Automobile product recommendation method and device, equipment, medium and program product

Abstract

The invention provides an automobile product recommendation method, device and equipment, a medium and a program product, relates to the field of artificial intelligence, and can improve prediction accuracy and recommendation effectiveness in a scene with relatively sparse information related to automobile purchase. The automobile product recommendation method comprises the steps of obtaining data of a first user in a bank system under the condition that authorization of the first user is obtained to obtain first initial data, then screening out M pieces of feature data most related to automobile purchase willingness of the user from the first initial data to obtain a first feature set, utilizing the first feature set to construct a first prompt instruction represented by a natural language; the first prompt instruction is input to a user car purchase intention recognition model, a first car purchase intention recognition result output by the user car purchase intention recognition model is obtained, and the user car purchase intention recognition model is obtained by fine tuning training of a pre-trained large language model; and when the first car purchase willingness identification result meets the potential car purchase condition, recommending a car product to the first user.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and can be used in the field of financial technology or other fields. More specifically, it relates to a method, device, equipment, medium and program product for recommending automobile products. Background Art

[0002] With the popularization of automobile consumption and the improvement of economic level, automobile installment products have become the choice of more and more consumers when buying cars. At the same time, with the development of artificial intelligence, the prediction and recommendation of automobile products through machine learning models have also been adopted by more and more companies.

[0003] In the process of realizing the concept of the present invention, the inventors found that the related art has the following defects when predicting automobile products through machine learning models: the related art usually uses user car purchase behavior data (such as users browsing car models, asking prices, etc.) to train models to predict the automobile products that users may purchase, and cannot effectively predict users who have not yet shown obvious car purchase behavior or lack relevant behavior data. In reality, the number of users who have not yet purchased a car or have not shown obvious car purchase behavior or lack relevant behavior data is much larger than the amount of data of users who have purchased a car or have car purchase behavior. In other words, when users who have not yet purchased a car or have not shown obvious car purchase behavior or lack relevant behavior data are also included in the prediction range, the information related to car purchase that the model can learn will become very sparse, resulting in inaccurate model predictions. Summary of the invention

[0004] In view of the above problems, the present invention provides a method and device for recommending automobile products that can effectively recommend automobile products in scenarios where information related to purchasing automobiles is relatively sparse, a method and device for training a user car purchasing intention recognition model based on a large language model, and an electronic device, a medium and a program product.

[0005] According to a first aspect of an embodiment of the present invention, a method for recommending automobile products is provided. The method includes: obtaining authorization from a first user to use the data in a bank system; obtaining the data of the first user in the bank system after obtaining the authorization from the first user to obtain first initial data; filtering out M feature data corresponding to M features from the first initial data to obtain a first feature set, wherein the M features are a set of features most relevant to the user's car purchase intention filtered out by a feature engineering algorithm, wherein M is an integer greater than or equal to 2; constructing a first prompt instruction expressed in natural language using the first feature set; inputting the first prompt instruction into a user's car purchase intention recognition model, and obtaining a first car purchase intention recognition result output by the user's car purchase intention recognition model, wherein the user's car purchase intention recognition model is obtained after fine-tuning a pre-trained large language model; and recommending automobile products to the first user when the first car purchase intention recognition result meets the potential car purchase condition.

[0006] The second aspect of the embodiments of the present invention provides a method for training a user car purchase intention recognition model based on a large language model. The method includes: obtaining the authorization of a second user to use the data in his bank system, wherein the second user is a user who has purchased a car installment product in the bank system or a user who has not purchased a car but has not used any car-related products in the bank system; after obtaining the authorization of the second user, obtaining the data of the second user in the bank system to obtain second initial data; filtering out M feature data corresponding to M features from the second initial data to obtain a second feature set, wherein the M features are a set of features that are most relevant to the user's car purchase intention selected by a feature engineering algorithm, wherein M is an integer greater than or equal to 2; when the When the second user is the user who has purchased the car, the second feature set is labeled with the information of the purchased car as a label to obtain a positive sample data; and when the second user is the user who has not purchased the car, the second feature set is labeled with the information of the non-purchased car as a label to obtain a negative sample data; after obtaining equal amounts of the positive sample data and the negative sample data, for each of the positive sample data and each of the negative sample data, a prompt instruction expressed in natural language is constructed to obtain an instruction fine-tuning data set; and the large language model is fine-tuned and trained using the instruction fine-tuning data set to obtain the user car purchase intention recognition model.

[0007] According to a third aspect of an embodiment of the present invention, a device for recommending automobile products is provided. The device includes: a first acquisition module, a feature engineering and feature analysis module, a car purchase intention prediction module, and an automobile product recommendation module. The first acquisition module is used to obtain the authorization of the first user to use the data in the bank system, and obtain the data of the first user in the bank system after obtaining the authorization of the first user, so as to obtain the first initial data. The feature engineering and feature analysis module is used to filter out M feature data corresponding to M features from the first initial data to obtain a first feature set, wherein the M features are a set of features most relevant to the user's car purchase intention selected by the feature engineering algorithm, wherein M is an integer greater than or equal to 2. The car purchase intention prediction module is used to construct a first prompt instruction expressed in natural language using the first feature set, and input the first prompt instruction to a user's car purchase intention recognition model, and obtain a first car purchase intention recognition result output by the user's car purchase intention recognition model, wherein the user's car purchase intention recognition model is obtained after fine-tuning the pre-trained large language model. The automobile product recommendation module is used to recommend automobile products to the first user when the first car purchase intention recognition result meets the potential car purchase condition.

[0008] In a fourth aspect of an embodiment of the present invention, a training device for a user car purchase intention recognition model based on a large language model is provided. The training device includes a second acquisition module, a feature engineering and feature analysis module, and a fine-tuning training module. The second acquisition module is used to obtain the authorization of the second user to use the data in his bank system, and after obtaining the authorization of the second user, obtain the data of the second user in the bank system to obtain the second initial data, wherein the second user is a user who has purchased a car installment product in the bank system or a user who has not purchased a car but has not used any car-related products in the bank system. The feature engineering and feature analysis module is used to: filter out M feature data corresponding to M features from the second initial data to obtain a second feature set, wherein the M features are a set of features most relevant to the user's car purchase intention filtered out by a feature engineering algorithm, wherein M is an integer greater than or equal to 2; when the second user is the user who has purchased a car, the second feature set is labeled with the information of the purchased car as a label to obtain a positive sample data; and when the second user is the user who has not purchased a car, the second feature set is labeled with the information of the non-purchased car as a label to obtain a negative sample data. The fine-tuning training module is used to construct prompt instructions expressed in natural language for each of the positive sample data and each of the negative sample data after obtaining equal amounts of the positive sample data and the negative sample data, so as to obtain an instruction fine-tuning data set, and to use the instruction fine-tuning data set to fine-tune the large language model to obtain the user car purchase intention recognition model.

[0009] According to a fifth aspect of the embodiments of the present invention, an electronic device is provided. The electronic device includes: one or more processors and a memory. The memory is used to store one or more computer programs. The one or more processors execute the one or more computer programs to implement the steps of the method provided in the first aspect or the second aspect.

