Vehicle insurance quotation information pushing method and device, storage medium and computer device

Abstract

The application discloses a vehicle insurance quotation information pushing method and device, a storage medium and computer equipment, relates to the field of financial technology, and mainly aims to solve the problem of poor effectiveness of existing vehicle insurance quotation information pushing. The method comprises the following steps: acquiring multi-modal vehicle data of an insured vehicle, and performing coding and fusion on the multi-modal vehicle data based on a Bayesian multi-modal fusion encoder to obtain fusion features; predicting the fusion features based on a quotation prediction model that has completed model training to obtain quotation action prediction information and quotation uncertainty prediction information, wherein the quotation prediction model is obtained by performing model training on a Bayesian neural network based on an uncertainty penalty term and an information gain reward term; and generating quotation information corresponding to the quotation action prediction information and the quotation uncertainty prediction information based on a target quotation strategy, and pushing the quotation information.

Description

Technical Field

[0001] This application relates to the field of financial technology, and in particular to a method, apparatus, storage medium, and computer equipment for pushing vehicle insurance quote information. Background Technology

[0002] Vehicle insurance pricing is a core aspect of vehicle insurance business operations, risk pricing, and customer conversion. Traditional vehicle insurance pricing often relies on fixed rate tables and manual experience-based rules, resulting in a rigid pricing model that cannot adapt to the differentiated insurance needs of different customers and vehicle models. With the increasing application of artificial intelligence decision-making algorithms in financial risk control, Deep Reinforcement Learning (DRL), with its advantages of autonomous interactive learning and dynamic decision optimization, is gradually being applied to intelligent dynamic pricing information push scenarios in vehicle insurance. This enables autonomous iteration of pricing strategies and personalized pricing, effectively compensating for the lack of flexibility in traditional rule-based pricing.

[0003] Currently, existing vehicle insurance quote information delivery typically employs a classic deep reinforcement learning framework for decision-making and delivery. For example, it uses customer attributes, vehicle model risk, and market conditions as environmental inputs, and specific insurance quotes as continuous outputs. A reward function is built based on environmental feedback, and the strategy network is continuously iterated and optimized to ultimately deliver the optimal dynamic quote. However, using customer attributes, vehicle model risk, and market conditions as environmental inputs ignores the role of environmental and cognitive uncertainty in decision modeling. Furthermore, in complex environments, this leads to significant fluctuations and slow convergence in the early stages of the decision model's learning, and it is highly prone to converging to suboptimal local strategies, such as always quoting high or low prices, greatly reducing the effectiveness of vehicle insurance quote information delivery. Summary of the Invention

[0004] In view of this, this application provides a method, apparatus, storage medium, and computer equipment for pushing vehicle insurance quote information, with the main purpose of solving the problem of poor effectiveness of existing vehicle insurance quote information push.

[0005] According to one aspect of this application, a method for pushing vehicle insurance quote information is provided, comprising: Acquire multimodal vehicle data of insured vehicles, and encode and fuse the multimodal vehicle data based on a Bayesian multimodal fusion encoder to obtain fused features; Based on the completed model training of the price prediction model, the fused features are predicted to obtain price action prediction information and price uncertainty prediction information. The price prediction model is obtained by training a Bayesian neural network based on an uncertainty penalty term and an information gain reward term. Based on the target pricing strategy, pricing information corresponding to the pricing action prediction information and the pricing uncertainty prediction information is generated and pushed out.

[0006] Furthermore, before the quoted price prediction model, based on the completed model training, predicts the fused features to obtain quoted price action prediction information and quoted price uncertainty prediction information, the method further includes: A Bayesian neural network is constructed based on a policy network, wherein the policy of the Bayesian neural network is composed of the mean of the bidding action and the uncertainty of the action; Acquire multimodal feature samples, which include state samples labeled with different actions and different uncertainties; The Bayesian neural network is trained based on the multimodal feature samples to obtain the price prediction model. The reward parameter of the price prediction model includes an uncertainty penalty term and an information gain reward term. The uncertainty penalty term is determined based on the action covariance, and the information gain reward term is determined based on the mutual information between the state and the action.

[0007] Furthermore, the step of training the Bayesian neural network based on the multimodal feature samples to obtain the price prediction model includes: The optimal policy parameters are determined by Bayesian policy gradient, and the posterior distribution is determined by variational inference. The loss function is determined based on the optimal policy parameters, the posterior distribution, and the preset hyperparameters. The loss function is then used to constrain the training of the Bayesian neural network, thereby completing the model training of the Bayesian neural network using the multimodal feature samples.

[0008] Furthermore, before training the Bayesian neural network based on the multimodal feature samples to obtain the price prediction model, the method further includes: Training task information is generated through an environment simulator. The training task information includes risk weights and reward conversion weights for different insurance risks. The step of training the Bayesian neural network based on the multimodal feature samples is then performed based on the training task information. After training the Bayesian neural network based on the multimodal feature samples to obtain the price prediction model, the method includes: Obtain the model training performance index of the Bayesian neural network, and update the training task information based on the model training performance index, so as to re-execute the step of training the Bayesian neural network based on the multimodal feature samples based on the updated training task information.

