Flow distribution method and device based on behavior data

Abstract

The invention discloses a flow distribution method and device based on behavior data. A server side method comprises the following steps: collecting historical behavior data of each user in an automobile financial platform; obtaining multiple types of system records of each user to obtain current user state information; inputting the historical behavior data and the current user state information into a preset flow distribution plug-in, and outputting a probability value of each preset intention label corresponding to each user; the preset flow distribution plug-in comprises a pre-trained customer intention prediction model generated based on a machine learning algorithm; and taking the preset intention label corresponding to the maximum probability value as a final intention label of each user, distributing each user to a preset flow pool corresponding to the final intention label of each user, and continuing to execute the step of collecting the historical behavior data of each user in the automobile financial platform. According to the invention, the method can accurately recognize the customer with the vehicle purchase financing intention, reduces the invalid flow distribution, and improves the resource utilization efficiency.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a traffic allocation method and apparatus based on behavioral data. Background Technology

[0002] In an environment conducive to boosting consumption, financial institutions face a massive potential customer base, including many with a strong demand for auto finance products. For example, a customer might browse detailed configuration information for several popular car models on an auto finance platform and frequently communicate with customer service online, demonstrating a high level of interest in car financing. However, accurately identifying customers with genuine intent to finance their car purchases from this large pool of potential clients has become a critical issue that financial institutions urgently need to address.

[0003] In related technologies, auto finance companies commonly use traditional SCRM (Social Customer Relationship Management) systems for customer management and traffic allocation. These systems rely on manually set rules, such as simple sorting and filtering based on basic data like registration time and pageview count, to identify potential high-intent customers. However, manually set rules lack flexibility and fail to fully consider the complexity and diversity of customer behavior data, resulting in a large amount of invalid traffic being allocated and causing significant waste of company resources. Summary of the Invention

[0004] This application provides a traffic allocation method and apparatus based on behavioral data. To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general description, nor is it intended to identify key / important components or describe the scope of protection of these embodiments. Its sole purpose is to present some concepts in a simple form as a prelude to the detailed description that follows.

[0005] In a first aspect, embodiments of this application provide a traffic allocation method based on behavioral data, applied to a server, the method comprising: Historical behavioral data of each user on the auto finance platform is collected according to a preset cycle; It interacts in real time with the telemarketing system, the post-loan system, and the treasury system to obtain multiple system records for each user and to obtain the current user status information. Input historical behavior data and current user status information into the preset traffic allocation plugin, and output the probability value of each preset intention tag corresponding to each user; the preset traffic allocation plugin includes a pre-trained customer intention prediction model generated based on machine learning algorithms; The preset intention tag corresponding to the maximum probability value is used as the final intention tag for each user, and each user is assigned to a preset traffic pool corresponding to each user's final intention tag. The process of collecting historical behavior data of each user in the auto finance platform continues.

[0006] Optionally, the preset traffic allocation plugin includes a preprocessing component, a feature acquisition component, a feature fusion component, and a probability value generation component, with the pre-trained customer intention prediction model integrated into the probability value generation component; Historical behavior data and current user status information are input into a pre-trained customer intent prediction model, which outputs the probability values ​​of each user's corresponding preset intent tags, including: The preprocessing component preprocesses historical behavior data and user status information; The feature acquisition component extracts historical behavioral features and current user status features related to customer intent from preprocessed historical behavioral data and user status information; The feature fusion component fuses historical behavioral features and current user state features to obtain a comprehensive feature vector; The probability value generation component generates probability values ​​for each user's preset intention tags based on a comprehensive feature vector and a pre-trained customer intention prediction model.

[0007] Optionally, the comprehensive feature vector is a multi-dimensional array, where each dimension corresponds to a feature value; the pre-trained customer intent prediction model includes an input layer, a quantization layer, and an output layer; Based on the comprehensive feature vector and combined with a pre-trained customer intent prediction model, probability values ​​for each preset intent tag corresponding to each user are generated, including: The input layer receives the comprehensive feature vector. The total number of neurons in the input layer is the same as the dimension of the comprehensive feature vector, and each neuron receives one feature value of the comprehensive feature vector. The quantization layer loads pre-learned weight matrices and bias vectors, which are used to describe the relationship between features and intention labels. The quantization layer uses a weight matrix and a bias vector to perform a linear transformation on the multidimensional array received by the input layer, and obtains the linear transformation result. The output layer uses a preset Softmax function to convert the linear transformation result into a probability distribution, outputting the probability value of each preset intention tag for each user; the preset Softmax function is:

[0008] in, It is the first The probability value of a preset intention tag. This is the total number of preset intention tags. It is the first The result of linear transformation of each neuron.

[0009] Optionally, a pre-trained customer intent prediction model may be generated by following these steps: The system collects browsing history, search keywords, and online communication records of registered users within a preset historical time period from the auto finance platform to obtain sample behavioral data for each registered user. Collect various system record information from the telemarketing system, loan post-loan system, and treasury system to obtain sample user status information for each registered user. Based on a pre-defined set of intention tags, sample behavior data of each registered user, and sample user status information, multiple model training samples are generated. A first customer intention prediction model is created using machine learning algorithms, and a model structure is defined in the first customer intention prediction model to obtain a second customer intention prediction model. Based on the training samples of each model and the second customer intention prediction model, generate the model loss value; If the model loss value reaches its minimum, a pre-trained customer intention prediction model is generated; or if the model loss value does not reach its minimum, the process of inputting each model training sample into the second customer intention prediction model continues until the model loss value reaches its minimum.

[0010] Optionally, the preset intention tag set contains each preset intention tag and the priority of each preset intention tag; Based on a pre-defined set of intention tags, sample behavior data of each registered user, and sample user status information, multiple model training samples are generated, including: Establish a positive mapping strategy for each preset intention label; Based on the priority of each preset intention tag, the positive mapping strategies of each preset intention tag are sorted from high to low priority to obtain a strategy tree; Extract each field name and its attribute value from the sample behavior data and sample user status information of each registered user; Expand each field name and its attribute value into an atomic field dictionary to obtain the atomic field dictionary for each registered user; Based on the atomic field dictionary and strategy tree of each registered user, determine the optimal preset intention tag for each registered user; The optimal preset intention label for each registered user is labeled on a tuple consisting of sample behavior data and sample user status information of each registered user, resulting in multiple model training samples.

[0011] Optionally, based on the atomic field dictionary and strategy tree of each registered user, determine the optimal preset intention tag for each registered user, including: A positive mapping strategy is traversed in the strategy tree according to priority order for a predefined intention label. The positive mapping strategy includes a hard condition for assigning the highest confidence score, multiple sub-conditions and their weights, and a target threshold. Determine if the atomic field dictionary for each registered user contains an attribute value that meets a hard condition; If so, use the preset intention tags traversed as the optimal preset intention tags for each registered user and stop traversing; If not, determine each target sub-condition from multiple sub-conditions that satisfies the atomic field dictionary of each registered user; sum the weights corresponding to each target sub-condition; if the summation result is greater than or equal to a preset threshold, take the traversed preset intention label as the optimal preset intention label for each registered user and stop traversing; or if the summation result is less than the preset threshold, continue executing the steps of traversing the forward mapping strategy of a preset intention label in the strategy tree according to priority order until the optimal preset intention label for each registered user is determined and stop traversing.