[0010] A sixth aspect of an embodiment of the present invention is a computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions, when executed by a processor, implement the steps of the method provided in the first aspect or the second aspect above.

[0011] According to a seventh aspect of the embodiments of the present invention, a computer program product includes a computer program or instructions, which, when executed by a processor, implement the steps of the method provided in the first or second aspect above.

[0012] One or more of the above embodiments have the following advantages or beneficial effects: important features are first screened out from the initial data, and then potential car purchasing users are identified using a user car purchasing intention recognition model built based on a large language model. With the help of feature extraction of feature engineering algorithms and the general understanding ability of large language models, the problem of sparse car purchasing related data caused by the number of users who have not purchased a car or have not yet shown obvious car purchasing behavior or lack relevant behavior data is effectively solved, which is much larger than the amount of data of users who have purchased a car or have car purchasing behavior. Therefore, the understanding and prediction capabilities of users' car purchasing intentions are improved, and the pertinence of automobile product recommendations is improved. In addition, the embodiments of the present invention recommend to users when the results of user car purchasing intention recognition meet potential car purchasing conditions, thereby expanding the range of predicted populations while reducing unnecessary disturbances to users, thereby improving user experience and purchase conversion rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The above contents and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings, in which:

[0014] Figure 1 Schematically illustrates the conceptual framework of a method and device for recommending automobile products according to an embodiment of the present invention;

[0015] Figure 2 The application scenarios of the method, device, equipment, medium and program product for recommending automobile products according to the embodiments of the present invention are schematically shown;

[0016] Figure 3 A flowchart schematically shows a method for recommending automobile products according to an embodiment of the present invention;

[0017] Figure 4A flowchart schematically shows a method for training a user car purchase intention recognition model according to an embodiment of the present invention, wherein the user car purchase intention recognition model is obtained by fine-tuning a pre-trained large language model;

[0018] Figure 5 The overall flow chart of each model in the training phase in one embodiment of the present invention is schematically shown;

[0019] Figure 6 A flowchart schematically illustrating a method for recommending an automobile product according to another embodiment of the present invention;

[0020] Figure 7 A block diagram schematically shows a device for recommending automobile products according to an embodiment of the present invention;

[0021] Figure 8 A block diagram schematically shows a training device for a user car purchase intention recognition model according to an embodiment of the present invention; and

[0022] Fig. 9 A block diagram of an electronic device according to an embodiment of the present invention is schematically shown. DETAILED DESCRIPTION

[0023] Below, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of concepts of the present invention.

[0024] In the technical solution of the present invention, the user information involved (including but not limited to the user's data in the banking system, user personal information, contract records, user image information, user device information, such as location information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) are all information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with relevant laws, regulations and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0025] In the scenario of using personal information for automated decision-making, the methods, devices, and systems provided by the embodiments of the present invention provide users with corresponding operation portals for users to choose to agree or reject the automated decision-making results; if the user chooses to reject, the expert decision-making process will be entered. The expression "automated decision-making" here refers to the activity of automatically analyzing and evaluating an individual's behavioral habits, interests and hobbies, or economic, health, credit status, etc. through computer programs, and making decisions. The expression "expert decision-making" here refers to the activity of making decisions by people who specialize in a certain field, have specialized experience, knowledge and skills, and have reached a certain level of professionalism.

[0026] The embodiments of the present invention provide a method, device, equipment, medium and program product for recommending automobile products. When obtaining the user's authorization to use the user's data in the bank system, the user's car purchase intention can be analyzed based on the user's data in the bank system, and when it is determined that the user has a high car purchase intention, automobile products can be recommended to the user in a targeted manner. Figure 1 The conceptual framework of the method and device for recommending automobile products according to an embodiment of the present invention is schematically shown.

[0027] refer to Figure 1 Considering that the user's data in the bank system may lack data related to car purchasing behavior (such as browsing car models, asking prices), and there may also be feature redundancy, which interferes with the accuracy of the prediction, the embodiment of the present invention will first filter out feature data with a high correlation with the car purchasing intention from the user's data obtained in the bank system through feature engineering, thereby reducing the interference of factors such as feature redundancy on the prediction results when directly using the user's data in the bank system.

[0028] Furthermore, considering that the information related to car purchases by users in the banking system mainly comes from the contract records of automobile installment products, the information related to car purchases by users in the banking system is significantly smaller than the scale and data volume of car purchase users in the entire market, resulting in the problem of sparse data volume that can be learned by the model. In this embodiment of the present invention, a user car purchase intention recognition model obtained by fine-tuning the pre-trained large language model is used to identify the user's car purchase intention. In this way, the large language model can be used to convert prompt instructions expressed in natural language into high-dimensional semantic representations, further improve the expression ability of features, capture the context and emotional information therein, thereby improving the understanding and prediction ability of users' car purchase intentions, and more accurately mining potential car purchase users.

[0029] It can be seen that after screening out important features, the embodiments of the present invention can effectively solve the problems of feature extraction and combination, nonlinear relationship modeling, and sparse car purchase-related data in the banking system when identifying potential car purchasing customer groups using the user car purchase intention recognition model built based on the large language model, thereby improving the understanding and prediction capabilities of users' car purchase intentions, thereby improving the pertinence of automobile product recommendations, and improving user experience and purchase conversion rate.

[0030] Figure 2 The application scenarios of the method, apparatus, device, medium and program product for recommending automobile products according to the embodiments of the present invention are schematically illustrated.

[0031] like Figure 2 As shown, the application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 is used to provide a medium for a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired, wireless communication links, or optical fiber cables, etc.

[0032] The user can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 through the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as a mobile banking client.

[0033] The server 105 may be a server that provides various services, such as a background management server that provides support for websites browsed by users using the first terminal device 101 , the second terminal device 102 , and the third terminal device 103 (only as an example).

[0034] It should be noted that the method for recommending automobile products and / or the method for training a user's car purchase intention recognition model provided in the embodiments of the present invention can generally be executed by the server 105. Accordingly, the device for recommending automobile products and / or the device for training a user's car purchase intention recognition model provided in the embodiments of the present invention can generally be set in the server 105.

[0035] It should be understood that Figure 2 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0036] The following will be based on Figure 2 The scenario described is used to describe the automotive product method and device of the embodiments of the present invention in detail.

[0037] Figure 3 The flowchart of the method for recommending automobile products according to an embodiment of the present invention is schematically shown.

[0038] like Figure 3 As shown, the method for recommending automobile products in this embodiment may include operations S301 to S306 , and the method may be executed by the server 105 .

[0039] In operation S301 , the first user's authorization to use the data in the banking system is obtained. For example, the server 105 requests the user's authorization by interacting with the first terminal device 101 , the second terminal device 102 , or the third terminal device 103 .

[0040] In operation S302, after obtaining authorization from the first user, the data of the first user in the bank system is obtained to obtain first initial data, wherein the data of the user in the bank system may be, for example, basic information of the user, transaction flow and other data.

[0041] In operation S303, M feature data corresponding to M features are screened out from the first initial data to obtain a first feature set. The M features are a set of features most relevant to the user's car purchase intention screened out by a feature engineering algorithm, where M is an integer greater than or equal to 2.