[0009] Furthermore, after the push is performed, the method further includes: Collect operation information on the quoted information, including transaction operation information and session operation information for vehicle insurance; The key pricing information in the transaction operation information and the session operation information is parsed, and the multimodal feature samples and the training task information are updated based on the key pricing information to retrain the Bayesian neural network model.

[0010] Furthermore, the Bayesian multimodal fusion encoder performs encoding and fusion on the multimodal vehicle data to obtain fused features, including: The probability distribution is determined based on the multimodal vehicle data and the potential state representation; The multimodal vehicle data and the probability distribution are encoded and fused based on the fusion network to obtain fused features. The loss function of the fusion network is constructed based on the encoder network, the decoder network, and KL divergence.

[0011] Furthermore, the step of generating pricing information corresponding to the pricing action prediction information and the pricing uncertainty prediction information based on the target pricing strategy includes: The pricing strategy engine is invoked. The pricing strategy engine has multiple preset target pricing strategies. Different target pricing strategies are used to characterize the rules for generating pricing information according to the pricing action based on different pricing uncertainty prediction information. Based on the pricing uncertainty prediction information, at least one target pricing strategy is determined, and reference pricing information corresponding to the pricing action prediction information is generated according to the target pricing strategy. If multiple identical reference price information exist or the reference price information is unique, then the reference price information is determined as the price information to be pushed. If all the reference price information is different, then a price information to be pushed is generated based on the multiple reference price information.

[0012] According to another aspect of this application, a vehicle insurance quote information push device is provided, comprising: The acquisition module is used to acquire multimodal vehicle data of insured vehicles, and to encode and fuse the multimodal vehicle data based on a Bayesian multimodal fusion encoder to obtain fused features; The prediction module is used to predict the fused features based on the completed model training of the price prediction model, and to obtain price action prediction information and price uncertainty prediction information. The price prediction model is obtained by training a Bayesian neural network based on an uncertainty penalty term and an information gain reward term. The push module is used to generate and push quotation information corresponding to the quotation action prediction information and the quotation uncertainty prediction information based on the target quotation strategy.

[0013] Furthermore, the device also includes: The training module is used to construct a Bayesian neural network based on a policy network. The policy of the Bayesian neural network is represented by the mean of the bidding action and the uncertainty of the action. It acquires multimodal feature samples, which include state samples labeled with different actions and different uncertainties. Based on the multimodal feature samples, it trains the Bayesian neural network to obtain the bidding prediction model. The reward parameter of the bidding prediction model introduces an uncertainty penalty term and an information gain reward term. The uncertainty penalty term is determined based on the action covariance, and the information gain reward term is determined by the mutual information between the state and the action.

[0014] Furthermore, The training module is specifically used to determine the optimal policy parameters through the Bayesian policy gradient and determine the posterior distribution through variational inference; determine the loss function based on the optimal policy parameters, the posterior distribution, and preset hyperparameters, and apply training constraints to the Bayesian neural network through the loss function to complete the model training of the Bayesian neural network using the multimodal feature samples.

[0015] Furthermore, the device also includes: A generation module is used to generate training task information through an environment simulator. The training task information includes risk weights and reward conversion weights for different insurance risks, so as to perform a step of training the Bayesian neural network based on the multimodal feature samples based on the training task information. An update module is used to obtain the model training performance index of the Bayesian neural network and update the training task information based on the model training performance index, so as to re-execute the step of training the Bayesian neural network based on the multimodal feature samples based on the updated training task information.

[0016] Furthermore, the update module is also used to collect operation information on the quotation information, including transaction operation information and session operation information for vehicle insurance; parse the key quotation information in the transaction operation information and the session operation information, and update the multimodal feature samples and the training task information based on the key quotation information, so as to retrain the Bayesian neural network model.

[0017] Furthermore, the acquisition module is specifically used to determine the probability distribution based on the multimodal vehicle data and the potential state representation; and to encode and fuse the multimodal vehicle data and the probability distribution based on the fusion network to obtain fused features. The loss function of the fusion network is constructed based on the encoder network, the decoder network, and KL divergence.

[0018] Furthermore, the push module is specifically used to invoke a pricing strategy engine, which has multiple preset target pricing strategies. Different target pricing strategies are used to characterize the rules for generating pricing information according to pricing actions based on different pricing uncertainty prediction information. Based on the pricing uncertainty prediction information, at least one target pricing strategy is determined, and reference pricing information corresponding to the pricing action prediction information is generated according to the target pricing strategy. If multiple identical reference pricing information exists or the reference pricing information is unique, the reference pricing information is determined as the pricing information to be pushed. If multiple reference pricing information are different, the pricing information to be pushed is generated based on the multiple reference pricing information.

[0019] According to another aspect of this application, a storage medium is provided, wherein at least one executable instruction is stored therein, the executable instruction causing a processor to perform an operation corresponding to the above-described vehicle insurance quotation information push method.

