Potential new energy vehicle owner prediction method and system based on user portrait and semi-supervised learning, and storage medium

Abstract

The present application provides a potential new energy vehicle owner prediction method and system based on user portrait and semi-supervised learning, and a storage medium. Starting from traditional vehicle owner data, the actual purchase and purchase intention of users are distinguished and marked to find target users. A self-training method using semi-supervised learning is proposed at the data level to solve the problem of insufficient data. Finally, the user intention data is introduced into the model iteration link, the data results are associated with the actual business, the accuracy and business explanation of the model prediction results are improved, and the potential new energy target vehicle owners of traditional vehicle enterprises are effectively mined. The present application can overcome the problems that the customer resources of traditional vehicles and new energy vehicles are difficult to interoperate, the small scale of new energy vehicle owners leads to data skew, and the long purchase cycle is difficult to cope with model verification.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a technology for predicting potential new energy vehicle owners based on user profiles and semi-supervised learning. Background Technology

[0002] For automakers transitioning from traditional gasoline-powered vehicles to new energy vehicles, the ability to accurately identify potential new energy vehicle customers among traditional car owners directly impacts the sales volume of new energy products and the company's profitability. Therefore, forecasting new energy vehicle owners based on traditional vehicle data is a key focus for traditional automakers undergoing this transformation.

[0003] In existing technologies, traditional methods for identifying potential customers mainly rely on sales and lead generation information. These methods collect information proactively recorded by users to find potential customers, which is costly but results in relatively accurate reach. Newer methods, on the other hand, rely primarily on internet technology to analyze user profiles to find potential customers. This is less costly but results in less accurate reach. Clearly, the inherent difficulties in data collection and inaccurate user tagging significantly impact predictions.

[0004] Therefore, how to propose a method that can more efficiently combine traditional information with new technologies to screen potential customers for new energy vehicles from traditional car owners has become a key issue that new energy vehicle companies urgently need to solve. Summary of the Invention

[0005] To overcome the challenges of limited interoperability between traditional and new energy vehicle customer resources, data bias due to the small scale of new energy vehicle owners, and difficulties in model validation due to long purchase cycles, this invention proposes a method for predicting potential new energy vehicle owners based on user profiling and semi-supervised learning. Starting with traditional vehicle owner data, the method differentiates and labels users based on their actual purchases and purchase intentions to identify target users. At the data level, it employs a semi-supervised learning self-training method to address insufficient data. Finally, user intention data is incorporated into the model iteration process, linking the data results with actual business operations to improve the accuracy and interpretability of the model's predictions. This effectively helps traditional automakers identify potential new energy vehicle target owners.

[0006] The technical solution of the present invention is as follows:

[0007] In its first aspect, this invention proposes a method for predicting new energy vehicle owners based on user profiles and a semi-supervised learning model, which mainly includes the following steps:

[0008] Step 1, User Profile: Collect traditional car owner information to create user profiles, and obtain profile indicators from three dimensions: basic information, car usage scenarios, and car purchase intention, which are used as features x.

[0009] Through this step, the present invention collects multi-dimensional user profiles from existing traditional car company data. The user profiles are comprehensive and can basically answer questions such as who the user is, how they use the car, and what kind of car they use.

[0010] In this step, "traditional car owners" refers to non-new energy vehicle owners. The entire method is based on the user pool of traditional car owners, predicting the probability that those traditional car owners who have not yet purchased or are unaware of purchasing new energy vehicles will purchase them in the future.

[0011] Step 2: Confirm User Tags: Match traditional car owner information and new energy vehicle owner information to obtain a binary label y indicating whether traditional car owners have purchased new energy vehicles. This label is divided into labeled data L(x, y) and labeled data U(x, nan). This step, based on the dual conditions of "whether they have already purchased" and "whether they have the intention to purchase," effectively filters out source data with training value. The user tags are clear and can distinguish different types of car buyers.

