Underwriting probability prediction method based on artificial intelligence algorithm

Abstract

The present disclosure provides an underwriting probability prediction method based on an artificial intelligence algorithm, which comprises: obtaining historical main insurance underwriting data and additional insurance underwriting data, matching and associating the underwriting object sample of the underwritten main insurance with the corresponding additional insurance underwriting data based on a preset association field, and giving a corresponding binary classification label based on whether the underwriting object sample further underwrites additional insurance; extracting multi-dimensional features affecting the insurance behavior of a target object and converting them into a structured feature vector; constructing a probability prediction model containing multiple layers of feature interaction encoding layers, inputting the structured feature vector into the probability prediction model to learn the feature interaction information, and training in combination with the binary classification label; outputting the prediction probability of the to-be-predicted object based on the trained probability prediction model, and determining the object with a prediction probability greater than a preset threshold as a potential target object. Thus, the accuracy and objectivity of the target object underwriting probability prediction can be greatly improved.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, specifically to a method for predicting underwriting probability based on artificial intelligence algorithms. Background Technology

[0002] Motor vehicle insurance business is divided into two categories of coverage: primary insurance and supplementary insurance. Primary insurance provides basic coverage, including vehicle damage insurance and third-party liability insurance, while supplementary insurance can only be underwritten if primary insurance has already been insured. In actual business processes, the processing of supplementary driver's accident insurance, which is bundled with vehicle damage insurance, usually occurs after the target customer has completed vehicle damage insurance coverage. Business operators typically decide whether to recommend driver's accident insurance to the target customer based on their subjective understanding of the target customer's risk profile, their manual processing experience, and their habits.

[0003] Existing business systems typically support the function of simultaneously processing driver's accident insurance after the target object completes vehicle damage insurance underwriting, allowing operators to process supplementary insurance underwriting for the target object in the same business process.

[0004] In related technologies, the decision-making process for configuring driver accident insurance heavily relies on the human experience and judgment of operators. Key decisions such as whether to recommend driver accident insurance to the target group, the strength of the recommendation, and the processing priority are all made by subjective human judgment. Standards vary among different personnel, and the actual implementation effect is greatly affected by differences in individual experience. Secondly, existing solutions do not fully utilize the inherent data and process dependencies between the main vehicle damage insurance and supplementary insurance. Summary of the Invention

[0005] This disclosure aims to at least partially address one of the technical problems in the related art.

[0006] Therefore, the purpose of this disclosure is to propose an underwriting probability prediction method, device, electronic device and storage medium based on artificial intelligence algorithms. By establishing an end-to-end artificial intelligence prediction framework, the implicit relationship between the main and supplementary insurance is transformed into a binary classification supervised learning task, and a multi-layer feature interaction coding layer is used to automatically capture high-order feature relationships between data, thereby significantly improving the accuracy and objectivity of the underwriting probability prediction of the target object.

[0007] To achieve the above objectives, the underwriting probability prediction method based on artificial intelligence algorithms proposed in the first aspect of this disclosure includes:

[0008] The system acquires historical underwriting data for the main insurance and supplementary insurance. Based on preset association fields, it matches and associates the samples of the underwritten main insurance with the corresponding supplementary insurance underwriting data. It also assigns corresponding binary labels based on whether the samples of the underwritten main insurance are further underwritten for the supplementary insurance. Extract multidimensional features that influence the insurance purchase behavior of the target object and transform them into structured feature vectors; A probabilistic prediction model containing multiple layers of feature interaction coding is constructed. The structured feature vector is input into the probabilistic prediction model to learn feature interaction information, and the model is trained in combination with the binary classification labels. Based on the trained probability prediction model, the predicted probability of the object to be predicted is output, and objects with a predicted probability greater than a preset threshold are identified as potential target objects.

[0009] To achieve the above objectives, the underwriting probability prediction device based on artificial intelligence algorithms proposed in the second aspect of this disclosure includes: The acquisition module is used to acquire historical main insurance underwriting data and supplementary insurance underwriting data. Based on preset association fields, it matches and associates the underwriting object samples of the main insurance with the corresponding supplementary insurance underwriting data, and assigns corresponding binary labels based on whether the underwriting object samples further underwrite the supplementary insurance. The feature extraction module is used to extract multidimensional features that affect the insurance purchase behavior of the target object and transform them into structured feature vectors; The model building module is used to build a probabilistic prediction model containing multiple layers of feature interaction coding. The structured feature vector is input into the probabilistic prediction model to learn feature interaction information, and the model is trained in combination with the binary classification labels. The prediction module is used to output the predicted probability of the object to be predicted based on the trained probability prediction model, and to identify objects with a predicted probability greater than a preset threshold as potential target objects.

