Object risk assessment method and apparatus, electronic device, and storage medium

Abstract

The application relates to the technical field of intelligent decision-making, and discloses an object risk assessment method and device, electronic equipment and a storage medium, which comprise the following steps: receiving an object processing request, extracting target scene context features and collecting target multi-source heterogeneous data; obtaining a target heterogeneous sub-model set based on target scene context feature matching; performing feature mapping and confidence assessment on the target multi-source heterogeneous data by using the target heterogeneous sub-model set, and obtaining a multi-dimensional initial feature tensor and a multi-dimensional output confidence; performing joint quantitative mapping on the multi-dimensional output confidence by combining the target scene context features to obtain a dynamic fusion weight; performing weighted fusion on the multi-dimensional initial feature tensor based on the dynamic fusion weight to obtain a comprehensive risk representation vector, and then performing risk assessment to obtain a target object processing strategy. The application can be applied to the financial technology scene and the health care scene, and improves the accuracy of object risk assessment in a variable scene.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence technology and intelligent decision-making technology, and can be applied to financial technology scenarios and healthcare scenarios. In particular, it relates to a method, device, electronic device and storage medium for object risk assessment. Background Technology

[0002] Object risk assessment is used to quantify and predict the characteristics of dynamic entities involved in various complex business scenarios from multiple dimensions, enabling accurate decision-making and planning in uncertain environments. This avoids system instability or resource allocation errors caused by delayed risk prediction. Taking intelligent risk control and dynamic credit granting in the financial sector as an example, object risk assessment can achieve end-to-end multi-source heterogeneous feature fusion, from static user identity profiles, real-time transaction behavior stream data, unstructured behavioral images to external macroeconomic environment data. This provides a reliable data foundation for dynamic security authentication and automatic policy distribution in financial systems, improving the efficiency of risk interception and reducing platform operational risks. Similarly, in the field of intelligent monitoring and intervention guidance in healthcare, object risk assessment can classify and extract various multimodal data, such as historical electronic medical records, real-time vital sign streams from wearable devices, and CT medical image text reports, and perform adaptive risk rating. This provides clear and orderly quantitative support for medical terminals to quickly generate personalized intervention strategies and accurately predict health risks, ensuring the accuracy and efficiency of medical data processing.

[0003] Currently, common object risk assessment methods typically begin by acquiring single-source historical structured data of the target object. This data is then input into a pre-defined, fixed assessment model for feature extraction, yielding single-dimensional risk features. Finally, these single-dimensional risk features are mapped and calculated based on fixed, preset weighting rules to obtain the final risk management strategy. However, because this object risk assessment method uses a fixed assessment model, its rigid processing is completely decoupled from the specific business context when facing complex and ever-changing real-world business requests. It cannot adaptively match and dynamically fuse heterogeneous models according to the object's current scenario, leading to distortion in the extracted target risk feature vector across different scenarios. Consequently, the generated risk management strategy cannot accurately match the current actual risk state, resulting in reduced accuracy in object risk assessment. Therefore, improving the accuracy of object risk assessment in variable scenarios has become a pressing technical problem. Summary of the Invention

[0004] This invention provides an artificial intelligence-based object risk assessment method, apparatus, computer equipment, and medium, aiming to solve the technical problem that existing object risk assessment methods, due to the use of static and fixed models, are decoupled from the specific business scenario context, resulting in distorted risk characterization in variable scenarios, thereby improving the accuracy of object risk assessment in variable scenarios.

[0005] Firstly, an object risk assessment method is provided, which is applied to an object risk assessment system, including: Receive an object processing request and extract features from the object processing request to obtain target scene context features; Based on the object processing request, multimodal data acquisition is performed to obtain target multi-source heterogeneous data; Based on the target scene context features, model matching is performed to obtain a set of target heterogeneous sub-models; Based on the target heterogeneous sub-model set, modal feature mapping is performed on the target multi-source heterogeneous data to obtain a multi-dimensional initial feature tensor. Based on the target heterogeneous sub-model set, the model confidence of the target multi-source heterogeneous data is evaluated to obtain multi-dimensional output confidence. Based on the target scene context features, the multidimensional output confidence is jointly quantized and mapped to obtain dynamic fusion weights; Based on the dynamic fusion weights, the multi-dimensional initial feature tensor is weighted and fused to obtain a comprehensive risk representation vector. Based on the comprehensive risk representation vector, an object risk assessment is performed to obtain the target object processing strategy.

[0006] Secondly, a device for assessing object risk is provided, comprising: The request feature extraction module is used to receive object processing requests and extract features from the object processing requests to obtain target scene context features; The data acquisition module is used to perform multimodal data acquisition based on the object processing request to obtain target multi-source heterogeneous data; The model matching module is used to perform model matching based on the target scene context features to obtain a set of target heterogeneous sub-models; The feature mapping module is used to perform modal feature mapping on the target multi-source heterogeneous data based on the target heterogeneous sub-model set to obtain a multi-dimensional initial feature tensor; The confidence assessment module is used to assess the model confidence of the target multi-source heterogeneous data based on the target heterogeneous sub-model set, and obtain multi-dimensional output confidence. The quantization mapping module is used to perform joint quantization mapping on the multidimensional output confidence based on the target scene context features to obtain dynamic fusion weights; The feature fusion module is used to perform weighted feature fusion on the multi-dimensional initial feature tensor based on the dynamic fusion weights to obtain a comprehensive risk representation vector; The risk assessment module is used to perform object risk assessment based on the comprehensive risk characterization vector to obtain the target object processing strategy.

[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described object risk assessment method.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described object risk assessment method.

[0009] The aforementioned object risk assessment method, device, computer equipment, and storage medium can first extract target scene context features and collect target multi-source heterogeneous data by receiving object processing requests, breaking the limitation of single data dimension in traditional assessment schemes. Subsequently, based on the target scene context features, a set of heterogeneous target sub-models is dynamically matched, effectively solving the defect of decoupling between traditional static fixed models and complex and variable scenarios at the underlying architecture level. Next, the set of heterogeneous target sub-models is used to perform submodal feature mapping and model confidence assessment respectively. After obtaining dynamic fusion weights by quantizing the joint target scene context features, weighted feature fusion is performed on the multi-dimensional initial feature tensors to obtain a comprehensive risk representation vector. This enables the object risk assessment system to adaptively strengthen high-reliability risk features and suppress interference noise from mismatched scenarios, greatly improving the three-dimensionality and accuracy of underlying risk characterization. Finally, a target object processing strategy is generated based on this comprehensive risk representation vector, opening up a closed loop from real-time risk perception to dynamic strategy issuance, improving the adaptive decision-making efficiency and accuracy of the object risk assessment system in multi-mode business scenarios. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1This is a schematic diagram of an application environment for an object risk assessment method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an object risk assessment method according to an embodiment of the present invention; Figure 3 yes Figure 2 A schematic diagram of a specific implementation method for step S20; Figure 4 yes Figure 2 A schematic diagram of a specific implementation method for step S30; Figure 5 yes Figure 2 A schematic diagram of a specific implementation of step S40; Figure 6 This is a schematic diagram of the structure of an object risk assessment device according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 8 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] The object risk assessment method provided in this invention can be applied to, for example... Figure 1In this application environment, the client communicates with the server via a network. The server can receive object processing requests from the client, extract features from the requests to obtain target scene context features; based on the object processing requests, it performs multimodal data acquisition to obtain target multi-source heterogeneous data; based on the target scene context features, it performs model matching to obtain a set of target heterogeneous sub-models; based on the target heterogeneous sub-model set, it performs submodal feature mapping on the target multi-source heterogeneous data to obtain a multi-dimensional initial feature tensor; based on the target heterogeneous sub-model set, it performs model confidence evaluation on the target multi-source heterogeneous data to obtain a multi-dimensional output confidence score; based on the target scene context features, it performs joint quantization mapping on the multi-dimensional output confidence score to obtain dynamic fusion weights; based on the dynamic fusion weights, it performs weighted feature fusion on the multi-dimensional initial feature tensor. A comprehensive risk representation vector is obtained. Finally, based on the comprehensive risk representation vector, object risk assessment is performed to obtain the target object processing strategy. In this invention, for insurance intelligent pricing business entities, multimodal data collection and dynamic model matching are used to generate dynamic fusion weights by combining scenario context and model confidence. This transforms massive heterogeneous data into a comprehensive risk representation vector. Finally, based on this comprehensive risk representation vector, target object processing strategies such as adaptive intelligent pricing are generated, making the final processing strategy highly correlated with multidimensional real-time risks. This solves the problems of staticity, scenario insensitivity, and insufficient data utilization in traditional insurance intelligent pricing systems, improves the scientificity and accuracy of object risk assessment and adaptive pricing, and improves the efficiency of insurance business flow and user experience. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a separate server or a server cluster composed of multiple servers. The invention will be described in detail below through specific embodiments.

