribute regret tendency rating vector, attribute competitiveness rating vector, and attribute motivation rating vector. Using a large language model, based on all video comment data for each vehicle model, calculate the video sentiment score vector, video attribute competitiveness score vector, and video motivation score vector for each vehicle model; Based on the video sentiment score vector, video attribute competitiveness score vector, and video motivation score vector of each vehicle model, calculate the video recommendation index for each vehicle model. Based on the review recommendation index and video recommendation index of all car models, personalized car model recommendations are given to users to be recommended. The comment text can be either a satisfactory comment or a dissatisfied comment. The method of using a large language model to calculate the attribute satisfaction score vector, attribute regret tendency score vector, attribute competitiveness score vector, and attribute motivation score vector for each vehicle model based on all user review data includes: For each of the aforementioned vehicle models, the following steps shall be performed: For each user review data point of the aforementioned vehicle model, the credibility of the user review data is calculated based on user characteristic data and ratings of multiple vehicle attributes; Based on a large language model, the attribute mentions of each satisfaction review text for the vehicle model are identified. Based on the attribute mentions of all satisfaction review texts and the credibility of all satisfaction review texts, the attribute satisfaction score vector for the vehicle model is calculated. The attribute mentions of the satisfaction review texts are used to describe the vehicle attributes mentioned in the satisfaction review texts and the emotional intensity of the vehicle attributes. The attribute satisfaction score vector includes the attribute satisfaction score of each vehicle attribute of the vehicle model. Based on a large language model, the attribute mentions of each unsatisfactory review text of the vehicle model are identified. Based on the attribute mentions of all unsatisfactory review texts of the vehicle model and the credibility of all unsatisfactory review texts, the attribute regret tendency score vector of the vehicle model is calculated. The attribute mentions of the unsatisfactory review texts are used to describe the car attributes mentioned in the unsatisfactory review texts and the emotional intensity of the car attributes. The attribute regret tendency score vector includes the attribute regret tendency score of each car attribute of the vehicle model. Based on a large language model, the comparative attribute mentions of each review text for the vehicle model are identified, and the attribute competitiveness score vector of the vehicle model is calculated based on all comparative attribute mentions and all credibility levels. The comparative attribute mentions are used to describe the car attributes mentioned in the review text when compared with other car models, as well as the sentiment intensity of the mentioned car attributes. The attribute competitiveness score vector includes the attribute competitiveness score of each car attribute of the vehicle model. Based on a large language model, the purchase motivation attribute of each review text of the vehicle model is identified, and the attribute motivation score vector of the vehicle model is calculated based on all purchase motivation attributes; the purchase motivation attribute is one of multiple vehicle attributes, and the attribute motivation score vector includes the attribute motivation score of each vehicle attribute of the vehicle model.

2. The personalized car recommendation method according to claim 1, characterized in that, The user feature data includes the authentication status of the corresponding user account, the recommendation level of the corresponding user comment data, and the interaction index of the corresponding user comment data; The process of calculating the credibility of user review data based on user characteristic data and ratings of multiple car attributes includes: Through the formula: ； Calculate user review data Credibility ; in, , This represents a set of IDs representing user review data for a particular car model. Indicates the first The weight of each credibility indicator, Represents user comment data The The value of a reliable indicator: ； = ； ； ； in, Represents user comment data The value of the first credibility indicator, Represents user comment data The value of the second credibility indicator, Represents user comment data The value of the third credibility indicator, Represents user comment data The value of the fourth credibility indicator, A set of numbers representing vehicle attributes. This indicates the car attributes in all user review data for the aforementioned car model. The average rating, Represents user comment data Regarding car attributes The rating, Represents user comment data The interaction index.

3. The personalized car recommendation method according to claim 2, characterized in that, The step of calculating the attribute satisfaction rating vector for the vehicle model based on the attribute mentions in all satisfactory review texts and the corresponding credibility of all satisfactory review texts includes: Through the formula: ； ； ； ； Calculate the attribute satisfaction rating vector of vehicle models ; in, This represents the satisfaction score of attribute 1 of the vehicle model. Indicates the vehicle attributes of the model. Attribute satisfaction rating A set of numbers representing vehicle attributes. The set of numbers representing the texts of satisfactory comments. This indicates the transpose operation. Indicates the vehicle attributes of the model. Attribute satisfaction rating Text indicating satisfaction Regarding car attributes The rating mentioned Text indicating satisfaction Regarding car attributes The intensity of emotions, express Aggregate weights, Text indicating satisfaction Corresponding credibility Transformation function: ； The method for calculating the attribute competitiveness score vector of the vehicle model based on all comparative attribute mentions and all credibility levels includes: Through the formula: ； ； ； ； ； Calculate the attribute competitiveness score vector of the vehicle model ; in, This represents the attribute competitiveness score of vehicle attribute 1 for the aforementioned vehicle model. This indicates the vehicle attributes of the vehicle model. Attribute competitiveness score This indicates the vehicle attributes of the vehicle model. Attribute competitiveness score express Aggregate weights, Comment text When comparing with other models, mention the car's attributes. The intensity of emotions, Comment text When comparing with other car models, the car's attributes The rating mentioned.

