Method and system for dynamically grading transaction prediction probability based on user behavior data

Abstract

The embodiment of the invention discloses a transaction prediction probability dynamic grading method and system based on user behavior data. According to the embodiment of the invention, firstly, user attributes, behaviors, vehicle source attributes and sales interaction data are collected to construct a multi-dimensional feature set; performing semantic analysis on the call content by using an LLM model, extracting features of multiple dimensions such as transaction, vehicle condition, time and the like, and generating a normalized score; fusing the semantic features and the behavior features, and constructing a sample by taking whether a transaction is made as a target variable; an XGBoost model is adopted for training, and a user transaction probability is output; and finally, automatically determining a dynamic grading threshold value through inflection point detection based on probability distribution, and dividing the users into high, medium, low and unintentional grades. According to the method, the defects of high subjectivity, single feature dimension and rigid grading standard in the prior art are overcome, objective, accurate and self-adaptive customer intention evaluation is realized, and the sales resource distribution efficiency is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a dynamic classification method and system for transaction prediction probability based on user behavior data. Background Technology

[0002] In the used car retail industry, customer transaction prediction and intent grading are core aspects of sales management and lead follow-up. Currently, the industry's commonly used customer intent assessment methods mainly rely on manual grading based on experience-based rules. This means salespeople make judgments based on subjective experience such as communication content with customers, commitment to purchase time, and price range of the models they are interested in. In addition, some companies use rule-based scoring systems based on static indicators. These systems use preset weights to calculate the weighted average of structured behavioral data such as user browsing frequency, contact information, and store visits, and then categorize intent levels according to fixed score ranges. Some platforms also attempt to use simple machine learning models such as logistic regression or decision trees, using user behavioral characteristics as input to predict the probability of a transaction and grading based on manually set probability thresholds. However, these existing methods generally suffer from problems such as strong subjectivity in customer grading, reliance on limited structured data and a lack of semantic-level behavioral information mining, and fixed grading rules that are difficult to adapt to market changes. This results in insufficient prediction accuracy and poor grading stability, making it difficult to effectively support the precise allocation of sales resources and improve conversion efficiency. Summary of the Invention

[0003] To address this, embodiments of the present invention provide a dynamic grading method and system for predicting transaction probabilities based on user behavior data, in order to solve the technical problems of inaccurate predictions and rigid grading caused by existing used car customer intention assessment methods, which rely on subjective experience, lack the utilization of semantic features, and use fixed thresholds.

[0004] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0005] According to a first aspect of the present invention, a dynamic grading method for predicting transaction probability based on user behavior data is provided, the method comprising:

[0006] Collect user behavior data and construct a multi-dimensional feature set that includes user attribute features, behavior features, vehicle source attribute features, and sales interaction features;

[0007] Semantic analysis of the sales call content text is performed using a large language model (LLM), extracting semantic features from multiple preset dimensions and generating a normalized score for each dimension.

[0008] The multidimensional feature set is fused with the semantic features, aggregated according to the user dimension to form training samples, and whether a transaction is completed is used as the target variable;

[0009] The training samples are trained using a machine learning model to obtain a transaction prediction model, and the model is used to output the probability of a user completing a transaction.

[0010] Based on the transaction probability distribution of all users, the classification threshold is automatically determined by inflection point detection or business indicator analysis, and users are divided into multiple intention levels according to the classification threshold.

[0011] The transaction prediction model is deployed in the business system for real-time prediction and classification, and the model and classification thresholds are updated regularly.

[0012] Furthermore, user behavior data is collected, and a multi-dimensional feature set is constructed, including user attribute features, behavioral features, vehicle source attribute features, and sales interaction features, including:

[0013] The behavioral characteristics include the number of times browsing, the number of times leaving inquiries, the method of leaving inquiries, the concentration of behavior, and the number of car sources followed.

[0014] Furthermore, user behavior data is collected, and a multi-dimensional feature set is constructed that includes user attribute features, behavioral features, vehicle source attribute features, and sales interaction features. This also includes:

[0015] The sales interaction features include the number of follow-ups, communication duration, and call content text.

[0016] Furthermore, semantic analysis of the sales call content text is performed using a Large Language Model (LLM), extracting semantic features across multiple preset dimensions, and generating a normalized score for each dimension, including:

[0017] The multiple preset dimensions extracted by semantic analysis using the Large Language Model (LLM) include at least seven dimensions from the following: transaction-related, vehicle condition-related, time-related, location-related, subsequent behavior-related, negative features, and sentiment features.

[0018] Furthermore, the machine learning model is the XGBoost model.