[0012] Step 3: Self-training to achieve semi-supervised learning: Input the features x and labels y into the semi-supervised learning model and train it using a self-training algorithm.

[0013] This step is to train the model because the label y can only mark those "users who have already made a purchase = 1" and "users who have no intention to make a purchase = 0". The remaining large number of users are actually unknown whether they have made a purchase or not, so y is unknown, which is the unlabeled data Ubel mentioned later.

[0014] This invention trains data using a semi-supervised learning model, which is well adapted to the actual situation of traditional car companies with small scale and poor data quality of labeled data. Furthermore, it implements semi-supervised learning in the form of a self-training algorithm. Compared with other semi-learning algorithms, the self-training algorithm has a higher utilization rate of existing labels and can fully explore the value of existing labels.

[0015] Step 4: Iterate the classifier repeatedly: Introduce an ensemble algorithm as a semi-supervised learning classifier C(x) into the self-training algorithm, combining labeled data L(x, y) and reliable unlabeled data U. bel The input model (x, y′) is iterated repeatedly.

[0016] Step 5: Determine credibility: Introduce user intention data to determine the credibility of the prediction results, remove outliers, and optimize the model.

[0017] In this step, the intent information in the user profile is introduced into the credibility judgment. On the one hand, this avoids the precision and recall problems that are easy to occur when using user profiles as white-box rules. On the other hand, it prioritizes the extraction of records with business value, corrects the classifier effect, and improves the robustness of the model. Furthermore, by considering both the intent time and intent type dimensions, it can more accurately explore the user's car purchase potential.

[0018] Step 6: Output prediction results: All data that pass the confidence level judgment U bel (x, y′) is output as the prediction result, yielding the probability y′ of each user purchasing a new energy vehicle. For the XGBoost model, it can output not only the 0-1 classification but also the prior probabilities of the 0-1 classification.

[0019] Furthermore, in step 2, the binary label for whether a traditional car owner has purchased a new energy vehicle includes:

[0020] Traditional car owners who have already purchased new energy vehicles and traditional car owners who have not purchased and have no intention of purchasing new energy vehicles are categorized into labeled data L(x, y).

[0021] Traditional car owners who have not purchased new energy vehicles, those who intend to purchase new energy vehicles, and those whose intention to purchase new energy vehicles is unknown are classified as unlabeled data U(x, nan).

[0022] Furthermore, the specific method for achieving semi-supervised learning through self-training in step 3 is as follows:

[0023] Step 3.1: Train labeled data L(x, y) to obtain a classifier C(x);

[0024] Step 3.2: Use classifier C(x) to predict the unlabeled data U(x, nan) to obtain the predicted label y′, and then use the unlabeled data U with higher confidence level. bel (x, y′) is assigned to the labeled data L(x, y) and expanded into a labeled dataset L′(x, y);

[0025] Step 3.3: After relearning the labeled dataset L′(x, y), a new classifier C′(x) is obtained and then predicted and filtered; repeat this process until the unlabeled data U(x, nan) and the predicted label y′ are assigned to the labeled data L(x, y).

[0026] The above three steps illustrate how to achieve self-training: train the model with samples of known y → use the trained model to predict samples of unknown y, obtain their y', select the more reliable samples to expand the samples of known y → train with the new samples of known y → ... repeat until stopped.

[0027] Furthermore, step 5, determining the credibility, includes:

[0028] Step 5.1, set the intention time coefficient ω time Intended vehicle model coefficient ω type As a confidence adjustment coefficient with a threshold of [0, 1], the predicted label y is reordered:

[0029]

[0030] Among them, the intention time coefficient ω time The desired vehicle model coefficient ω is obtained by normalization using the preset year as the maximum value. type The participation of motor power in new energy types decreases arithmetically; if there is missing data or unknown information in the source indicators of the two coefficients, the coefficient is set to 0.