[0014] Please see Figure 2 As shown, Figure 2 This is a flowchart illustrating an object risk assessment method provided in an embodiment of the present invention, which can be applied to an object risk assessment system and includes the following steps: S10: Receive object processing requests and extract features from the object processing requests to obtain target scene context features; S20: Based on object processing requests, perform multimodal data acquisition to obtain target multi-source heterogeneous data; S30: Based on the contextual features of the target scene, perform model matching to obtain a set of heterogeneous sub-models for the target; S40: Based on the target heterogeneous sub-model set, perform modal feature mapping on the target multi-source heterogeneous data to obtain a multi-dimensional initial feature tensor; S50: Based on the target heterogeneous sub-model set, evaluate the model confidence of the target multi-source heterogeneous data to obtain multi-dimensional output confidence; S60: Based on the contextual features of the target scene, the multi-dimensional output confidence is jointly quantized and mapped to obtain dynamic fusion weights; S70: Based on dynamic fusion weights, weighted feature fusion is performed on multi-dimensional initial feature tensors to obtain a comprehensive risk representation vector; S80: Based on the comprehensive risk representation vector, conduct object risk assessment to obtain the target object processing strategy.

[0015] Steps S10 to S80 of this embodiment first extract the target scene context features and collect multi-source heterogeneous data of the target by receiving the object processing request, breaking the limitation of the single data dimension of the traditional assessment scheme. Then, based on the target scene context features, the target heterogeneous sub-model set is dynamically matched, which effectively solves the defect of decoupling the traditional static fixed model from the complex and variable scenario at the underlying architecture level. Next, the target heterogeneous sub-model set is used to perform submodal feature mapping and model confidence evaluation respectively. After obtaining the dynamic fusion weight by quantizing the joint target scene context features, the multi-dimensional initial feature tensor is weighted feature fusion to obtain a comprehensive risk representation vector. This enables the object risk assessment system to adaptively strengthen the high reliability risk features and suppress the interference noise of non-matching scenarios, which greatly improves the three-dimensional dimension and accuracy of the underlying risk characterization. Finally, the target object processing strategy is generated based on the comprehensive risk representation vector, which opens up the entire closed loop from real-time risk perception to dynamic strategy issuance, and improves the adaptive decision-making efficiency and accuracy of the object risk assessment system in multi-mode business scenarios.

[0016] The object risk assessment method provided by this invention can be applied to the back-end servers of various adaptive intelligent pricing systems in the insurance technology field. These back-end servers can receive multimodal data from insured customers and external IoT sensor data in real time. For example, in the field of dynamic auto insurance pricing, insurance institutions often need to provide personalized auto insurance quotes for car owners under specific scenarios or time periods. This often requires the back-end server to integrate the insured's historical background, real-time behavioral data, and macroeconomic environmental data in real time, in order to accurately perceive the customer's actual risk shift in different scenarios and ensure the scientific nature of premium pricing and the comprehensiveness of risk coverage.

[0017] For example, a car owner plans a three-day long-distance road trip to a mountainous area in another province during the peak rainy season and submits a request to temporarily apply for long-distance road trip insurance through an insurance app. Upon receiving this request, the backend server first extracts features to determine the target scenario context features: long-distance travel across provinces, severe weather, and mountainous road conditions. Next, the backend server immediately and concurrently collects multi-source heterogeneous data, including the car owner's historical claims records, real-time operation flow data from the vehicle's OBD device, and real-time weather and traffic data for the target area. Subsequently, based on the aforementioned long-distance rainstorm scenario context features, the backend server accurately matches and activates a set of heterogeneous target sub-models, including a time-series behavior sub-model and an external perception sub-model, which are adept at handling dynamic risks. These sub-models are then used to map the collected data into multi-dimensional initial feature tensors representing driving habits and environmental impacts. Furthermore, the backend server evaluates the model confidence of each sub-model's output, finding that the external environment perception model has extremely high confidence. Combined with the scenario features, a dynamic fusion weight is generated that assigns a higher proportion of meteorological and environmental risk. Finally, the backend server performs weighted feature fusion on the multi-dimensional initial feature tensor based on the dynamic fusion weight to obtain a comprehensive risk representation vector. Based on this, it evaluates the comprehensive risk rating of this high-risk mountain self-driving trip and then automatically calculates and outputs the target object processing strategy with an adaptive insurance premium of 185 yuan for this self-driving task. At the same time, it continuously monitors the vehicle's location and warning events during the policy's effective period so as to provide timely feedback and update the assessment strategy when extremely high risks are triggered.

[0018] In step S10 of some embodiments, the object processing request refers to computer communication data that initiates a risk assessment to obtain relevant decision instructions. For example, in the application scenario of dynamic pricing for Internet auto insurance, when a user submits basic vehicle information through a mobile terminal and triggers the premium calculation logic, the object processing request may be real-time quotation calculation communication data containing the vehicle identification code and the policyholder's identifier. In the application scenario of triage for acute and critical care in regional medical centers, if the on-board equipment of the ambulance synchronizes vital sign data to the triage station, the object processing request may be communication data containing the risk assessment of comprehensive complications of patients, including heart rate and blood oxygen concentration.

[0019] Target scenario context features refer to high-dimensional semantic variables extracted from object processing requests to characterize the external environment and business state of the object processing request. For example, in the application scenario of highway travel insurance underwriting during holidays, when the object risk assessment system parses that the current time is a statutory holiday and the location is located at an inter-provincial highway toll station, the target scenario context features can be high-dimensional spatiotemporal context variables characterizing long-distance fatigue driving and the risk state of weather in other places. In the application scenario of initial screening for influenza during the peak season in respiratory clinics, if the object risk assessment system detects a surge in regional influenza cases in the current month and patients show signs of acute high fever, the target scenario context features can be business context variables characterizing the state of highly infectious clustered disease.

[0020] In this embodiment, the object risk assessment system can first receive an object processing request initiated by an external terminal through a communication interface. Then, it activates the scene recognizer in the built-in meta-learning fusion engine to parse the object processing request data, extract the request source, trigger time, device type and product attribute basic fields in the object processing request, and then use the semantic mapping network to convert the basic fields into high-dimensional features that can characterize the specific environmental state of the object processing request, thereby outputting the target scene context features.

[0021] In step S20 of some embodiments, the target multi-source heterogeneous data refers to a preprocessed set of underlying data used for risk assessment that spans different data dimensions and structure types. For example, in the application scenario of insurance underwriting for new energy ride-hailing vehicles, when the target risk assessment system concurrently accesses the vehicle network platform and the traffic management bureau database, the target multi-source heterogeneous data may be a set of data including the driver's historical violation records and the real-time sequence of battery charging and discharging. In the application scenario of remote monitoring and early warning for cardiovascular chronic diseases, if the target risk assessment system is connected to the patient's home smart wearable device and the hospital's electronic medical record database at the same time, the target multi-source heterogeneous data may be a set of data including the continuous dynamic time-series flow of electrocardiograms and the text reports of previous echocardiograms.