[0012] Optionally, based on each model training sample and the second customer intention prediction model, a model loss value is generated, including: Preprocess each model training sample; Extract sample behavioral features and sample user status features related to customer intent from each preprocessed model training sample; The sample behavior features and sample user state features are fused to obtain the sample comprehensive feature vector of each model training sample; Input the sample's comprehensive feature vector into the second customer intention prediction model and output the model loss value.

[0013] Optionally, the model structure of the second customer intention prediction model includes an input layer, a quantization layer, and an output layer; The sample composite feature vector is input into the second customer intention prediction model, and the model loss value is output, including: The input layer receives the comprehensive feature vector of the samples; The quantization layer uses a pre-learned weight matrix and bias vector to describe the relationship between features and intention labels to perform a linear transformation on the sample integrated feature vector received by the input layer, and obtains the sample linear transformation result. The output layer uses a preset Softmax function to convert the linear transformation result of the samples into a probability distribution, thereby obtaining the probability value of each model training sample for each preset intention label; The output layer uses a loss function and the probability value of each training sample for each preset intention label to calculate and output the model loss value; the expression for the loss function is:

[0014] in, This represents the model loss value. It is the total number of training samples for the model. This is the total number of preset intention tags. It is the first One-Hot encoding of the true intention labels of the model training samples, if the model training samples Belongs to the If there are preset intention tags, then Otherwise, it is 0. It is the first The training samples of the model belong to the first... The probability value of a preset intention tag.

[0015] Optionally, a preset traffic allocation plugin can be generated by following these steps: Create a preprocessing component for preprocessing historical behavior data and user status information; Create a feature acquisition component to obtain historical behavioral characteristics and current user status characteristics related to customer intent; Create a feature fusion component that combines historical behavioral features and current user state features; Create a probability value generation component that generates probability values ​​for each preset intention tag corresponding to each user; Obtain a pre-trained customer intent prediction model; The pre-trained customer intent prediction model is integrated into the probability value generation component to obtain the final probability value generation component. The final probability value generation component, preprocessing component, feature acquisition component, and feature fusion component are integrated to obtain the preset traffic allocation plugin.

[0016] Secondly, embodiments of this application provide a traffic allocation device based on behavioral data, the device comprising: The historical behavior data collection module is used to collect historical behavior data of each registered user in the auto finance platform according to a preset period. The current user status information acquisition module is used to interact with the telemarketing system, loan post-loan system and financial system in real time to obtain multiple system records for each user and obtain the current user status information. The probability value output module is used to input historical behavior data and current user status information into the preset traffic allocation plugin and output the probability value of each preset intention tag corresponding to each user; the preset traffic allocation plugin includes a pre-trained customer intention prediction model generated based on machine learning algorithms; The traffic allocation module is used to take the preset intention tag corresponding to the maximum probability value as the final intention tag of each user, and allocate each user to the preset traffic pool corresponding to each user's final intention tag, and continue to execute the step of collecting historical behavior data of each registered user in the auto finance platform.

[0017] The technical solutions provided in this application embodiment may include the following beneficial effects: In this embodiment, on the one hand, a pre-trained customer intent prediction model generated using machine learning algorithms is employed. This model not only considers historical behavioral data of customers on the auto finance platform but also integrates current user status information from the telemarketing system, post-loan system, and treasury system in real time. This enables the system to accurately allocate traffic based on the complexity and diversity of user data, reducing the allocation of invalid traffic and thus improving the utilization rate of enterprise resources. On the other hand, by labeling each registered user's optimal preset intent tag onto a tuple composed of sample behavioral data and sample user status information, the optimal preset intent tag is determined by introducing a positive mapping strategy, a strategy tree, and an atomic field dictionary. This labeling method is highly accurate, avoids dirty data caused by manual labeling, provides high-quality training samples for model training, and thus improves the model's prediction accuracy.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] Figure 1 This is a schematic diagram of a traffic allocation method based on behavioral data provided in an embodiment of this application; Figure 2 This is a structural diagram of a preset traffic allocation plugin provided in an embodiment of this application; Figure 3 This is a schematic block diagram illustrating a traffic allocation process based on behavioral data, provided in an embodiment of this application. Figure 4 This is a flowchart illustrating a customer intent prediction model training method provided in an embodiment of this application; Figure 5 This is a structural diagram of the model structure of a second customer intention prediction model provided in an embodiment of this application; Figure 6 This is a schematic flowchart of a model training process provided in this application; Figure 7This is a schematic diagram of the structure of a traffic allocation device based on behavioral data provided in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0021] The following description and accompanying drawings fully illustrate specific embodiments of this application to enable those skilled in the art to practice them.

[0022] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0023] In the following description, when referring to the accompanying drawings, the same numbers in different drawings denote the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0024] In the description of this application, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances. Furthermore, in the description of this application, unless otherwise stated, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0025] Currently, auto finance companies generally use traditional SCRM (Social Customer Relationship Management) systems for customer management and traffic allocation.

[0026] The applicant of this application recognizes that the system relies on manually set rules, such as simple sorting and filtering based on basic data like customer registration time and number of visits, to identify potential high-intent customer groups. However, manually set rules lack flexibility and fail to fully consider the complexity and diversity of customer behavior data, resulting in a large amount of invalid traffic being allocated and causing serious waste of enterprise resources.

[0027] To address the aforementioned issues, this application provides a traffic allocation method and apparatus based on behavioral data to resolve the problems existing in the related technologies. In the embodiments of this application, on the one hand, a pre-trained customer intention prediction model generated using a machine learning algorithm is employed. This model not only considers historical behavioral data of customers on the auto finance platform but also integrates current user status information from the telemarketing system, post-loan system, and treasury system in real time. This enables the system to achieve accurate traffic allocation by combining the complexity and diversity of user data, reducing the allocation of invalid traffic and thus improving the utilization rate of enterprise resources. On the other hand, by labeling each registered user's optimal preset intention tag onto a tuple composed of sample behavioral data and sample user status information, the optimal preset intention tag is determined by introducing a positive mapping strategy, a strategy tree, and an atomic field dictionary. This labeling method has high accuracy, avoids dirty data caused by manual labeling, provides high-quality training samples for model training, and thus improves the model's prediction accuracy. Exemplary embodiments are described in detail below.

[0028] The following will be combined with the appendix Figure 1 - Appendix Figure 6 This application provides a detailed description of the traffic allocation method based on behavioral data, as provided in its embodiments. This method can be implemented using a computer program and can run on a traffic allocation device based on behavioral data and the von Neumann architecture. The computer program can be integrated into an application or run as a standalone utility application.

[0029] Please see Figure 1 This is a flowchart illustrating a traffic allocation method based on behavioral data, provided in an embodiment of this application, applied to the server side. For example... Figure 1 As shown, the method in this application embodiment includes the following steps: S101 collects historical behavior data of each user on the auto finance platform according to a preset cycle; The preset period refers to a pre-defined time interval, which can be determined according to actual needs, such as every ten minutes, daily, or weekly. The preset period ensures the frequency and timeliness of data updates, guaranteeing the system can promptly obtain the latest user behavior information for traffic allocation. The auto finance platform is an online platform providing auto finance services. Users can apply for auto loans, inquire about financial products, and make repayments through this platform. Each user is one of all users registered and using the services on the auto finance platform. The platform records users' past behavior, including browsing history, search keywords, submitted forms, clicked links, and inquiry records.