[0019] Furthermore, based on the transaction probability distribution of all users, a grading threshold is automatically determined through inflection point detection or business indicator analysis. Users are then divided into multiple intention levels according to this grading threshold, including:

[0020] The automatic determination of the classification threshold through inflection point detection specifically involves: sorting users' transaction probabilities, plotting a cumulative revenue curve, and determining the probability threshold for segmenting different preference levels by analyzing the inflection points of the curve.

[0021] The multiple levels of intent include high-intent users, medium-intent users, low-intent users, and no-intent users.

[0022] Furthermore, the transaction prediction model is deployed in the business system for real-time prediction and classification, and the model and classification thresholds are updated periodically, including:

[0023] The periodic update of the model and the grading threshold specifically includes: adding newly generated transaction samples to the training set, retraining the transaction prediction model, and re-executing the dynamic grading step based on the new transaction probability distribution output by the model to update the grading threshold.

[0024] Furthermore, the transaction prediction model is deployed in the business system for real-time prediction and classification, and the model and classification thresholds are updated periodically, which also includes:

[0025] The calculated user conversion probability and intent level tags are synchronized to the customer relationship management (CRM) system so that sales staff can allocate resources and formulate follow-up strategies.

[0026] Furthermore, before performing semantic analysis on the sales call content text using a Large Language Model (LLM) to extract semantic features from multiple preset dimensions and generate a normalized score for each dimension, the process also includes: performing speech recognition on the sales call recording and converting it into call content text that can be processed by the Large Language Model (LLM).

[0027] According to a second aspect of the present invention, a dynamic hierarchical system for predicting transaction probability based on user behavior data is provided, the system comprising:

[0028] The data acquisition and feature construction module is used to collect user behavior data and construct a multi-dimensional feature set that includes user attribute features, behavior features, vehicle source attribute features, and sales interaction features.

[0029] The call content semantic feature extraction module is used to perform semantic analysis on the sales call content text using a large language model (LLM), extract semantic features in multiple preset dimensions, and generate a normalized score for each dimension.

[0030] The feature fusion and sample construction module is used to fuse the multidimensional feature set with the semantic features, aggregate them according to the user dimension to form training samples, and use whether a transaction is completed as the target variable.

[0031] The transaction probability prediction module is used to train the training samples using a machine learning model to obtain a transaction prediction model, and to output the user's transaction probability using the model.

[0032] The dynamic grading module is used to automatically determine the grading threshold based on the transaction probability distribution of all users, through inflection point detection or business indicator analysis, and divide users into multiple intention levels according to the grading threshold.

[0033] The application and update module is used to deploy the transaction prediction model in the business system for real-time prediction and classification, and to update the model and classification thresholds periodically.

[0034] The embodiments of the present invention have the following advantages:

[0035] This invention first collects user attributes, behaviors, vehicle attributes, and sales interaction data to construct a multi-dimensional feature set; then, it uses an LLM model to perform semantic analysis on the call content, extracting features from multiple dimensions such as transaction, vehicle condition, and time, and generating normalized scores; finally, it fuses these semantic features with behavioral features, constructing samples with whether a transaction is completed as the target variable; an XGBoost model is used for training, outputting the user's transaction probability; and finally, based on the probability distribution, it automatically determines dynamic grading thresholds through inflection point detection, classifying users into high, medium, low, and no-intent levels. This invention overcomes the shortcomings of existing technologies, such as strong subjectivity, single feature dimensions, and rigid grading standards, achieving objective, accurate, and adaptive customer intent assessment, effectively improving the efficiency of sales resource allocation. Attached Figure Description

[0036] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0037] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0038] Figure 1 A schematic diagram of the logical structure of a dynamic hierarchical system for predicting transaction probability based on user behavior data, provided in an embodiment of the present invention;

[0039] Figure 2 This is a flowchart illustrating a dynamic grading method for predicting transaction probability based on user behavior data, provided in an embodiment of the present invention. Detailed Implementation

[0040] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] In the used car retail industry, customer transaction forecasting and intent classification are crucial aspects of sales management and lead follow-up. Currently, the mainstream approaches mainly include the following categories:

[0042] 1. Manual grading method based on empirical rules

[0043] Sales staff manually categorize customer intentions based on their subjective experience, including communication with customers, commitment to purchase timelines, and interest in specific price ranges. For example, a sales consultant might use established rules such as "expecting to purchase within 3 days" as a high level of intention and "no budget at the moment" as a low level of intention to assess customer interest.