[0020] According to another aspect of this application, a computer device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described vehicle insurance quotation information push method.

[0021] By employing the above technical solutions, the technical solutions provided in the embodiments of this application have at least the following advantages: This application provides a method, apparatus, storage medium, and computer device for pushing vehicle insurance quote information. The embodiments of this application acquire multimodal vehicle data of the insured vehicle and encode and fuse the multimodal vehicle data based on a Bayesian multimodal fusion encoder to obtain fusion features. Based on a pre-trained quote prediction model, the fusion features are predicted to obtain quote action prediction information and quote uncertainty prediction information. The quote prediction model is obtained by training a Bayesian neural network based on an uncertainty penalty term and an information gain reward term. Quote information corresponding to the quote action prediction information and the quote uncertainty prediction information is generated based on a target quote strategy and pushed out. By using Bayesian decision-making for quote prediction, the errors and rigid wording of manual calculations are avoided, greatly improving the decision-making ability of the quote prediction model and meeting the high-accuracy prediction requirements for quote information under complex data dimensions, thereby improving the efficiency of vehicle insurance quote push.

[0022] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This paper illustrates a flowchart of a method for pushing vehicle insurance quote information according to an embodiment of this application. Figure 2 This illustration shows a block diagram of a vehicle insurance quote information push device according to an embodiment of this application; Figure 3 A schematic diagram of the structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0027] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0028] Based on this, in one embodiment, this application provides a method for pushing vehicle insurance quote information. Taking the application of this method to computer devices such as servers as an example, the server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0029] This application provides a method for pushing vehicle insurance quote information, such as... Figure 1 As shown, the method includes: 101. Obtain multimodal vehicle data of the insured vehicle, and encode and fuse the multimodal vehicle data based on a Bayesian multimodal fusion encoder to obtain fused features.

[0030] In this embodiment, the current execution terminal, acting as the executing entity, can be a processor or server of a financial institution to acquire multimodal vehicle data of insured vehicles. This multimodal vehicle data can include image data, text data, voice data, and time-series behavioral data. For example, text data may include policy information, vehicle information, and user information; voice data may include call recordings and voicemail messages; image data may include ID photos and vehicle photos; and time-series behavioral data may include policy claim timelines, price comparison behavior, and operation times. This embodiment does not impose specific limitations on these data. The multimodal vehicle data in this embodiment can be retrieved based on data cached in the processor or server. This embodiment does not impose specific limitations on these data.

[0031] It should be noted that, in order to be applicable to the model input of the price prediction model, the current execution end performs encoding and fusion of multimodal vehicle data to obtain fused features. At this point, a Bayesian multimodal fusion encoder can be used for encoding and fusion. The Bayesian multimodal fusion encoder, based on Bayesian probabilistic inference, performs probabilistic feature extraction on multi-source heterogeneous modal data such as text, speech, and behavioral temporal data, and achieves cross-modal feature fusion based on Bayesian posterior probability. The output fused features are no longer a single fixed feature vector, but a latent variable distribution with mean and variance, which serves as the prediction input parameters for the price prediction model.

[0032] 102. Based on the completed model training, the fused features are predicted to obtain bidding action prediction information and bidding uncertainty prediction information.

[0033] In this embodiment, the price prediction model is obtained by training a Bayesian neural network based on an uncertainty penalty term and an information gain reward term. The price prediction model uses a deep learning policy network as its foundation to construct a Bayesian neural network, and uses the fusion features obtained in step 101 as model input to make predictions, thus obtaining the price prediction result. The price prediction result can include price action prediction information and price uncertainty prediction information. Price action prediction information is used to characterize the determined price value in the auto insurance scenario, which can be a premium discount rate, a base premium increase or decrease coefficient, or the final policy price amount; it belongs to the decision-making action itself. Price uncertainty prediction information is used to characterize the risk level of the current price value and represents whether the price prediction model is confident in the current pricing result; it is a confidence quantification indicator.

[0034] In one embodiment, the pricing action prediction information is input into the pricing prediction model from the mean portion of the Bayesian multimodal fusion features, and the output consists of continuous actions constituting the pricing action prediction information. The pricing uncertainty prediction information has two layers of meaning. The first layer is accidental uncertainty or environmental uncertainty, that is, the risk brought about by random fluctuations in customer purchasing behavior. The second layer is cognitive uncertainty or model uncertainty, that is, the decision-making risk brought about by the sparse data of new car models and new customers and the lack of experience of the pricing prediction model.

[0035] 103. Generate pricing information corresponding to the pricing action prediction information and the pricing uncertainty prediction information based on the target pricing strategy, and push it out.

[0036] In this embodiment, the target pricing strategy is used to characterize the rules for generating pricing information based on different pricing uncertainty prediction information according to the pricing action. This serves as the basis for generating the second pricing, generating pricing information corresponding to the pricing action prediction information and the pricing uncertainty prediction information, and then pushing it out. At this time, the pricing information may include, but is not limited to, the final pricing values ​​under different times and scenarios; this embodiment does not impose specific limitations.