[0022] In this embodiment, after receiving an object processing request, the object risk assessment system synchronously schedules a multi-source heterogeneous data acquisition and preprocessing module to concurrently pull relevant data from various internal databases and external application interfaces. The pulled data includes customer static identity data, wearable device real-time behavior stream data, unstructured appearance photos, and traffic and meteorological external environment data. Subsequently, the object risk assessment system performs missing value imputation and time-domain alignment preprocessing operations on these data streams spanning different modalities, thereby converting the pulled data into target multi-source heterogeneous data that can be uniformly called by the internal algorithm of the object risk assessment system.

[0023] In detail, such as Figure 3As shown, step S20, which involves multimodal data acquisition based on the object processing request to obtain target multi-source heterogeneous data, includes the following steps: S21: Based on object processing requests, perform concurrent capture of multi-source data to obtain heterogeneous raw data streams; S22: Perform time-domain interpolation on the heterogeneous raw data stream to obtain a time-aligned data stream; S23: Based on the preset feature space transformation matrix, perform dimensional alignment mapping on the time-series aligned data stream to obtain the target multi-source heterogeneous data.

[0024] In step S21 of some embodiments, the heterogeneous raw data stream refers to the initial continuous data that the object risk assessment system concurrently captures based on object processing requests, spanning different data dimensions and structure types and without alignment processing. For example, in the application scenario of insurance underwriting for new energy ride-hailing vehicles, when the object risk assessment system concurrently accesses the vehicle network platform and the traffic management bureau database, the heterogeneous raw data stream may be the initial continuous data containing driver heart rate and vital signs sequences with different sampling frequencies and vehicle on-board automatic diagnostic system status sequences. In the application scenario of remote monitoring and early warning for cardiovascular chronic diseases, if the object risk assessment system is simultaneously connected to the patient's home smart wearable device and hospital monitor, the heterogeneous raw data stream may be the initial continuous data containing dynamic electrocardiogram sequences and non-invasive blood pressure monitoring sequences with large differences in acquisition intervals.

[0025] In this embodiment, after receiving an object processing request, the object risk assessment system triggers its internal multi-source heterogeneous data acquisition and preprocessing module. It synchronously pulls data from different internal business databases and external environment application interfaces through asynchronous concurrent requests. For example, it simultaneously acquires customer static identity data, real-time behavioral stream data uploaded by wearable devices, and traffic and meteorological external environment data. Since the source channels of these data are completely different from the underlying file structure, the object risk assessment system aggregates them in parallel into the memory buffer, thereby obtaining a heterogeneous raw data stream that has not been uniformly formatted.

[0026] In step S22 of some embodiments, the time-aligned data stream refers to continuous data that remains consistent at the timestamp node after the heterogeneous original data stream has undergone time-dimension interpolation processing. For example, in the above-mentioned application scenario of insurance underwriting for new energy ride-hailing vehicles, when the object risk assessment system uses an interpolation algorithm to perform time-domain completion on the low-frequency vital sign sequence, the time-aligned data stream can be continuous data of the driver's heart rate vital sign sequence and the vehicle's on-board automatic diagnostic system status sequence with the same sampling rate. In the above-mentioned application scenario of remote monitoring and early warning for cardiovascular chronic diseases, if the object risk assessment system performs polynomial interpolation on the interrupted blood pressure sequence, the time-aligned data stream can be continuous data of the ECG sequence and blood pressure sequence with one-to-one timestamp.

[0027] In this embodiment of the application, the acquisition frequency of each sub-data stream in the heterogeneous original data stream has objective differences or there is data packet loss due to network transmission. Therefore, the object risk assessment system can use the built-in time-series processing unit to perform time-domain interpolation processing on the heterogeneous original data stream. Specifically, the object risk assessment system can use linear interpolation or spline interpolation algorithms to complete the missing timestamp node data, so that the originally asynchronous and high- and low-frequency mixed continuous sequence data are completely aligned under the same time base, thereby outputting a time-aligned data stream without time misalignment error.

[0028] In step S23 of some embodiments, the preset feature space transformation matrix refers to a mathematical matrix vector pre-set in the object risk assessment system to uniformly map data with different dimensional attributes to a specific high-dimensional space. For example, in the application scenario of vehicle property insurance survey during rainstorm weather, when it is necessary to align external environmental data and internal business data, the preset feature space transformation matrix can be a mathematical matrix vector used to uniformly map radar precipitation latitude and longitude coordinates and vehicle damage location coordinates to a standard spatial plane. In the application scenario of interdisciplinary joint consultation assistance for complex rare diseases, if the object risk assessment system needs to integrate biochemical indicators and gene sequences, the preset feature space transformation matrix can be a mathematical matrix vector used to uniformly map discrete biochemical test values ​​and high-dimensional gene codes to a standardized medical feature space.

[0029] In this embodiment, after completing the alignment in the time dimension, the object risk assessment system calls a pre-configured feature space transformation matrix in memory to perform a linear or nonlinear projection transformation of the spatial dimension on the time-aligned data stream. This maps the time-aligned data streams, which originally had different physical dimensions and huge differences in data dimensions, to a normalized feature dimension space that can be seamlessly recognized by the underlying algorithm of the object risk assessment system. After completing this dimension alignment mapping operation, the object risk assessment system finally outputs target multi-source heterogeneous data with a unified structure and completely aligned semantics, so that it can be directly distributed to specific heterogeneous computing networks for feature assessment.

[0030] Steps S21 to S23 of this embodiment first involve the object risk assessment system concurrently capturing multi-source data based on object processing requests to obtain heterogeneous raw data streams. This significantly expands the dimensional richness of the underlying input data for risk assessment and the efficiency of real-time data acquisition across all dimensions. Subsequently, the object risk assessment system performs temporal interpolation on the heterogeneous raw data streams to obtain time-aligned data streams. This effectively eliminates the time misalignment errors caused by inconsistent sampling frequencies of multi-source sensors and packet loss during transmission, ensuring strong consistency of multimodal data at the temporal baseline level. Finally, the object risk assessment system performs dimensional alignment mapping on the time-aligned data streams based on a preset feature space transformation matrix to obtain target multi-source heterogeneous data. This further eliminates the differences in physical dimensions and spatial dimensional barriers between cross-modal underlying data. This progressive preprocessing process, from concurrent capture to temporal alignment to spatial dimensional mapping, not only completely breaks the silo limitations of traditional single-source static data but also provides a high-quality, standardized, and spatiotemporally aligned data foundation for subsequent heterogeneous model feature extraction and dynamic fusion. This improves the robustness and accuracy of the object risk assessment system in perceiving all-dimensional risks in complex scenarios.

[0031] In step S30 of some embodiments, the target heterogeneous sub-model set refers to a model combination consisting of at least two evaluation networks with different algorithm architectures, dynamically extracted from a preset model library according to the specific scenario context. For example, in the application scenario of vehicle property insurance survey under heavy rain, when the target scenario context features indicate that the external environment is under extremely severe weather conditions, the target heterogeneous sub-model set can be a model combination consisting of an external perception network that processes radar precipitation data and a convolutional neural network that processes vehicle damage images. In the application scenario of interdisciplinary joint consultation assistance for complex rare diseases, if the target scenario context features indicate that the patient's symptoms involve multiple organ failure, the target heterogeneous sub-model set can be a model combination consisting of a deep temporal network that processes gene sequencing sequences and a graph neural network that processes pathological slides.

[0032] In this embodiment, the object risk assessment system is pre-configured with a polymorphic risk assessment model library. The model scheduler in the meta-learning fusion engine receives the aforementioned target scene context features and, based on the specific scene semantics indicated by the target scene context features, dynamically retrieves at least two heterogeneous computing networks from the polymorphic risk assessment model library that best match the target scene context features through a gating routing strategy. This forms a set of target heterogeneous sub-models dedicated to this risk calculation, thereby achieving adaptive dynamic decoupling between the underlying computing network and the target scene context features.