[0030] In some embodiments of this application, the system periodically extracts each user's behavior records (such as browsing pages, searching keywords, submitting forms, consultation records, etc.) from the database of the auto finance platform over a period of time.

[0031] For example, if the preset cycle is weekly, then from midnight last Monday to midnight this Monday, all user activity records for these 7 days will be extracted from the auto finance platform's database, including: page browsing history, search keywords, and online communication records.

[0032] S102 interacts in real time with the telemarketing system, the post-loan system, and the financial system to obtain multiple system records for each user and get the current user status information. The telemarketing system records user actions such as inquiries, appointments, and purchases made via telephone. The post-loan system manages loan-related matters, recording user repayment history, overdue payments, and loan status. The treasury system manages finances and funds, recording user funding needs, cash flow, and account balances. By interacting with these three systems in real time, the server can obtain multi-dimensional user information, gaining a comprehensive understanding of user behavior and status.

[0033] Among them, various system records include telephone inquiry records, repayment records, and funding needs.

[0034] It should be noted that the acquisition of privacy data in the telemarketing system, post-loan system, and treasury system complies with relevant laws and regulations and can only be carried out with the authorization of the platform.

[0035] In some embodiments of this application, the system interacts in real time with the telemarketing system, post-loan system, and treasury system to obtain each user's telephone consultation records, repayment records, and funding needs information, thereby forming comprehensive status information for each user. This information includes multi-dimensional data such as the user's concerns, credit status, and funding needs, providing comprehensive data support for subsequent traffic allocation and customer intention prediction.

[0036] For example, the server interacts with the telemarketing system to retrieve user A's phone inquiry record at 3 PM the previous day, which concerned the interest rate and repayment method of a certain car loan. It can also retrieve user B's phone inquiry record at 10 AM the previous day, which concerned details of car insurance. The server interacts with the loan management system to retrieve user A's repayment history for the past 12 months, showing 11 months of on-time payments and 1 month of 1-day delinquency. It can also retrieve user B's repayment history for the past 6 months, showing 5 months of on-time payments and 1 month of 3-day delinquency. Finally, the server interacts with the financial system to retrieve user A's funding needs information, showing an application for a 100,000 yuan car loan with a 3-year term. It can also retrieve user B's funding needs information, showing an application for a 50,000 yuan car loan with a 2-year term.

[0037] At this moment, User A's current status information includes browsing high-end auto loan product pages, calling to inquire about loan interest rates and repayment methods, having a good repayment history, and applying for a 100,000 yuan auto loan. User B's current status information includes browsing used car loan product pages, calling to inquire about car insurance details, having one overdue payment in their repayment history, and applying for a 50,000 yuan auto loan.

[0038] S103, input historical behavior data and current user status information into the preset traffic allocation plugin, and output the probability value of each preset intention tag corresponding to each user; the preset traffic allocation plugin includes a pre-trained customer intention prediction model generated based on machine learning algorithms; Among them, for example Figure 2 As shown, the preset traffic allocation plugin includes a preprocessing component, a feature acquisition component, a feature fusion component, and a probability value generation component. The pre-trained customer intention prediction model is integrated into the probability value generation component.

[0039] In some embodiments of this application, the specific process of inputting historical behavior data and current user status information into a pre-trained customer intention prediction model and outputting the probability values ​​of each preset intention tag corresponding to each user includes: a preprocessing component preprocessing historical behavior data and user status information; a feature acquisition component acquiring historical behavior features and current user status features related to customer intention from the preprocessed historical behavior data and user status information; a feature fusion component fusing the historical behavior features and current user status features to obtain a comprehensive feature vector; and a probability value generation component generating the probability values ​​of each preset intention tag corresponding to each user based on the comprehensive feature vector and combined with the pre-trained customer intention prediction model.

[0040] The preprocessing component ensures that the data input to the model is in a consistent format, improving the model's accuracy and efficiency. Preprocessing operations within the preprocessing component include data cleaning and formatting. The feature extraction component reduces data dimensionality by extracting key features, improving model training efficiency and predictive performance. The feature fusion component provides a more comprehensive user profile by fusing features from different sources, helping the model more accurately predict user behavior and intentions. The probability value generation component generates probability values ​​for different intention labels for each user based on the model's predictions. The pre-trained customer intention prediction model is a pre-trained AI model, specifically a machine learning or deep learning model, used to predict user intentions based on the input feature vector.

[0041] In one possible implementation, historical behavioral data for each user is collected from the database of the auto finance platform. Through real-time interaction with the telemarketing system, post-loan system, and treasury system, each user's telephone consultation records, repayment records, and funding needs information are obtained. A preprocessing component is used to clean and format the collected data, removing invalid data, filling missing values, standardizing the data format, and normalizing the data. A feature acquisition component is used to extract historical behavioral features and current user status features related to customer intent from the preprocessed data. For example, the frequency of a user browsing auto loan products, the concentration of search keywords, the topic of the most recent telephone consultation, and repayment status over the past month are extracted. A feature fusion component is used to fuse the historical behavioral features and current user status features to obtain a comprehensive feature vector. This comprehensive feature vector is input into a pre-trained customer intent prediction model. The model outputs the probability value of different preset intent labels for each user.

[0042] For example, User A's historical behavioral data includes: 10 visits to auto loan product pages and 5 searches for the keyword "low-interest loan." Current status information includes: phone inquiry records (inquiries about the interest rate and repayment method of a certain auto loan), repayment history (good repayment record for the past 12 months), and funding needs (application for a 100,000 yuan auto loan). User A's features are extracted as follows: 10 visits to auto loan product pages, 5 searches for the keyword "low-interest loan," phone inquiries about loan interest rates and repayment methods, good repayment history, and a funding need of 100,000 yuan. User A's features are fused into a comprehensive feature vector [10, 5, 1, 1, 10], where 1 represents a good phone inquiry and repayment history, and 0 represents overdue payments. User A's comprehensive feature vector [10, 5, 1, 1, 10] is input into a pre-trained customer intention prediction model, and the model outputs the probability values ​​for different preset intention labels.

[0043] The comprehensive feature vector is a multi-dimensional array, where each dimension corresponds to a feature value. The pre-trained customer intent prediction model includes an input layer, a quantization layer, and an output layer, for example... Figure 5 As shown.

[0044] In some embodiments of this application, the specific process of generating probability values ​​for each preset intention tag corresponding to each user based on a comprehensive feature vector and a pre-trained customer intention prediction model includes: the input layer receiving a comprehensive feature vector, wherein the total number of neurons in the input layer is the same as the dimension of the comprehensive feature vector and each neuron receives a feature value of the comprehensive feature vector; the quantization layer loading a pre-learned weight matrix and bias vector, wherein the weight matrix and bias vector are used to describe the relationship between features and intention tags; the quantization layer using the weight matrix and bias vector to perform a linear transformation on the multidimensional array received by the input layer to obtain a linear transformation result; and the output layer using a preset Softmax function to convert the linear transformation result into a probability distribution and outputting the probability values ​​for each preset intention tag corresponding to each user.