[0031] Step 5.2, y′ and y′ bel If both criteria are met, unlabeled data U with higher reliability can be filtered out. bel (x, y′) are assigned to the labeled data L(x, y) for the next round of classifier training.

[0032] In a second aspect, the present invention provides a potential new energy vehicle owner prediction system based on user profiles and semi-supervised learning, comprising a memory, a processor, and a potential new energy vehicle owner prediction program based on user profiles and semi-supervised learning stored in the memory and executable on the processor. When the program is executed by the processor, it implements the steps of the potential new energy vehicle owner prediction method based on user profiles and semi-supervised learning as described in the first aspect of the present invention.

[0033] In a third aspect, the present invention provides a storage medium storing a potential new energy vehicle owner prediction program based on user profiles and semi-supervised learning, wherein when the program is executed by the processor, it implements the steps of the potential new energy vehicle owner prediction method based on user profiles and semi-supervised learning as described in the first aspect of the present invention.

[0034] The advantages of this invention are as follows:

[0035] 1. This invention collects multi-dimensional user profiles from existing traditional car company data. The user profiles are comprehensive and can basically answer questions such as who the user is, how they use the car, and what kind of car they use.

[0036] 2. This invention uses a dual criterion of "whether a purchase has already been made" and "whether there is an intention to purchase" to differentiate and label users, effectively filtering out source data with training value. The user labels are clear and can distinguish different types of car buyers.

[0037] 3. This invention uses a semi-supervised learning model to train data, which is fully adapted to the actual situation of traditional car companies with small scale and poor data quality of labeled data; among them, the semi-supervised form is implemented by a self-training algorithm. Compared with other semi-learning algorithms, the self-training algorithm has a higher utilization of existing labels and can fully explore the value of existing labels.

[0038] 4. This invention incorporates intent information from user profiles into credibility judgment. On the one hand, it avoids the precision-recall problem that is easily caused by using user profiles as white-box rules. On the other hand, it prioritizes the extraction of records with business value, corrects the classifier effect, and improves the robustness of the model. Furthermore, by considering both intent time and intent type dimensions, it can more accurately uncover users' car purchase potential and has practical business value. Attached Figure Description

[0039] Figure 1 : This is a logic block diagram of the method of the present invention;

[0040] Figure 2 This is a logic diagram of the traditional vehicle owner information collection and user profiling method of the present invention.

[0041] Figure 3 This is a logic diagram of the user tag verification process in this invention.

[0042] Figure 4 This is a logic diagram of the self-training algorithm used in this invention for model training;

[0043] Figure 5 This is a logic diagram of the classifier iterative process in this invention.

[0044] Figure 6 : This is the logic diagram for the credibility judgment of this invention. Detailed Implementation

[0045] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0046] See Figure 1 This embodiment details the specific steps of the new energy vehicle owner prediction method based on user profiles and semi-supervised learning models.

[0047] 1. Traditional car owner information collection and user profiling, see [link to relevant documentation]. Figure 2 :

[0048] In this step, we need to obtain the purchase information, contact information, usage information, and news information of traditional car owners within a fixed period, and statistically analyze the profile indicators of three dimensions: "basic information", "usage scenario" and "purchase intention". These indicators are stored in the database and updated at the user level. Except for the intention indicator (i.e., the indicator of "purchase intention"), all other indicators are used as features x in the model.

[0049] Among the portrait metrics:

[0050] Basic information includes the user's gender, place of birth, age, type of work, and city of residence, representing both natural person information and social information.

[0051] The vehicle usage scenario refers to the user's driving behavior, which is a statistical analysis of the total amount, frequency, consistency, short-term fluctuation range, extreme values, and other indicators of user driving big data collected within a fixed period.