[0033] In detail, such as Figure 4 As shown, step S30, which involves model matching based on the target scene context features to obtain a set of target heterogeneous sub-models, includes the following steps: S31: Based on the contextual features of the target scene, perform multidimensional semantic clustering analysis to obtain multimodal scene semantic clusters; S32: Based on the multimodal scene semantic clusters and the preset model performance benchmark topology, perform gated route calculation to obtain the activation probability of candidate sub-model routes; S33: Based on the candidate sub-model routing activation probability and the preset activation threshold, perform feature mask filtering to obtain the target activation mask; S34: Based on the target activation mask, extract models from the preset polymorphic risk assessment model library to obtain a set of target heterogeneous sub-models.

[0034] In step S31 of some embodiments, the multimodal scene semantic cluster refers to the structured semantic data obtained after clustering analysis of the contextual features of the target scene, which is used to finely distinguish the types of multidimensional business risks. For example, in the application scenario of dynamic pricing of Internet auto insurance, when the object risk assessment system performs unsupervised aggregation of a large number of environmental features, the multimodal scene semantic cluster can be aggregated semantic data that represents the frequent occurrence of extreme and severe weather accompanied by traffic congestion. In the application scenario of triage of critical and severe patients in regional medical centers, if the object risk assessment system performs joint clustering of the patient's basic medical history and the season of onset, the multimodal scene semantic cluster can be aggregated semantic data that represents the symptoms of sudden onset of cardiovascular underlying diseases in winter.

[0035] In this embodiment, after receiving the target scene context features extracted in the preceding process, the meta-learning fusion engine of the object risk assessment system uses a built-in unsupervised clustering algorithm to perform multi-dimensional semantic clustering analysis on the current high-dimensional target scene context features. This aggregates scene variables with similar environmental factors, time patterns, or business attributes onto specific high-dimensional semantic space nodes, thereby transforming discrete context variables into structured data that can finely distinguish business risk types. In other words, it outputs multimodal scene semantic clusters. For example, the object risk assessment system can use this to identify that the current request belongs to a high-risk composite scene cluster of extreme weather and congested road conditions.

[0036] In step S32 of some embodiments, the preset model performance benchmark topology map refers to the network topology data pre-constructed within the object risk assessment system for mapping the relationship between nodes that calculate performance in different historical scenarios of various underlying heterogeneous networks. For example, in the application scenario of monitoring the risk of heavy truck fleet driving at night in freight logistics, when it is necessary to match the optimal underlying algorithm, the preset model performance benchmark topology map may be the node mapping data that records the highest benchmark score of the time-series behavior network in the night fatigue monitoring scenario. In the application scenario of predicting postoperative infection complications in the intensive care unit, if it is necessary to schedule the disease prediction network, the preset model performance benchmark topology map may be the node mapping data that maps the graph neural network to the node with the highest evaluation performance on complex medical record text.

[0037] The candidate submodel routing activation probability refers to the probability value calculated by the gating network to quantify the degree of adaptation of the candidate evaluation network to the semantics of the current specific scenario. For example, in the application scenario of insurance underwriting for new energy ride-hailing vehicles, when the scenario semantics point to high-frequency abnormal battery discharge, the candidate submodel routing activation probability can be a probability value with extremely high fit for the output of the battery time series analysis model. In the application scenario of cross-disciplinary joint consultation assistance for complex rare diseases, if the scenario involves multi-gene mutation association analysis, the candidate submodel routing activation probability can be a probability value with high fit for the output of the gene sequence map network.

[0038] In this embodiment, after the object risk assessment system obtains the multimodal scene semantic clusters, it performs association calculations with the preset model performance benchmark topology graph pre-configured in memory. Since the preset model performance benchmark topology graph records the performance distribution of all underlying heterogeneous networks in different historical scenarios, the model scheduler of the object risk assessment system performs matrix multiplication and gated routing calculations on the multimodal scene semantic clusters and the graph structure through a gated network. This quantifies the expected fit of each candidate evaluation network in processing the current specific clustering scenario, and then outputs the candidate sub-model routing activation probability of each candidate network.

[0039] In step S33 of some embodiments, the preset activation threshold refers to the probability baseline value set in the object risk assessment system to determine whether the candidate assessment network is qualified to start. For example, in the application scenario of vehicle property insurance investigation during rainstorm weather, when the object risk assessment system schedules network computing power, the preset activation threshold can be set to 80% to filter out inefficient network nodes. In the application scenario of intelligent assisted diagnosis of imaging in grassroots township health centers, if the underlying computing power resources of the object risk assessment system are extremely limited, the preset activation threshold can be set to an extremely high standard to ensure that only the most core diagnostic network is activated.

[0040] A target activation mask is a binary state vector data generated by filtering probabilities based on a preset activation threshold, used to accurately extract target networks from a model library. For example, in the agricultural crop index insurance pricing application scenario during a sudden cold wave, when the probability of the meteorological sensing network exceeds the threshold while the historical yield network does not, the target activation mask can be binary state vector data that only sets the corresponding position of the meteorological sensing network to an effective state. In the initial assessment application scenario of a sudden public health event, if only the dynamic prediction model meets the activation requirements, the target activation mask can be binary state vector data that only allows the extraction of that specific network.

[0041] In this embodiment of the application, in order to ensure the accurate allocation of computing resources and eliminate redundant interference models, the object risk assessment system compares the calculated activation probabilities of candidate sub-model routes of each candidate network with the preset activation thresholds built into the system for verification, and performs feature mask filtering accordingly. When the activation probability of a candidate network exceeds the preset activation threshold, the object risk assessment system sets its corresponding position in the mask vector to the activation allowed state; otherwise, it sets it to the blocking filtering state. Through this hard filtering and binary state conversion, the object risk assessment system finally generates a target activation mask for accurately addressing the target network.

[0042] In step S34 of some embodiments, the preset polymorphic risk assessment model library refers to a resource repository that is persistently stored in the object risk assessment system, which contains multiple heterogeneous underlying networks for risk assessment in different dimensions. For example, in the application scenario of dynamic prediction of ride-hailing accident risk during the morning rush hour in cities, when the object risk assessment system starts risk calculation, the preset polymorphic risk assessment model library can be an algorithm resource repository that stores basic risk models and behavioral dynamic models. In the application scenario of comprehensive assessment of drug resistance to targeted drugs for malignant tumors, if the object risk assessment system needs to comprehensively assess the clinical risk of patients, the preset polymorphic risk assessment model library can be an algorithm resource repository that stores clinical rule engines and pathological image convolutional networks.

[0043] In this embodiment, the object risk assessment system uses the aforementioned target activation mask as an addressing and distribution instruction to directly initiate a call request to a preset polymorphic risk assessment model library that stores various basic risk models and related graph models. Based on the distribution of allowed states in the target activation mask, the system automatically and accurately extracts the corresponding underlying network architecture that is in the allowed activation state from the preset polymorphic risk assessment model library. By combining these extracted heterogeneous networks into instances, the object risk assessment system finally completes the construction of a dedicated computing network array adapted to the current scenario, thus obtaining a set of target heterogeneous sub-models.