[0010] The electronic device proposed in the third aspect of this disclosure includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the underwriting probability prediction method based on artificial intelligence algorithm proposed in the first aspect of this disclosure.

[0011] The fourth aspect of this disclosure provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the underwriting probability prediction method based on an artificial intelligence algorithm as proposed in the first aspect of this disclosure.

[0012] The fifth aspect of this disclosure provides a computer program product that, when executed by a processor, performs an underwriting probability prediction method based on an artificial intelligence algorithm as proposed in the first aspect of this disclosure.

[0013] The present disclosure provides an artificial intelligence algorithm-based method, device, electronic device, and storage medium for predicting underwriting probability. This method acquires historical primary insurance underwriting data and supplementary insurance underwriting data. Based on preset association fields, it matches and associates samples of underwritten primary insurance with corresponding supplementary insurance underwriting data, assigning corresponding binary classification labels based on whether the underwritten sample will further underwrite supplementary insurance. It extracts multi-dimensional features influencing the target object's insurance behavior and transforms them into structured feature vectors. A probabilistic prediction model containing multiple layers of feature interaction coding is constructed. The structured feature vectors are input into the probabilistic prediction model to learn feature interaction information, and training is performed using binary classification labels. Based on the trained probabilistic prediction model, the predicted probability of the object to be predicted is output, and objects with predicted probabilities greater than a preset threshold are identified as potential target objects. Therefore, by establishing an end-to-end artificial intelligence prediction framework, the implicit relationship between primary and supplementary insurance is transformed into a binary classification supervised learning task, and the high-order feature associations between data are automatically captured using multiple layers of feature interaction coding, thereby significantly improving the accuracy and objectivity of predicting the underwriting probability of the target object.

[0014] Additional aspects and advantages of this disclosure will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this disclosure. Attached Figure Description

[0015] The above and / or additional aspects and advantages of this disclosure will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, in which: Figure 1 This is a flowchart illustrating an embodiment of the underwriting probability prediction method based on artificial intelligence algorithms proposed in this disclosure. Figure 2 This is a schematic diagram of the underwriting propensity probability model proposed in this disclosure; Figure 3 This is a schematic diagram of the structure of an underwriting probability prediction device based on an artificial intelligence algorithm according to an embodiment of this disclosure; Figure 4 This is a block diagram of an electronic device according to an embodiment of the present application. Detailed Implementation

[0016] Embodiments of this disclosure are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are used only to explain this disclosure, and should not be construed as limiting this disclosure. Rather, embodiments of this disclosure include all variations, modifications, and equivalents falling within the spirit and scope of the appended claims.

[0017] It should be noted that the examples in this application may involve user data, data acquisition, and / or use. These aspects all comply with relevant laws, regulations, and provisions. In the examples of this application, all data collection, acquisition, processing, manipulation, forwarding, and use are conducted with the user's knowledge and confirmation. Accordingly, when implementing each example, the type, scope of use, and usage scenarios of any data or information that may be involved should be communicated to the user and their authorization obtained through appropriate means in accordance with relevant laws and regulations. The specific methods of notification and / or authorization can vary depending on the actual situation and application scenario, and the scope of the solution is not limited in this regard.

[0018] In this manual and the sample solutions, any processing of personal information will be conducted only under legal grounds (such as obtaining the consent of the data subject or being necessary for the performance of a contract) and will only be carried out within the scope stipulated or agreed upon. A user's refusal to process personal information beyond what is necessary for basic functions will not affect the user's use of basic functions.

[0019] Figure 1 This is a flowchart illustrating an embodiment of the underwriting probability prediction method based on artificial intelligence algorithms proposed in this disclosure.

[0020] It should be noted that the execution subject of the underwriting probability prediction method based on artificial intelligence algorithm in this embodiment is the underwriting probability prediction device based on artificial intelligence algorithm. The device can be implemented by software and / or hardware. The device can be configured in an electronic device, which may include, but is not limited to, a terminal, a server, etc., such as a mobile phone, a PDA, etc.

[0021] like Figure 1 As shown, the underwriting probability prediction method based on artificial intelligence algorithms includes: S101: Obtain historical main insurance underwriting data and supplementary insurance underwriting data, match and associate the underwriting object samples of the underwritten main insurance with the corresponding supplementary insurance underwriting data based on the preset association fields, and assign corresponding binary labels based on whether the underwriting object samples are further underwritten with supplementary insurance.