[0045] The preset Softmax function is:

[0046] in, It is the first The probability value of a preset intention tag. This is the total number of preset intention tags. It is the first The result of linear transformation of each neuron.

[0047] The weight matrix and bias vector are parameters learned by the neural network during training, used to describe the relationship between input features and output labels. The linear transformation is a linear transformation performed by the quantization layer on the multidimensional array received by the input layer using the weight matrix and bias vector. The specific formula is as follows:

[0048] in, It is the result of the linear transformation of the i-th neuron. These are elements in the weight matrix, representing the weight of the j-th feature to the i-th neuron. It is the j-th eigenvalue in the composite eigenvector. It is the bias vector of the i-th neuron. It is the dimension of the comprehensive feature vector.

[0049] For example, for user A, the preset intention tag could be car purchase intention (1 indicates intention to buy a car, 0 indicates no intention to buy a car). User A's comprehensive feature vector might have three feature values: X1: number of times the user browsed the car page, X2: number of times the user interacted with customer service, and X3: number of times the user searched for "car purchase discounts". For example, user A's feature values ​​might be: X1=5; X2=3; X3=2.

[0050] The parameters of a pre-trained model are, for example: weight matrix :

[0051] The first row corresponds to a weight of 1 for a car purchase intention, and the second row corresponds to a weight of 0 for a car purchase intention.

[0052] bias vector :

[0053] The input layer receives the comprehensive feature vector; at this point, the input layer receives the comprehensive feature vector from user A. for:

[0054] Secondly, the quantization layer uses a weight matrix. and bias vector For the input vector Perform a linear transformation and calculate the output of each neuron:

[0055]

[0056] Therefore, the result after the linear transformation is:

[0057] Finally, the output layer uses the Softmax function to calculate the probability distribution, converting the linear transformation result into a probability distribution.

[0058] ; Finally, the model outputs the probability values ​​of each preset intention label for user A: the probability of purchase intention 1 (having the intention to purchase a car) is approximately 0.474; the probability of purchase intention 0 (having no intention to purchase a car) is approximately 0.526.

[0059] It should be noted that the above is merely a simple computing scenario used to illustrate the essence of the solution, and does not represent the actual business computing scenarios. In actual business scenarios, the number of preset intention tags is enormous and can be dynamically updated by the administrator.

[0060] In some embodiments of this application, the specific process of generating a preset traffic allocation plugin includes: creating a preprocessing component for preprocessing historical behavior data and user status information; creating a feature acquisition component for acquiring historical behavior features and current user status features related to customer intent; creating a feature fusion component for fusing historical behavior features and current user status features; creating a probability value generation component for generating probability values ​​of preset intent tags for each user; acquiring a pre-trained customer intent prediction model; integrating the pre-trained customer intent prediction model into the probability value generation component to obtain the final probability value generation component; and integrating the final probability value generation component, preprocessing component, feature acquisition component, and feature fusion component to obtain the preset traffic allocation plugin. The final integrated result is, for example... Figure 2 As shown.

[0061] S104, take the preset intention tag corresponding to the maximum probability value as the final intention tag of each user, assign each user to the preset traffic pool corresponding to each user's final intention tag, and continue to execute the step of collecting the historical behavior data of each user in the auto finance platform.

[0062] The preset traffic pools are pre-configured traffic allocation pools based on different intent tags. Each traffic pool corresponds to a specific marketing strategy or service. Even after a user has been assigned to a traffic pool, the system continues to collect the user's latest historical behavior data in order to update the user's intent tags and traffic allocation strategy in real time.

[0063] In some embodiments of this application, for user A, the probability values ​​of the intention label output by the model are as follows: the probability of a purchase intention of 1 (having the intention to purchase a car) is 0.474, and the probability of a purchase intention of 0 (not having the intention to purchase a car) is 0.526. In this case, the label with the highest probability is "purchase intention of 0" (not having the intention to purchase a car), with a probability of 0.526. Therefore, the user's final intention label is determined to be "not having the intention to purchase a car". For example, there are the following preset traffic pools: Traffic Pool 1 (for users with the intention to purchase a car) and Traffic Pool 2 (for users without the intention to purchase a car). Based on the user's final intention label "not having the intention to purchase a car", user A will be assigned to "Traffic Pool 2". At this time, user A will receive marketing information or services related to users without the intention to purchase a car.

[0064] Furthermore, while continuously acquiring user A's historical behavioral data in real time, user A's behavior changes, including an increased number of times browsing the car purchase page, frequent inquiries about car purchase discounts during interactions with customer service, and an increased number of searches for "car purchase discounts." This new behavioral data will be collected and input into the model, which will recalculate the user's intention tag probability value. For example, the probability of a purchase intention of 1 (having the intention to purchase a car) is 0.7, and the probability of a purchase intention of 0 (not having the intention to purchase a car) is 0.3. In this case, the user's final intention tag will be updated to "having the intention to purchase a car," and user A will be reassigned to "Traffic Pool 1" to receive marketing information or services related to users with the intention to purchase a car.

[0065] For example Figure 3 As shown, Figure 3 This application provides a schematic block diagram of a traffic allocation process based on behavioral data. The system initiates data collection and processing through periodic tasks (such as cron jobs) or message triggers. The system collects users' historical behavioral data, including browsing history, search history, and interaction behavior. It also obtains users' current status information, such as call records, repayment status, and funding needs. The collected historical behavioral data and current user status information are aligned and sent to a Kafka message queue as a unified data bus. The collected data is preprocessed, such as cleansing and formatting. Features related to user intent are extracted from the preprocessed data. The extracted features are fused to generate a comprehensive feature vector. The Softmax function is used to convert the comprehensive feature vector into a probability distribution, generating probability values ​​for each user's preset intent tags. The argmax function is used to select the intent tag corresponding to the highest probability value from the probability values ​​and allocates the user to the corresponding traffic pool. Based on the user's final intent tag, the user is allocated to the corresponding operational system, such as telesales, post-loan management, financial resource management, or marketing systems.

[0066] In this embodiment, on the one hand, a pre-trained customer intent prediction model generated using machine learning algorithms is employed. This model not only considers historical behavioral data of customers on the auto finance platform but also integrates current user status information from the telemarketing system, post-loan system, and treasury system in real time. This enables the system to accurately allocate traffic based on the complexity and diversity of user data, reducing the allocation of invalid traffic and thus improving the utilization rate of enterprise resources. On the other hand, by labeling each registered user's optimal preset intent tag onto a tuple composed of sample behavioral data and sample user status information, the optimal preset intent tag is determined by introducing a positive mapping strategy, a strategy tree, and an atomic field dictionary. This labeling method is highly accurate, avoids dirty data caused by manual labeling, provides high-quality training samples for model training, and thus improves the model's prediction accuracy.

[0067] Please see Figure 4 This is a flowchart illustrating a customer intent prediction model training method provided in an embodiment of this application. Figure 4 As shown, the method in this application embodiment may include the following steps: S201: Collect browsing history, search keywords, and online communication records of registered users within a preset historical time period from the auto finance platform to obtain sample behavioral data for each registered user; S202: Collect various system record information from the telemarketing system, post-loan system, and financial system for each registered user to obtain sample user status information for each registered user; S203, based on a preset set of intention tags, sample behavior data of each registered user, and sample user status information, generate multiple model training samples; The preset intention tag set contains each preset intention tag and the priority of each preset intention tag.