[0044] Steps S31 to S34 as illustrated in this embodiment first involve using an object risk assessment system to perform multi-dimensional semantic clustering analysis based on the contextual features of the target scene to obtain multimodal scene semantic clusters. This transforms discrete and complex scene variables into highly structured clustering semantic bases. Then, the object risk assessment system performs gated routing calculations based on the multimodal scene semantic clusters and a preset model performance benchmark topology to obtain the activation probability of candidate sub-model routes. By associating historical performance, the system objectively quantifies the underlying adaptation degree of all candidate networks to the current scene. Finally, the object risk assessment system uses a preset activation threshold to specifically evaluate the activation probability of candidate sub-model routes. The target activation mask is generated by filtering the target activation mask, thereby achieving hard control over computing resources and automatically eliminating inefficient and redundant evaluation networks. Finally, the object risk assessment system accurately extracts models from the preset polymorphic risk assessment model library based on the target activation mask to obtain a set of heterogeneous sub-models for the target. This progressive scheduling mechanism, from semantic clustering to gated routing calculation to accurate mask extraction, not only ensures that the retrieved evaluation network is highly compatible with the current specific business scenario, but also effectively blocks the consumption of computing power caused by invalid calculations at the system's lower level, greatly improving the accuracy of the heterogeneous network matching process and the overall operating efficiency of the object risk assessment system.

[0045] In step S40 of some embodiments, the multidimensional initial feature tensor refers to the matrix data that represents each single-dimensional risk attribute after various types of underlying data are mapped through the deep space of the target heterogeneous sub-model set. For example, in the application scenario of monitoring the risk of heavy truck fleet driving at night in freight logistics, when each evaluation network independently completes the deep analysis of the vehicle fatigue monitoring video stream and driving trajectory, the multidimensional initial feature tensor can be matrix data that respectively represents the abnormality of the driver's eye-closing frequency and the risk of route deviation. In the application scenario of predicting postoperative infection complications in the intensive care unit, if each evaluation network has completed the deep calculation of the variability of white blood cell indicators and sputum culture report, the multidimensional initial feature tensor can be matrix data that respectively represents the cytokine storm tendency and the probability of drug-resistant strain hit.

[0046] In this embodiment, the object risk assessment system routes and distributes the preprocessed target multi-source heterogeneous data to the corresponding dedicated analysis networks in the target heterogeneous sub-model set. Each dedicated analysis network in the target heterogeneous sub-model set leverages its specific underlying algorithm architecture to perform forward depth calculations. For example, the time-series behavior dedicated analysis network independently analyzes the target multi-source heterogeneous data to extract behavior offset features. Subsequently, the object risk assessment system aligns and concatenates the feature calculation results output by each dedicated analysis network in a unified vector space, and finally summarizes them to obtain a multi-dimensional initial feature tensor.

[0047] It should be noted that, in the embodiments of this application, the target heterogeneous sub-model set may include graph relation sub-models and time sequence behavior sub-models. Graph relation sub-models refer to deep learning computing models deployed within the object risk assessment system and built on a graph neural network architecture for analyzing the topological relationship network between entities. For example, in the application scenario of identifying complex car insurance gang fraud, when the object risk assessment system accesses the business relationship data between the policyholder and the repair shop, the graph relation sub-model can be a graph neural network computing model specifically used to mine the hidden fraud network between the policyholder and the target. In the application scenario of assisting in the epidemiological investigation of regional infectious diseases, if the object risk assessment system needs to analyze the contact trajectory between patients, the graph relation sub-model can be a graph neural network computing model specifically used to identify super-spreaders and cluster infection networks. The temporal behavior sub-model refers to the underlying algorithm network deployed within the object risk assessment system and built on a temporal network architecture for analyzing continuous time series state changes. For example, in the application scenario of monitoring the risk of heavy truck fleet driving at night in freight logistics, when the object risk assessment system receives continuous trajectory data from the vehicle-mounted equipment in real time, the temporal behavior sub-model can be a dynamic temporal network specifically used to extract the frequency of driver emergency braking and route deviation characteristics. In the application scenario of remote monitoring and early warning of cardiovascular chronic diseases, if the object risk assessment system continuously acquires data from the patient's smart wearable device at home, the temporal behavior sub-model can be a deep temporal network specifically used to analyze the deviation characteristics of continuous electrocardiogram waveforms.

[0048] Based on this, such as Figure 5 As shown, in step S40, that is, based on the target heterogeneous sub-model set, the target multi-source heterogeneous data is subjected to modal feature mapping to obtain a multi-dimensional initial feature tensor, including the following steps: S41: Based on the graph relation sub-model, perform random walks of heterogeneous nodes on the target multi-source heterogeneous data to obtain the node association interaction path sequence; S42: Based on the graph relation sub-model, graph features are aggregated on the node association interaction path sequence to obtain the association risk feature tensor; S43: Based on the temporal behavior sub-model, dynamic temporal feature extraction is performed on the target multi-source heterogeneous data to obtain the behavior offset feature tensor; S44: Perform vector space mapping on the associated risk feature tensor and the behavior offset feature tensor to obtain unified dimension features; S45: Integrate the features of the same dimension into a multimodal tensor to obtain a multidimensional initial feature tensor.

[0049] In step S41 of some embodiments, the node association interaction path sequence refers to the ordered list data that records the mutual associations and interaction trajectories between different entity nodes after performing a walk sampling operation on the graph structure network. For example, in the application scenario of identifying complex car insurance gang fraud, after the object risk assessment system completes graph sampling, the node association interaction path sequence may be an ordered sequence of data containing the policyholder's sequential association with a specific repair shop and a specific claims adjuster. In the application scenario of assisting in the epidemiological investigation of regional infectious diseases, if the object risk assessment system completes contact trajectory sampling, the node association interaction path sequence may be an ordered sequence of data containing the confirmed patient's sequential passage through a specific means of transportation and a specific shopping mall.

[0050] In this embodiment, after the object risk assessment system distributes the target multi-source heterogeneous data to the graph relation sub-model, the graph relation sub-model will start the underlying sampling algorithm from a specific target node based on the probability distribution of node and edge attributes in the graph structure. It will then perform a random walk exploration operation with a specific bias or restart mechanism in a complex network containing multiple types of entities such as humans, machines, and objects. By recording each entity node passed through during the walk and its associated edge type, the graph relation sub-model will finally generate a node association interaction path sequence that can represent the high-frequency local connection state and topological structure between entities. In this way, the object risk assessment system can effectively trace the hidden association interaction trajectory between the target entity and different surrounding entity nodes.

[0051] In step S42 of some embodiments, the associated risk feature tensor refers to the high-dimensional matrix data output after the graph relation sub-model aggregation calculation, which is used to quantify the degree of potential association between multiple entity nodes. For example, in the application scenario of identifying complex car insurance gang fraud, after the graph feature aggregation calculation is completed, the associated risk feature tensor can be matrix data that characterizes the degree of high risk of the insured participating in a specific gang fraud network. In the application scenario of assisting in the epidemiological investigation of regional infectious diseases, if the contact trajectory aggregation calculation is completed, the associated risk feature tensor can be matrix data that characterizes the degree of high risk of an individual being at the center of a cluster infection network.

[0052] In this embodiment, after successfully obtaining the node association interaction path sequence, the graph relation sub-model utilizes the message passing mechanism at the bottom layer of the graph neural network to transmit and summarize the attribute features and connection edge weights of each neighboring node to the core target node layer by layer along the topological connection direction indicated by the node association interaction path sequence. Subsequently, nonlinear feature aggregation calculation is performed through a multilayer perceptron and activation function. The graph relation sub-model finally maps the discrete topological network information into a high-dimensional dense matrix form, thereby outputting an association risk feature tensor used to accurately quantify the risk of entity aggregation or the high-risk degree of fraud gang networks.

[0053] In step S43 of some embodiments, the behavior offset feature tensor refers to the high-dimensional matrix data output by the temporal behavior sub-model after dynamic temporal feature extraction, which is used to quantify the degree to which the dynamic behavior of the object deviates from the normal safety benchmark. For example, in the application scenario of risk monitoring of heavy truck fleet driving at night in freight logistics, after the dynamic temporal feature extraction is completed, the behavior offset feature tensor can be matrix data that characterizes the degree to which the driver's fatigue driving behavior deviates from the safe driving benchmark. In the application scenario of remote monitoring and early warning of cardiovascular chronic diseases, if the electrocardiogram sequence analysis is completed, the behavior offset feature tensor can be matrix data that characterizes the degree to which the patient's heart rate waveform deviates from the normal physiological benchmark.

