Vehicle type preference crowd circling method based on user behavior and interest inference

Abstract

The invention discloses a vehicle type preference crowd selection method based on user behavior and interest inference, and the method comprises the steps: S1, carrying out the data collection through a multi-party data source, and carrying out the unified arrangement through a user ID; s2, carrying out behavior intensity quantification, interest label inference and mathematical vector representation on the uniformly arranged data, and then constructing a user portrait; s3, intentional users and clear and uninterested users are defined according to the user portraits, and preference probability scores of the users are predicted by using the trained model; and S4, setting a probability threshold of circling and selecting, and circling and selecting the users whose scores are higher than the threshold. According to the method, after offline store arrival and vertical medium deep inquiry are combined, the judgment accuracy of user intentions can be greatly improved; and in combination with an interest inference mode, the method uses an LDA model to predict potential topic preferences which are difficult to manually define in advance in the data, and an unexpected user preference mode can be mined, so that the generated user portrait is more accurate and richer.

Description

Technical Field

[0001] This invention relates to the field of data analysis and processing technology, and in particular to a method for selecting car model preference groups based on user behavior and interest inference. Background Technology

[0002] Audience targeting refers to the process in marketing and advertising of analyzing and identifying specific target audiences in order to more effectively deliver information and promote products or services. In the automotive sales process, this method can more efficiently identify potential users.

[0003] Existing methods for identifying car model preferences mostly rely solely on user website behavior data. While this approach provides a direct indication of customer intentions, it fails to consider user interests for targeted inference. Consequently, it overlooks users with genuine car-buying interest during the segmentation process, hindering timely follow-up by sales staff. Therefore, this solution proposes a car model preference segmentation method based on user behavior and interest inference to address these issues. Summary of the Invention

[0004] The purpose of this invention is to provide a method for selecting car model preference groups based on user behavior and interest inference, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for selecting car model preference groups based on user behavior and interest inference, comprising the following steps: S1. Data is collected from multiple sources, including website behavior data, customer relationship management data, offline data, automotive media data, and DMP data. The collected data from different channels are then organized using user IDs. S2, after the unified and organized data is quantified by behavior intensity, inferred by interest tags and represented by mathematical vectors, user profiles are constructed. S3. Define potential users and users with no interest based on user profiles to form positive and negative samples. Input the features of positive and negative samples into the gradient boosting tree model for training. Use the trained model to predict the user's preference probability score. S4 sets the probability threshold for selection, and selects users whose scores are higher than the threshold.

[0006] Preferably, the website behavior data is collected by tracking points to collect every user behavior on the official car manufacturer's website, car model app, and mini-program. The website behavior data includes browsing behavior, search behavior, and interaction behavior. The browsing behavior includes browsing vehicle model pages, configuration pages, review articles, vehicle review videos, and the duration of browsing; The search behavior refers to the search for keyword content; The interactive behaviors include whether the loan calculator was used, whether the car model was saved, and whether the test drive appointment was booked.

[0007] Preferably, the customer relationship management data includes registered member information, historical maintenance records, and customer service consultation records; The offline data includes in-store data and is matched with online identities; The automotive media data includes user behavior data on automotive media platforms; The DMP data includes demographic attributes and interest tags recorded by users on the data platform.

[0008] Preferably, the behavior intensity quantification involves recording the number of times a user visits the vehicle model page within 30 days, checking whether the user has browsed in-depth pages related to price and parameter configuration, and querying the record time and duration of the user's most recent browsing. The interest tags are inferred based on content, behavioral patterns, and user preferences. If a user frequently browses content related to new energy vehicles, a new energy intention tag is assigned. If a user repeatedly compares the configurations of luxury car models, a luxury brand mid-size sedan competition tag is assigned. The user's preferences are inferred by analyzing the different content topics browsed by the user through the LDA model, and the user is represented as a mathematical vector based on the preferences.

[0009] Preferably, the LDA model inference operation method includes the following steps: A1 first collects text information such as titles, body text, video descriptions, and tags of all content about cars that users have viewed; A2, perform text cleaning, including removing stop words, word segmentation, and removing low-frequency or high-frequency words; A3 combines keywords from content viewed by the user into a document, and then uses TF-IDF to vectorize the text data; A4, preset the number of topics K, and then input the vectorized data into the LDA model. K is determined with the help of perplexity or consistency score; A5 allows you to view the top 10-20 words with the highest weight under a preset topic and define the preferences represented by the topic.