[0042] The process of selecting M features by the feature engineering algorithm is completed before operation S301. In one embodiment, through the feature engineering algorithm, after obtaining the corresponding user authorization, based on the data of a large number of users accumulated in the bank system, combined with information such as whether the user purchases auto installment products, a corresponding model can be established to select important features related to the willingness to purchase auto installment products.

[0043] Methods for feature engineering algorithm to screen features include but are not limited to: feature importance based on tree model, feature importance based on linear model, feature importance based on feature selection, feature importance based on neural network and Shapley Additive exPlanations (SHAP) value, etc.

[0044] In one embodiment, the feature engineering algorithm is specifically constructed using a random forest model.

[0045] Accordingly, the screening process of M features can be as follows. First, after obtaining the user's authorization to use his data in the bank system, train the random forest model to predict the user's willingness to buy a car based on the user's data in the bank system. For example, use an equal number of training sets composed of data in the bank system consisting of users who have purchased a car (users who have purchased car installment products) and users who have not purchased a car (users who have not purchased car installment products) to train the random forest model, wherein users who have purchased a car and users who have not purchased a car are respectively labeled to train the random forest model to predict the probability of the user buying a car. Then, after the random forest model is trained, M features are screened out based on the importance of the input features of the random forest model and the Shapley additive interpretation SHAP value. In this way, a series of user features related to the user's willingness to buy a car are screened out through the feature importance and SHAP value in the random forest model, thereby improving the feature pertinence of the subsequent large language model processing and reducing the interference of redundant features on the large language model analysis.

[0046] Furthermore, in another embodiment, the feature engineering algorithm can be implemented by multiple random forest models, each random forest model learning a car purchase attribute (such as price, category, or energy type).

[0047] Accordingly, the screening process of M features can be as follows. First, obtain authorization from at least one user who has purchased a car to use the data in his bank system, where the user who has purchased a car is a user who has purchased a car installment product in the bank system. Then, after obtaining the authorization of the user who has purchased a car, obtain the data of the user who has purchased a car in the bank system, and obtain n labels of the user who has purchased a car based on the category information of n attributes of the car in the car installment product purchased by the user who has purchased a car, where n is an integer greater than or equal to 2. Next, train n random forest models, each of which is used to predict one of the n attributes, wherein when training each random forest model, the user who has purchased a car is labeled with a label corresponding to the attribute to be predicted by the random forest model, and the data of the user who has purchased a car in the bank system is used as input to train the random forest model. Finally, after the training of the n random forest models is completed, M features are screened out based on the importance of the input features and the Shapley additive explanation SHAP value of the n random forest models.

[0048] The advantage of using multiple random forest models to screen features, where each random forest model learns a car purchase attribute, is that, on the one hand, only the data of car-purchasing users in the banking system can be used to construct the random forest model, thus avoiding the problem of incomplete and inaccurate information when using the data of non-car-purchasing users in the banking system, because there are users who purchase cars through other channels instead of signing up for bank car installment products; on the other hand, since the data in the banking system is mainly the financial data of users who purchase installment products, it does not come from car purchase behavior data, and there is a problem of data sparsity, modeling and feature screening for different attributes of automobile products can improve the targetedness of modeling, enhance the accuracy and comprehensiveness of feature screening, increase the number and dimension of model learning, and alleviate or even overcome the defect of data sparsity to a certain extent.

[0049] Next, in operation S304, a first prompt instruction expressed in natural language is constructed using the first feature set. The prompt instruction can be constructed using a specially designed instruction framework. For example, the following prompt instruction can be constructed:

[0050] "Please assess the likelihood of this individual purchasing a car in the near future, presented on a scale of 0-1 and briefly explain the reason of this score: Gender: female; Age: 37; Education: Bachelor's degree..."

[0051] Among them, Gender, Education, Bachelor's degree, etc. are the M selected features, and the order of these features in the prompt instruction can be random.

[0052] Then, in operation S305, the first prompt instruction is input to the user car purchase intention recognition model, and a first car purchase intention recognition result output by the user car purchase intention recognition model is obtained, wherein the user car purchase intention recognition model is obtained after fine-tuning the pre-trained large language model. The first car purchase intention recognition result can be represented by, but not limited to, a score value, a probability value, a classification result, etc.

[0053] Finally, in operation S306, when the first car purchase intention recognition result meets the potential car purchase condition, the car product is recommended to the first user. The potential car purchase condition may be, for example, that the score or probability value in the car purchase intention recognition result exceeds a threshold, or is ranked as top N in a large number of user groups, or is classified into a high car purchase intention category, etc.

[0054] In one embodiment, accurate recommendations can be made by constructing an automobile product recommendation model. Specifically, the first initial data (which may also be the first feature set in some embodiments) can be input into the automobile product recommendation model, and the recommended category information of the automobile product output by the automobile product recommendation model can be obtained, and then the automobile product indicated by the recommended category information can be recommended to the first user. The recommended category information can, for example, include category information of multiple attributes (such as price, energy type, category, etc.). In this way, for users who meet the potential car purchase conditions, the types of cars that are more likely to be purchased can be specifically recommended, thereby improving the pertinence and conversion rate of automobile product recommendations and enhancing user experience. It also improves the effectiveness of the marketing and promotion of the bank's automobile installment products.

[0055] The automobile product recommendation method of the embodiment of the present invention can, when making automobile product recommendations, make use of the feature extraction of the feature engineering algorithm and the general understanding ability of the large language model to solve the technical problems such as feature extraction and combination, nonlinear relationship modeling, and data sparsity faced in the process of identifying potential car purchasing customers based on data in the banking system, thereby improving the understanding and prediction capabilities of users' car purchasing intentions, thereby improving user experience and purchase conversion rates.

[0056] The user car purchase intention recognition model used in the above operation S305 is obtained by fine-tuning the pre-trained large language model in the training phase before operation S301. Figure 4 The flowchart schematically shows a method for training a user car purchase intention recognition model according to an embodiment of the present invention.

[0057] refer to Figure 4 According to this embodiment, the process of fine-tuning the large language model to obtain a user car purchase intention recognition model may include operations S401 to S406.

[0058] First, in operation S401, the authorization of the second user to use the data in the bank system is obtained. The second user is a user who has purchased an automobile installment product in the bank system or a user who has not purchased a car but has not used any automobile-related products in the bank system.

[0059] When selecting users who have not purchased a car, we try to exclude as many features related to car purchase as possible. For example, not only are users required not to have purchased car installment products in the bank system, but users who have participated in car maintenance and car refueling transactions are also excluded through the bank transaction process, thereby ensuring that the selection of users who have not purchased a car is as objective and accurate as possible.

[0060] In operation S402, after obtaining authorization from the second user, the data of the second user in the bank system is acquired to obtain second initial data.

[0061] In operation S403, M feature data corresponding to the M features are screened out from the second initial data to obtain a second feature set, wherein the M features are a set of features most relevant to the user's car purchase intention screened out by the above-mentioned feature engineering algorithm.

[0062] In operation S404, when the second user is a user who has purchased a car, the second feature set is labeled with the information of the purchased car as a label to obtain a positive sample data; and when the second user is a user who has not purchased a car, the second feature set is labeled with the information of the non-purchased car as a label to obtain a negative sample data.