[0054] In this embodiment, at the same time as the aforementioned graph analysis, the object risk assessment system synchronously routes and distributes the continuous dynamic data stream from the target multi-source heterogeneous data to the temporal behavior sub-model. The temporal behavior sub-model uses a long short-term memory network or a one-dimensional temporal convolution operator to perform deep pattern recognition calculations on the data within the continuous time window, keenly capturing the abrupt changes and abnormal trend shifts that occur in the data stream over time. This allows the normal temporal fluctuation noise to be removed and the abnormal business behavior patterns to be focused on. After forward propagation and spatial mapping of the deep network, the temporal behavior sub-model finally outputs a behavior offset feature tensor that can quantitatively express the degree to which dynamic business behavior deviates from the safety benchmark.

[0055] In step S44 of some embodiments, the unified dimensional feature refers to the standardized feature data with completely consistent dimensional specifications output after vector space mapping of various feature data with different initial physical dimensions and tensor sizes. For example, in the application scenario of insurance underwriting for new energy ride-hailing vehicles, after the vector space mapping is completed, the unified dimensional feature can be driving behavior features and network graph feature data with consistent dimensional specifications. In the application scenario of remote monitoring and early warning of cardiovascular chronic diseases, if the spatial mapping is completed, the unified dimensional feature can be dynamic electrocardiogram features and static medical record feature data with consistent dimensional specifications.

[0056] In this embodiment, since the correlation risk feature tensor output by the graph relation sub-model and the behavior offset feature tensor output by the temporal behavior sub-model have significant differences in physical dimensions, semantic scale, and feature dimension, the feature alignment module of the object risk assessment system will use a pre-learned and optimized spatial transformation matrix to perform linear or non-linear projection transformation operations on the tensor matrices output by these two heterogeneous networks respectively. This will force the feature vectors that were originally in different high-dimensional subspaces to be mapped and projected into a preset standard shared latent feature space, thereby completely eliminating the dimensional barriers caused by different underlying assessment architectures and outputting unified dimensional features with standardized format and consistent semantic scale.

[0057] In step S45 of some embodiments, after obtaining the aligned unified dimensional features, the object risk assessment system immediately executes the high-dimensional tensor calculation logic of the multimodal fusion layer. By performing matrix concatenation or element-wise weighted superposition operations on the standardized features of the graph dimension and the standardized features of the temporal dimension in the channel dimension, a joint feature expression form containing rich underlying risk perspective information and with a completely coherent structure is constructed. After completing the tensor integration processing, the object risk assessment system can directly output a multi-dimensional initial feature tensor that simultaneously possesses the depth of the underlying network topology and the breadth of temporal inference, so that it can seamlessly enter the confidence assessment and dynamic weighted fusion stage later.

[0058] Steps S41 to S45 as illustrated in this embodiment first use a graph relation sub-model to perform random walks on heterogeneous nodes in the target multi-source heterogeneous data to obtain a sequence of node association interaction paths, accurately uncovering hidden topological interaction trajectories between heterogeneous entities in complex business networks. Then, the object risk assessment system aggregates graph features from the node association interaction path sequence based on the graph relation sub-model to obtain an association risk feature tensor, effectively transforming discrete connection trajectories into a high-dimensional group network risk characterization matrix that can be rapidly computed by a computer. Next, the object risk assessment system simultaneously performs dynamic temporal feature extraction on the target multi-source heterogeneous data based on a temporal behavior sub-model to obtain a behavior offset feature tensor, thereby keenly capturing individual abnormal behavior evolving over time at a micro-dynamic level. To address this risk, based on this multidimensional perception, the object risk assessment system performs vector space mapping on the associated risk feature tensor and the behavioral offset feature tensor to obtain unified dimensional features. This completely breaks down the dimensional barriers and semantic expression gaps caused by the outputs of different underlying architecture algorithms. Finally, the object risk assessment system integrates the unified dimensional features into multimodal tensors to obtain multidimensional initial feature tensors. This progressive calculation and fusion mapping process not only achieves deep complementary advantages between macro-level clustered graph features and micro-level temporal dynamic features, but also lays a very solid and high-precision data foundation for the subsequent generation of a comprehensive and dimension-free final risk assessment strategy. This significantly enhances the object risk assessment system's ability to perceive complex and ever-changing object risks in a three-dimensional way and improves the accuracy of its underlying characterization.

[0059] In step S50 of some embodiments, the multidimensional output confidence refers to the probability metric value that each evaluation network adaptively calculates based on the input data quality and distribution characteristics, representing the reliability of its own feature extraction results. For example, in the anti-fraud application scenario of used car insurance in remote mountainous areas, when the incoming vehicle exterior inspection photo has severe backlight blur and the positioning signal is discontinuous, the multidimensional output confidence can be the extremely low reliability value output by the image recognition sub-network and the relatively high reliability value output by the static identity rule network. In the application scenario of intelligent assisted diagnosis of images in grassroots township health centers, if the medical images uploaded by the grassroots have obvious artifacts that make it difficult for the evaluation network to fit historical clear samples, the multidimensional output confidence can be the extremely low accuracy metric value adaptively output by the image evaluation sub-network.

[0060] In this embodiment of the application, during the above feature mapping calculation process, each dedicated analysis network in the target heterogeneous sub-model set not only outputs a feature matrix, but also calculates an objective quantitative score at the end of the network based on the completeness of the allocated target multi-source heterogeneous data and the fitting similarity between the target multi-source heterogeneous data and the distribution of historical training samples. This objective quantitative score is used to characterize the degree of accuracy of the dedicated analysis network in its own output results. The object risk assessment system collects all the objective quantitative scores fed back by the target heterogeneous sub-model set and performs normalization processing to obtain the multidimensional output confidence level that characterizes the overall result reliability distribution.

[0061] In step S60 of some embodiments, dynamic fusion weight refers to coefficient distribution data that is jointly calculated by combining the target scenario context features and multidimensional output confidence to allocate importance proportions during feature fusion. For example, in the agricultural crop index insurance pricing application scenario under the sudden cold wave, when the object risk assessment system identifies an extreme temperature drop and the reliability of the meteorological data source is much greater than that of conventional historical yield data, the dynamic fusion weight can be coefficient distribution data that assigns extremely high importance to the meteorological perception network and significantly reduces the importance of the historical static network. In the initial assessment application scenario of a sudden public health event, if the reliability of the traditional empirical rule network calculation is extremely low due to unknown pathogens while the reliability of the infectious disease dynamics prediction network calculation increases, the dynamic fusion weight can be coefficient distribution data that significantly increases the allocation proportion of the infectious disease dynamics prediction network.

[0062] In this embodiment, the adaptive fusion unit of the object risk assessment system receives the target scene context features from external input and the multidimensional output confidence of the network self-feedback. Then, it adopts a meta-learning strategy to perform attention cross-operation and non-linear quantization allocation on the external scene requirement data and the internal network reliability score, thereby calculating the importance distribution ratio of each dedicated analysis network output feature under the corresponding target scene context features. This generates dynamic fusion weights that are completely tailored to the specific object processing request, thus completely overcoming the data distortion defects caused by fixed weight allocation rules.

[0063] In detail, based on the contextual features of the target scene, the multi-dimensional output confidence is jointly quantized and mapped to obtain dynamic fusion weights, including: Based on the contextual features of the target scene and the pre-acquired historical performance data of the model, context tensor mapping is performed to obtain the scene-aware feature matrix. Based on the scene-aware feature matrix, attention feature interaction is performed on the multi-dimensional output confidence to obtain the confidence residual feature; The confidence residual features are assigned nonlinear weights to obtain dynamic fusion weights.