[0044] 2. Rule-based scoring system based on static indicators

[0045] Some CRM systems use structured behavioral data such as pageview counts, number of inquiries, whether customers add each other on WeChat, and whether they visit the store, assigning fixed weights to calculate a comprehensive score. Customers are then categorized into high, medium, and low-intent customers based on their score range. These methods rely on preset thresholds and weighting formulas, resulting in fixed scoring standards.

[0046] 3. Prediction methods based on simple statistical or machine learning models

[0047] Some platforms attempt to use logistic regression or decision tree models, taking user behavior characteristics (such as page views and contact information) as input, outputting the probability of a transaction, and manually setting probability thresholds for grading. However, these models generally rely on limited structured data and lack the ability to understand behavior at the semantic level.

[0048] While existing tiered methods based on sales' subjective judgment or estimated sales cycle have some operability in actual business, they still have significant shortcomings:

[0049] 1. The customer segmentation process is highly subjective.

[0050] Traditional methods for assessing customer intent rely primarily on sales staff experience or manually set rules, lacking unified and objective quantitative standards. They are easily influenced by human factors, resulting in poor stability and repeatability of the results.

[0051] 2. Lack of multi-dimensional and semantic data utilization

[0052] Existing solutions mainly rely on structured behavioral data (number of views, methods of leaving inquiries, etc.), failing to fully explore the semantic information contained in the sales call content, such as the timing of car purchase, price sensitivity, and vehicle condition concerns, resulting in the predictive model failing to accurately reflect the user's true car purchase intention.

[0053] 3. The hierarchical rules lack adaptability.

[0054] Customer segmentation thresholds are usually fixed values ​​set by human experience and cannot be dynamically adjusted based on the model output distribution or the inflection point of actual transaction conversion. This makes it difficult to adapt to market changes and causes the segmentation results to deviate from the actual transaction probability.

[0055] To address the aforementioned technical problems of inaccurate predictions and rigid hierarchical structures caused by reliance on subjective experience, lack of semantic feature utilization, and fixed thresholds.

[0056] LLM (Large Language Model) is a natural language processing model trained on large-scale text data, capable of understanding, generating, and analyzing natural language. In this invention, LLM is used to perform semantic understanding of sales and user conversations, extracting seven dimensions of features: "transaction, vehicle condition, time, location, subsequent behavior, negative sentiment, and emotion," and generating corresponding normalized scores.

[0057] XGBoost (eXtreme Gradient Boosting) is an efficient gradient boosting tree algorithm suitable for classification and regression tasks. In this invention, XGBoost is used to construct a binary classification model, using "whether a transaction is completed" as the target variable, and combining multi-dimensional user features to predict the probability of a transaction.

[0058] CRM (Customer Relationship Management) is a software system used to manage interactions between businesses and customers, sales opportunities, communication records, and customer information. In this invention, the CRM system receives user conversion probabilities and rating labels output by the model, thereby automating sales resource allocation and follow-up strategies.

[0059] refer to Figure 1 This invention discloses a dynamic hierarchical system for predicting transaction probability based on user behavior data. The system includes: a data acquisition and feature construction module 1; a call content semantic feature extraction module 2; a feature fusion and sample construction module 3; a transaction probability prediction module 4; a dynamic hierarchical module 5; and an application and update module 6.

[0060] Corresponding to the aforementioned dynamic grading system for transaction prediction probability based on user behavior data, this invention also discloses a dynamic grading method for transaction prediction probability based on user behavior data. The following details a method for dynamic grading of transaction prediction probability based on user behavior data disclosed in this invention, in conjunction with the aforementioned dynamic grading system for transaction prediction probability based on user behavior data.

[0061] This invention discloses a dynamic grading method for transaction prediction probability based on user behavior data, aiming to address the problems of strong subjectivity, insufficient feature dimensions, and lack of adaptability in grading standards in existing technologies, in order to achieve the following technical objectives:

[0062] 1. Overcome the subjectivity of customer segmentation

[0063] By building machine learning models, we can achieve quantitative prediction of users' transaction intentions, reduce interference from human judgment, and improve the objectivity of the classification.

[0064] 2. Enrich the model's feature dimensions and introduce semantic understanding capabilities.

[0065] By integrating user attributes, behaviors, vehicle source characteristics, and semantic information from call content, and utilizing a Large Language Model (LLM) to extract multi-dimensional semantic features, the model can understand the user's true intention to purchase a car.

[0066] 3. Implement adaptive dynamic updates for customer segmentation rules.

[0067] By performing inflection point detection or business indicator analysis on the transaction probability distribution output by the model, the segmentation threshold is automatically determined, enabling dynamic optimization and adaptive updating of customer segmentation.