[0022] Among them, the main insurance and the supplementary insurance refer to the basic protection items (such as commercial vehicle damage insurance) configured for the target object, as well as the additional protection items (such as driving accident insurance) configured on the basis of them.

[0023] Here, the binary label refers to the target variable (Label) used for supervised learning. In this scheme, the insured object is labeled as one class (usually 1, i.e., positive sample) if it is covered by the supplementary insurance, and labeled as another class (usually 0, i.e., negative sample) if it is not covered.

[0024] Optionally, in some embodiments, the preset associated fields include at least one or more of the following: policy number, object identification, and vehicle identification; the step of converting them into a structured feature vector includes: The system performs aberration cleaning, standardization of field definitions, missing value handling, continuous variable standardization, and categorical variable encoding on multi-source heterogeneous historical underwriting and claims data to transform them into structured feature vectors.

[0025] Among them, the related field refers to the unique identifier used in the database to connect different data tables (such as policy serial number, encrypted identity number).

[0026] Among them, multi-source heterogeneous data refers to data (such as text, numerical values, and categories) from different business systems and in different formats.

[0027] Standardization of continuous variables refers to scaling numerical features with large differences in their range of values ​​(such as a certain underwriting quantitative indicator value of 10,000 versus the number of claims of 2) to the same order of magnitude (such as a mean of 0 and a variance of 1) in order to eliminate the influence of dimensions.

[0028] Categorical variable encoding refers to converting non-numerical categorical data (such as vehicle brand identifiers "Factory A" and "Factory B") into numerical vector representations that the model can recognize.

[0029] Based on the above embodiments, the quality and stability of the input data can be guaranteed. A standardized preprocessing pipeline eliminates the interference of dirty data and dimensional differences on the model gradient, accelerating the convergence speed of the neural network and enhancing the model's generalization ability.

[0030] S102: Extract multidimensional features that influence the insurance purchase behavior of the target object and transform them into structured feature vectors.

[0031] Among them, structured feature vectors refer to ordered numerical arrays formed by mathematically transforming original business data (such as text, numerical values, and dates), which can be directly calculated by computers and neural networks.

[0032] Optionally, in some embodiments, the main insurance is vehicle damage commercial insurance, and the supplementary insurance is driving accident insurance; the multidimensional features include: Contextual risk level characteristics, including main insurance risk level measurement indicators, main insurance underwriting quantitative indicators, and premium fluctuation coefficient; Based on basic human characteristics and historical behavioral preferences, including the number of times the main and supplementary insurance policies were underwritten, historical underwriting value metrics, and related policy information; Claims-related characteristics include the number of historical claims and quantitative indicators of historical payouts; The characteristics of a vehicle include its age, brand, usage, number of seats, tonnage, engine displacement, and energy type.

[0033] Among them, the contextual risk level feature can be used to describe the overall risk and system parameter status (such as quantitative indicators and fluctuation coefficients) of the target object when underwriting the main insurance.

[0034] Among them, the characteristics of human basic / historical behavioral preferences can be used to describe the inherent attributes of the target object and its past insurance preferences (such as the number of times it has been insured in the past).

[0035] Among them, claims-related features can be used to describe the target's past accidents and claims, reflecting its risk exposure level.

[0036] Among them, vehicle characteristics can be used to describe the physical attributes and usage of the vehicle itself (such as engine displacement and vehicle age).

[0037] Based on the above embodiments, this disclosure can construct a three-dimensional, multi-dimensional object feature profile. By introducing specific data features from four dimensions—"people, vehicles, history, and compensation"—it provides rich and strongly correlated input signals for the probabilistic prediction model, significantly improving the upper limit and accuracy of the model's predictions.

[0038] S103: Construct a probabilistic prediction model containing multiple layers of feature interaction coding. Input structured feature vectors into the probabilistic prediction model to learn feature interaction information, and train it in combination with binary classification labels.

[0039] Among them, the multi-layer feature interaction coding layer refers to the core processing module in the deep neural network, which is composed of multiple network layers stacked together. It is specifically used to learn the cross-combination and non-linear relationship between different features (such as "vehicle age" and "number of accidents").

