A method for automobile sales auxiliary decision-making based on large models

Abstract

The present invention discloses a method for assisting decision-making in automobile sales based on large models, belonging to the technical field of automobile sales market prediction, including multi-source heterogeneous data collection and storage, four-layer division modeling, BERT + BILSTM intention extraction, candidate vehicle model screening, scoring function sorting, generating recommendation reasons, and feedback training the model, which solves the technical problem of improving the accuracy of automobile sales decision-making assistance. The present invention realizes the efficient fusion of cross-source heterogeneous data through the "data lake + data warehouse" dual architecture and the time window alignment mechanism, constructs a four-level label system of macro - meso - meso-micro - micro, which can cover the full-chain prediction of policy transmission, regional competition, group migration, and individual behavior, and improves the decision-making accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of automotive sales market prediction, and particularly relates to a method for assisting decision-making in automotive sales based on a large model. Background Art

[0002] Currently, the technical frameworks mainly relied on in the field of automotive sales assistance decision-making include traditional data-driven systems, single-granularity prediction models, and interaction and recommendation technologies;

[0003] Traditional data-driven systems build statistical models based on structured data (such as CRM, order records), predict sales trends through time series analysis (such as ARIMA) or regression models, and rely on manual rules to generate customer portraits and recommendation logics.

[0004] The data sources of traditional data-driven systems are single, the models are difficult to capture the non-linear impact of external factors such as policies on sales, the update of customer portraits lags behind, and the recommendation results are out of touch with real-time demands.

[0005] Single-granularity prediction models include the macro level and the micro level:

[0006] At the macro level, industry-level economic indicators (GDP, oil price) are used to predict the overall market trend, but it often ignores regional competition and urban-level consumption differences;

[0007] At the micro level, collaborative filtering or click sequence is used to model individual preferences, but it often lacks the ability to model group behavior migration and policy conduction effects.

[0008] In interaction and recommendation technologies, natural language processing (NLP) technologies are mostly used to parse customer consultations through keyword matching or shallow semantic models (such as TF-IDF). Currently, recommendation systems rely on static rules (such as price range filtering) or single scoring models, and cannot integrate multi-level prediction results, resulting in serious homogenization of recommended models. Summary of the Invention

[0009] The purpose of the present invention is to provide a method for assisting decision-making in automotive sales based on a large model, which solves the technical problem of improving the accuracy of automotive sales assistance decision-making.

[0010] To achieve the above purpose, the present invention adopts the following technical solutions:

[0011] A method for assisting decision-making in automotive sales based on a large model, comprising the following steps:

[0012] Step 1: Deploy a data acquisition architecture at the data acquisition layer. The data acquisition architecture obtains various heterogeneous data from multiple data sources; establish a unified data management architecture for storing and integrating heterogeneous data;

[0013] Step 2: Construct a feature engineering process and a hierarchical label system in the data preprocessing layer; the feature engineering process includes preprocessing heterogeneous data and outputting four levels of feature sets; construct a hierarchical label system based on the feature sets, including macro labels, meso labels, meso-micro labels, and micro labels. Macro labels include national sales trends, oil price fluctuations, GDP growth rates, and purchase tax changes; meso labels include city-level advertising placement frequencies, competitive brand price adjustment records, and local 4S store densities; meso-micro labels include car purchase inclinations, hot-selling model preferences, and behavior heat indexes at the city-customer group dimension; micro labels include individual-level car purchase preferences, browsing paths, budget ranges, and past purchase behaviors.

[0014] Step 3: In the prediction modeling layer, construct a hierarchical prediction model system according to the hierarchical label system, including a macro layer model, a meso layer model, a meso-micro layer model, and a micro layer model:

[0015] Specifically, the meso-micro layer model uses meso-micro labels as inputs, constructs a customer group car purchase probability prediction model based on the city-customer group dimension, and generates a hot-selling heat map of the city-customer group.