[0068] In some embodiments of this application, the specific process of generating multiple model training samples based on a preset intention tag set, sample behavior data of each registered user, and sample user status information includes: establishing a forward mapping strategy for each preset intention tag; sorting the forward mapping strategies of each preset intention tag from high to low priority according to the priority of each preset intention tag to obtain a strategy tree; extracting each field name and its attribute value from the sample behavior data and sample user status information of each registered user; expanding each field name and its attribute value into an atomic field dictionary to obtain the atomic field dictionary of each registered user; determining the optimal preset intention tag for each registered user based on the atomic field dictionary and strategy tree of each registered user; and labeling the optimal preset intention tag of each registered user on a tuple composed of the sample behavior data and sample user status information of each registered user to obtain multiple model training samples.

[0069] Specifically, the process of determining the optimal preset intention label for each registered user based on their atomic field dictionary and policy tree includes: traversing the forward mapping strategy of a preset intention label in the policy tree according to priority order; the forward mapping strategy includes a hard condition for assigning the highest confidence score, multiple sub-conditions and their weights, and a target threshold; determining whether each registered user's atomic field dictionary contains an attribute value that satisfies the hard condition; if so, using the traversed preset intention label as the optimal preset intention label for each registered user and stopping the traversal; if not, determining each target sub-condition from the multiple sub-conditions that satisfies each registered user's atomic field dictionary; summing the weights corresponding to each target sub-condition; if the summation result is greater than or equal to the preset threshold, using the traversed preset intention label as the optimal preset intention label for each registered user and stopping the traversal; or if the summation result is less than the preset threshold, continuing to execute the step of traversing the forward mapping strategy of a preset intention label in the policy tree according to priority order until the optimal preset intention label for each registered user is determined, and then stopping the traversal.

[0070] It should be noted that if no optimal preset intention label is found after all iterations are completed, the fallback default label will be assigned to the tuple consisting of the registered user's sample behavior data and the sample user's status information.

[0071] The system comprises several key components: a predefined set of intent tags representing different user purchase intentions, such as "high intention," "medium intention," and "low intention." The strategy tree is a tree-like structure built based on the priority of intent tags, used to organize and traverse the positive mapping strategy. The atomic field dictionary is a collection of field names and their attribute values ​​extracted from user behavior data and status information. A hard condition is the single condition with the highest priority in the positive mapping strategy; if satisfied, it is directly assigned the highest confidence score. Sub-conditions are all conditions other than the hard condition, each with its own weight.

[0072] In one possible implementation, a set of intention labels is determined, and a priority is assigned to each label. Hard conditions and sub-conditions, as well as weights for the sub-conditions, are defined for each intention label. Based on the priority of the intention labels, the forward mapping strategy is organized into a policy tree. Field names and their attribute values ​​are extracted from user behavior data and state information to form an atomic field dictionary. The policy tree is traversed, and the hard conditions of each intention label are checked in priority order. If the hard condition is met, the label is directly taken as the optimal intention label. If not, the sub-conditions are checked, and a weighted sum is calculated. If the weighted sum is greater than or equal to a target threshold, the label is also taken as the optimal intention label. The determined optimal intention labels are labeled on the user behavior data and state information to form training samples.

[0073] For example, there are three intent labels: "High Intent," "Medium Intent," and "Low Intent," with priorities of 1, 2, and 3 respectively. The hard condition for High Intent is that the user has viewed the car page more than 10 times in the past week. The hard condition for Medium Intent is that the user has viewed the car page 3 to 10 times in the past week. Low Intent has no hard condition and serves as a fallback option. For user A, their atomic field dictionary is as follows: number of times they viewed the car page (15 times), phone inquiry records (2 times), and funding needs (none). According to the policy tree, we first check the hard condition for "High Intent." Because user A's browsing frequency meets the hard condition (more than 10 times), the optimal preset intent label for user A is "High Intent." We don't need to consider other sub-conditions or check the strategies for "Medium Intent" and "Low Intent." At this point, we obtain a model training sample: user A's behavioral data and state information, along with the labeled "High Intent."

[0074] S204, A first customer intention prediction model is created using a machine learning algorithm, and a model structure is defined in the first customer intention prediction model to obtain a second customer intention prediction model; Machine learning algorithms are algorithms used in the field of artificial intelligence to solve specific problems, such as machine learning and deep learning. The first customer intention prediction model is the initially created model. The second customer intention prediction model is a structured model defined based on the first model.

[0075] In some embodiments, a suitable algorithm, such as a decision tree, random forest, or neural network, is selected based on the nature of the problem. A preliminary model is created using the selected algorithm. Based on the characteristics of the data and the requirements of the prediction task, the model architecture is designed, including the number of layers, the number of neurons in each layer, and the activation function, resulting in a second customer intention prediction model.

[0076] S205, Based on each model training sample and the second customer intention prediction model, generate the model loss value; In some embodiments of this application, the specific process of generating model loss values ​​based on each model training sample and the second customer intention prediction model includes: preprocessing each model training sample; extracting sample behavior features and sample user state features related to customer intention from each preprocessed model training sample; fusing the sample behavior features and sample user state features to obtain a sample comprehensive feature vector for each model training sample; inputting the sample comprehensive feature vector into the second customer intention prediction model and outputting the model loss value.

[0077] Among them, for example Figure 5 As shown, the model structure of the second customer intention prediction model includes an input layer, a quantization layer, and an output layer.

[0078] In some embodiments of this application, the specific process of inputting the sample comprehensive feature vector into the second customer intention prediction model and outputting the model loss value includes: the input layer receiving the sample comprehensive feature vector; the quantization layer using a pre-learned weight matrix and bias vector to describe the relationship between features and intention labels to perform a linear transformation on the sample comprehensive feature vector received by the input layer, obtaining the sample linear transformation result; the output layer using a preset Softmax function to convert the sample linear transformation result into a probability distribution, obtaining the probability value of each model training sample for each preset intention label; and the output layer using a loss function and the probability value of each model training sample for each preset intention label to calculate and output the model loss value.

[0079] Specifically, the expression for the loss function is:

[0080] in, This represents the model loss value. It is the total number of training samples for the model. This is the total number of preset intention tags. It is the first One-Hot encoding of the true intention labels of the model training samples, if the model training samples Belongs to the If there are preset intention tags, then Otherwise, it is 0. It is the first The training samples of the model belong to the first... The probability value of a preset intention tag.

[0081] S206, if the model loss value reaches the minimum, generate a pre-trained customer intention prediction model; or if the model loss value does not reach the minimum, continue to execute the step of inputting each model training sample into the second customer intention prediction model until the model loss value reaches the minimum.