[0040] Optionally, in some embodiments, the probabilistic prediction model includes, from bottom to top, an input feature layer, a feature embedding layer, a multi-layer feature cross-coding layer, and a top-level output layer. Structured feature vectors are transformed into fully embedded features through a feature embedding layer; Each layer in the multi-layer feature interaction coding layer contains a factorization machine interaction module and a linear compression module, which are used to extract non-linear feature interaction information based on the full embedded features.

[0041] Among them, the feature embedding layer is a dimensionality reduction and densification technique that maps high-dimensional sparse class features into low-dimensional dense real number vectors.

[0042] Among them, the Factorization Machine (FM) interaction module refers to a network module specifically designed to extract second-order cross information between features (i.e., the influence produced by the pairwise combination of two features).

[0043] The linear compression module refers to the module that performs overall dimensionality reduction and linear weighting on all feature information, and is used to capture macroscopic information from a global perspective.

[0044] Based on the above embodiments, a network architecture highly adaptable to tabular business data can be designed. The embedding layer solves the curse of dimensionality problem caused by large-scale discrete categorical variables; while the combination of factorization machine and linear compression module enables the model to simultaneously take into account both "local fine-grained crossover" and "global macro information" between features.

[0045] Optionally, in some embodiments, the feature processing in the multi-layer feature cross-coding layer specifically includes: The second-order interactive information representation between the learnable projection matrix in the factorization machine interaction module and the features extracted by the multilayer perceptron; The global information representation is obtained by linearly compressing the full embedded features using a linear compression module. The second-order interactive information representation and the global information representation are concatenated and fused, and then mapped back to the original dimension through a linear transformation. After residual connection and layer normalization processing are introduced in the multi-layer stacked structure, the data is input to the next layer.

[0046] The second-order interactive information representation is the result vector after multiplying / combining the features pairwise, while the global information representation is the result vector after linearly adding all the features.

[0047] Residual connection refers to directly superimposing the input of a certain layer onto the output across layers (input + output of a specific layer), forming a short-circuit mechanism.

[0048] Layer normalization refers to the standardization of feature vectors within each layer of a neural network to maintain a stable distribution.

[0049] Based on the above embodiments, the model's deep expressive power can be enhanced and training crashes can be prevented. Residual connections effectively alleviate the gradient vanishing problem common in deep networks, allowing the network to be made deeper; layer normalization ensures the stability of signals during transmission through multiple layers of the network, improving the robustness of training.

[0050] Optionally, in some embodiments, the top output layer comprises a multilayer perceptron. The output and training process of the probabilistic prediction model specifically includes: The Sigmoid function is used as the activation function to map the representation after multi-layer feature interaction to a single output node to output the predicted probability; During model training, a binary cross-entropy function is used as the loss function. The parameters are updated iteratively through backpropagation algorithm and adaptive moment estimation optimizer. In addition, embedding layer regularization, network structure regularization and random deactivation mechanism are combined to suppress model overfitting.

[0051] The Sigmoid function is a non-linear activation function that maps any real number to the interval between 0 and 1, and is often used to output probabilities.

[0052] The binary cross-entropy function is a loss function that measures the difference between the predicted probability distribution of the model output and the true label (0 or 1).

[0053] Among them, the backpropagation and adaptive moment estimation optimizer (Adam) is an algorithm that calculates the gradient from the output layer to the input layer based on the error of the loss function and automatically adjusts the learning rate to update the network weights.

[0054] Regularization and Dropout are techniques that prevent the model from "memorizing" training data by adding a penalty term to the loss or randomly dropping some neurons during training.

[0055] Based on the above embodiments, it can be ensured that the model can correctly learn and output probability values ​​that are meaningful for decision-making. The combination of cross-entropy and the Adam optimizer ensures that the model can quickly and accurately find the optimal parameters; the regularization mechanism effectively prevents overfitting, ensuring that the model still has a high prediction accuracy when facing new test objects.

[0056] S104: Based on the trained probability prediction model, output the predicted probability of the object to be predicted, and determine the object with a predicted probability greater than a preset threshold as a potential target object.

[0057] The preset threshold is the critical cut-off point used to transform the continuous probability values ​​(between 0 and 1) output by the model into the final decision (whether to insure or not).

[0058] Optionally, in some embodiments, the step of determining the preset threshold includes: On the validation set, a confusion matrix is ​​constructed based on the true labels and predicted probabilities, and precision and recall are calculated under different candidate thresholds. Based on the preset comprehensive evaluation function, the threshold that optimizes the comprehensive evaluation function result is selected from multiple candidate thresholds as the preset threshold.