[0016] Step 4: In the real-time interaction layer, access the real-time behavior data of customers, use the BERT+BILSTM joint model to decode the natural language of customers, extract the current car purchase intention, generate a semantic embedding vector, and dynamically correct the short-term preference labels in the customer portrait.

[0017] Step 5: In the recommendation generation layer, obtain candidate models according to the results of Step 3 and Step 4, construct a scoring function, use the scoring function to score and rank the candidate models, select the top 3 ranking results, and then use the template filling method to generate personalized recommendation reasons and output.

[0018] Step 6: In the training layer, establish a feedback mechanism, collect the feedback data of the prediction results of each layer model and the scoring function in the actual sales scenario, return the feedback data to the training sample set, and use the training sample set to optimize and train each layer model and the scoring function.

[0019] Preferably, when performing Step 1, the heterogeneous data specifically includes structured data, semi-structured data, and unstructured data;

[0020] Structured data includes customer basic information, order records, car purchase history, vehicle model parameters, power types, drive modes, inventory status, price information, GDP data, purchase tax adjustment records, macro indicators such as oil price time series, and the quantity and geographical distribution of sales service points in each city.

[0021] Semi-structured data includes official website and App buried point logs, test drive frequencies, model preferences, advertising display frequencies and time distributions in each city, and customer portraits;

[0022] Unstructured data includes customer review texts, car purchase forum posts, policy announcements, news articles, and voice call transcription texts;

[0023] The unified data management architecture specifically includes a data access layer, a data storage layer, a metadata management layer, and a time window unification layer:

[0024] The data access layer is responsible for regularly collecting heterogeneous data by means of API calls, database scraping, and log transmission;

[0025] The data storage layer includes a data lake and a data warehouse with a parallel structure. The data warehouse is used to store structured data, and the data lake is used to store semi-structured data and unstructured data;

[0026] The metadata management layer is responsible for uniformly extracting and tagging key meta-information such as customer ID, timestamp, geographical location, and model ID;

[0027] The time window unification layer is responsible for time-aligning all heterogeneous data according to the daily-weekly-monthly dimensions.

[0028] Preferably, when performing step 3, the macro-layer model adopts a trend decomposition model and a deep time series correction model, uses the macro tags as the model input, and predicts the overall impact of the national total sales volume and policy changes;

[0029] The mid-level model uses mid-level tags as the model input, adopts a graph neural network GNN to model the competition and cooperation relationships between cities, and predicts the model sales volume at the city level;

[0030] The mid-micro layer model uses mid-micro tags as the model input, adopts a population clustering model and a softmax model to generate a city-customer group hot sales heat map;

[0031] The micro-layer model uses micro tags as the input, uses a reinforcement learning model to model the customer click behavior sequence, and predicts the individualized car purchase probability and configuration preference.

[0032] Preferably, when performing step 4, the real-time interaction layer real-time accesses the behavior data of customers' clicks, comparisons, or collections on the APP, mini-program, or official website, uses these behavior data as the input of the BERT+BILSTM joint model, extracts the current car purchase intention, and obtains the intention recognition result;

[0033] The customer portrait tags are dynamically adjusted through a rule engine and a behavior weight update mechanism.

[0034] Preferably, when performing step 5, it specifically includes the following steps:

[0035] Step 5-1: Construct a candidate vehicle model set according to the prediction results of each layer of models in Step 3, which specifically includes the following steps:

[0036] Step 5-1-1: Obtain the output result of the macro-layer model , and preferentially retain the vehicle models whose increment reaches the preset threshold within the future preset period;

[0037] Step 5-1-2: Obtain the output result of the meso-layer model , and in the city , select the vehicle model set ranked top 10 in terms of sales volume prediction, and conduct the first screening on the vehicle model set according to the vehicle models retained in Step 5-1-1;

[0038] Step 5-1-3: Obtain the result output by the meso-micro layer model, and conduct a secondary screening on the vehicle model set according to the city-customer group hot sales heat map and the customer group purchase probability , and select the vehicle models that meet the following conditions:

[0039] and ;

[0040] wherein, is the preset threshold, respectively represent the city, customer group, and vehicle model;

[0041] Step 5-1-4: Obtain the output result of the micro-layer model, and conduct another screening on the vehicle models obtained in Step 5-1-3 according to the individualized purchase probability and configuration preference to obtain the candidate vehicle model set;

[0042] Step 5-2: Construct a scoring function to score and rank the candidate vehicle models in the candidate vehicle model set to obtain a candidate list. The scoring function is as follows:

[0043] ;

[0044] wherein, is the vehicle model score value, respectively represent the customer, vehicle model, and city, represents the preference matching score, represents the macro sales volume growth factor, represents the city heat score, represents the intention adjustment factor obtained by the method in Step 4, are all weights;

[0045] Step 5-3: Select the top 3 vehicle models in the candidate list as the recommended results, generate personalized recommendation reasons through template filling, and output.

[0046] Preferably, when performing step 6, a feedback server is deployed in the training layer. The feedback server automatically records whether the customer accepts the recommendation, the reasons for clicking / rejecting, and feeds them back to the training samples.

[0047] The feedback server compares the preset indicators of recommendation hit rate, conversion rate, and customer satisfaction, and tunes or rolls back each model.

[0048] A method for assisting automobile sales decision-making based on a large model according to the present invention solves the technical problem of improving the accuracy of automobile sales decision-making assistance. The present invention realizes the efficient fusion of cross-source heterogeneous data through the "data lake + data warehouse" dual architecture and the time window alignment mechanism, constructs a four-level label system of macro-meso-micro-micro, and can cover the whole chain prediction of policy transmission, regional competition, group migration, and individual behavior, improving the decision-making accuracy. Brief Description of the Drawings

[0049] Figure 1 is the main flowchart of the present invention;

[0050] Figure 2 is the flowchart of step 5 of the present invention;

[0051] Figure 3 is a system architecture diagram in actual use in this embodiment;

[0052] Figure 4 is a schematic diagram of the functional division of each level of the present invention. Detailed Embodiments

[0053] As shown by Figures 1 - 4 a method for assisting automobile sales decision-making based on a large model includes the following steps:

[0054] Step 1: Deploy a data collection architecture in the data collection layer. The data collection architecture obtains various heterogeneous data from multiple data sources; establish a unified data management architecture for storing and integrating heterogeneous data;

[0055] When performing step 1, the heterogeneous data specifically includes structured data, semi-structured data, and unstructured data.

[0056] The structured data includes:

[0057] basic customer information, order records, purchase history, etc. collected by the CRM system;

[0058] model parameters, power type, drive mode, inventory status, price information collected by the model configuration table and inventory price system;

[0059] Macroeconomic indicators such as GDP data, purchase tax adjustment records, and oil price time series publicly available on the platforms of relevant national departments;

[0060] The quantity and geographical distribution of sales service points in each city collected from the 4S store network database.

[0061] Semi-structured data includes:

[0062] Buried point behaviors such as user click, browsing, and bounce behavior sequences collected from the official website and App buried point logs;

[0063] Test drive frequencies and vehicle model preferences collected from the test drive system logs;

[0064] The advertisement display frequencies and time distributions in each city collected from the third-party advertisement placement platform;

[0065] Structured tags such as customer interests, income estimates, and family structures in the customer portrait system, as well as the update history.

[0066] Unstructured data includes:

[0067] Customer review texts and car purchase forum posts for sentiment analysis and opinion extraction;

[0068] Policy announcements and news articles for data extraction of policy changes and competitor behaviors;

[0069] Purchase intention, budget expression, and preference keywords contained in the transcribed text of voice calls.