[0037] In another embodiment of this application, for further definition and explanation, before the step of predicting the fused features based on the completed model training of the price prediction model to obtain price action prediction information and price uncertainty prediction information, the method further includes: Constructing a Bayesian neural network based on a policy network; Obtain multimodal feature samples; The Bayesian neural network is trained based on the multimodal feature samples to obtain the price prediction model.

[0038] To achieve reinforcement learning of bidding actions and thus improve the accuracy of bid prediction, the current execution end constructs a Bayesian neural network based on a policy network. All network weights of the Bayesian neural network are modeled as Gaussian probability distributions, and the policy is represented by the mean of the bidding action and the uncertainty of the action. Furthermore, the multimodal feature samples include state samples labeled with different actions and different uncertainties. After model training, the output of the Bayesian neural network is no longer a single, definitive bid, but rather a probability distribution of bidding actions. Simultaneously, it outputs the bid mean and variance, i.e., the predicted action value and the predicted uncertainty value. This embodiment does not impose specific limitations on these aspects.

[0039] During model training, after constructing a Bayesian Neural Network (BNN) based on a deep reinforcement learning policy network, the weight parameters of each layer of the policy network are modeled as Gaussian distributions, and network training is completed through variational approximation. The Bayesian Neural Network outputs a probability distribution of bidding actions. Bidding action prediction information is obtained from the mean of the distribution, and bidding uncertainty prediction information is obtained from the variance of the distribution. This bidding uncertainty prediction information is then used to impose constraints on the bidding action range. For example, the policy network... (a|s) is also constructed as a Bayesian neural network (BNN), with the input sample being the state ( , The output is the action. The probability distribution p(a|s) is usually assumed to be a Gaussian distribution N( , The strategy can be represented as: ;in, This is the average of all pricing actions (such as rate coefficients and discount rates). For action uncertainty, combined with state uncertainty Given the uncertainty of the parameters of the policy network itself, at this time, The larger the value, the greater the uncertainty of the pricing prediction model in this state, meaning it doesn't know how to price.

[0040] In a specific embodiment, the reward parameters of the price prediction model can also incorporate an uncertainty penalty term and an information gain reward term. In this case, the uncertainty penalty term is determined based on the action covariance, and the information gain reward term is determined by the mutual information between the state and the action. Specifically, the improved reward parameters... Represented as: ; in, The uncertainty penalty term is calculated based on the action covariance matrix and is proportional to... If the penalty is high, then the decision-making method with uncertainty is selected, which encourages the price prediction model to take conservative actions (such as moving closer to the benchmark price) when there is uncertainty, so as to improve the safety of decision-making. This is the information gain reward term, calculated based on the mutual information between state and action. Maximizing this term encourages the predictive pricing model to take actions that minimize state uncertainty, i.e., explore offers that provide the most information. For example, for a customer with ambiguous risk, a moderate tentative offer (e.g., a possible deal or no deal) provides more information than an extremely low or extremely high offer (e.g., a customer's behavior is easily predictable). , To balance hyperparameters.

[0041] In another embodiment of this application, for further definition and explanation, the step of training the Bayesian neural network based on the multimodal feature samples to obtain the price prediction model includes: The optimal policy parameters are determined by Bayesian policy gradient, and the posterior distribution is determined by variational inference. The loss function is determined based on the optimal policy parameters, the posterior distribution, and the preset hyperparameters. The loss function is then used to constrain the training of the Bayesian neural network, thereby completing the model training of the Bayesian neural network using the multimodal feature samples.

[0042] To enhance the decision-making ability of the price prediction model, a reward is introduced to improve the accuracy of the price quotes. Specifically, when training the Bayesian neural network model at the current execution end, the optimal policy parameters are first determined through the Bayesian policy gradient, and then the posterior distribution, i.e., the posterior distribution p(θ|τ) of the optimal policy parameters, is determined through variational inference, where τ represents the trajectory data. Furthermore, variational inference is used to determine the posterior distribution in order to learn an approximate posterior. The distribution is preferably Gaussian. The loss function includes policy performance such as cumulative reward and KL divergence with the prior: ; in, The optimal policy parameters are determined so that the expected cumulative reward can be improved through the policy π_θ. , These are preset hyperparameters that control the complexity and fit of a model. Let KL divergence be denoted as KL divergence.

[0043] In another embodiment of this application, for further definition and explanation, before the step of training the Bayesian neural network based on the multimodal feature samples to obtain the price prediction model, the method further includes: Training task information is generated using an environment simulator; To better adapt the training process to the vehicle insurance pricing scenario and improve the learning effect of the pricing prediction model, the current execution end generates training task information through an environment simulator. This environment simulator can be a model training simulator built based on a course learning scheduler, used to simulate the triggering of training tasks to dynamically adjust the difficulty of the training environment and the task objectives. Furthermore, the training task information includes risk weights and reward conversion weights for different insurance risks, enabling the execution of the step of training the Bayesian neural network based on the multimodal feature samples.