[0059] The confusion matrix is ​​a table that records the cross-classification between the model's predicted results and the actual results (including true positives, false positives, true negatives, and false negatives).

[0060] Precision and recall refer to the fact that precision represents "how many of the predicted potential targets are actually positive samples" and recall represents "how many of the actual positive samples are successfully detected by the model".

[0061] The comprehensive evaluation function refers to the overall score calculated by assigning different weights to precision and recall based on specific business needs (such as prioritizing control system processing resources or expanding recall coverage).

[0062] Based on the above embodiments, the alignment between technical indicators and practical application goals has been achieved. By transforming rigid probability values ​​into flexibly adjustable judgment tools, the system can find the optimal balance between "processing resource input (precision)" and "target object discovery (recall)".

[0063] In this embodiment, by acquiring historical primary insurance underwriting data and supplementary insurance underwriting data, the insured object samples of the primary insurance have been matched and associated with the corresponding supplementary insurance underwriting data based on preset association fields. A corresponding binary classification label is assigned based on whether the insured object sample will further underwrite supplementary insurance. Multidimensional features influencing the object's insurance behavior are extracted and transformed into structured feature vectors. A probabilistic prediction model containing multiple layers of feature interaction coding is constructed. The structured feature vectors are input into the probabilistic prediction model to learn feature interaction information, and training is performed in conjunction with binary classification labels. Based on the trained probabilistic prediction model, the predicted probability of the object to be predicted is output, and objects with predicted probabilities greater than a preset threshold are identified as potential target objects. Thus, by establishing an end-to-end artificial intelligence prediction framework, the implicit association between primary and supplementary insurance is transformed into a binary classification supervised learning task, and the high-order feature associations between data are automatically captured using multiple layers of feature interaction coding, thereby significantly improving the accuracy and objectivity of the target object's underwriting probability prediction.

[0064] Optionally, in some embodiments, the trained probabilistic prediction model can be exported as a general model data exchange format and deployed to the target inference service environment to build a prediction service interface; the prediction data generated by the actual application can be sent back to the database, and the probabilistic prediction model can be iteratively optimized periodically.

[0065] The general model data exchange format refers to an open standard format for representing deep learning models, which facilitates the migration of models between different computing frameworks and hardware.

[0066] The target inference service environment is the production server environment where the model is deployed, which is used to receive requests from the processing system in real time and return prediction results.

[0067] Closed-loop management refers to collecting the actual feedback results generated after the model's prediction and feeding them back into the model as new training data.

[0068] Based on the above embodiments, the general model data exchange format ensures high-performance computing and inference of the model; the closed-loop management mechanism enables the model to self-evolve and iterate as time goes by and data distribution changes, maintaining long-term vitality and prediction accuracy.

[0069] In summary, this disclosure defines the question of whether to further insure supplementary insurance for an entity already insured for vehicle damage commercial insurance as a probability prediction and propensity screening decision problem based on risk characteristics. It constructs an underwriting probability prediction model to achieve accurate identification of potential entities. The technical solution of this disclosure uses entities already insured for vehicle damage insurance as the basic candidate set. Based on long-term accumulated historical underwriting and claims data, it learns the mapping relationship between risk characteristics and supplementary insurance underwriting behavior. Through artificial intelligence algorithms, it performs end-to-end modeling of underwriting propensity, outputs the probability of underwriting a single entity, and combines a threshold decision mechanism to transform continuous probabilities into a final list of potential entities.

[0070] The technical solution proposed in this disclosure includes data integration and sample construction, artificial intelligence algorithm model training, threshold decision-making, and deployment stages. First, it integrates vehicle damage insurance and supplementary insurance underwriting data to construct supervised learning samples and define binary classification labels. Then, it constructs a feature library based on multi-dimensional factors such as risk level, historical preferences, and vehicle attributes. In the prediction stage, it employs a deep neural network model modified based on a self-attention mechanism, learning complex nonlinear relationships through a multi-layer feature interaction encoding structure to output the probability value of the object being covered by supplementary insurance. Finally, in the decision-making stage, it analyzes the performance under different thresholds based on a confusion matrix, selects the optimal threshold in conjunction with resource objectives, and transforms the predicted probability into a target label.

[0071] This disclosed technical solution, based on large-scale historical data accumulated over a long period, constructs a data processing and prediction model. It transforms decision-making problems that originally relied on human experience into quantifiable and optimizable probabilistic prediction problems, achieving risk-driven precision screening. By combining probability output with threshold strategies, the system can improve screening accuracy while balancing computational resource investment and detection efficiency, thereby enhancing overall processing performance.