[0070] The unified data management architecture specifically includes a data access layer, a data storage layer, a metadata management layer, and a time window unification layer:

[0071] The data access layer is responsible for regularly collecting heterogeneous data by means of API calls, database scraping, and log transmission;

[0072] The data storage layer includes a data lake (such as S3 / HDFS) and a data warehouse (such as Snowflake) with a parallel structure. The data warehouse is used to store structured data, and the data lake is used to store semi-structured data and unstructured data;

[0073] The metadata management layer is responsible for uniformly extracting and marking key metadata such as customer ID, timestamp, geographical location, and vehicle model ID;

[0074] The time window unification layer is responsible for time-aligning all heterogeneous data according to the dimensions of day-week-month.

[0075] Step 2: Construct a feature engineering process and a hierarchical label system in the data preprocessing layer; the feature engineering process includes preprocessing heterogeneous data and outputting a feature set at 4 levels;

[0076] The preprocessing includes preprocessing such as text structuring, time series standardization, missing value filling, outlier screening, and text cleaning.

[0077] In this embodiment, the four-level feature sets specifically include:

[0078] Macro feature set: Extract features from policy announcements, oil price statistics, and GDP growth data, including oil price trend curves, purchase tax adjustment cycles, and year-on-year GDP growth rates;

[0079] Mesoscopic feature set: Extract features from advertising logs, store data, and news comments, including urban advertising placement frequencies, competitor price adjustment heat indices, and 4S store densities;

[0080] Mesoscopic and microscopic feature set: Extract features from order records, buried point behaviors, and customer portraits, including city-customer group model preference matrices, behavior heat indices, and budget distribution heat maps;

[0081] Micro feature set: Extract features from user click behavior sequences, test drive records, and voice intent transcriptions, including access path sequences, budget ranges, and interest brand preference scores.

[0082] Construct a hierarchical label system based on the feature sets, including macro labels, mesoscopic labels, mesoscopic and microscopic labels, and micro labels:

[0083] Macro labels include national sales trends, oil price fluctuations, GDP growth rates, and purchase tax changes;

[0084] Mesoscopic labels include urban-level advertising placement frequencies, competitive brand price adjustment records, and local 4S store densities;

[0085] Mesoscopic and microscopic labels include car purchase tendencies, popular model preferences, and behavior heat indices in the city-customer group dimension;

[0086] Micro labels include individual-level car purchase preferences, browsing paths, budget ranges, and past purchase behaviors;

[0087] Step 3: In the prediction modeling layer, construct a hierarchical prediction model system based on the hierarchical label system, including a macro layer model, a mesoscopic layer model, a mesoscopic and microscopic layer model, and a micro layer model:

[0088] When executing Step 3, the macro layer model uses a trend decomposition model and a deep time series correction model, takes the macro labels as model inputs, and predicts the overall impact of the national total sales volume and policy changes;

[0089] In this embodiment, the macro layer model is specifically as follows:

[0090] Trend decomposition model and deep time series correction model:

[0091]

[0092] in, Represents the predicted future Heaven Sales volume of vehicle models, For the LSTM model, Indicates that from Day to Day Sky type National sales series, is the time step, Indicates that from Day to Day The year-on-year GDP growth rate series of the day, Indicates that from Day to Day The oil price series of the day, Indicates that from Day to Day The purchase tax adjustment status sequence of the day, Indicates that from Day to Day The policy announcement text score sequence of the day, Obtained by analyzing policy announcements using the BERT model and its sentiment module;

[0093] The calculation formula of the macro sales growth factor is as follows:

[0094] ;

[0095] in, Indicates the Sky-type National sales volume;

[0096] The meso-level model uses meso-level labels as model input, and uses graph neural networks (GNNs) to model the competitive and cooperative relationships between cities and predict vehicle sales at the city level.

[0097] The graph neural network GNN model is as follows:

[0098] ;

[0099] in, They all represent city nodes, representing different cities. represents the attention weight, which specifically represents the similarities between cities, such as geographical proximity and consumption habits. The larger it is, the higher the similarity between cities is; Represents a city node In the The hidden state of the layer, Represents a city node At the Hidden layer state of the layer; Represents the Learnable weight matrix of the layer, Is the activation function; Represents the city node Set of neighbor nodes.