[0063] In operation S405, after obtaining equal amounts of positive sample data and negative sample data, a prompt instruction expressed in natural language is constructed for each positive sample data and each negative sample data to obtain an instruction fine-tuning data set. The instruction framework for constructing fine-tuning instructions is basically the same when fine-tuning the large language model and when using the user car purchase intention recognition model in the application stage.

[0064] In operation S406, the large language model is fine-tuned and trained using the instruction fine-tuning data set to obtain a user car purchase intention recognition model.

[0065] The embodiment of the present invention can fine-tune the pre-trained large language model to enable the large language model to fully learn the relationship between each feature and the car purchase intention from the important features related to the car purchase intention of users who have purchased a car and users who have not purchased a car. Specifically, it can be manifested in three aspects: effective extraction and combination of semantic features, learning and modeling of non-linear relationships between features and car purchase intention, and effective response to data sparsity:

[0066] (1) Make full use of the general understanding ability of the large language model to effectively extract and combine semantic features. The large language model can convert the text information expressed in natural language in the instruction fine-tuning dataset into a high-dimensional semantic representation, capturing the context and emotional information. This representation can be used as a rich feature to predict car purchase intention. In addition, the large language model can also automatically learn the correlation and interaction between features, further improving the expressiveness of features.

[0067] (2) Learning and modeling the nonlinear relationship between features and car purchase intention: The formation of car purchase intention is affected by many factors, among which complex nonlinear relationships are often difficult to accurately model using traditional linear models. Large language models can extract high-level abstract representations of data layer by layer through deep neural network structures, thereby better understanding the complex relationships in the data; through nonlinear transformations of activation functions and weight matrices, large language models can learn nonlinear patterns in the data, thereby improving the accuracy of predicting car purchase intention.

[0068] (3) Ability to make accurate predictions in scenarios with data sparsity: As mentioned above, when predicting car purchase intention based on data in the banking system, the input data is relatively sparse, including user behavior data, preference information, and other aspects of the data are either insufficient or missing. Traditional models may perform poorly when processing sparse data, and large language models have certain advantages in this regard. By fine-tuning the pre-trained large language model and training on large-scale data, the large language model can learn the distribution characteristics and potential structure in the data, improve the modeling ability of sparse data, and thus better utilize sparse data to explore the deep relationship between features and car purchase intention. In addition, the large language model can also focus on important features through technologies such as attention mechanisms to further improve the accuracy of predictions.

[0069] It can be seen that the embodiment of the present invention obtains a user car purchase intention recognition model by fine-tuning the pre-trained large language model, which can effectively solve technical problems such as semantic feature extraction and combination, nonlinear relationship modeling, and data sparsity processing, and improve the understanding and prediction capabilities of user car purchase intentions based on data in the bank system. Therefore, the car purchase behavior of users who have not yet shown obvious car purchase behavior or lack relevant data can be effectively mined, so that the prediction of automobile products is not limited to the user group that can obtain relevant car purchase behavior data, expanding the range of predicted people who purchase automobile products and improving the purchase conversion rate of automobile products.

[0070] The embodiments of the present invention can effectively overcome the problems that when using the data of the banking system to predict automobile products, a lot of data does not come from car purchasing behavior (such as browsing car models, inquiring prices), and the information related to user car purchases in the banking system is relatively sparse (such as the proportion of users who purchase automobile installment products is small, the information is not comprehensive, etc.), and thus can effectively improve the mining of potential car purchasing behaviors of users who have not yet purchased a car or have not shown obvious car purchasing behaviors or lack relevant behavior data, and improve the ability to identify and recommend potential car purchasing users.

[0071] The following is passed Figure 5 and Figure 6 The overall process framework of the model training phase and the application phase of the automobile product recommendation method according to an embodiment of the present invention is described in detail. It should be understood that the following description is only exemplary and does not constitute a limitation of the present invention.

[0072] Figure 5 The overall flow chart of each model in the training stage in one embodiment of the present invention is schematically shown, and the overall calculation framework of the training stage is described.

[0073] In this embodiment, the models that need to be constructed in the training phase include: a random forest model, a user car purchase intention recognition model, and an automobile product recommendation model. Among them, the training random forest model is used to construct feature engineering and feature analysis modules, the purpose of which is to filter out the M features most relevant to the user's car purchase intention from the user's initial data in the bank system according to the random forest model. The user car purchase intention recognition model is obtained by fine-tuning the pre-trained large language model and is used to identify the user's car purchase intention. The automobile product recommendation model is used to predict the type information of automobile products that the user may purchase.

[0074] In this embodiment, the feature engineering and feature analysis modules are first used to obtain the most important features related to the user's car purchase intention from the input existing user car analysis records, and then these high-importance features are used to construct prompt instructions to obtain a fine-tuning data set, and then these fine-tuning data sets are used to fine-tune the pre-trained large language model to obtain a user car purchase intention recognition model. Finally, the relevant data in the input existing user car analysis records are used to train the car product recommendation model.

[0075] First, construct the feature engineering and feature analysis modules through the following steps (1) and (2).

[0076] Step (1) is to obtain the user's authorization to use his / her data and obtain the contract record of the user who has signed the auto installment product.

[0077] Assume that the series of records obtained are In the form of Represents the user, Represents the car purchased by the user. By associating the basic personal information registered by each user in the bank system and the transaction-related features generated by the transaction, a complete car installment data set is obtained: ,in is the feature of the jth user, including basic user features, financial features, and non-financial features. There can be three attributes of the feature: price, energy type and category.

[0078] In step (2), a random forest model is constructed based on each of the above three attributes, thus constructing a total of three random forest models. The three random forest models all use all the features contained in the complete car installment data set compiled in step (1) as input, and use the price range, energy type and car category purchased by the user as prediction targets, respectively, and use the complete car installment data set in step (1) for supervised training.

[0079] In one embodiment, the values ​​of the car price range, energy type and car category can be set as follows. The car price range can be divided into 11 categories, including "0-50,000", "50,000-80,000", "80,000-120,000", "120,000 to 150,000", "150,000-180,000", "180,000-250,000", "250,000-350,000", "350,000-500,000", "500,000-800,000", "800,000-1 million" and "more than 1 million". The car energy type can be divided into 2 categories, including "traditional energy car" and "new energy car". The car category can be divided into 6 categories, including "mid-size car", "compact car", "compact SUV", "mid-size SUV", "medium and large car" and "others".

[0080] After the three random forest models were trained, features were screened based on the importance of the input features and the Shapley additivity interpretation SHAP value of each random forest model.

[0081] Among them, the feature importance can be calculated based on the average reduction impurity method. The average reduction impurity calculation formula is:

[0082]

[0083] in The characteristics that need to be measured include: basic characteristics such as user age and gender, financial characteristics such as transaction flow, and non-financial characteristics such as user education and marital status.

[0084] is the number of trees in the random forest model, It is Gini impurity of a node before splitting the tree. It is in Features in a tree The weighted sum of the Gini impurities of the two child nodes after the split (the weight is the proportion of the child nodes). Therefore, Indicates the reduction in Gini impurity after the node split. The higher the sum of the Gini impurity reduction values ​​of all nodes, the greater the influence of this feature on the decision.