[0044] In a specific embodiment, the generation of training task information may include an imitation learning phase, a balanced learning phase, and a robustness and exploration phase. In the imitation learning phase, the environment simulator primarily simulates training scenarios for quoting simple insurance policies with historically high conversion rates and low risk, such as common new family cars. Reward function. Initially, the conversion weight α is very high, while the risk weight γ is low. During the equilibrium learning phase, the environment simulator gradually introduces more diverse, moderate-risk policies and adjusts... The weights of α and γ are adjusted to train the quote prediction model to achieve a balance between conversion and risk.

[0045] For the robustness and exploration phases, the environment simulator introduces edge cases, such as rare car models, high-risk drivers, and adversarial scenarios simulating competitor price fluctuations. Simultaneously, improvement rewards are increased during this phase. Information gain reward The weights are assigned to encourage the pricing prediction model to explore and learn more deeply.

[0046] Correspondingly, after training the Bayesian neural network based on the multimodal feature samples to obtain the price prediction model, the method includes: Obtain the model training performance index of the Bayesian neural network, and update the training task information based on the model training performance index, so as to re-execute the step of training the Bayesian neural network based on the multimodal feature samples based on the updated training task information.

[0047] To enable cyclical dynamic adjustments during model training and improve training effectiveness, the current execution end can acquire model training performance metrics after training and update the training task information based on these metrics. These metrics may include average reward, risk control achievement rate, etc., but are not specifically limited in this embodiment.

[0048] In a specific example, when M exceeds the threshold T for N consecutive periods, and the policy is uncertain... When the value falls below the threshold U, the training task information is updated, and the model enters the next round of training.

[0049] In another embodiment of this application, for further definition and explanation, after the push step, the method further includes: Collect operation information on the quoted price information; The key pricing information in the transaction operation information and the session operation information is parsed, and the multimodal feature samples and the training task information are updated based on the key pricing information to retrain the Bayesian neural network model.

[0050] To meet the need for accurate learning of the predictive pricing model and improve the deep learning effect of the pricing strategy, the current execution terminal collects customer operation information related to the final pricing information after pushing it out. This operation information includes vehicle insurance transaction operation information and conversation operation information. Transaction operation information refers to the information generated when a user purchases vehicle insurance according to the pricing, while conversation operation information refers to the information generated when a user interacts with the user via voice to inquire or negotiate a price based on the pricing. This embodiment of the application does not impose specific limitations on these aspects.

[0051] Furthermore, the current execution end parses the data format of transaction operation information and session operation information, such as performing natural language semantic parsing or speech recognition parsing, to obtain key pricing information in the transaction operation information and session operation information, such as pricing keywords like "agree," "okay," and "a bit expensive," etc. This embodiment of the application does not impose specific limitations. Finally, based on the key pricing information, the multimodal feature samples and training task information are updated to retrain the Bayesian neural network model. This update includes updating the pricing values ​​or uncertainties in the multimodal feature samples, and may also include updating the risk weights or reward conversion weights in the training task information, such as increasing or decreasing them. This can be done based on a preset update mapping relationship, which is not specifically limited in this embodiment of the application.

[0052] In another embodiment of this application, for further definition and explanation, the step of encoding and fusing the multimodal vehicle data based on a Bayesian multimodal fusion encoder to obtain fused features includes: The probability distribution is determined based on the multimodal vehicle data and the potential state representation; The multimodal vehicle data and the probability distribution are encoded and fused based on the fusion network to obtain fused features.

[0053] To fuse multimodal data and meet the uncertainty recognition requirements of subsequent price prediction models, thereby improving the accuracy of price prediction, the current execution end first determines the probability distribution based on multimodal vehicle data and latent state representations during encoding fusion. This probability distribution is represented as p(z|X), where z is the latent state representation and X is the multimodal vehicle data. Preferably, X = { , , The data consists of image, speech, and text vehicle data, respectively. Then, a fusion network is used to encode and fuse the multimodal vehicle data and probability distributions to obtain fused features. The loss function of the fusion network is constructed based on the encoder and decoder networks and KL divergence.

[0054] In a specific embodiment, z follows a Gaussian distribution N( , The fusion process can be represented as: ; The loss function can be expressed as: ; in, For encoder networks, the parameters are: This is used to estimate the parameters of this distribution. For a Bayesian neural network with weights that are distributed rather than point estimates, each weight in the network has a mean and a variance, and its forward propagation involves sampling, thus affecting its output. and It also has probabilistic significance. Let z be the mean vector of the potential state z, serving as the state representation. . Let be the standard deviation vector of the latent state z, representing cognitive uncertainty. When the input X is a new car model not seen by the model or a blurry image... The value will increase significantly. Let z be a decoder network with parameter ψ, which attempts to reconstruct the input X from z. For the KL divergence, the posterior distribution is forced to... Approximate the prior p(z), preferably the standard normal distribution. β is a hyperparameter that controls the balance between reconstruction accuracy and distribution constraints. In this case, β can be increased to make the learned z representation more decoupled and more interpretable.