[0072] Step 1: Based on historical vehicle damage insurance data and supplementary insurance underwriting data, relevant feature data that may affect whether an object is actually insured is integrated and processed. Through association fields such as policy number, object identity identifier, and vehicle identifier, the object samples insured for the main insurance are matched and associated with the corresponding supplementary insurance information to construct the training sample set required for supervised learning. Binary classification labels are defined, where LABEL=1 indicates that both vehicle damage insurance and driver's accident insurance are insured, and LABEL=0 indicates that vehicle damage insurance is insured but driver's accident insurance is not insured. This formalizes the business problem into a binary classification prediction problem that can be used for model training. The modeling objective is:

[0073] in, The factor representing whether or not driver's accident insurance is covered, and f represents the probability of whether or not driver's accident insurance is covered and the aforementioned features fitted by the constructed deep neural network model. The complex nonlinear relationship between them.

[0074] Building upon this foundation, a unified approach is taken to construct and process the multidimensional features influencing insurance behavior. These features include contextual risk level features, basic human characteristics, historical behavioral preference features, claims-related features, and vehicle characteristics. Contextual risk level features include vehicle damage insurance risk measurement indicators, underwriting quantification indicators, and premium fluctuation coefficients, used to characterize the risk situation in the current scenario. Historical behavioral preference features include the number of past insurance policies, historical underwriting quantification indicators, and related policy information, used to represent past habits. Claims-related features include the number of past claims and historical payout quantification indicators, used to reflect the degree of risk exposure and payout characteristics. Vehicle characteristics include vehicle age, vehicle brand, usage type, number of seats, tonnage, engine displacement, and energy type, used to characterize the inherent attributes of the vehicle. All of these features are preprocessed using methods such as cleaning outlier samples, standardizing field definitions, handling missing values, standardizing continuous variables, and encoding categorical variables. This transforms multi-source heterogeneous data into structured feature vectors, which serve as input to the model, providing a stable and generalizable feature foundation for subsequent training.

[0075] Step 2: This disclosure employs a multi-layer feature interaction deep neural network as the core algorithm of the underwriting propensity probability model. Through supervised learning, it fits the complex nonlinear relationship between the LABEL constructed in step 1 and the multi-dimensional features, thereby outputting the probability of the object being insured.

[0076] like Figure 2 As shown, Figure 2This is a schematic diagram of the underwriting propensity probability model proposed in this disclosure. The overall structure of the model, from bottom to top, includes: an original input feature layer, a feature embedding layer, a multi-layer feature interaction coding layer, and a top-level linear output layer. The bottom layer input is a structured multi-dimensional feature vector, which is transformed into embedding features. Then, it enters the multi-layer feature interaction layer for deep modeling. Each interaction layer consists of a factorization machine interaction module and a linear compression module: the former efficiently extracts the second-order interaction information between features and outputs the interaction representation through a learnable projection matrix and a multi-layer perceptron; the linear compression module linearly compresses the full embedding features to obtain a global information representation. The two are concatenated and mapped back to the original dimension through a linear transformation. Residual connections and layer normalization are introduced to achieve stable training and deep representation. Based on the output of the interaction layer, the feature representation is further mapped to a single output node through the top-level multi-layer perceptron, using the sigmoid function as the activation function to output the predicted value of each object sample. During training, a binary cross-entropy function is used as the loss function. All parameters are updated iteratively through backpropagation and the Adam optimizer. Overfitting is suppressed by combining mechanisms such as embedding regularization, network regularization, and dropout.

[0077] Step 3: After the model is trained and the predicted probability of each object is output, a threshold decision mechanism converts continuous probability values ​​into 0 or 1 labels. On the validation set, a confusion matrix is ​​constructed based on the true labels and predicted probabilities. Precision, recall, and other metrics are calculated at different thresholds, and a comprehensive evaluation function is set based on the objective (e.g., giving higher weight to recall to prioritize finding potential positive samples, or increasing the weight of precision to control processing resource costs). By selecting the threshold with the optimal comprehensive metrics from multiple candidate thresholds as the decision boundary, objects with predicted probabilities greater than this threshold are identified as potential target objects, thus forming the final screening list.

[0078] Step 4: After determining the threshold strategy, the trained deep learning model is exported into a general model data exchange format for real-time inference and deployed to the target inference service environment. A real-time or near-real-time prediction service interface is built through a hardware-accelerated computing platform. By embedding the model into the actual business flow, recording the actual inference data after the model goes live and sending it back to the database, and regularly optimizing and iterating the model, closed-loop management is achieved, thereby improving the efficiency of automated identification of potential objects and the accuracy of data processing.