[0085] The SHAP value is calculated as follows: First, a portion of samples is randomly selected from the complete automobile installment data set for SHAP value calculation (for example, 10,000 samples are used). Then, for each sample, each feature in the decision tree is traversed, and the contribution of each feature to the prediction of a specific sample is evaluated by considering all possible feature subsets. Finally, the path of the feature in the decision tree is calculated, that is, the path from the root node to the leaf node, which determines how the feature affects the final prediction. After the calculation of each sample is completed, the SHAP value of each tree in the random forest is averaged to obtain the overall SHAP value.

[0086] Based on the feature importance and the absolute value of SHAP value, the most relevant factors for car purchase intention are selected. features (e.g. ), as the input features for fine-tuning the instructions of the large language model. For example, the weighted values ​​of feature importance and SHAP value can be set as the basis for screening. Among them, when screening from the three random forest models, it can be that the same number of features are screened out according to each random forest model, and then summarized and deduplicated; or the decision trees in the three random forest models can be combined to screen out M features. The specific screening method is not limited in the present invention.

[0087] Next, we build a user car purchase intention recognition model based on the large language model through the following steps (3) to (6).

[0088] In step (3), after obtaining the user's authorization to use their data, randomly select the same number of users who have not signed up for the auto installment product (users who have not purchased a car) as the number of users who have signed up for the auto installment product in step (1) from among the users in the banking system. When selecting users who have not signed up for the auto installment product (users who have not purchased a car), users with other characteristics related to car purchase (such as car maintenance consumption and car refueling consumption) are excluded as much as possible, so as to ensure that the selected users of the auto installment product do not include as few users who have actually purchased a car as possible.

[0089] Then, the initial data of the users who signed up for the auto installment product obtained in step (1) and the newly acquired users who did not sign up for the auto installment product in the bank system are sorted according to the selected data in step (2). The features are extracted or filtered to obtain the labeled data set shown below:

[0090]

[0091] in, For randomly selected users of non-contracted auto installment products, These are the important features screened in step (2). In the last dimension, “1” indicates that there is a car purchase record, and “0” indicates that there is no car purchase record.

[0092] In step (4), use the labeled dataset obtained in step (3) to construct the instruction fine-tuning dataset. The construction method example is as follows.

[0093] Hypothetical Sample The corresponding user characteristics are (gender: male; age: 27; education: undergraduate; ...), then the prompt instructions designed in natural language can be: = "Please estimate the likelihood of this individual purchasing a car, with a value ranging from 0 to 1: Gender: Male; Age: 27; Education: Bachelor's degree...", the corresponding instruction fine-tuning sample is: ( , 1), is a positive sample data. It should be noted that when constructing the instruction fine-tuning sample, the order in which each feature appears in the prompt instruction can be random to improve the robustness of the fine-tuned large language model.

[0094] All samples of the labeled dataset can be constructed in a similar way to obtain the instruction fine-tuning dataset shown below:

[0095] .

[0096] In step (5), the instruction fine-tuning dataset obtained in step (4) is used to fine-tune the instruction of the large language model. Specifically, the low-rank adaptive model fine-tuning method (Low-Rank Adaptation, LoRA) based on low-rank matrix decomposition technology can be used to fine-tune the instruction of the large language model.

[0097] The core idea of ​​LoRA is to introduce two low-rank matrices into the weight matrix of the pre-trained large language model to achieve fine-tuning of the model. , LoRA is updated in the following ways:

[0098]

[0099] in, and are two low-rank matrices whose product is used to adjust the original weight matrix .because and The rank is much smaller than , so this method only needs to introduce a small number of additional parameters, thereby reducing the storage requirements and computational cost of the model.

[0100] During fine-tuning training, first load the pre-trained large language model, including its weights and configuration, and then initialize two low-rank matrices for each weight matrix in the large language model and Next, use the instructions obtained in step (4) to fine-tune the dataset and use the back-propagation algorithm to fine-tune the large language model. During the training process, keep the original weight parameters of the large language model unchanged and only update the low-rank matrix and This preserves the semantic understanding ability and context and sentiment analysis capabilities of the large language model, allowing the large language model to specifically learn the distribution characteristics and potential structures in the data, and improve the model's prediction ability when the input data is relatively sparse with respect to car purchases.

[0101] After completing the fine-tuning training, the updated model weights are saved to obtain a user car purchase intention recognition model for subsequent potential car purchase user identification.

[0102] In step (6), the large language model fine-tuned in step (5) is used as a model for identifying users’ willingness to buy a car, and a series of prompt instructions are designed to identify potential users who are willing to buy a car.

[0103] Specifically, assuming that the basic information of user A is as follows: {age: 37 years old; gender, female; education: undergraduate; ...}, the following prompt instructions can be constructed in a manner similar to the construction of instruction fine-tuning samples used when fine-tuning a large language model:

[0104] "Please assess the likelihood of this individual purchasing a car in the near future, presented on a scale of 0-1 and briefly explain the reason of this score: Gender: female; Age: 37; Education: Bachelor's degree...". The prompt instruction is input to the user car purchase intention recognition model, and the user car purchase intention recognition model will output a score in the range of 0-1 and the corresponding scoring basis.

[0105] Similarly, after the user car purchase intention identification model outputs the scores of all users to be analyzed, the N users with the highest scores can be screened out through sorting, or users with scores greater than a threshold can be screened out as the target customer group for the marketing of the auto installment product.

[0106] Then, build the automobile product recommendation model through the following steps (7) to (9).

[0107] Step (7): Build a car product recommendation model.

[0108] According to some embodiments of the present invention, when building a car product recommendation model, in order to achieve accurate recommendation according to the three attributes of price range, energy type and category when recommending cars to users, a car attribute classifier can be built according to price, energy type and category respectively, which is specifically used for the classification of corresponding attributes. Among them, each attribute of the car product is recommended using a special car attribute classifier, which can improve the accuracy of the recommendation and overcome the problem that the relevant information of the user's car purchase in the bank system is relatively sparse and affects the accuracy of the recommendation.

[0109] Furthermore, the automobile attribute classifier may also include multiple sub-classifiers, wherein different sub-classifiers are constructed using different algorithm models.

[0110] For example, for each car attribute classifier, a sub-classifier based on gradient boosting decision tree (GBDT), an adaptive boosting (AdaBoost) sub-classifier and a sub-classifier based on multi-layer perceptron (MLP) are constructed respectively. Each sub-classifier learns the mapping relationship between user features and car product attributes. In this way, when all sub-classifiers are trained, a voting mechanism can be used to integrate the category information of the attributes output by multiple sub-classifiers contained in each car attribute classifier, and determine an attribute of the car that the user may choose to buy, further overcoming the problem that the relevant information of users' car purchases in the banking system is relatively sparse and affects the accuracy of recommendations, and further improves the accuracy of recommendations through the integration of multiple sub-classifier results, improves user experience, and reduces the information interference caused by inaccurate recommendations to users. In this way, for the three classification tasks of price range, fuel type, and category, three sub-classifiers can be constructed respectively, so that a total of nine sub-classifiers are constructed.