[0052] Typically, clustering or actual rules are used to determine indicators based on changes in the area, mileage, and speed of vehicle use. For example, clustering is based on indicators such as "whether the user frequently uses the car within a certain period of time", "preferred driving time", and "gentle or aggressive driving". Clustering results include "daily work driving", "weekend travel driving", "short-distance gentle trips", and "long-distance aggressive trips".

[0053] Purchase intention is measured by short-term behavioral indicators derived from user app activity. These indicators, such as clicks, likes, favorites, and comments, are statistically analyzed periodically, along with conversion rates, to create a funnel-shaped metric. The main goal is to determine if users are consistently paying attention to new energy vehicle information; if so, they are considered more likely to make a purchase.

[0054] 2. User tag confirmation, such as Figure 3 As shown:

[0055] By matching traditional car owner information and new energy vehicle owner information in the database, a binary label of "whether traditional car owners have purchased new energy vehicles" is obtained, which is used as the label y entering the model.

[0056] 2.1 Traditional car owners who have already purchased new energy vehicles and those who have not purchased and have no intention of purchasing new energy vehicles belong to the "categorized" data L(x, y). Due to the low ownership rate of new energy vehicles and the difficulty in collecting intention information, this group of users will inevitably have a small scale and data bias.

[0057] 2.2 Among traditional car owners who have not purchased new energy vehicles, those who intend to purchase new energy vehicles and those whose intention to purchase are unknown belong to the "unlabeled" data U(x, nan). This type of user is the target of prediction in this method.

[0058] 3. Use a self-training algorithm for model training, such as... Figure 4 As shown:

[0059] Input the above features x and labels y into the semi-supervised learning model and proceed with the self-training algorithm. The specific method is as follows:

[0060] 3.1 Train labeled data L(x, y) to obtain a classifier C(x).

[0061] 3.2 Use classifier C(x) to predict the unlabeled data U(x, nan) to obtain the predicted label y′, and then use the unlabeled data U with higher confidence level. bel (x, y′) is incorporated into the labeled data L(x, y) and expanded into a labeled dataset L′(x, y).

[0062] 3.3 After relearning the labeled dataset L′(x, y), a new classifier C′(x) is obtained and then used for prediction and filtering. This process is repeated until the unlabeled data U(x, nan) and the predicted label y′ are assigned to the labeled data L(x, y).

[0063] 4. The classifier is iterated repeatedly, such as... Figure 5 As shown:

[0064] In this embodiment, the ensemble algorithm XGBoost is introduced into the self-training algorithm as a semi-supervised learning classifier C(x), which combines labeled data L(x, y) and reliable unlabeled data U. bel The input model (x, y′) is iterated repeatedly.

[0065] 4.1 The labeled data L(x, y) is input into the model for training. The tree structure score is calculated based on the second-order gradient of the left and right sub-leaf. The model is stabilized when the score decreases to a preset threshold. The model is then used to predict the total score of the unlabeled data U(x, nan). After activation by the sigmoid function, the predicted label y′ is obtained, which represents the probability of a user purchasing a new energy vehicle.

[0066] 4.2 Select a portion of unlabeled data with high reliability U bel (x, y′) are assigned to labeled data L(x, y) and input into the model for repeated training.

[0067] 5. Credibility assessment, such as Figure 6 As shown:

[0068] The self-training algorithm requires extracting highly reliable unlabeled data U. bel (x, y′) Retrain the classifier C(x). This invention proposes to design a credibility judgment algorithm by "introducing prior data to judge the credibility of prediction results". By introducing user intention data, the model is optimized and credibility judgment is performed.

[0069] 5.1 Set the intention time coefficient ω time Intended vehicle model coefficient ω type As a confidence adjustment coefficient with a threshold of [0, 1], the predicted label y is reordered.

[0070]

[0071] Among them, the intention time coefficient ωtime The desired model exhibits a non-linear decreasing trend, and is obtained by normalization using a preset year as the maximum value; the intended model coefficient ω type The motor power participation decreases arithmetically according to the type of electric vehicle; if there is missing data or unknown information in the source indicators of the two coefficients, the coefficient is set to 0.