[0079] Figure 3 This is a schematic diagram of the structure of an underwriting probability prediction device based on an artificial intelligence algorithm proposed in an embodiment of this disclosure.

[0080] like Figure 3As shown, the underwriting probability prediction device 30 based on artificial intelligence algorithms includes: The acquisition module 301 is used to acquire historical main insurance underwriting data and supplementary insurance underwriting data. Based on the preset association fields, it matches and associates the underwriting object samples of the underwritten main insurance with the corresponding supplementary insurance underwriting data, and assigns corresponding binary labels based on whether the underwriting object samples are further underwritten with supplementary insurance. The feature extraction module 302 is used to extract multi-dimensional features that affect the insurance purchase behavior of the target object and convert them into structured feature vectors; The model building module 303 is used to build a probabilistic prediction model containing multiple layers of feature interaction coding. The structured feature vector is input into the probabilistic prediction model to learn feature interaction information, and the model is trained in combination with binary classification labels. The prediction module 304 is used to output the predicted probability of the object to be predicted based on the trained probability prediction model, and to identify objects with a predicted probability greater than a preset threshold as potential target objects.

[0081] It should be noted that the foregoing explanation of the underwriting probability prediction method based on artificial intelligence algorithms also applies to the underwriting probability prediction device based on artificial intelligence algorithms in this embodiment, and will not be repeated here.

[0082] In this embodiment, by acquiring historical primary insurance underwriting data and supplementary insurance underwriting data, the insured object samples of the primary insurance have been matched and associated with the corresponding supplementary insurance underwriting data based on preset association fields. A corresponding binary classification label is assigned based on whether the insured object sample will further underwrite supplementary insurance. Multidimensional features influencing the object's insurance behavior are extracted and transformed into structured feature vectors. A probabilistic prediction model containing multiple layers of feature interaction coding is constructed. The structured feature vectors are input into the probabilistic prediction model to learn feature interaction information, and training is performed in conjunction with binary classification labels. Based on the trained probabilistic prediction model, the predicted probability of the object to be predicted is output, and objects with predicted probabilities greater than a preset threshold are identified as potential target objects. Thus, by establishing an end-to-end artificial intelligence prediction framework, the implicit association between primary and supplementary insurance is transformed into a binary classification supervised learning task, and the high-order feature associations between data are automatically captured using multiple layers of feature interaction coding, thereby significantly improving the accuracy and objectivity of the target object's underwriting probability prediction.

[0083] According to embodiments of this application, this application also provides an electronic device and a readable storage medium.

[0084] Figure 4 This is a block diagram of an electronic device according to an embodiment of the present application.

[0085] like Figure 4 As shown, the electronic device includes: The memory 401, the processor 402, and the computer instructions stored in the memory 401 and executable on the processor 402.

[0086] When processor 402 executes instructions, it implements the underwriting probability prediction method based on artificial intelligence algorithms provided in the above embodiments.

[0087] Furthermore, electronic devices also include: Communication interface 403 is used for communication between memory 401 and processor 402.

[0088] The memory 401 is used to store computer instructions that can be executed on the processor 402.

[0089] Memory 401 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0090] The processor 402 is used to implement the underwriting probability prediction method based on artificial intelligence algorithm described in the above embodiment when executing the program.

[0091] If the memory 401, processor 402, and communication interface 403 are implemented independently, then the communication interface 403, memory 401, and processor 402 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0092] Optionally, in a specific implementation, if the memory 401, processor 402, and communication interface 403 are integrated on a single chip, then the memory 401, processor 402, and communication interface 403 can communicate with each other through an internal interface.

[0093] Processor 402 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0094] This application also proposes a computer program product that, when executed by an instruction processor, implements the underwriting probability prediction method based on artificial intelligence algorithms according to the embodiments of this application.

[0095] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0096] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0097] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0098] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0099] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0100] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0101] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0102] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for predicting underwriting probability based on artificial intelligence algorithms, characterized in that, include: The system acquires historical underwriting data for primary insurance and supplementary insurance. Based on preset association fields, it matches and associates samples of insured objects that have been underwritten for the primary insurance with the corresponding supplementary insurance underwriting data. It also assigns corresponding binary labels based on whether the samples of insured objects are further underwritten for supplementary insurance. Extract multidimensional features that influence the insurance purchase behavior of the target object and transform them into structured feature vectors; A probabilistic prediction model containing multiple layers of feature interaction coding is constructed. The structured feature vector is input into the probabilistic prediction model to learn feature interaction information, and the model is trained in combination with the binary classification labels. Based on the trained probability prediction model, the predicted probability of the object to be predicted is output, and objects with a predicted probability greater than a preset threshold are identified as potential target objects.