[0082] For example Figure 6 As shown, Figure 6This application provides a schematic flowchart of a model training process, which involves collecting historical user behavior data on the platform, such as browsing history, search history, and purchase history. It also involves obtaining the user's current state information, such as call records, repayment records, and funding needs. All possible fields are extracted from the collected data, and a unique identifier is generated for each field, forming a dictionary. A policy tree is used to automatically label each user's intention tag. The policy tree defines hard conditions (conditions that must be met) and sub-conditions (optional conditions), as well as their weights and thresholds. Based on the labeling results of the policy tree, sample data for training the model is generated. Each sample contains the user's behavior data, state information, and corresponding intention tag. The structure of a neural network (NN) is defined, and a Softmax function is used in the output layer. The neural network is used to learn the relationship between user behavior and state information and intention tags. The forward propagation algorithm is used to calculate the model's loss value L. The loss value measures the difference between the model's predicted result and the actual label. It is then checked whether the current loss value L has been minimized. If it has not been minimized, backpropagation continues to update the weights and biases; if it has been minimized, the pre-trained model is output. If the loss value L is not minimized, the backpropagation algorithm is used to update the weights (W) and biases (b) of the neural network to reduce the loss value. When the loss value L is minimized, the trained model is output to predict the behavioral intentions of new users.

[0083] In this embodiment, on the one hand, a pre-trained customer intent prediction model generated using machine learning algorithms is employed. This model not only considers historical behavioral data of customers on the auto finance platform but also integrates current user status information from the telemarketing system, post-loan system, and treasury system in real time. This enables the system to accurately allocate traffic based on the complexity and diversity of user data, reducing the allocation of invalid traffic and thus improving the utilization rate of enterprise resources. On the other hand, by labeling each registered user's optimal preset intent tag onto a tuple composed of sample behavioral data and sample user status information, the optimal preset intent tag is determined by introducing a positive mapping strategy, a strategy tree, and an atomic field dictionary. This labeling method is highly accurate, avoids dirty data caused by manual labeling, provides high-quality training samples for model training, and thus improves the model's prediction accuracy.

[0084] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0085] Please see Figure 7This illustration shows a schematic diagram of a traffic allocation device based on behavioral data provided in an exemplary embodiment of this application. This traffic allocation device based on behavioral data can be implemented as all or part of an electronic device through software, hardware, or a combination of both. The device 1 includes a historical behavioral data collection module 10, a current user status information acquisition module 20, a probability value output module 30, and a traffic allocation module 40.

[0086] The historical behavior data collection module 10 is used to collect historical behavior data of each registered user in the auto finance platform according to a preset cycle. The current user status information acquisition module 20 is used to interact with the telemarketing system, the post-loan system and the financial system in real time to obtain multiple system records for each user and obtain the current user status information. The probability value output module 30 is used to input historical behavior data and current user status information into the preset traffic allocation plugin and output the probability value of each preset intention tag corresponding to each user; the preset traffic allocation plugin includes a pre-trained customer intention prediction model generated based on machine learning algorithms; The traffic allocation module 40 is used to take the preset intention tag corresponding to the maximum probability value as the final intention tag of each user, allocate each user to the preset traffic pool corresponding to the final intention tag of each user, and continue to perform the step of collecting historical behavior data of each registered user in the auto finance platform.

[0087] It should be noted that the traffic allocation device based on behavioral data provided in the above embodiments is only illustrated by the division of the above functional modules when executing the traffic allocation method based on behavioral data. In practical applications, the above functional allocation can be completed by different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the traffic allocation device based on behavioral data provided in the above embodiments and the traffic allocation method embodiments based on behavioral data belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0088] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0089] In this embodiment, on the one hand, a pre-trained customer intent prediction model generated using machine learning algorithms is employed. This model not only considers historical behavioral data of customers on the auto finance platform but also integrates current user status information from the telemarketing system, post-loan system, and treasury system in real time. This enables the system to accurately allocate traffic based on the complexity and diversity of user data, reducing the allocation of invalid traffic and thus improving the utilization rate of enterprise resources. On the other hand, by labeling each registered user's optimal preset intent tag onto a tuple composed of sample behavioral data and sample user status information, the optimal preset intent tag is determined by introducing a positive mapping strategy, a strategy tree, and an atomic field dictionary. This labeling method is highly accurate, avoids dirty data caused by manual labeling, provides high-quality training samples for model training, and thus improves the model's prediction accuracy.

[0090] This application also provides a computer-readable medium having program instructions stored thereon, which, when executed by a processor, implement the traffic allocation method based on behavior data provided in the above-described method embodiments.

[0091] This application also provides a computer program product containing instructions that, when run on a computer, cause the computer to execute the traffic allocation method based on behavioral data from the various method embodiments described above.

[0092] Please see Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 8 As shown, the electronic device 1000 may include: at least one processor 1001, at least one network interface 1004, a user interface 1003, a memory 1005, and at least one communication bus 1002.

[0093] The communication bus 1002 is used to realize the connection and communication between these components.

[0094] The user interface 1003 may include a display screen and a camera. Optionally, the user interface 1003 may also include a standard wired interface and a wireless interface.

[0095] The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0096] The processor 1001 may include one or more processing cores. The processor 1001 connects to various parts within the electronic device 1000 using various interfaces and lines. It executes various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 1005, and by calling data stored in the memory 1005. Optionally, the processor 1001 may be implemented using at least one hardware form selected from Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor 1001 may integrate one or more of the following: a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), and a modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip, without being integrated into the processor 1001.

[0097] The memory 1005 may include random access memory (RAM) or read-only memory. Optionally, the memory 1005 may include a non-transitory computer-readable storage medium. The memory 1005 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 1005 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 1005 may also be at least one storage system located remotely from the aforementioned processor 1001. Figure 8 As shown, the memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a traffic allocation application based on behavioral data.

[0098] exist Figure 8In the illustrated electronic device 1000, the user interface 1003 is mainly used to provide an input interface for the user and to obtain user input data; while the processor 1001 can be used to call the traffic allocation application based on behavior data stored in the memory 1005, and specifically perform the following operations: Historical behavioral data of each user on the auto finance platform is collected according to a preset cycle; It interacts in real time with the telemarketing system, the post-loan system, and the treasury system to obtain multiple system records for each user and to obtain the current user status information. Input historical behavior data and current user status information into the preset traffic allocation plugin, and output the probability value of each preset intention tag corresponding to each user; the preset traffic allocation plugin includes a pre-trained customer intention prediction model generated based on machine learning algorithms; The preset intention tag corresponding to the maximum probability value is used as the final intention tag for each user, and each user is assigned to a preset traffic pool corresponding to each user's final intention tag. The process of collecting historical behavior data of each user in the auto finance platform continues.

[0099] In one embodiment, when the processor 1001 executes the operation of inputting historical behavior data and current user state information into a pre-trained customer intention prediction model and outputting the probability values ​​of each preset intention label corresponding to each user, it specifically performs the following operations: The preprocessing component preprocesses historical behavior data and user status information; The feature acquisition component extracts historical behavioral features and current user status features related to customer intent from preprocessed historical behavioral data and user status information; The feature fusion component fuses historical behavioral features and current user state features to obtain a comprehensive feature vector; The probability value generation component generates probability values ​​for each user's preset intention tags based on a comprehensive feature vector and a pre-trained customer intention prediction model.