[0064] In some embodiments, historical model performance data refers to statistical data stored in advance in the object risk assessment system to record the prediction accuracy and loss convergence of each heterogeneous sub-model in various real business scenarios. For example, in the application scenario of predicting accidents during the morning rush hour of urban ride-hailing, when the evaluation network completes historical backtesting, the historical model performance data can be statistical data recording the extremely low prediction deviation rate of the time-series behavior network under congested road conditions. In the application scenario of predicting postoperative infection complications in the intensive care unit, if the object risk assessment system completes monthly business review, the historical model performance data can be historical statistical data recording the high recall rate of the graph relationship network in the prediction of multidrug-resistant bacteria.

[0065] A scene-aware feature matrix refers to high-dimensional feature data that is output after the context features of the target scene have been mapped to higher dimensions and integrated with historical performance experience. This data can deeply perceive and quantify the current business environment status. For example, in the application scenario of predicting accidents during the morning rush hour in urban ride-hailing services, after the tensor mapping is completed, the scene-aware feature matrix can be high-dimensional feature data that quantifies the high-pressure environment during the morning rush hour in rainy weather and its corresponding model fit. In the application scenario of predicting postoperative infection complications in intensive care units, if the tensor mapping is completed, the scene-aware feature matrix can be high-dimensional feature data that quantifies the ward environment of seasonal influenza outbreaks and its corresponding model fit.

[0066] The confidence residual feature refers to the high-dimensional feature data extracted after interactive computation through the attention mechanism, which is used to quantify the deviation and fluctuation details between the current output confidence of the underlying network and its historical benchmark performance. For example, in the application scenario of predicting accidents during the morning rush hour for urban ride-hailing services, after the interactive computation is completed, the confidence residual feature can be the difference deviation data that expresses the difference between the current image model's confidence and the historical benchmark due to the obstruction caused by heavy rain. In the application scenario of predicting postoperative infection complications in intensive care units, if the interactive computation is completed, the confidence residual feature can be the difference deviation data that expresses the difference between the current rule model's confidence and the historical benchmark due to the sudden drop in confidence caused by a new variant strain.

[0067] In this embodiment, the meta-learning fusion engine of the object risk assessment system first receives the previously extracted target scene context features and simultaneously retrieves pre-acquired model historical performance data from the built-in meta-database. Then, the object risk assessment system concatenates the target scene context features with the model historical performance data and inputs it into a multilayer perceptron to perform a context tensor mapping operation. Through multilayer nonlinear transformation, the discrete environmental variables are upgraded to a continuous high-dimensional latent space, resulting in a scene perception feature matrix that can deeply characterize the correspondence between the current external environment and historical experience. Subsequently, the object risk assessment system uses a self-attention mechanism module to use this scene perception feature matrix as a query vector and uses the multidimensional output confidence scores of each underlying network as key-value vectors for attention feature interaction calculation, thereby deeply mining the current features. To ensure the rationality and credibility of the confidence scores of each network under a given scenario environment perception, the object risk assessment system further calculates the difference between the multidimensional output confidence scores and the interaction results in the current specific scenario and the preset historical benchmark expectation. It extracts the confidence score residual features that can accurately quantify the degree of deviation in network performance. Finally, the object risk assessment system sends the confidence score residual features containing rich scenario interaction details and performance deviation fluctuations into the normalized activation function calculation layer. Through the activation function, a strict nonlinear weight allocation operation is performed to automatically amplify the weight allocation ratio of heterogeneous networks that are highly consistent with the current high-risk scenario and have stable confidence scores. At the same time, a nonlinear decay mechanism is used to strictly suppress the network weights that have drastic confidence score fluctuations or deviate too much from the historical performance benchmark. Finally, it outputs dynamic fusion weights that are accurate and used to guide subsequent multimodal feature fusion.

[0068] This application first uses an object risk assessment system to perform context tensor mapping based on the target scene context features and pre-acquired historical model performance data to obtain a scene-aware feature matrix. This successfully elevates discrete scene environment parameters and past experience into a high-dimensional, computable, continuous, and dense feature matrix. Then, based on the scene-aware feature matrix, attention feature interaction is performed on the multi-dimensional output confidence to obtain confidence residual features. By introducing an attention mechanism, the deep coupling relationship between external environmental factors and internal model computation reliability is deeply explored, and residual details of dynamic fluctuations in model performance are successfully extracted. Finally, nonlinear weight allocation is performed on the confidence residual features to obtain dynamic fusion weights. This progressive calculation process completely abandons the rigid and scene-insensitive static weight combination mechanism in traditional risk assessment, enabling the object risk assessment system to perceive sudden changes in the external business environment and minor deviations in internal model performance in real time. It adaptively and extremely accurately allocates the fusion weight distribution that best fits the current business context to each underlying heterogeneous computing network, significantly improving the robustness of multimodal feature fusion and the interpretability of dynamic decision-making in the face of complex, volatile, and extreme abnormal risk environments.

[0069] In step S70 of some embodiments, the comprehensive risk representation vector refers to the global low-level risk expression data that is highly condensed and integrated with multi-source modal information after weighted modulation and dimensionality reduction. For example, in the application scenario of dynamic prediction of ride-hailing accident risk during the morning rush hour in cities, when the object risk assessment system completes the weighted matrix multiplication of road network congestion features and driver emergency braking features and removes redundant information, the comprehensive risk representation vector can be the global low-level risk expression data that deeply couples the high-pressure environment during the morning rush hour and the driver's aggressive driving behavior. In the application scenario of comprehensive assessment of drug resistance to targeted drugs for malignant tumors, if the object risk assessment system performs a nonlinear smoothing compression operation on gene mutation features and previous drug use frequency features, the comprehensive risk representation vector can be the global low-level risk expression data that deeply couples the microscopic target mutation risk and the macroscopic clinical drug resistance risk.

[0070] In this embodiment, the object risk assessment system uses the calculated dynamic fusion weights to perform element-wise multiplication on each modal feature matrix corresponding to the multidimensional initial feature tensor, thereby significantly amplifying the matrix feature components with high confidence and high fit to the target scene context features. At the same time, it strictly suppresses or masks out feature noise interference caused by irrelevant modalities. Finally, the object risk assessment system performs nonlinear smoothing and high-dimensional space compression operations on the modulated modal feature matrix to remove redundant information and output a highly condensed comprehensive risk representation vector.

[0071] In detail, based on dynamic fusion weights, weighted feature fusion is performed on the multi-dimensional initial feature tensor to obtain a comprehensive risk representation vector, including: Based on dynamic fusion weights, a weighted feature projection is performed on the multi-dimensional initial feature tensor to obtain a multi-dimensional weighted feature tensor. A nonlinear cross-mapping is performed on the multidimensional weighted feature tensor to obtain a dimension-reduced fusion feature matrix; The reduced-dimensionality fusion feature matrix is ​​vectorized and smoothed to obtain the comprehensive risk representation vector.

[0072] In some embodiments, a multidimensional weighted feature tensor refers to a high-dimensional feature matrix that highlights key high-risk factors and suppresses redundant interference information after the initial multidimensional features are weighted and projected. For example, in the application scenario of dynamic pricing for new energy vehicle insurance, the multidimensional weighted feature tensor can be a high-dimensional feature matrix that strengthens the battery spontaneous combustion risk indicator and weakens the minor scratch indicator. In the application scenario of comprehensive risk assessment of chronic disease complications, the multidimensional weighted feature tensor can be a high-dimensional feature matrix that strengthens the arrhythmia warning indicator and weakens the abnormal indicators of ordinary physical examination.

[0073] Dimensionality reduction fusion feature matrix refers to matrix data with lower dimensionality but denser semantic information output after multidimensional weighted feature tensors are cross-combined and calculated and duplicate or invalid information is removed. For example, in the application scenario of joint modeling for anti-fraud in auto insurance, the dimension reduction fusion feature matrix can be matrix data that removes noise from normal claims behavior and condenses the core gang's motives. In the application scenario of cross-departmental auxiliary diagnosis of complex and difficult critical illnesses, the dimension reduction fusion feature matrix can be matrix data that removes background noise from healthy tissue and condenses the core pathology of multiple complications.