2. The method as described in claim 1, characterized in that, The main insurance is commercial vehicle damage insurance, and the supplementary insurance is driver accident insurance; the multi-dimensional features include: Contextual risk level characteristics, including main insurance risk level measurement indicators, main insurance underwriting quantitative indicators, and premium fluctuation coefficient; Based on basic human characteristics and historical behavioral preferences, including the number of times the main and supplementary insurance policies were underwritten, historical underwriting value metrics, and related policy information; Claims-related characteristics include the number of historical claims and quantitative indicators of historical payouts; The characteristics of a vehicle include its age, brand, usage, number of seats, tonnage, engine displacement, and energy type.

3. The method as described in claim 1, characterized in that, The preset associated fields include at least one or more of the following: policy number, object identification, and vehicle identification; the step of converting them into a structured feature vector includes: The system performs aberration cleaning, standardization of field definitions, missing value handling, continuous variable standardization, and categorical variable encoding on multi-source heterogeneous historical underwriting and claims data to transform them into structured feature vectors.

4. The method as described in claim 1, characterized in that, The probability prediction model, from bottom to top, includes an input feature layer, a feature embedding layer, a multi-layer feature interactive coding layer, and a top-level output layer. The structured feature vector is transformed into fully embedded features through the feature embedding layer; Each layer in the multi-layer feature interaction coding layer includes a factorization machine interaction module and a linear compression module, which are used to extract nonlinear feature interaction information based on the full embedded features.

5. The method as described in claim 4, characterized in that, The feature processing in the multi-layer feature cross-coding layer is as follows: The second-order interactive information representation between the learnable projection matrix in the factorization machine interaction module and the features extracted by the multilayer perceptron; The global information representation is obtained by linearly compressing the full embedded features based on the linear compression module. The second-order interactive information representation and the global information representation are concatenated and fused, then mapped back to the original dimension through a linear transformation. After residual connection and layer normalization processing are introduced in the multi-layer stacked structure, the data is input to the next layer.

6. The method as described in claim 4, characterized in that, The top-level output layer contains a multilayer perceptron. The output and training process of the probability prediction model includes: The Sigmoid function is used as the activation function to map the representation after multi-layer feature interaction to a single output node to output the predicted probability; During model training, a binary cross-entropy function is used as the loss function. The parameters are updated iteratively through backpropagation algorithm and adaptive moment estimation optimizer. In addition, embedding layer regularization, network structure regularization and random deactivation mechanism are combined to suppress model overfitting.

7. The method as described in claim 1, characterized in that, The preset threshold is determined based on the following method: On the validation set, a confusion matrix is ​​constructed based on the real labels and the predicted probabilities, and the precision and recall are calculated under different candidate thresholds. By combining a preset comprehensive evaluation function, the threshold that optimizes the result of the comprehensive evaluation function is selected from multiple candidate thresholds as the preset threshold.

8. The method as described in claim 1, characterized in that, The method further includes: The trained probability prediction model is exported as a general model data exchange format and deployed to the target inference service environment to build a prediction service interface. The predicted data generated from the actual application is sent back to the database, and the probability prediction model is periodically iterated and optimized.

9. An underwriting probability prediction device based on artificial intelligence algorithms, characterized in that, include: The acquisition module is used to acquire historical main insurance underwriting data and supplementary insurance underwriting data. Based on preset association fields, it matches and associates the underwriting object samples of the main insurance with the corresponding supplementary insurance underwriting data, and assigns corresponding binary labels based on whether the underwriting object samples are further underwritten for supplementary insurance. The feature extraction module is used to extract multidimensional features that affect the insurance purchase behavior of the target object and transform them into structured feature vectors; The model building module is used to build a probabilistic prediction model containing multiple layers of feature interaction coding. The structured feature vector is input into the probabilistic prediction model to learn feature interaction information, and the model is trained in combination with the binary classification labels. The prediction module is used to output the predicted probability of the object to be predicted based on the trained probability prediction model, and to identify objects with a predicted probability greater than a preset threshold as potential target objects.

10. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

11. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, in, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-8.

12. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-8.

Images