[0010] Preferably, the positive samples represent highly interested users, including users who have completed a test drive, users who have made in-depth inquiries about prices on automotive vertical media, and users who have visited the same model page multiple times and stayed for more than 5 minutes. The negative samples represent users who have no clear interest, including users who only visit a single page and users who browse content unrelated to automobiles.

[0011] Preferably, the training of the gradient boosting tree model includes the following steps: B1 first filters feature data based on the importance of tree model features, with the filtering targets including positive and negative samples; B2, the selected feature data is divided into a training set and a test set, with the training set accounting for 70%–80% and the test set accounting for 20%–30%; B3 uses the training set data to train a gradient boosting tree model, and then uses the test set to evaluate the model. The evaluation metrics include accuracy, recall (F1-score), and ROC-AUC. B4 uses the trained model to predict the preference probability score for new users and outputs the preference probability score for each user.

[0012] Preferably, the training set for training the model includes the following steps: B301, the initial prediction value is a constant value, where the constant value is the mean of the target variable; B302, iteratively add trees, and fit the current residual each time; B303, in each iteration, obtains the direction of improvement for the current model by calculating the negative gradient of the loss function.

[0013] Preferably, the preference probability score is between 0 and 1, and the preference probability score indicates the probability that the user belongs to the high-intent group of the target car model.

[0014] Preferably, the probability threshold for selection is a probability score > 0.8, and high-scoring user leads are assigned to the sales team for follow-up.

[0015] The technical effects and advantages of this invention are as follows: This invention's targeting method integrates and unifies offline in-store data, vertical media inquiry data, and online browsing behavior data through user IDs. By merging online and offline data, it reduces the misjudgment rate of analyzing only online behavior. Furthermore, by combining offline in-store data and in-depth vertical media inquiries, the accuracy of judging user intentions is significantly improved. In addition, this method combines interest inference with the use of LDA models to predict potential, manually difficult-to-predefine thematic preferences in the data, which can uncover unexpected user preference patterns, making the generated user profiles more accurate and richer. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the operational steps of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This invention provides, for example Figure 1 The method for selecting car model preference groups based on user behavior and interest inference, as shown, includes the following steps: S1 collects data from multiple sources, including website behavior data, customer relationship management data, offline data, automotive media data, and DMP data. The collected data from different channels is organized in a unified manner by user ID. The richer and more accurate the data, the more accurate the subsequent inferences will be. Specifically, website behavior data is collected through event tracking of user behavior on official car manufacturer websites, car model apps, and mini-programs. Website behavior data includes browsing behavior, search behavior, and interaction behavior. Browsing behavior includes browsing vehicle model pages, configuration pages, review articles, vehicle test videos, and the duration of browsing; The search behavior refers to the search keywords, which include tagged content such as "MPV with a range of 500km", "B-class sedan priced between 250,000 and 300,000 RMB" and "the top ten best-selling sedans in monthly sales". Interactive behaviors include whether the loan calculator was used, whether the car model was saved, and whether the test drive appointment was booked.

[0019] Customer relationship management data includes registered member information, historical maintenance records, and customer service consultation records; Offline data includes in-store data and is matched with online identity. In-store data is captured via in-store Wi-Fi and Bluetooth, and user authorization is required to obtain this data. It should be noted that the aforementioned website behavior data, customer relationship management data, and offline data can directly reflect user behavior. Detailed tracking points need to be set up on the website or app to ensure that all important user interactions are captured.

[0020] Automotive vertical media data includes user behavior data on automotive media platforms, such as which car reviews they viewed and which models they compared. This type of data is used to infer user intentions. DMP data includes user demographic attributes and interest tags recorded on the data platform. Demographic attributes include age and income, while interest tags include user hobbies such as interest in finance, enjoyment of outdoor activities, and love of watching movies.

[0021] It should be noted that data from different channels can be integrated through user IDs (mobile phone numbers, device IDs, WeChat OpenIDs) to form a unified customer data platform (CDP), ensuring a comprehensive view of users.