[0068] refer to Figure 2 This invention discloses a dynamic grading method for transaction prediction probability based on user behavior data, comprising:

[0069] S1. Data Acquisition and Feature Construction

[0070] Data was collected from user behavior logs, CRM systems, and sales call records of used car retail platforms to construct a multi-dimensional feature set, including:

[0071] User attribute characteristics include: channel source, registration time, region, device type, etc.

[0072] Behavioral characteristics include the number of views, the number of inquiries left, the method of leaving inquiries, the concentration of behavior (i.e. the degree of concentration of visits during the active period), and the number and characteristics of the vehicles followed.

[0073] Vehicle source attributes based on user behavior: Brand, price range, age, and model class of the vehicles that users are interested in or inquiring about;

[0074] Sales interaction characteristics include the number of follow-ups, communication duration, and call content.

[0075] S2. Call content feature extraction

[0076] After converting the sales call recordings to text using speech recognition, a Large Language Model (LLM) is used to perform semantic analysis on the text, extracting seven semantic dimension features from the call text and generating a normalized score (range [0,1]) for each dimension:

[0077] Transaction-related features include: transaction process, payment methods, required materials, price-related aspects, and transfer-related aspects.

[0078] Vehicle condition-related characteristics: including vehicle details, historical condition, and warranty issues;

[0079] Time-related characteristics include arrival time and when the user is available.

[0080] Location-related features: These include store address, distance, and accessibility.

[0081] Subsequent behavioral characteristics: such as whether they indicated they were visiting the store, or added contact information;

[0082] Negative characteristics: such as expressions like "already purchased a car" or "not considering it for the time being";

[0083] Emotional characteristics: Analyze customer emotions (positive, neutral, negative).

[0084] The above features, after being standardized, are used as part of the model input features.

[0085] S3, Feature Fusion and Sample Construction

[0086] User behavior features and semantic features are aggregated according to user dimensions to form samples, and training sets and validation sets are constructed, with "whether a transaction is completed" as the target variable (Y=1 completed, Y=0 not completed).

[0087] S4. Modeling and Transaction Probability Prediction

[0088] The XGBoost binary classification model is used to train the samples, and the tree model is used to automatically learn the nonlinear relationships and interaction effects of various features. The model outputs the transaction probability P(deal)∈[0,1) for each user.

[0089] S5, Dynamic Hierarchical Strategy

[0090] The probability of conversion for all users is ranked. Combining the business conversion rate and cumulative revenue curve, an inflection point detection algorithm or other distribution analysis methods are used to determine the level classification thresholds (e.g., P1, P2, P3), and users are then divided into:

[0091] High-intent users (P ≥ P1);

[0092] Interested users (P2 ≤ P < P1);

[0093] Low-intent users (P3 ≤ P < P2);

[0094] Uninterested users (P < P3).

[0095] The threshold can be automatically updated periodically based on model training results or business performance, achieving dynamic adaptation of the grading standard.

[0096] S6, Real-time Prediction and Applications

[0097] Deploy the trained model to the business system;

[0098] Receive real-time updates on user behavior and the latest interactions.

[0099] Automatically calculate the probability of a sale and the corresponding rating label, and synchronize them to the CRM system;

[0100] Sales personnel can allocate resources and follow-up strategies based on their priority level.

[0101] S7, Performance Monitoring and Iterative Optimization

[0102] Regularly monitor the conversion rate, ROI, and other metrics at each level;

[0103] The latest transaction samples are used for model retraining, forming a closed-loop mechanism for continuous optimization.

[0104] The embodiments of the present invention have the following advantages:

[0105] 1) Improve the accuracy of transaction prediction: By utilizing multi-dimensional features and implicit information extracted from large models, the prediction effect is significantly improved.

[0106] 2) Data-driven user segmentation: Replacing subjective experience with scientific and quantifiable user segmentation.

[0107] 3) Improve sales efficiency: accurately identify high-intent customers, rationally allocate sales resources, and reduce time and cost waste.

[0108] 4) High scalability: The model and features can be iterated and updated as the business develops, supporting more feature dimensions and new behavioral data.

[0109] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A dynamic hierarchical method for predicting transaction probability based on user behavior data, characterized in that, The method includes: Collect user behavior data and construct a multi-dimensional feature set that includes user attribute features, behavior features, vehicle source attribute features, and sales interaction features; Semantic analysis of the sales call content text is performed using a large language model (LLM), extracting semantic features from multiple preset dimensions and generating a normalized score for each dimension. The multidimensional feature set is fused with the semantic features, aggregated according to the user dimension to form training samples, and whether a transaction is completed is used as the target variable; The training samples are trained using a machine learning model to obtain a transaction prediction model, and the model is used to output the user's transaction probability. Based on the transaction probability distribution of all users, the classification threshold is automatically determined by inflection point detection or business indicator analysis, and users are divided into multiple intention levels according to the classification threshold. The transaction prediction model is deployed in the business system for real-time prediction and classification, and the model and classification thresholds are updated regularly.