4. The personalized car recommendation method according to claim 3, characterized in that, The calculation of the attribute motivation score vector for the vehicle model based on all purchase motivation attributes includes: Through the formula: ； ； ； Calculate the attribute motivation score vector of the vehicle model. ; in, This represents the attribute motivation score for vehicle attribute 1 of the aforementioned vehicle model. This indicates the vehicle attributes of the vehicle model. Attribute motivation rating This indicates the vehicle attributes of the vehicle model. Attribute motivation rating, Comment text With car attributes The motivational relationship, if the car attributes For comment text Purchase motivation attributes, If the car attributes Not for comment text Purchase motivation attributes, .

5. The personalized car recommendation method according to claim 4, characterized in that, The calculation of the review recommendation index for each vehicle model based on its attribute satisfaction rating vector, attribute regret tendency rating vector, attribute competitiveness rating vector, and attribute motivation rating vector includes: Through the formula: ； ； ； ； ； Calculate the review and recommendation index of a car model ; in, Indicates the competitiveness weight. This indicates the vehicle attributes of the vehicle model. The attribute of regret tendency score, , This represents the attribute regret tendency rating vector. Represents group preference value. Represents the normalized attribute motivation rating vector , This indicates the unweighted recommendation index. This represents the user's personal preference value. Indicates car attributes Personalized weights, This indicates the strength of the user's personalized preference group attributes.

6. The personalized car recommendation method according to claim 1, characterized in that, The method utilizes a large language model to calculate, based on all video comment data for each vehicle model, a video sentiment score vector, a video attribute competitiveness score vector, and a video motivation score vector for each vehicle model, including: For each vehicle model, perform the following steps: For each video comment data point of the aforementioned vehicle model, the video credibility of the video comment data is calculated based on the video popularity data within the video comment data. Based on a large language model, the attribute mentions of each video text content are identified, and a video sentiment score vector is calculated based on the attribute mentions of all video text content and the credibility of all videos. The attribute mentions of the video text content are used to describe the car attributes mentioned in the video text content, as well as the sentiment intensity of the car attributes. The video sentiment score vector includes the video sentiment score for each car attribute of the car model. Based on a large language model, the video comparison mentions of each video text content are identified, and a video attribute competitiveness score vector is calculated based on all video comparison mentions and all video credibility. The video comparison mentions are used to describe the car attributes mentioned in the video text when comparing other car models, as well as the emotional intensity of the mentioned car attributes. The video attribute competitiveness score vector includes the video attribute competitiveness score of each car attribute of the car model. Based on a large language model, the video purchase motivation attribute of each video text content is identified, and a video motivation score vector is calculated based on all video purchase motivation attributes and all video credibility; the video purchase motivation attribute is one of multiple car attributes, and the video motivation score vector includes the video motivation score of each car attribute of the car model.

7. The personalized car recommendation method according to claim 6, characterized in that, The video popularity data includes the number of likes, comments, favorites, and shares of the video; The step of calculating the video credibility of the video comment data based on the video popularity data in the video comment data includes: Through the formula: ； ； Calculate video comment data Video credibility ; in, , This represents the set of IDs for video comment data. Indicates the first The weight of each popularity data point Represents video comment data The The score of each popularity data point Represents video comment data The The value of a heat data point, when At that time, the first The popularity metric is the number of likes. At that time, the first The popularity metric is the number of comments, when... At that time, the first The popularity data is the number of collections, when At that time, the first The popularity data is the number of reposts.

8. The personalized car recommendation method according to claim 7, characterized in that, The video recommendation index for each vehicle model is calculated based on its video sentiment score vector, video attribute competitiveness score vector, and video motivation score vector, including: Through the formula: ； ； ； ； ； Calculate the video recommendation index for car models ; in, Indicates the competitiveness weight. This indicates the vehicle attributes of the vehicle model. Video sentiment rating This indicates the vehicle attributes of the vehicle model. Video attribute competitiveness score This indicates the vehicle attributes of the vehicle model. Video motivation rating, Represents the normalized video motivation rating vector , This indicates the unweighted video user recommendation index. This represents the video group's preference value. This indicates the individual's preference value for the video. Indicates car attributes Video personalization weight, This indicates the strength of the user's personalized video group attributes.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the car personalization recommendation method based on a large language model as described in any one of claims 1 to 8.

Images