[0022] S2. After uniformly organizing the data, construct user portraits by quantifying behavior intensity, inferring interest tags, and representing them as mathematical vectors, and process the original data into features that can be processed by subsequent models. Specifically, behavior intensity quantification records the number of times a user visits the vehicle model page within a month (frequency behavior), checks whether the user has browsed in-depth pages related to prices and parameter configurations (depth behavior), and queries the time and duration of the user's most recent browsing record (recency behavior). Through the recency behavior, the user's recent behavior can be grasped, so this most recent time needs to be within a month. Interest tags are inferred based on content, behavior patterns, and user preferences. If a user frequently browses content related to new energy vehicles, including content on battery technology and new energy vehicle charging pile installation, then mark a new energy intention tag. If a user repeatedly compares the configurations of luxury models, then mark a luxury brand mid-size sedan competition circle tag. Analyze the different content themes browsed by users through the LDA model to infer preferences, including cost performance, performance handling, sense of technology, and family space. Represent the user as a mathematical vector according to the preferences, expressed as: Luxury car interest score: 0.95, New energy interest score: 0.2, Family SUV interest score: 0.4, Income level: 8, Age: 28. Through this vector, a user's interest preferences can be comprehensively represented.

[0023] Furthermore, the operation method for inferring by the LDA model includes the following steps: A1. First, collect the title, text, video description, and tag-like text information of all content related to cars that the user has browsed. A2. Clean the collected text, including removing stop words, word segmentation, and removing low-frequency or high-frequency words. General words include meaningless words such as "de", "le", "ne", and "ma". Word segmentation is to use the jieba-like Chinese word segmentation library to split sentences into words. Low-frequency or high-frequency words are those that appear too few or too many. Those that appear too few are directly deleted, and those that appear too many are filtered and duplicate ones are deleted. A3. Merge the content keywords browsed by the user into a document. After merging the documents, one user corresponds to one document. Then use TF-IDF to vectorize the text data, generating a matrix of "number of users x vocabulary size". The values in the matrix represent the importance of a certain word in a certain user's document. A4. Preset the number of topics K, and then input the vectorized data into the LDA model. K is assisted in determination by perplexity or coherence score. The output of the model includes a list of representative words for each topic and the topic probability distribution for each user. A5 allows you to view the top 10-20 words with the highest weight under a preset topic and define the preferences represented by the topic.

[0024] S3. Based on the user profile, define the intended users and users with no clear interest to form positive and negative samples. Input the features of the positive and negative samples into the gradient boosting tree model for training. Use the trained model to predict the user's preference probability score. Through machine learning, identify the features of seed users with clear interests and use the features of these users to find similar users in the full user base. Specifically, positive samples represent highly interested users, including users who have completed test drives, users who have made in-depth inquiries about prices on automotive media, and users who have visited the same model page multiple times and stayed for more than 5 minutes. Negative samples represent users who have no clear interest, including users who only make a single page visit and users who browse content unrelated to automobiles.

[0025] Specifically, training a gradient boosting tree model involves the following steps: B1 first filters feature data based on the importance of tree model features. The filtering targets include positive and negative samples, and the data is preprocessed, including missing value handling, outlier detection, and feature scaling. B2, the selected feature data is divided into a training set and a test set, with the training set accounting for 70%–80% and the test set accounting for 20%–30%; B3 uses the training set data to train a gradient boosting tree model. The model parameters can be adjusted as needed, including the learning rate, tree depth, and the number of weak learners per tree. Then, the model is evaluated using a test set. Evaluation metrics include accuracy, recall, F1-score, and ROC-AUC. Based on the evaluation results, hyperparameters are tuned to improve the model's performance. Furthermore, training the model on the training set includes the following steps: B301, initialize the predicted value as a constant value, and the constant value is the mean of the target variable; B302, iteratively add trees, and fit the current residual each time; B303, in each iteration, the direction of improvement of the current model is obtained by calculating the negative gradient of the loss function; B4. Use the trained model to predict the preference probability score of new users, output the preference probability score of each user, save the trained model for later use, and save the model's parameters and structure.

[0026] Specifically, the preference probability score ranges from 0 to 1, indicating the likelihood that a user belongs to the high-intent group for the target car model.

[0027] S4. Set a probability threshold for selection. Select users whose scores are higher than the threshold. The probability threshold is a probability score > 0.8. Alternatively, select the top 10% of users and assign high-scoring user leads to the sales team for follow-up. At the same time, synchronize the selected user groups to the advertising platform for precise push of WeChat Moments and information flow ads. The ad content is specific car models that they are interested in.