[0100] The input of the graph neural network GNN model is the city 's advertisement display frequency, the city 's 4S store density.

[0101] The city - vehicle model sales prediction model is as follows:

[0102] ;

[0103] Among them, Represents the city 's Predicted sales volume of the vehicle model, Represents the node features obtained from the last layer of the graph neural network GNN model, Represents the city In the Competitive product price adjustment frequency of the vehicle model; Is a multi - layer perceptron model, used for non - linear transformation and prediction of input features.

[0104] The calculation formula for the city heat score is as follows:

[0105] ;

[0106] Among them, Represents the heat score of the In the Vehicle model, Represents except for the Vehicle model; other vehicle models; Is the sum value of other vehicle models.

[0107] The LSTM model, the graph neural network GNN model, and the multi - layer perceptron model are all existing technologies, so they will not be described in detail.

[0108] In the micro - layer model, the micro - layer label is used as the model input, and the population clustering model and the softmax model are used to generate the city - customer group hot sales heat map;

[0109] The population clustering model is specifically as follows:

[0110] ;

[0111] where, is the clustering label to which user u belongs, is the center vector of the k-th customer group, is the behavior embedding of user u, extracted from order records and buried-point behaviors;

[0112] The purchase probability of the customer group (customer population) is calculated using the softmax model as follows:

[0113] ;

[0114] where, represents the probability of customer group g purchasing vehicle model m in city c, is the frequency of clicks or test drives of vehicle model m by customer group g in city c.

[0115] The micro heat score in is calculated as follows:

[0116] ;

[0117] where, is the normalization operation.

[0118] The micro-layer model takes micro labels as input and uses a reinforcement learning model to model the customer click behavior sequence, predicting the individualized purchase probability and configuration preference.

[0119] The reinforcement learning model is specifically a behavior-configuration reinforcement learning model, and the specific formula is as follows:

[0120] ;

[0121] where, represents the click behavior sequence of user u, represents the test drive frequency of user u for vehicle model m, represents the semantic embedding vector of the real-time intention, obtained from step 4, represents the configuration characteristics of vehicle model m.

[0122] The individual purchase probability of customer u for vehicle model m is calculated as follows:

[0123] ;

[0124] where, is the activation function.

[0125] The calculation formula of the configuration preference matching score is as follows:

[0126] ;

[0127] represents the cosine similarity function represents the individual interest vector of user u represents the attribute feature vector of vehicle model m, such as brand, price, configuration, etc.

[0128] Step 4: Access the real-time behavior data of the customer in the real-time interaction layer, use the BERT+BILSTM joint model to decode the natural language of the customer, extract the current car purchase intention, generate a semantic embedding vector, and dynamically correct the short-term preference tags in the customer portrait;

[0129] When executing Step 4, the real-time interaction layer accesses in real time the behavior data of the customer's clicks, comparisons, or collections on the APP, mini-program, or official website, uses these behavior data as the input of the BERT+BILSTM joint model, extracts the current car purchase intention, and obtains the intention recognition result;

[0130] Dynamically adjust the customer portrait tags through the rule engine and the behavior weight update mechanism.

[0131] Step 5: In the recommendation generation layer, according to the results obtained in Step 3 and Step 4, obtain candidate vehicle models, construct a scoring function, use the scoring function to score and rank the candidate vehicle models, select the top 3 ranking results, and then use the method of template filling to generate personalized recommendation reasons and output;

[0132] When executing Step 5, it specifically includes the following steps:

[0133] Step 5-1: According to the prediction results of each layer model in Step 3, construct a candidate vehicle model set, which specifically includes the following steps:

[0134] Step 5-1-1: Obtain the output result of the macro layer model , and preferentially retain the vehicle models whose increment reaches the preset threshold within the future preset period;

[0135] Step 5-1-2: Obtain the output result of the middle layer model , in the city , select the vehicle model set with the top 10 sales predictions, and conduct the first screening of the vehicle model set according to the vehicle models retained in Step 5-1-1;