[0100] In one embodiment, when the processor 1001 generates probability values ​​for each preset intention tag corresponding to each user based on a comprehensive feature vector and a pre-trained customer intention prediction model, it specifically performs the following operations: The input layer receives the comprehensive feature vector. The total number of neurons in the input layer is the same as the dimension of the comprehensive feature vector, and each neuron receives one feature value of the comprehensive feature vector. The quantization layer loads pre-learned weight matrices and bias vectors, which are used to describe the relationship between features and intention labels. The quantization layer uses a weight matrix and a bias vector to perform a linear transformation on the multidimensional array received by the input layer, and obtains the linear transformation result. The output layer uses a preset Softmax function to convert the linear transformation result into a probability distribution, and outputs the probability value of each preset intention tag corresponding to each user.

[0101] In one embodiment, when the processor 1001 executes the generation of a pre-trained customer intent prediction model, it specifically performs the following operations: The system collects browsing history, search keywords, and online communication records of registered users within a preset historical time period from the auto finance platform to obtain sample behavioral data for each registered user. Collect various system record information from the telemarketing system, loan post-loan system, and treasury system to obtain sample user status information for each registered user. Based on a pre-defined set of intention tags, sample behavior data of each registered user, and sample user status information, multiple model training samples are generated. A first customer intention prediction model is created using machine learning algorithms, and a model structure is defined in the first customer intention prediction model to obtain a second customer intention prediction model. Based on the training samples of each model and the second customer intention prediction model, generate the model loss value; If the model loss value reaches its minimum, a pre-trained customer intention prediction model is generated; or if the model loss value does not reach its minimum, the process of inputting each model training sample into the second customer intention prediction model continues until the model loss value reaches its minimum.

[0102] In one embodiment, when the processor 1001 generates multiple model training samples based on a preset intention label set, sample behavior data of each registered user, and sample user status information, it specifically performs the following operations: Establish a positive mapping strategy for each preset intention label; Based on the priority of each preset intention tag, the positive mapping strategies of each preset intention tag are sorted from high to low priority to obtain a strategy tree; Extract each field name and its attribute value from the sample behavior data and sample user status information of each registered user; Expand each field name and its attribute value into an atomic field dictionary to obtain the atomic field dictionary for each registered user; Based on the atomic field dictionary and strategy tree of each registered user, determine the optimal preset intention tag for each registered user; The optimal preset intention label for each registered user is labeled on a tuple consisting of sample behavior data and sample user status information of each registered user, resulting in multiple model training samples.

[0103] In one embodiment, when the processor 1001 determines the optimal preset intention tag for each registered user based on the atomic field dictionary and policy tree for each registered user, it specifically performs the following operations: A positive mapping strategy is traversed in the strategy tree according to priority order for a predefined intention label. The positive mapping strategy includes a hard condition for assigning the highest confidence score, multiple sub-conditions and their weights, and a target threshold. Determine if the atomic field dictionary for each registered user contains an attribute value that meets a hard condition; If so, use the preset intention tags traversed as the optimal preset intention tags for each registered user and stop traversing; If not, determine each target sub-condition from multiple sub-conditions that satisfies the atomic field dictionary of each registered user; sum the weights corresponding to each target sub-condition; if the summation result is greater than or equal to a preset threshold, take the traversed preset intention label as the optimal preset intention label for each registered user and stop traversing; or if the summation result is less than the preset threshold, continue executing the steps of traversing the forward mapping strategy of a preset intention label in the strategy tree according to priority order until the optimal preset intention label for each registered user is determined and stop traversing.

[0104] In one embodiment, when the processor 1001 generates model loss values ​​based on each model training sample and the second customer intent prediction model, it specifically performs the following operations: Preprocess each model training sample; Extract sample behavioral features and sample user status features related to customer intent from each preprocessed model training sample; The sample behavior features and sample user state features are fused to obtain the sample comprehensive feature vector of each model training sample; Input the sample's comprehensive feature vector into the second customer intention prediction model and output the model loss value.

[0105] In one embodiment, when processor 1001 inputs the sample integrated feature vector into the second customer intention prediction model and outputs the model loss value, it specifically performs the following operations: The input layer receives the comprehensive feature vector of the samples; The quantization layer uses a pre-learned weight matrix and bias vector to describe the relationship between features and intention labels to perform a linear transformation on the sample integrated feature vector received by the input layer, and obtains the sample linear transformation result. The output layer uses a preset Softmax function to convert the linear transformation result of the samples into a probability distribution, thereby obtaining the probability value of each model training sample for each preset intention label; The output layer uses a loss function and the probability value of each training sample of the model for each preset intention label to calculate and output the model loss value.

[0106] In one embodiment, when the processor 1001 executes the generation of the preset traffic allocation plugin, it specifically performs the following operations: Create a preprocessing component for preprocessing historical behavior data and user status information; Create a feature acquisition component to obtain historical behavioral characteristics and current user status characteristics related to customer intent; Create a feature fusion component that combines historical behavioral features and current user state features; Create a probability value generation component that generates probability values ​​for each preset intention tag corresponding to each user; Obtain a pre-trained customer intent prediction model; The pre-trained customer intent prediction model is integrated into the probability value generation component to obtain the final probability value generation component. The final probability value generation component, preprocessing component, feature acquisition component, and feature fusion component are integrated to obtain the preset traffic allocation plugin.

[0107] In this embodiment, on the one hand, a pre-trained customer intent prediction model generated using machine learning algorithms is employed. This model not only considers historical behavioral data of customers on the auto finance platform but also integrates current user status information from the telemarketing system, post-loan system, and treasury system in real time. This enables the system to accurately allocate traffic based on the complexity and diversity of user data, reducing the allocation of invalid traffic and thus improving the utilization rate of enterprise resources. On the other hand, by labeling each registered user's optimal preset intent tag onto a tuple composed of sample behavioral data and sample user status information, the optimal preset intent tag is determined by introducing a positive mapping strategy, a strategy tree, and an atomic field dictionary. This labeling method is highly accurate, avoids dirty data caused by manual labeling, provides high-quality training samples for model training, and thus improves the model's prediction accuracy.

[0108] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The traffic allocation program based on behavioral data can be stored in a computer-readable storage medium. When executed, the program can include the processes of the embodiments of the above methods. The storage medium for the traffic allocation program based on behavioral data can be a magnetic disk, optical disk, read-only memory, or random access memory, etc.

[0109] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A traffic distribution method based on behavior data, characterized by, Applied to the server side, the method comprises: Collecting historical behavior data of each user in an auto finance platform according to a preset period; Real-time interaction with an electric sales system, a post-loan system and a financial system to obtain multi-type system records of each user and obtain current user state information; Inputting the historical behavior data and the current user state information into a preset traffic distribution plug-in to output probability values of each preset intention label corresponding to each user; the preset traffic distribution plug-in comprises a pre-trained customer intention prediction model generated based on a machine learning algorithm; Taking a preset intention label corresponding to a maximum probability value as a final intention label of each user, and distributing each user to a preset traffic pool corresponding to the final intention label of each user, and continuing to perform the step of collecting historical behavior data of each user in the auto finance platform.