[0055] In another embodiment of this application, for further definition and explanation, the step of generating quotation information corresponding to the quotation action prediction information and the quotation uncertainty prediction information based on the target quotation strategy includes: Retrieve the pricing strategy engine; Based on the pricing uncertainty prediction information, at least one target pricing strategy is determined, and reference pricing information corresponding to the pricing action prediction information is generated according to the target pricing strategy. If multiple identical reference price information exist or the reference price information is unique, then the reference price information is determined as the price information to be pushed. If all the reference price information is different, then a price information to be pushed is generated based on the multiple reference price information.

[0056] To generate a final effective quote value and improve quote validity, the current execution end first invokes the quote strategy engine when generating quote information. This engine has multiple preset target quote strategies, each representing a rule for generating quote information based on different quote uncertainty predictions. Then, based on the quote uncertainty predictions in the quote prediction results, at least one target quote strategy capable of generating quote information is determined, thereby generating reference quote information corresponding to the quote action prediction information according to the target quote strategy. For example, if the uncertainty prediction information corresponding to quote 'a' is less than 10% or relatively uncertain, target quote strategy 1 and target quote strategy 2 can be matched. Specifically, quote 'a' is increased by 10% according to strategy 1 and increased by 20% according to target quote strategy 2, respectively, as reference quote information. If multiple identical reference quote information or a unique reference quote information exists, it indicates that the quote is relatively stable; therefore, this reference quote information is determined as the quote information to be pushed. If multiple reference quote information are different, it indicates that the quote is relatively unstable; therefore, quote information to be pushed is generated based on multiple reference quote information.

[0057] In a training example for a price prediction model, the sample is the price quote for a modified, older sports car—a typical high-risk, data-sparse scenario. The inputs are vehicle model data (old sports car), exterior images (obvious modifications), and text records (multiple speeding tickets). Because the combination of "modified car" and "multiple speeding tickets" is rare in the training data, the cognitive uncertainty of the Bayesian fusion encoder output is significant. It will be significantly higher than that of ordinary vehicles. Status Information such as "high-risk, young car owner preference" was encoded, but This is represented as "the reliability of this state judgment is low". During Bayesian DRL decision training, the policy network... take over( , Due to high input uncertainty and the rarity of such samples in policy network training, the uncertainty of the network's own parameters is also activated, resulting in a very large uncertainty σ_a in its output action. At this point, the huge... This will lead to The penalty term has a large absolute value. To prevent the total reward from being too high... If the value is too low, the policy network will be forced to tend to output a certain value. Smaller, or more "certain," actions, such as outputting a quote that is close to prior knowledge or historical success stories, like an actuarial benchmark price. For example, a higher base rate with a markup factor, such as 1.8, and without offering large coupons, achieves robust decision-making under uncertainty. Meanwhile, if the training task information is in the robust and exploratory phase, the environment simulator might set an exploratory objective for such high-risk samples, allowing the pricing prediction model to give a price deviating from the robust strategy with a certain probability, such as a moderate rate bundled with a discount on advanced driver monitoring equipment, to collect I( ; Information gain reward. If the price prediction model performs poorly on such samples in the long run, such as with a low average reward for performance parameters, the environment simulator may temporarily revert it to stage two, consolidate its basic strategy with more medium-risk samples, and then challenge the high-risk scenario again.

[0058] In a sample price quote push, a high-rate, low-discount quote for a recommended in-vehicle monitoring device is generated and pushed as a structured card. The server generates a prompt message indicating that the customer's risk profile is significant, the system provides a high-premium quote with moderate confidence, and suggests emphasizing the risk binding and the value of the equipment discount. At this point, regardless of whether the customer accepts, rejects, or negotiates, their feedback is recorded. Customer acceptance reinforces the high-risk, high-pricing strategy; rejection followed by the customer switching to a competitor serves as a negative reward, prompting the strategy to fine-tune prices or add-ons in similar situations in the future. All this interaction data is used to continuously update the posterior distribution q(θ) of the Bayesian policy network, making the strategy increasingly robust and accurate.

[0059] This application provides a method for pushing vehicle insurance quote information. This method acquires multimodal vehicle data of the insured vehicle and encodes and fuses the multimodal vehicle data using a Bayesian multimodal fusion encoder to obtain fusion features. It then predicts the fusion features using a pre-trained quote prediction model to obtain quote action prediction information and quote uncertainty prediction information. The quote prediction model is obtained by training a Bayesian neural network based on an uncertainty penalty term and an information gain reward term. Finally, it generates quote information corresponding to the quote action prediction information and the quote uncertainty prediction information based on a target quote strategy and pushes it to the insured vehicle insurance company. By using Bayesian decision-making to predict quotes, this method avoids the errors and rigid wording of manual calculations, greatly improving the decision-making ability of the quote prediction model and meeting the need for high-accuracy quote prediction under complex data dimensions, thereby improving the efficiency of vehicle insurance quote push.