[0111] Next, in this embodiment, the complete automobile installment data set obtained in step (1) can be used to train each sub-classifier. During training, the complete automobile installment data set can be divided by time and quantity. For example, 20% of the data with a relatively recent time is used as an inter-period validation set, and the remaining data is used as a training set. The generalization degree of the model for new trends is evaluated by verifying the performance of the trained model on the inter-period validation set, and then the hyperparameters of the model are adjusted. For parameters related to model training, for example, the maximum number of iterations of the GBDT model and the AdaBoost model can be set to 100, the number of layers of the MLP is 2, and a random linear rectifier unit (ReLU) with leakage is used as the activation function. The hidden layer dimension can be set to 256, and the maximum number of iterations can be set to 20 times. After training on the training set is completed, the hyperparameter selection of the model is evaluated by performance on the inter-period validation set, and the best performing set of hyperparameters is fixed. The inter-period validation set is also used for model training to obtain the corresponding sub-classifiers. At the same time, the accuracy, recall, precision, and F1-score of each model on the inter-period validation set are recorded.

[0112] In step (8), the target customer group obtained in step (5) is associated with the input features of the automobile product recommendation model trained in step (7) to obtain the target customer group dataset.

[0113] Input the target customer data set into the nine sub-classifiers trained in step (7) respectively, and collect the prediction results of each sub-classifier. For each attribute, a voting mechanism is used to combine the prediction results of the three sub-classifiers to achieve joint prediction. For example, for an attribute, if the prediction results of two or three sub-classifiers are the same, then the prediction result is used as the prediction result of the attribute; if the results output by the three sub-classifiers are all different, the prediction result of the sub-classifier model with the highest F1-score in the cross-period validation set can be selected as the prediction result of the attribute. In addition, by continuously monitoring the recommendation effect, the weight of each model in the voting mechanism can be dynamically adjusted to ensure the accuracy and timeliness of the recommendation results.

[0114] Step (9) obtains the prediction results of the three attributes of the automobile products predicted for each user in step (8), matches them in the car model database, and selects automobile products that meet the requirements. For example, the information of the five models with the highest market scores and / or the five models with the highest market sales is pushed to the corresponding users as the results of personalized recommendations.

[0115] In actual applications, the effectiveness of each model is evaluated based on business feedback, and model parameters and recommendation strategies are continuously optimized.

[0116] Figure 6A flowchart schematically shows a method for recommending an automobile product according to another embodiment of the present invention, wherein: Figure 6 It is the overall flow chart of the model application phase, describing the overall computational framework of the application phase.

[0117] refer to Figure 6 In the application stage, the user car purchase intention identification model can be used to identify the car purchase intention scores of the target users whose car purchase intention needs to be evaluated, and then the users with high car purchase intention can be screened out according to the potential car purchase conditions. Then, the car product recommendation model can be used to predict the specific car purchase tendency of users with high car purchase intention and make recommendations.

[0118] Among them, the initial input in the application phase is the initial data of the target user in the bank system. According to the high-importance features screened by the feature engineering and feature analysis modules in the training phase, the corresponding features of the target user are extracted from the initial data, and then prompt instructions are generated. Then the prompt instructions are input to the user's car purchase intention recognition model, and the user's car purchase intention recognition model gives the car purchase intention recognition result (such as a score). Then, when the target user's car purchase intention recognition result meets the potential car purchase conditions, the target user is determined to be a user with high car purchase intention. Next, the user's feature data can be sorted according to the input data features of the automobile product recommendation model, and then input into the automobile product recommendation model to obtain a specific car purchase tendency.

[0119] Figure 7 The block diagram of the device for recommending automobile products according to an embodiment of the present invention is schematically shown.

[0120] like Figure 7 As shown, the automobile product recommendation device 700 of this embodiment may include a first acquisition module 710 , a feature engineering and feature analysis module 720 , a car purchase intention prediction module 730 and an automobile product recommendation module 740 .

[0121] The first acquisition module 710 is used to obtain the authorization of the first user to use the data in the bank system, and obtain the data of the first user in the bank system after obtaining the authorization of the first user to obtain the first initial data. In one embodiment, the first acquisition module 710 can perform the operations S301 and S302 described above.

[0122] The feature engineering and feature analysis module 720 is used to filter out M feature data corresponding to M features from the first initial data to obtain a first feature set, wherein the M features are a set of features most relevant to the user's car purchase intention filtered out by the feature engineering algorithm, wherein M is an integer greater than or equal to 2. In one embodiment, the feature engineering and feature analysis module 720 may perform the operation S303 described above.

[0123] The car purchase intention prediction module 730 is used to construct a first prompt instruction expressed in natural language using the first feature set, input the first prompt instruction to the user car purchase intention recognition model, and obtain the first car purchase intention recognition result output by the user car purchase intention recognition model, wherein the user car purchase intention recognition model is obtained by fine-tuning the pre-trained large language model. In one embodiment, the car purchase intention prediction module 730 can perform the operations S304 and S305 described above.

[0124] The automobile product recommendation module 740 is used to recommend automobile products to the first user when the first automobile purchase intention recognition result meets the potential automobile purchase condition. In one embodiment, the automobile product recommendation module 740 may perform the operation S306 described above.

[0125] In one embodiment, the feature engineering and feature analysis module 720 is specifically used to train a random forest model to predict the user's car purchase intention based on the user's data in the bank system after obtaining the user's authorization to use the data in the bank system; and after the random forest model training is completed, based on the importance of the input features of the random forest model and the Shapley additive interpretation SHAP value, select M features. Among them, the feature engineering algorithm is constructed using the random forest model.

[0126] In one embodiment, the feature engineering and feature analysis module 720 is further used to: obtain authorization from at least one user who has purchased a car to use the data in his / her bank system, where the user who has purchased a car is a user who has purchased a car installment product in the bank system; after obtaining the authorization from the user who has purchased a car, obtain the data of the user who has purchased a car in the bank system, and obtain n labels of the user who has purchased a car based on the category information of n attributes of the car in the car installment product purchased by the user who has purchased a car, where n is an integer greater than or equal to 2; train n random forest models, each of which is used to predict one of the n attributes; wherein, when training each random forest model, the user who has purchased a car is labeled with a label corresponding to the attribute to be predicted by the random forest model, and the data of the user who has purchased a car in the bank system is used as input to train the random forest model; after the training of the n random forest models is completed, M features are screened out based on the importance and Shapley additive explanation SHAP value of the input features in the n random forest models.

[0127] In one embodiment, the automobile product recommendation module 740 is also used to input the first initial data into the automobile product recommendation model, and obtain the recommended category information of the automobile product output by the automobile product recommendation model; wherein the automobile product recommendation model is trained using the data of the users who have purchased the car in the bank system after obtaining the authorization of at least one user who has purchased the car to use the data in his bank system, wherein the users who have purchased the car are users who have purchased the automobile installment product in the bank system; and the automobile product indicated by the recommended category information is recommended to the first user.