2. The method of claim 1, wherein, The preset traffic distribution plug-in comprises a preprocessing component, a feature acquisition component, a feature fusion component and a probability value generation component, and the pre-trained customer intention prediction model is integrated in the probability value generation component; The inputting the historical behavior data and the current user state information into the pre-trained customer intention prediction model to output probability values of each preset intention label corresponding to each user comprises: The preprocessing component preprocesses the historical behavior data and the user state information; The feature acquisition component acquires historical behavior features and current user state features related to customer intention from the preprocessed historical behavior data and user state information; The feature fusion component performs feature fusion on the historical behavior features and the current user state features to obtain a comprehensive feature vector; The probability value generation component generates probability values of each preset intention label corresponding to each user based on the comprehensive feature vector and in combination with the pre-trained customer intention prediction model.

3. The method of claim 2, wherein, The comprehensive feature vector is a multidimensional array, and each dimension in the multidimensional array corresponds to a feature value; the pre-trained customer intention prediction model comprises an input layer, a quantization layer and an output layer; The generating the probability values of each preset intention label corresponding to each user based on the comprehensive feature vector and in combination with the pre-trained customer intention prediction model comprises: The input layer receives the comprehensive feature vector, and the total number of neurons of the input layer is the same as the dimension of the comprehensive feature vector, and each neuron receives a feature value of the comprehensive feature vector; The quantization layer loads pre-learned weight matrices and bias vectors, and the weight matrices and bias vectors are used to describe the relationship between features and intention labels; The quantization layer performs linear transformation on the multidimensional array received by the input layer using the weight matrices and bias vectors to obtain a linear transformation result; The output layer converts the linear transformation result into a probability distribution using a preset Softmax function to output the probability values of each preset intention label corresponding to each user; the preset Softmax function is: in, It is the first The probability value of a preset intention tag. This is the total number of preset intention tags. It is the first The result of linear transformation of each neuron.

4. The method according to any one of claims 1 to 3, characterized in that, The pre-trained customer intention prediction model is generated according to the following steps, comprising: Collecting browsing records, search keywords, online communication records of registered users in a preset historical time period from an automobile finance platform to obtain sample behavior data of each registered user; Collecting multiple types of system record information of each registered user from an electric sales system, a post-loan system and a financial system to obtain sample user state information of each registered user; Generating multiple model training samples based on a preset intention label set, sample behavior data of each registered user and sample user state information; Creating a first customer intention prediction model by using a machine learning algorithm and defining a model structure in the first customer intention prediction model to obtain a second customer intention prediction model; Generating a model loss value based on each model training sample and the second customer intention prediction model; In the case that the model loss value reaches a minimum, a pre-trained customer intention prediction model is generated; or in the case that the model loss value does not reach a minimum, the step of inputting each model training sample into the second customer intention prediction model is continuously executed until the model loss value reaches a minimum.

5. The method of claim 4, wherein, Each preset intention label and a priority of each preset intention label exist in a preset intention label set; The generating of multiple model training samples based on a preset intention label set, sample behavior data of each registered user and sample user state information comprises: Establishing a positive mapping strategy of each preset intention label; Based on the priority of each preset intention label, the positive mapping strategy of each preset intention label is sorted in descending order of priority to obtain a strategy tree; Extracting each field name and its attribute value from sample behavior data of each registered user and sample user state information; Expanding each field name and its attribute value into an atomic field dictionary to obtain an atomic field dictionary of each registered user; Determining an optimal preset intention label of each registered user according to the atomic field dictionary of each registered user and the strategy tree; Labeling the optimal preset intention label of each registered user on a binary tuple composed of sample behavior data of each registered user and sample user state information to obtain multiple model training samples.

6. The method of claim 5, wherein, The determining of an optimal preset intention label of each registered user according to the atomic field dictionary of each registered user and the strategy tree comprises: Traversing a positive mapping strategy of a preset intention label in the strategy tree in a priority order; the positive mapping strategy comprises a hard condition for giving a highest confidence score, multiple sub-conditions and their weights, and a target threshold value; Judging whether the atomic field dictionary of each registered user has an attribute value satisfying the hard condition; If yes, the traversed preset intention label is taken as the optimal preset intention label of each registered user, and the traversal is stopped; If no, from the plurality of sub-conditions, determine each target sub-condition satisfying the atomic field dictionary of each registered user; sum the weight corresponding to each target sub-condition; in the case where the sum is greater than or equal to a preset threshold, stop traversing and take the traversed preset intention label as the optimal preset intention label of each registered user; or in the case where the sum is less than the preset threshold, continue to execute the step of traversing the positive mapping strategy of one preset intention label in the strategy tree according to the priority order until the optimal preset intention label of each registered user is determined, and stop traversing.

7. The method of claim 4, wherein, The second customer intention prediction model is used for generating a model loss value based on each model training sample, and includes the following steps: preprocessing each model training sample; extracting sample behavior features and sample user state features related to customer intention from each preprocessed model training sample; performing feature fusion on the sample behavior features and the sample user state features to obtain a sample comprehensive feature vector of each model training sample; inputting the sample comprehensive feature vector into the second customer intention prediction model to output a model loss value.

8. The method of claim 7, wherein, The model structure of the second customer intention prediction model includes an input layer, a quantization layer, and an output layer. The input layer receives the sample comprehensive feature vector. The quantization layer adopts a pre-learned weight matrix and bias vector for describing the relationship between features and intention labels to perform linear transformation on the sample comprehensive feature vector received by the input layer to obtain a sample linear transformation result. The output layer adopts a preset Softmax function to convert the sample linear transformation result into a probability distribution to obtain a probability value of each preset intention label of each model training sample. The output layer adopts a loss function and the probability value of each preset intention label of each model training sample to calculate and output a model loss value; wherein the expression of the loss function is: The following steps are used to generate a preset traffic distribution plug-in, including: in, This represents the model loss value. It is the total number of training samples for the model. This is the total number of preset intention tags. It is the first One-Hot encoding of the true intention labels of the model training samples, if the model training samples Belongs to the If there are preset intention tags, then Otherwise, it is 0. It is the first The training samples of the model belong to the first... The probability value of a preset intention tag.

9. The method according to any one of claims 1 to 3, characterized in that, creating a preprocessing component for preprocessing historical behavior data and user state information; creating a feature acquisition component for acquiring historical behavior features and current user state features related to customer intention; creating a feature fusion component for performing feature fusion on historical behavior features and current user state features; creating a probability value generation component for generating a probability value of each preset intention label corresponding to each user; acquiring a pre-trained customer intention prediction model; integrating the pre-trained customer intention prediction model into the probability value generation component to obtain a final probability value generation component; integrating the final probability value generation component, the preprocessing component, the feature acquisition component, and the feature fusion component to obtain a preset traffic distribution plug-in. The device includes:

10. A traffic distribution apparatus based on behavior data, characterized by, a historical behavior data collection module configured to collect historical behavior data of each registered user in an automobile finance platform according to a preset period; ​ The current user state information acquisition module is configured to interact with the electric call system, the post-loan system and the financial system in real time to acquire multi-type system records of each user and obtain current user state information. The probability value output module is configured to input the historical behavior data and the current user state information into a preset traffic distribution plug-in to output probability values of each preset intention label corresponding to each user. The preset traffic distribution plug-in includes a pre-trained customer intention prediction model generated based on a machine learning algorithm. The traffic distribution module is configured to take a preset intention label corresponding to the maximum probability value as a final intention label of each user, assign each user to a preset traffic pool corresponding to the final intention label of each user, and continue to execute the step of collecting historical behavior data of each registered user in the automobile finance platform.

Images