[0060] Furthermore, as a response to the above Figure 1 The implementation of the method shown in this application provides a vehicle insurance quote information push device, such as... Figure 2 As shown, the device includes: The acquisition module 21 is used to acquire multimodal vehicle data of the insured vehicle, and to encode and fuse the multimodal vehicle data based on a Bayesian multimodal fusion encoder to obtain fused features; Prediction module 22 is used to predict the fused features based on the completed model training of the quotation prediction model to obtain quotation action prediction information and quotation uncertainty prediction information. The quotation prediction model is obtained by training a Bayesian neural network based on an uncertainty penalty term and an information gain reward term. The push module 23 is used to generate and push quotation information corresponding to the quotation action prediction information and the quotation uncertainty prediction information based on the target quotation strategy.

[0061] Furthermore, the device also includes: The training module is used to construct a Bayesian neural network based on a policy network. The policy of the Bayesian neural network is represented by the mean of the bidding action and the uncertainty of the action. It acquires multimodal feature samples, which include state samples labeled with different actions and different uncertainties. Based on the multimodal feature samples, it trains the Bayesian neural network to obtain the bidding prediction model. The reward parameter of the bidding prediction model introduces an uncertainty penalty term and an information gain reward term. The uncertainty penalty term is determined based on the action covariance, and the information gain reward term is determined by the mutual information between the state and the action.

[0062] Furthermore, The training module is specifically used to determine the optimal policy parameters through the Bayesian policy gradient and determine the posterior distribution through variational inference; determine the loss function based on the optimal policy parameters, the posterior distribution, and preset hyperparameters, and apply training constraints to the Bayesian neural network through the loss function to complete the model training of the Bayesian neural network using the multimodal feature samples.

[0063] Furthermore, the device also includes: A generation module is used to generate training task information through an environment simulator. The training task information includes risk weights and reward conversion weights for different insurance risks, so as to perform a step of training the Bayesian neural network based on the multimodal feature samples based on the training task information. An update module is used to obtain the model training performance index of the Bayesian neural network and update the training task information based on the model training performance index, so as to re-execute the step of training the Bayesian neural network based on the multimodal feature samples based on the updated training task information.

[0064] Furthermore, the update module is also used to collect operation information on the quotation information, including transaction operation information and session operation information for vehicle insurance; parse the key quotation information in the transaction operation information and the session operation information, and update the multimodal feature samples and the training task information based on the key quotation information, so as to retrain the Bayesian neural network model.

[0065] Furthermore, the acquisition module is specifically used to determine the probability distribution based on the multimodal vehicle data and the potential state representation; and to encode and fuse the multimodal vehicle data and the probability distribution based on the fusion network to obtain fused features. The loss function of the fusion network is constructed based on the encoder network, the decoder network, and KL divergence.

[0066] Furthermore, the push module is specifically used to invoke a pricing strategy engine, which has multiple preset target pricing strategies. Different target pricing strategies are used to characterize the rules for generating pricing information according to pricing actions based on different pricing uncertainty prediction information. Based on the pricing uncertainty prediction information, at least one target pricing strategy is determined, and reference pricing information corresponding to the pricing action prediction information is generated according to the target pricing strategy. If multiple identical reference pricing information exists or the reference pricing information is unique, the reference pricing information is determined as the pricing information to be pushed. If multiple reference pricing information are different, the pricing information to be pushed is generated based on the multiple reference pricing information.

[0067] This application provides a vehicle insurance quote information push device. This embodiment acquires multimodal vehicle data of the insured vehicle and encodes and fuses the multimodal vehicle data based on a Bayesian multimodal fusion encoder to obtain fusion features. Based on a pre-trained quote prediction model, the fusion features are predicted to obtain quote action prediction information and quote uncertainty prediction information. The quote prediction model is obtained by training a Bayesian neural network based on an uncertainty penalty term and an information gain reward term. Quote information corresponding to the quote action prediction information and the quote uncertainty prediction information is generated based on a target quote strategy and pushed out. By using Bayesian decision-making for quote prediction, the device avoids the errors and rigid wording of manual calculations, greatly improving the decision-making ability of the quote prediction model and meeting the high-accuracy prediction requirements for quote information under complex data dimensions, thereby improving the efficiency of vehicle insurance quote push.

[0068] According to one embodiment of this application, a storage medium is provided, the storage medium storing at least one executable instruction, which can execute the vehicle insurance quotation information push method in any of the above method embodiments.

[0069] Figure 3 The diagram shows a structural schematic of a computer device according to one embodiment of the present application. The specific embodiments of the present application do not limit the specific implementation of the computer device.

[0070] like Figure 3 As shown, the computer device may include: a processor 302, a communications interface 304, a memory 306, and a communications bus 308.

[0071] The processor 302, communication interface 304, and memory 306 communicate with each other via communication bus 308.

[0072] Communication interface 304 is used to communicate with other network elements such as clients or other servers.

[0073] The processor 302 is used to execute program 310, specifically to execute the relevant steps in the above-described embodiment of the vehicle insurance quotation information push method.

[0074] Specifically, program 310 may include program code that includes computer operation instructions.