[0128] In one embodiment, the automobile product recommendation module 740 is also used for: when the automobile product recommendation model includes n automobile attribute classifiers corresponding one-to-one to n attributes, and each automobile attribute classifier includes multiple sub-classifiers, and different sub-classifiers are constructed using different algorithm models, a voting mechanism is used to integrate the category information of the attributes output by the multiple sub-classifiers contained in each automobile attribute classifier to obtain the category information of the attributes corresponding to each automobile attribute classifier.

[0129] The recommendation device 700 can perform reference Figure 3~Figure 6 The automotive product recommendation method described in is specifically introduced in the previous article and will not be repeated here.

[0130] Figure 8 The block diagram schematically shows a device for training a user's car purchase intention recognition model according to an embodiment of the present invention.

[0131] like Figure 8 As shown, the training device 800 according to this embodiment may include a second acquisition module 810 , a feature engineering and feature analysis module 720 , and a fine-tuning training module 830 .

[0132] The second acquisition module 810 is used to obtain the second user's authorization to use the data in the bank system, and after obtaining the second user's authorization, obtain the second user's data in the bank system to obtain the second initial data, wherein the second user is a user who has purchased a car installment product in the bank system or a user who has not purchased a car but has not used any car-related products in the bank system. In one embodiment, the second acquisition module 810 can perform the operations S401 and S402 described above.

[0133] In the training device 800, the feature engineering and feature analysis module 720 is specifically used to: filter out M feature data corresponding to the M features from the second initial data to obtain a second feature set; when the second user is a user who has purchased a car, label the second feature set with the information of the purchased car as a label to obtain a positive sample data; and when the second user is a user who has not purchased a car, label the second feature set with the information of the non-purchased car as a label to obtain a negative sample data. The M features are a set of features that are most relevant to the user's willingness to purchase a car, which are filtered out by the feature engineering algorithm, and M is an integer greater than or equal to 2. In one embodiment, the feature engineering and feature analysis module 720 can perform the operations S403 and S404 described above.

[0134] The fine-tuning training module 830 is used to construct prompt instructions expressed in natural language for each positive sample data and each negative sample data after obtaining equal amounts of positive sample data and negative sample data, so as to obtain an instruction fine-tuning data set, and to perform fine-tuning training on the large language model using the instruction fine-tuning data set, so as to obtain a user car purchase intention recognition model. In one embodiment, the fine-tuning training module 830 can perform the operations S405 and S406 described above.

[0135] In one embodiment, the training device 800 may be integrated into the recommendation device 700. In other embodiments, the training device 800 and the recommendation device 700 may be deployed separately, wherein both the recommendation device 700 and the training device 800 may call the feature engineering and feature analysis module 720 to perform corresponding operations.

[0136] According to an embodiment of the present invention, the training device 800 may execute a reference Figure 4~Figure 6 The training method of the user car purchase intention recognition model described in . Please refer to the previous introduction for details and will not be repeated here.

[0137] According to an embodiment of the present invention, any multiple modules of the first acquisition module 710, the feature engineering and feature analysis module 720, the car purchase intention prediction module 730, the automobile product recommendation module 740, the second acquisition module 810 and the fine-tuning training module 830 can be combined in one module, or any one of them can be split into multiple modules. Alternatively, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module. According to an embodiment of the present invention, at least one of the first acquisition module 710, the feature engineering and feature analysis module 720, the car purchase intention prediction module 730, the automobile product recommendation module 740, the second acquisition module 810 and the fine-tuning training module 830 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application specific integrated circuit (ASIC), or can be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or by any one of the three implementation methods of software, hardware and firmware or by a suitable combination of any of them. Alternatively, at least one of the first acquisition module 710, the feature engineering and feature analysis module 720, the car purchase intention prediction module 730, the automobile product recommendation module 740, the second acquisition module 810 and the fine-tuning training module 830 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is executed.

[0138] Fig. 9 A block diagram of an electronic device according to an embodiment of the present invention is schematically shown.

[0139] like Fig. 9 As shown, the electronic device 900 according to an embodiment of the present invention includes a processor 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage part 908 to a random access memory (RAM) 903. The processor 901 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or a related chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 901 may also include an onboard memory for caching purposes. The processor 901 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0140] In RAM 903, various programs and data required for the operation of electronic device 900 are stored. Processor 901, ROM 902 and RAM 903 are connected to each other via bus 904. Processor 901 performs various operations of the method flow according to the embodiment of the present invention by executing the program in ROM 902 and / or RAM 903. It should be noted that the program can also be stored in one or more memories other than ROM 902 and RAM 903. Processor 901 can also perform various operations of the method flow according to the embodiment of the present invention by executing the program stored in the one or more memories.

[0141] According to an embodiment of the present invention, the electronic device 900 may further include an input / output (I / O) interface 905, which is also connected to the bus 904. The electronic device 900 may further include one or more of the following components connected to the input / output (I / O) interface 905: an input portion 906 including a keyboard, a mouse, etc.; an output portion 907 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage portion 908 including a hard disk, etc.; and a communication portion 909 including a network interface card such as a LAN card, a modem, etc. The communication portion 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the input / output (I / O) interface 905 as needed. A removable medium 911, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 910 as needed, so that a computer program read therefrom is installed into the storage portion 908 as needed.

[0142] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiment; or may exist independently without being assembled into the device / apparatus / system. The above computer-readable storage medium carries one or more programs, and when the above one or more programs are executed, the method according to the embodiment of the present invention is implemented.

[0143] The embodiment of the present invention also includes a computer program product, which includes a computer program, and the computer program contains program code for executing the method shown in the flowchart. When the computer program product is run in a computer system, the program code is used to enable the computer system to implement the method provided by the embodiment of the present invention.

[0144] The computer program executes the above functions defined in the system / device of the embodiment of the present invention when it is executed by the processor 901. According to the embodiment of the present invention, the system, device, module, unit, etc. described above can be implemented by a computer program module.

[0145] The embodiments of the present invention are described above. However, these embodiments are only for the purpose of illustration, and are not intended to limit the scope of the present invention. Although each embodiment is described above, it does not mean that the measures in each embodiment cannot be used in combination advantageously. Without departing from the scope of the present invention, those skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present invention.

Claims

1. A method for recommending automobile products, characterized in that: The method comprises: Obtaining the first user's authorization to use his / her data in the banking system; After obtaining authorization from the first user, obtaining the data of the first user in the bank system to obtain first initial data; Screening out M feature data corresponding to M features from the first initial data to obtain a first feature set, wherein the M features are a set of features most relevant to the user's car purchase intention screened out by a feature engineering algorithm, wherein M is an integer greater than or equal to 2; constructing a first prompt instruction expressed in natural language using the first feature set; Inputting the first prompt instruction to a user car purchase intention recognition model, and obtaining a first car purchase intention recognition result output by the user car purchase intention recognition model, wherein the user car purchase intention recognition model is obtained by fine-tuning a pre-trained large language model; and When the first car purchase intention identification result meets the potential car purchase conditions, a car product is recommended to the first user.