[0136] Step 5-1-3: Obtain the output result of the middle and micro layer model, and according to the city-customer group hot sales heat map and the customer group car purchase probability , conduct a secondary screening of the vehicle model set, and select the vehicle models that meet the following conditions:

[0137] and ;

[0138] Among them, is a preset threshold value, respectively represent the city, customer group, and vehicle model;

[0139] Step 5-1-4: Obtain the output result of the micro-layer model, and re-screen the vehicle models obtained in Step 5-1-3 according to the individualized car purchase probability and configuration preference to obtain a candidate vehicle model set;

[0140] Step 5-2: Construct a scoring function to score and rank the candidate vehicle models in the candidate vehicle model set to obtain a candidate list. The scoring function is as follows:

[0141] ;

[0142] Among them, is the vehicle model score value, respectively represent the customer, vehicle model, and city, represents the preference matching score, represents the macro sales growth factor, represents the city popularity score, represents the intention adjustment factor obtained by the method in Step 4, are all weights.

[0143] In this embodiment, it can be calculated by using the cosine similarity function, such as: , among which, is the intention embedding vector output by the BERT model, such as new energy preference, high budget intention, etc., represents the vehicle model embedding, such as whether to choose a new energy vehicle, price, etc.

[0144] Step 5-3: Select the top 3 models in the candidate list as the recommendation results, generate personalized recommendation reasons through template filling, and output.

[0145] In this embodiment, multiple templates are preset, and the recommendation results can be directly filled in different templates, such as:

[0146] Template A: "Based on the [brand] you have recently frequently paid attention to and the [vehicle model category] you have test-driven, we recommend [specific vehicle model] for you. This vehicle model is very popular in [city where it is located] and meets your [budget range] budget."

[0147] Step 6: Establish a feedback mechanism in the training layer, collect the feedback data of the prediction results of each layer model and the scoring function in the actual sales scenario, return the feedback data to the training sample set, and use the training sample set to optimize and train each layer model and the scoring function.

[0148] When performing step 6, a feedback server is deployed in the training layer. The feedback server automatically records whether the customer accepts the recommendation and the reasons for clicking / rejecting, and feeds them back to the training samples;

[0149] The feedback server compares the preset metrics of recommendation hit rate, conversion rate, and customer satisfaction, and tunes or rolls back each model.

[0150] In this embodiment, in the actual application process, the following system architecture can be adopted:

[0151] In the data collection layer, a data collection architecture is deployed, specifically including:

[0152] An ETL server, responsible for heterogeneous data extraction (API calls, database scraping), transformation (format standardization), and loading (writing to the data lake / warehouse);

[0153] A log collection server, responsible for real-time collection of semi-structured data such as official website / App buried point logs and test drive records, supporting high-speed writing and retrieval;

[0154] An API gateway server, responsible for unified management of external data interfaces (such as policy announcement APIs, competitor price APIs);

[0155] An object storage server, responsible for storing unstructured data (customer comments, voice transcription texts).

[0156] A data preprocessing server is deployed in the data preprocessing layer. In the data preprocessing server, the following are established:

[0157] A data processing service module, used to perform data preprocessing or feature engineering, etc.;

[0158] A metadata management module, used to manage metadata such as customer ID, timestamp, and geographical location, and provide retrieval;

[0159] A time synchronization module, used to ensure that all data is aligned by day-week-month.

[0160] A model server is deployed in the prediction modeling layer. In the model server, the following are established:

[0161] A macro model module, used to deploy a trend decomposition model and a deep time series model;

[0162] A meso model module, used to run a GNN model;

[0163] A meso-micro model module, used to run a crowd clustering model and a softmax model;

[0164] A micro model module, used to run a reinforcement learning model.