[0072] 5.2 The model output results and the ranking results after coefficient adjustment tend to be homogeneous, which is considered a high degree of reliability. y′ and y′ bel If any one of the judgment conditions is met, unlabeled data U with a high degree of reliability can be filtered out. bel (x, y′) are assigned to the labeled data L(x, y) for the next round of classifier training.

[0073] Condition 1: The actual value of the predicted label y′ is greater than or equal to a threshold of 0.5 and the confidence level of the predicted label y′ is also greater than or equal to a threshold of 0.5. bel Within the top 25% of the ascending order.

[0074] Condition 2: The actual value of the predicted label y′ is less than the 0.5 threshold and the confidence level of the predicted label y′ is also higher. bel Within the next 25% of the orthogonal sequence.

[0075] 5.3 According to the statistical threshold, if the number of self-training iterations exceeds X% of the size of the unlabeled data U(x, nan), the unlabeled data U still cannot pass the screening. unbel (x, y′) are considered outliers that the model cannot predict and are therefore removed.

[0076] Here, X% is set to 75%. This percentage is variable. The 75% setting here is based on statistical experience and can be modified later according to business requirements.

[0077] 6. Output prediction results

[0078] The semi-supervised learning model terminates after outlier users are removed in step 5.3. All the aforementioned data U determined through credibility criteria... bel (x, y′) is the output of the prediction result. Based on the probability y′ of each user purchasing a new energy vehicle, precise marketing can be carried out, thereby realizing the entire process of new energy vehicle owner prediction.

[0079] Table 1: Explanation of English Abbreviations

[0080]

[0081]

[0082]

[0083] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various simple modifications can be made to the technical solution of the present invention, and these simple modifications all fall within the protection scope of the present invention.

[0084] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

Claims

1. A method for predicting potential new energy vehicle owners based on user profiles and semi-supervised learning, characterized in that, Including the following steps: Step 1, User Profiling: Collect traditional car owner information to create user profiles, obtaining profile indicators from three dimensions: basic information, car usage scenarios, and purchase intention, which serve as features. ; Step 2: Confirm User Tags: Match traditional car owner information with new energy vehicle owner information to obtain a binary tag indicating whether traditional car owners have purchased new energy vehicles, which will be used as the tag. Data is divided into labeled data. and unlabeled data ; Step 3: Self-training to achieve semi-supervised learning: Features... and tags Input a semi-supervised learning model and train it using a self-training algorithm; Step 4: Repeatedly iterate the classifier: Introduce the ensemble algorithm XGBoost as a semi-supervised learning classifier into the self-training algorithm. There will be tagged data and trustworthy unlabeled data The input model is iterated repeatedly; Step 5: Determine Credibility: Introduce user intent data to determine the credibility of the prediction results, remove outliers, and optimize the model; including: Step 5.1, Set the intention time coefficient Intended vehicle model coefficient As a confidence adjustment coefficient with a threshold of [0,1], it is used for the predicted label. Implement reordering: Among them, the intention time coefficient The desired vehicle model coefficient is obtained by normalization using the preset year as the maximum value. The participation of motor power in new energy types decreases arithmetically; if there is missing or unknown information in the source indicators of the two coefficients, the coefficient is set to 0. Step 5.2, and If either condition 1 or condition 2 is met simultaneously, then unlabeled data with a high degree of reliability can be filtered out. categorized into labeled data In the next round of classifier training, the judgment conditions include: Judgment condition 1: Predicted label Actual value >= 0.5 threshold and predicted label confidence Within the top 25% of the positive order; Condition 2: Predicted label Actual value < 0.5 threshold and predicted label confidence Within the next 25% of the ascending order; Step 6: Output prediction results: All unlabeled data with high confidence levels as determined by the confidence level assessment. As the prediction result output, the probability of each user purchasing a new energy vehicle is obtained. .