2. The method according to claim 1, characterized in that The process of fine-tuning the pre-trained large language model to obtain the user car purchase intention recognition model includes: Obtaining authorization from a second user for using data in the bank system, where the second user is a user who has purchased a car installment product in the bank system or a user who has not purchased a car but has not used any car-related products in the bank system; After obtaining authorization from the second user, acquiring the data of the second user in the bank system to obtain second initial data; Filtering out M feature data corresponding to the M features from the second initial data to obtain a second feature set; When the second user is the user who has purchased a car, the second feature set is labeled with the information of the purchased car as a label to obtain a positive sample data; and when the second user is the user who has not purchased a car, the second feature set is labeled with the information of the non-purchased car as a label to obtain a negative sample data; After obtaining an equal amount of the positive sample data and the negative sample data, constructing a prompt instruction expressed in natural language for each piece of the positive sample data and each piece of the negative sample data, respectively, to obtain an instruction fine-tuning data set; and The large language model is fine-tuned and trained using the instruction fine-tuning data set to obtain the user car purchase intention recognition model.

3. The method according to claim 2, characterized in that The fine-tuning training of the large language model using the instruction fine-tuning dataset includes: The large language model is fine-tuned using a low-rank matrix factorization method.

4. The method according to claim 1, characterized in that The feature engineering algorithm is constructed using a random forest model, wherein the screening process of the M features includes: After obtaining the user's authorization to use the user's data in the bank system, train a random forest model to predict the user's car purchase intention based on the user's data in the bank system; and After the random forest model is trained, the M features are screened out based on the importance of the input features of the random forest model and the Shapley additivity interpretation SHAP value.

5. The method according to claim 4, characterized in that The screening process of the M features specifically includes: Obtaining authorization from at least one car-buying user for using data in his / her bank system, wherein the car-buying user is a user who has purchased a car installment product in the bank system; After obtaining authorization from the user who has purchased the car, the data of the user who has purchased the car in the bank system is obtained, and n tags of the user who has purchased the car are obtained based on the category information of n attributes of the car in the car installment product purchased by the user who has purchased the car, where n is an integer greater than or equal to 2; Training n random forest models, each random forest model is used to predict one of the n attributes; wherein when training each random forest model, the user who has purchased the car is labeled with a label corresponding to the attribute to be predicted by the random forest model, and the data of the user who has purchased the car in the bank system is used as input to train the random forest model; After the training of the n random forest models is completed, the M features are screened out based on the importance of the input features and the Shapley additivity interpretation SHAP value of the n random forest models.

6. The method according to claim 1, characterized in that When the first car purchase intention identification result meets the potential car purchase condition, recommending a car product to the first user includes: The first initial data is input into an automobile product recommendation model, and recommended category information of automobile products output by the automobile product recommendation model is obtained; wherein the automobile product recommendation model is trained using the data of at least one automobile purchasing user in the bank system after obtaining authorization from the at least one automobile purchasing user to use the data in the bank system, wherein the automobile purchasing user is a user who has purchased an automobile installment product in the bank system; The automobile product indicated by the recommendation category information is recommended to the first user.

7. The method according to claim 6, characterized in that The recommendation category information includes category information of n attributes, and the automobile product recommendation model includes n automobile attribute classifiers corresponding to the n attributes one by one, wherein n is an integer greater than or equal to 2.

8. The method according to claim 7, characterized in that Each of the automobile attribute classifiers includes a plurality of sub-classifiers, wherein different sub-classifiers are constructed using different algorithm models, wherein the step of obtaining the recommended category information of automobile products output by the automobile product recommendation model includes: The category information of the attributes output by the multiple sub-classifiers included in each of the automobile attribute classifiers is synthesized by a voting mechanism to obtain the category information of the attributes corresponding to each of the automobile attribute classifiers.

9. The method according to claim 5 or 7, characterized in that: The n attributes include at least one of the following: price, category, and energy type.

10. A training method for a user car purchase intention recognition model based on a large language model, characterized in that: The method comprises: Obtaining authorization from a second user for using data in the bank system, where the second user is a user who has purchased a car installment product in the bank system or a user who has not purchased a car but has not used any car-related products in the bank system; After obtaining authorization from the second user, acquiring the data of the second user in the bank system to obtain second initial data; Screening out M feature data corresponding to the M features from the second initial data to obtain a second feature set, wherein the M features are a set of features most relevant to the user's car purchase intention screened out by a feature engineering algorithm, wherein M is an integer greater than or equal to 2; When the second user is the user who has purchased a car, the second feature set is labeled with the information of the purchased car as a label to obtain a positive sample data; and when the second user is the user who has not purchased a car, the second feature set is labeled with the information of the non-purchased car as a label to obtain a negative sample data; After obtaining an equal amount of the positive sample data and the negative sample data, constructing a prompt instruction expressed in natural language for each piece of the positive sample data and each piece of the negative sample data, respectively, to obtain an instruction fine-tuning data set; and The large language model is fine-tuned and trained using the instruction fine-tuning data set to obtain the user car purchase intention recognition model.

11. A device for recommending automobile products, characterized in that: The device comprises: A first acquisition module, used to obtain the authorization of the first user to use the data in the bank system, and to obtain the data of the first user in the bank system after obtaining the authorization of the first user, so as to obtain first initial data; A feature engineering and feature analysis module, configured to filter out M feature data corresponding to the M features from the first initial data to obtain a first feature set, wherein the M features are a set of features most relevant to the user's car purchase intention filtered out by a feature engineering algorithm, wherein M is an integer greater than or equal to 2; a car purchase intention prediction module, configured to construct a first prompt instruction expressed in natural language using the first feature set, input the first prompt instruction to a user car purchase intention recognition model, and obtain a first car purchase intention recognition result output by the user car purchase intention recognition model, wherein the user car purchase intention recognition model is obtained by fine-tuning a pre-trained large language model; and The automobile product recommendation module is used to recommend an automobile product to the first user when the first automobile purchase intention recognition result meets the potential automobile purchase conditions.

12. A training device for a user car purchase intention recognition model based on a large language model, characterized in that: The training device comprises: A second acquisition module is used to obtain the second user's authorization to use the data in the bank system, and after obtaining the second user's authorization, obtain the second user's data in the bank system to obtain second initial data; wherein the second user is a user who has purchased a car installment product in the bank system or a user who has not purchased a car but has not used any car-related products in the bank system; Feature engineering and feature analysis modules for: M feature data corresponding to the M features are screened out from the second initial data to obtain a second feature set; wherein the M features are a set of features most relevant to the user's car purchase intention screened out by a feature engineering algorithm, wherein M is an integer greater than or equal to 2; When the second user is the user who has purchased a car, the second feature set is labeled with the information of the purchased car as a label to obtain a positive sample data; and when the second user is the user who has not purchased a car, the second feature set is labeled with the information of the non-purchased car as a label to obtain a negative sample data; The fine-tuning training module is used to construct prompt instructions expressed in natural language for each of the positive sample data and each of the negative sample data after obtaining equal amounts of the positive sample data and the negative sample data, so as to obtain an instruction fine-tuning data set, and to use the instruction fine-tuning data set to fine-tune the large language model to obtain the user's car purchase intention recognition model.

13. An electronic device, comprising: one or more processors; a memory for storing one or more computer programs, It is characterized in that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 10.

14. A computer-readable storage medium having a computer program or instruction stored thereon, characterized in that: When the computer program or instruction is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

15. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

Images