[0165] In the real-time interaction layer, the following are deployed:

[0166] A real-time data processing server for real-time access to customer click and comparison behavior data;

[0167] An NLP inference server for deploying the BERT + BILSTM model;

[0168] A rule engine server for dynamically adjusting customer portraits, formulating rules, and building a rule engine;

[0169] Deploy a recommendation generation server, a load balancing server, and a result caching server in the recommendation generation layer. In the recommendation generation server, establish:

[0170] A recommendation engine module for performing candidate vehicle model screening, scoring and ranking, and performing similarity calculations, etc.;

[0171] A template rendering module for presetting templates, filling in templates, and generating personalized recommendations;

[0172] In the training layer, there are a feedback server and a model training server. The feedback server is responsible for collecting customer acceptance / rejection behavior data and storing feedback metrics through a time-series database; the model training server is responsible for training the models of each layer.

[0173] All servers of each layer communicate with each other through the Internet.

[0174] A method for assisting in automobile sales decision-making based on a large model according to the present invention solves the technical problem of improving the accuracy of automobile sales decision-making assistance. The present invention realizes the efficient fusion of cross-source heterogeneous data through the "data lake + data warehouse" dual architecture and the time window alignment mechanism, constructs a four-level label system of macro - meso - meso - micro - micro, which can cover the full-chain prediction of policy transmission, regional competition, group migration, and individual behavior, and improves the decision-making accuracy.

Claims

1. An auxiliary decision-making method for automobile sales based on a large model, characterized in that: It includes the following steps: Step 1: Deploy a data collection architecture at the data collection layer. The data collection architecture obtains various heterogeneous data from multiple data sources; establish a unified data management architecture for storing and integrating heterogeneous data; Step 2: Build a feature engineering process and a hierarchical label system at the data preprocessing layer; The feature engineering process includes preprocessing heterogeneous data and outputting feature sets at 4 levels; Build a hierarchical label system based on the feature sets, including macro labels, meso labels, meso-micro labels, and micro labels. Macro labels include national sales trends, oil price fluctuations, GDP growth rates, and purchase tax changes; meso labels include city-level advertising placement frequencies, competitive brand price adjustment records, and local 4S store densities; meso-micro labels include car purchase inclinations, popular model preferences, and behavior heat indexes at the city-customer group dimension; micro labels include individual-level car purchase preferences, browsing paths, budget ranges, and past purchase behaviors; Step 3: At the prediction modeling layer, build a hierarchical prediction model system according to the hierarchical label system, including a macro layer model, a meso layer model, a meso-micro layer model, and a micro layer model: Specifically, the meso-micro layer model uses meso-micro labels as inputs, constructs a customer group car purchase probability prediction model based on the city-customer group dimension, and generates a popularity heat map of the city-customer group; Step 4: At the real-time interaction layer, access the real-time behavior data of customers, use the BERT+BILSTM joint model to decode the natural language of customers, extract the current car purchase intention, generate a semantic embedding vector, and dynamically correct the short-term preference labels in the customer portrait; Step 5: At the recommendation generation layer, obtain candidate models according to the results of Steps 3 and 4, construct a scoring function, use the scoring function to score and rank the candidate models, select the top 3 ranking results, and then use the template filling method to generate personalized recommendation reasons and output; Step 6: At the training layer, establish a feedback mechanism, collect the feedback data of the prediction results of each layer model and the scoring function in the actual sales scenario, return the feedback data to the training sample set, and use the training sample set to optimize and train each layer model and the scoring function.