2. The method for predicting potential new energy vehicle owners based on user profiles and semi-supervised learning according to claim 1, characterized in that, The traditional car owner information includes purchase information, contact information, vehicle usage information, and news information of traditional car owners within a fixed period; the traditional car owner refers to non-new energy vehicle owners.

3. The method for predicting potential new energy vehicle owners based on user profiles and semi-supervised learning according to claim 1 or 2, characterized in that, Among the portrait metrics Basic information includes the user's gender, place of birth, age, type of work, and city of residence, representing both natural person information and social information; The vehicle usage scenario represents the user's driving behavior, which is a statistical analysis of the total amount, frequency, consistency, short-term fluctuation range, extreme values, and other indicators of user driving big data collected within a fixed period. Car purchase intention is a short-term behavioral indicator obtained from user app behavior. It is calculated periodically based on the frequency and conversion rate of clicks, likes, favorites, comments, etc., forming a funnel indicator.

4. The method for predicting potential new energy vehicle owners based on user profiles and semi-supervised learning according to claim 1, characterized in that, In step 2, the binary label for whether a traditional car owner has purchased a new energy vehicle includes: Traditional car owners who have already purchased new energy vehicles and those who have not purchased and have no intention of purchasing new energy vehicles are categorized into those with tagged data. ; Traditional car owners who have not yet purchased new energy vehicles, those who intend to purchase new energy vehicles, and those whose intention to purchase is unknown are categorized as unlabeled data. .

5. The method for predicting potential new energy vehicle owners based on user profiles and semi-supervised learning according to claim 1, characterized in that, The specific method for model training in step 3 is as follows: Step 3.1, training with labeled data To obtain a classifier ; Step 3.2, using a classifier For unlabeled data Perform predictions to obtain predicted labels. Unlabeled data with high credibility Included in labeled data Expanded into a labeled dataset ; Step 3.3, Relearn the labeled dataset Then, a new classifier is obtained. Then perform prediction and filtering; repeat this process until the unlabeled data is removed. and prediction tags Included in labeled data until.

6. The method for predicting potential new energy vehicle owners based on user profiles and semi-supervised learning according to claim 1, characterized in that, The repeated iterations of the classifier in step 4 include: Step 4.1 will have tag data. The semi-supervised learning model is trained by inputting data, and the tree structure score is calculated based on the second-order gradients of the left and right sub-leaf leaves. A stable model is output when the score decreases to a preset threshold. The stable model is then used to predict unlabeled data. The total score, after activation by the sigmoid function, yields the predicted label. This represents the probability that a user will purchase a new energy vehicle. Step 4.2: Select a portion of unlabeled data with high reliability. Included in labeled data The input model is then repeatedly trained.

7. The method for predicting potential new energy vehicle owners based on user profiles and semi-supervised learning according to claim 1, characterized in that, This also includes step 5.3, outlier removal: based on a statistical threshold, outliers are removed when the number of self-training iterations exceeds the number of unlabeled data. Even at X% of the scale, unlabeled data still cannot pass the filter. Users deemed outliers that the model cannot predict are removed; X% is set based on statistical experience.

8. A potential new energy vehicle owner prediction system based on user profiling and semi-supervised learning, characterized in that, The method includes a memory, a processor, and a program for predicting potential new energy vehicle owners based on user profiles and semi-supervised learning, which is stored in the memory and can run on the processor. When the program is executed by the processor, it implements the steps of the method for predicting potential new energy vehicle owners based on user profiles and semi-supervised learning as described in any one of claims 1 to 7.

9. A storage medium, characterized in that, The storage medium stores a potential new energy vehicle owner prediction program based on user profiles and semi-supervised learning. When the program is executed by a processor, it implements the steps of the potential new energy vehicle owner prediction method based on user profiles and semi-supervised learning as described in any one of claims 1 to 7.

Images