2. The dynamic grading method for transaction prediction probability based on user behavior data as described in claim 1, characterized in that, Collect user behavior data and construct a multi-dimensional feature set containing user attribute features, behavioral features, vehicle source attribute features, and sales interaction features, including: The behavioral characteristics include the number of times browsing, the number of times leaving inquiries, the method of leaving inquiries, the concentration of behavior, and the number of car sources followed.

3. The dynamic grading method for transaction prediction probability based on user behavior data as described in claim 2, characterized in that, Collect user behavior data and construct a multi-dimensional feature set that includes user attribute features, behavioral features, vehicle source attribute features, and sales interaction features, and also includes: The sales interaction features include the number of follow-ups, communication duration, and call content text.

4. The dynamic grading method for transaction prediction probability based on user behavior data as described in claim 1, characterized in that, Semantic analysis of the sales call content text was performed using a Large Language Model (LLM), extracting semantic features across multiple preset dimensions and generating a normalized score for each dimension, including: The multiple preset dimensions extracted by semantic analysis using the Large Language Model (LLM) include at least seven dimensions from the following: transaction-related, vehicle condition-related, time-related, location-related, subsequent behavior-related, negative features, and sentiment features.

5. The dynamic grading method for transaction prediction probability based on user behavior data as described in claim 1, characterized in that, The machine learning model is the XGBoost model.

6. The dynamic grading method for transaction prediction probability based on user behavior data as described in claim 1, characterized in that, Based on the transaction probability distribution of all users, a tiered threshold is automatically determined through inflection point detection or business indicator analysis. Users are then divided into multiple intention levels according to this threshold, including: The automatic determination of the classification threshold through inflection point detection specifically involves: sorting users' transaction probabilities, plotting a cumulative revenue curve, and determining the probability threshold for segmenting different preference levels by analyzing the inflection points of the curve. The multiple levels of intent include high-intent users, medium-intent users, low-intent users, and no-intent users.

7. The dynamic grading method for transaction prediction probability based on user behavior data as described in claim 1, characterized in that, The transaction prediction model is deployed in the business system for real-time prediction and classification, and the model and classification thresholds are updated periodically, including: The periodic update of the model and the grading threshold specifically includes: adding newly generated transaction samples to the training set, retraining the transaction prediction model, and re-executing the dynamic grading step based on the new transaction probability distribution output by the model to update the grading threshold.

8. The dynamic grading method for transaction prediction probability based on user behavior data as described in claim 7, characterized in that, The transaction prediction model is deployed in the business system for real-time prediction and classification, and the model and classification thresholds are updated regularly. The system also includes: The calculated user conversion probability and intent level tags are synchronized to the customer relationship management (CRM) system so that sales staff can allocate resources and formulate follow-up strategies.

9. The dynamic grading method for transaction prediction probability based on user behavior data as described in claim 1, characterized in that, Before performing semantic analysis on the sales call content text using a Large Language Model (LLM), extracting semantic features from multiple preset dimensions, and generating a normalized score for each dimension, the process further includes: performing speech recognition on the sales call recording and converting it into call content text that can be processed by the Large Language Model (LLM).

10. A dynamic hierarchical system for predicting transaction probability based on user behavior data, characterized in that, The system includes: The data acquisition and feature construction module is used to collect user behavior data and construct a multi-dimensional feature set that includes user attribute features, behavior features, vehicle source attribute features, and sales interaction features. The call content semantic feature extraction module is used to perform semantic analysis on the sales call content text using a large language model (LLM), extract semantic features in multiple preset dimensions, and generate a normalized score for each dimension. The feature fusion and sample construction module is used to fuse the multidimensional feature set with the semantic features, aggregate them according to the user dimension to form training samples, and use whether a transaction is completed as the target variable. The transaction probability prediction module is used to train the training samples using a machine learning model to obtain a transaction prediction model, and to output the user's transaction probability using the model. The dynamic grading module is used to automatically determine the grading threshold based on the transaction probability distribution of all users, through inflection point detection or business indicator analysis, and divide users into multiple intention levels according to the grading threshold. The application and update module is used to deploy the transaction prediction model in the business system for real-time prediction and classification, and to update the model and classification thresholds periodically.

Images