2. The method for assisting decision-making in automobile sales based on a large model according to claim 1, wherein: When implementing Step 1, the heterogeneous data specifically includes structured data, semi-structured data, and unstructured data; Structured data includes customer basic information, order records, car purchase history, vehicle model parameters, power types, drive modes, inventory status, price information, GDP data, purchase tax adjustment records, macro indicators of oil price time series, and the quantity and geographical distribution of sales service points in each city; Semi-structured data includes website and App buried point logs, test drive frequencies, vehicle model preferences, advertising display frequencies and time distributions in each city, and customer portraits; Unstructured data includes customer review texts, car purchase forum posts, policy announcements, news articles, and voice call transcription texts; The unified data management architecture specifically includes a data access layer, a data storage layer, a metadata management layer, and a time window unification layer: The data access layer is responsible for regularly collecting heterogeneous data by means of API calls, database scraping, and log transmission; The data storage layer includes a data lake and a data warehouse with a parallel structure. The data warehouse is used to store structured data, and the data lake is used to store semi-structured and unstructured data; The metadata management layer is responsible for uniformly extracting and marking the key meta-information of customer ID, timestamp, geographical location, and vehicle model ID; The time window unification layer is responsible for time-aligning all heterogeneous data according to the dimensions of day-week-month.

3. The method for assisting in making decisions on automobile sales based on a large model according to claim 2, wherein: When executing step 3, the macro layer model adopts a trend decomposition model and a deep time series correction model, takes the macro tags as the model input, and predicts the overall impact of the total national sales volume and policy changes; The meso layer model takes the meso tags as the model input, uses the graph neural network GNN to model the competition and cooperation relationships between cities, and predicts the vehicle model sales volume at the city level; The meso-micro layer model takes the meso-micro tags as the model input, uses a population clustering model and a softmax model to generate a city-customer group hot sales heat map; The micro layer model takes the micro tags as the input, uses a reinforcement learning model to model the customer click behavior sequence, and predicts the individualized car purchase probability and configuration preference.

4. The method for assisting in making automotive sales decisions based on a large model according to claim 3, wherein: When executing step 4, the real-time interaction layer real-time accesses the behavior data of customers' clicks, comparisons, or collections on the APP, mini-program, or official website, takes these behavior data as the input of the BERT+BILSTM joint model, extracts the current car purchase intention, and obtains the intention recognition result; The customer portrait tags are dynamically adjusted through a rule engine and a behavior weight update mechanism.

5. The automotive sales assistance decision-making method based on a large model according to claim 4, wherein: When executing step 5, it specifically includes the following steps: Step 5-1: According to the prediction results of each layer model in step 3, construct a candidate vehicle model set, which specifically includes the following steps: Step 5-1-1: Obtain the output result of the macro-level model , and retain the vehicle models with an increment reaching the preset threshold within the preset future period; Step 5-1-2: Obtain the output result of the mesoscopic layer model , in the city , select the set of the top 10 models in terms of sales volume prediction, and conduct the first screening of the model set according to the models retained in Step 5-1-1; Step 5-1-3: Obtain the results output by the mid- and micro-layer model, and based on the city-customer group hot sales heat map and the customer group's car purchase probability , conduct a secondary screening of the vehicle model set, and select the vehicle models that meet the following conditions: and ; wherein, is a preset threshold value, respectively represent city, customer group, and vehicle model; is the micro and macro heat score indicated in the city-customer group hot sales heat map; Step 5-1-4: Obtain the output result of the micro layer model, and re-screen the vehicle models obtained in step 5-1-3 according to the individualized car purchase probability and configuration preference to obtain the candidate vehicle model set; Step 5-2: Construct a scoring function, score and rank the candidate vehicle models in the candidate vehicle model set to obtain a candidate list. The scoring function is as follows: ； Among them, is the vehicle model score value, respectively represent the customer, vehicle model, and city, represents the preference matching score, represents the macro sales growth factor, represents the city popularity score, represents the intention adjustment factor obtained by the method in step 4, are all weights; Step 5-3: Select the top 3 models in the candidate list as the recommendation results, generate personalized recommendation reasons through template filling, and output.

6. The automotive sales assistance decision-making method based on a large model according to claim 4, characterized in that: When executing step 6, a feedback server is deployed in the training layer. The feedback server automatically records whether the customer accepts the recommendation, the reasons for clicking / rejecting, and returns them to the training samples; The feedback server compares the preset indexes of recommendation hit rate, conversion rate, and customer satisfaction, and tunes or rolls back each model.

Images