[0075] Processor 302 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. The computer device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0076] Memory 306 is used to store program 310. Memory 306 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0077] Specifically, program 310 can be used to cause processor 302 to perform the following operations: Acquire multimodal vehicle data of insured vehicles, and encode and fuse the multimodal vehicle data based on a Bayesian multimodal fusion encoder to obtain fused features; Based on the completed model training of the price prediction model, the fused features are predicted to obtain price action prediction information and price uncertainty prediction information. The price prediction model is obtained by training a Bayesian neural network based on an uncertainty penalty term and an information gain reward term. Based on the target pricing strategy, pricing information corresponding to the pricing action prediction information and the pricing uncertainty prediction information is generated and pushed out.

[0078] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0079] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for pushing vehicle insurance quote information, characterized in that, include: Acquire multimodal vehicle data of insured vehicles, and encode and fuse the multimodal vehicle data based on a Bayesian multimodal fusion encoder to obtain fused features; Based on the completed model training of the price prediction model, the fused features are predicted to obtain price action prediction information and price uncertainty prediction information. The price prediction model is obtained by training a Bayesian neural network based on an uncertainty penalty term and an information gain reward term. Based on the target pricing strategy, pricing information corresponding to the pricing action prediction information and the pricing uncertainty prediction information is generated and pushed out.

2. The method according to claim 1, characterized in that, Before the quoted price prediction model, based on the completed model training, predicts the fused features to obtain quoted price action prediction information and quoted price uncertainty prediction information, the method further includes: A Bayesian neural network is constructed based on a policy network, wherein the policy of the Bayesian neural network is composed of the mean of the bidding action and the uncertainty of the action; Acquire multimodal feature samples, which include state samples labeled with different actions and different uncertainties; The Bayesian neural network is trained based on the multimodal feature samples to obtain the price prediction model. The reward parameter of the price prediction model includes an uncertainty penalty term and an information gain reward term. The uncertainty penalty term is determined based on the action covariance, and the information gain reward term is determined based on the mutual information between the state and the action.

3. The method according to claim 2, characterized in that, The step of training the Bayesian neural network based on the multimodal feature samples to obtain the price prediction model includes: The optimal policy parameters are determined by Bayesian policy gradient, and the posterior distribution is determined by variational inference. The loss function is determined based on the optimal policy parameters, the posterior distribution, and the preset hyperparameters. The loss function is then used to constrain the training of the Bayesian neural network, thereby completing the model training of the Bayesian neural network using the multimodal feature samples.

4. The method according to claim 3, characterized in that, Before training the Bayesian neural network based on the multimodal feature samples to obtain the price prediction model, the method further includes: Training task information is generated through an environment simulator. The training task information includes risk weights and reward conversion weights for different insurance risks. The step of training the Bayesian neural network based on the multimodal feature samples is then performed based on the training task information. After training the Bayesian neural network based on the multimodal feature samples to obtain the price prediction model, the method includes: Obtain the model training performance index of the Bayesian neural network, and update the training task information based on the model training performance index, so as to re-execute the step of training the Bayesian neural network based on the multimodal feature samples based on the updated training task information.

5. The method according to claim 3, characterized in that, After the push is performed, the method further includes: Collect operation information on the quoted information, including transaction operation information and session operation information for vehicle insurance; The key pricing information in the transaction operation information and the session operation information is parsed, and the multimodal feature samples and the training task information are updated based on the key pricing information to retrain the Bayesian neural network model.

6. The method according to claim 1, characterized in that, The Bayesian multimodal fusion encoder encodes and fuses the multimodal vehicle data to obtain fused features, including: The probability distribution is determined based on the multimodal vehicle data and the potential state representation; The multimodal vehicle data and the probability distribution are encoded and fused based on the fusion network to obtain fused features. The loss function of the fusion network is constructed based on the encoder network, the decoder network, and KL divergence.

7. The method according to claim 1, characterized in that, The generation of pricing information based on the target pricing strategy, corresponding to the pricing action prediction information and the pricing uncertainty prediction information, includes: The pricing strategy engine is invoked. The pricing strategy engine has multiple preset target pricing strategies. Different target pricing strategies are used to characterize the rules for generating pricing information according to the pricing action based on different pricing uncertainty prediction information. Based on the pricing uncertainty prediction information, at least one target pricing strategy is determined, and reference pricing information corresponding to the pricing action prediction information is generated according to the target pricing strategy. If multiple identical reference price information exist or the reference price information is unique, then the reference price information is determined as the price information to be pushed. If all the reference price information is different, then a price information to be pushed is generated based on the multiple reference price information.

8. A vehicle insurance quote information push device, characterized in that, include: The acquisition module is used to acquire multimodal vehicle data of insured vehicles, and to encode and fuse the multimodal vehicle data based on a Bayesian multimodal fusion encoder to obtain fused features; The prediction module is used to predict the fused features based on the completed model training of the price prediction model, and to obtain price action prediction information and price uncertainty prediction information. The price prediction model is obtained by training a Bayesian neural network based on an uncertainty penalty term and an information gain reward term. The push module is used to generate and push quotation information corresponding to the quotation action prediction information and the quotation uncertainty prediction information based on the target quotation strategy.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

Images