[0028] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for selecting car model preference groups based on user behavior and interest inference, characterized in that, The selection method includes the following steps: S1. Data is collected from multiple sources, including website behavior data, customer relationship management data, offline data, automotive media data, and DMP data. The collected data from different channels are then organized using user IDs. S2, after the unified and organized data is quantified by behavior intensity, inferred by interest tags and represented by mathematical vectors, user profiles are constructed. S3. Define potential users and users with no interest based on user profiles to form positive and negative samples. Input the features of positive and negative samples into the gradient boosting tree model for training. Use the trained model to predict the user's preference probability score. S4 sets the probability threshold for selection, and selects users whose scores are higher than the threshold.

2. The method for selecting car model preference groups based on user behavior and interest inference according to claim 1, characterized in that, The website behavior data is collected by tracking points to collect every user behavior on the official car manufacturer's website, car model app, and mini-program. The website behavior data includes browsing behavior, search behavior, and interaction behavior. The browsing behavior includes browsing vehicle model pages, configuration pages, review articles, vehicle review videos, and the duration of browsing; The search behavior refers to the search for keyword content; The interactive behaviors include whether the loan calculator was used, whether the car model was saved, and whether the test drive appointment was booked.

3. The method for selecting car model preference groups based on user behavior and interest inference according to claim 1, characterized in that, The customer relationship management data includes registered member information, historical maintenance records, and customer service consultation records; The offline data includes in-store data and is matched with online identities; The automotive media data includes user behavior data on automotive media platforms; The DMP data includes demographic attributes and interest tags recorded by users on the data platform.

4. The method for selecting car model preference groups based on user behavior and interest inference according to claim 1, characterized in that, The behavior intensity quantification involves recording the number of times a user visits the vehicle model page within 30 days, checking whether the user has browsed in-depth pages related to price and parameter configuration, and querying the record time and duration of the user's most recent visit. The interest tags are inferred based on content, behavioral patterns, and user preferences; if a user frequently browses content related to new energy vehicles, a new energy intention tag is assigned. If a user repeatedly compares the configurations of luxury car models, then the user is tagged with a "competitive zone" label for luxury mid-size sedans. The user's preferences are inferred by analyzing the different content topics they browse using the LDA model, and the user is represented as a mathematical vector based on these preferences.

5. The method for selecting car model preference groups based on user behavior and interest inference according to claim 4, characterized in that, The LDA model inference operation method includes the following steps: A1 first collects text information such as titles, body text, video descriptions, and tags of all content about cars that users have viewed; A2, perform text cleaning, including removing stop words, word segmentation, and removing low-frequency or high-frequency words; A3 combines keywords from content viewed by the user into a document, and then uses TF-IDF to vectorize the text data; A4, preset the number of topics K, and then input the vectorized data into the LDA model. K is determined with the help of perplexity or consistency score; A5 allows you to view the top 10-20 words with the highest weight under a preset topic and define the preferences represented by the topic.

6. The method for selecting car model preference groups based on user behavior and interest inference according to claim 1, characterized in that, The positive samples represent high-intent users, including users who have completed test drives, users who have made in-depth inquiries about prices on automotive vertical media, and users who have visited the same model page multiple times and stayed for more than 5 minutes. The negative samples represent users who have no clear interest, including users who only visit a single page and users who browse content unrelated to automobiles.

7. The method for selecting car model preference groups based on user behavior and interest inference according to claim 1, characterized in that, The training of the gradient boosting tree model includes the following steps: B1 first filters feature data based on the importance of tree model features, with the filtering targets including positive and negative samples; B2, the selected feature data is divided into a training set and a test set, with the training set accounting for 70%–80% and the test set accounting for 20%–30%; B3 uses the training set data to train a gradient boosting tree model, and then uses the test set to evaluate the model. The evaluation metrics include accuracy, recall (F1-score), and ROC-AUC. B4 uses the trained model to predict the preference probability score for new users and outputs the preference probability score for each user.

8. The method for selecting car model preference groups based on user behavior and interest inference according to claim 7, characterized in that, The training set for training the model includes the following steps: B301, the initial prediction value is a constant value, where the constant value is the mean of the target variable; B302, iteratively add trees, and fit the current residual each time; B303, in each iteration, obtains the direction of improvement for the current model by calculating the negative gradient of the loss function.

9. The method for selecting car model preference groups based on user behavior and interest inference according to claim 7, characterized in that, The preference probability score is between 0 and 1, and it represents the probability that the user belongs to the high-intent group for the target car model.

10. The method for selecting car model preference groups based on user behavior and interest inference according to claim 1, characterized in that, The probability threshold for selection is a probability score > 0.8, and high-scoring user leads are assigned to the sales team for follow-up.

Images