[0074] In this embodiment, the adaptive fusion unit of the object risk assessment system first receives the dynamic fusion weights and multi-dimensional initial feature tensors output from the preceding sequence. Based on the dynamic fusion weights, it performs weighted feature projection on the multi-dimensional initial feature tensor. Through tensor multiplication, the weight coefficients of different dimensions are precisely applied to each feature channel of the multi-dimensional initial feature tensor to amplify the core risk features highly relevant to the current scene and suppress irrelevant edge interference information, thereby outputting a multi-dimensional weighted feature tensor. Then, the object risk assessment system performs nonlinear cross-mapping on the multi-dimensional weighted feature tensor. Using a multilayer perceptron or feature cross-network, it promotes high-order feature combination and information collision of data from different heterogeneous sub-models in a unified latent space, deeply mining potential coupled risks that cannot be represented by a single modality. Subsequently, redundant dimensional space is stripped away to obtain a dimensionality-reduced fusion feature matrix with significantly improved information density. Finally, the object risk assessment system performs vectorization smoothing on the dimensionality-reduced fusion feature matrix. Through layer normalization technology and pooling flattening operation, it eliminates the local severe jitter and extreme outliers remaining in the underlying feature calculation process, and finally outputs a comprehensive risk representation vector with compact dimensions and extremely stable semantic expression.

[0075] This application first obtains a multidimensional weighted feature tensor by weighted feature projection of the multidimensional initial feature tensor based on dynamic fusion weights. In the initial stage of feature fusion, it achieves accurate reshaping and selection of the importance of risk information of different modalities. Then, it performs nonlinear cross-mapping on the multidimensional weighted feature tensor to obtain a dimension-reduced fusion feature matrix. This deeply explores the complex cross-coupling correlation between heterogeneous data and effectively removes redundant noise that may cause computational disasters. Finally, it performs vectorization and smoothing on the dimension-reduced fusion feature matrix to obtain a comprehensive risk representation vector. This further eliminates local jitter and extreme anomaly deviations in the high-order multimodal combination process. This layer-by-layer feature extraction and noise reduction fusion mechanism not only completely breaks down the last barrier from discrete heterogeneous features to a globally unified risk characterization, but also significantly improves the expression purity and robust stability of the final risk representation result. This provides the most reliable data kernel foundation for the subsequent output of high-accuracy and highly interpretable insurance intelligent pricing and decision-making strategies.

[0076] In step S80 of some embodiments, the target object processing strategy refers to the control parameter instructions generated based on the final risk assessment results that can be directly applied to terminal management or business execution. For example, in the application scenario of dynamic billing by mileage for Internet auto insurance, when the decision network of the object risk assessment system completes the boundary constraint calculation of the comprehensive risk representation vector, the target object processing strategy may be the single trip premium increase instruction data issued for illegal driving behavior under specific severe weather conditions. In the application scenario of organ transplant donor-recipient matching in large hospitals, if the decision network determines that the comprehensive risk representation vector exceeds the rejection reaction safety threshold, the target object processing strategy may be the medical management instruction data that blocks the organ matching process and triggers the secondary immune indicator review.

[0077] In this embodiment, the object risk assessment system directly inputs the deeply fused comprehensive risk representation vector into the backend decision generation network. The decision generation network combines preset business optimization rules, resource cost constraints, and expected profit control target parameters, and performs boundary constraint evaluation and optimization on the comprehensive risk representation vector through a nonlinear probability mapping algorithm. This transforms the comprehensive risk representation vector into a combination of control parameters that can be directly executed and returned to the requesting client. For example, in the smart product pricing application scenario, it can calculate an accurate and adaptive premium adjustment amount and issue it as a target object processing strategy.

[0078] In some embodiments, the object risk quantification index refers to a specific numerical value or rating score output after probability mapping to intuitively measure the severity of the potential risk of the target object. For example, in the financial credit approval scenario, the object risk quantification index can be a quantitative credit score that characterizes the borrower's default risk level, while in the medical outpatient triage scenario, the object risk quantification index can be a quantitative rating index that characterizes the severity of the patient's current condition.

[0079] Resource cost constraints refer to hard constraints such as financial bottom lines, manpower limitations, or physical equipment limits that must be considered when executing business decisions and generating strategies. For example, in the scenario of quoting financial insurance products, resource cost constraints may be the minimum underwriting profit margin or the bottom line of claims reserve funds set by the insurance company in advance. In the scenario of intelligent dispatching of medical ambulances, resource cost constraints may be the number of idle ambulances that the emergency center can dispatch and the upper limit of the time for medical staff to go out.

[0080] Dynamic business scheduling strategies refer to business execution rules that are pre-set and can be flexibly adjusted based on the macro market environment, corporate strategic goals, or real-time system load. For example, in the scenario of financial insurance promotion, dynamic business scheduling strategies can be market expansion rules that dynamically reduce the premium increase ratio in order to seize market share. In the scenario of medical fever clinic resource allocation, dynamic business scheduling strategies can be response rules that dynamically adjust the number of nucleic acid testing channels open based on the daily number of infections.

[0081] The feasible domain for strategy optimization refers to the set of all candidate processing solutions that meet all resource constraints and comply with dynamic business rules after boundary condition filtering. For example, in the adaptive pricing scenario of financial auto insurance, the feasible domain for strategy optimization can be the set of multiple candidate pricing gradients that simultaneously meet the company's profit bottom line and the customer's historical premium fluctuation limit. In the medical surgery scheduling scenario, the feasible domain for strategy optimization can be the set of all legal scheduling combinations that simultaneously meet the surgeon's idle time and the available time slots in the operating room.

[0082] In this embodiment, the dynamic pricing generation module of the object risk assessment system first receives the comprehensive risk representation vector output by the previous calculation, and performs nonlinear probability mapping based on the comprehensive risk representation vector. Using a preset activation function, the high-dimensional and abstract risk characteristics are converted into intuitive risk probability values, thereby obtaining a quantitative indicator of object risk that can clearly measure the level of risk of the current insurance request. Then, the system simultaneously introduces the pre-acquired resource cost constraints representing the bottom line cost of the product and the dynamic business scheduling strategy representing the current macro market objectives. Based on this, the system performs strict boundary constraint calculations on the above-mentioned quantitative indicator of object risk, aiming to accurately eliminate those extremely ineffective decision intervals that would lead to losses or violate market regulatory rules. This allows the system to derive a feasible domain for strategy optimization that meets the bottom line of business operations and compliance requirements. Finally, based on this feasible domain, the system uses an optimization algorithm to perform dynamic strategy optimization. Under the premise of ensuring risk coverage and controllable cost and profit, the system accurately calculates the optimal pricing parameters that can maximize business benefits and finally outputs a target object processing strategy containing targeted premium amounts to return to the client.

[0083] This application first obtains the object risk quantification index by performing nonlinear probability mapping based on the comprehensive risk representation vector. This successfully transforms the black-box high-dimensional assessment features into a standard probability metric that can be directly understood and applied by the business side. Then, based on resource cost constraints and a preset dynamic business scheduling strategy, the boundary constraint calculation of the object risk quantification index is performed to obtain the feasible domain for strategy optimization. This cleverly integrates and effectively filters the underlying model prediction results with the complex business cost bottom line and dynamic market rules in the real world. Finally, based on the feasible domain for strategy optimization, dynamic strategy optimization is performed to obtain the target object processing strategy. This progressive transformation and optimization mechanism completely solves the core pain point that traditional risk assessment models are often separated from the final business pricing decision. This enables the object risk assessment system to not only accurately characterize multi-dimensional risks, but also adaptively generate the optimal business response plan that balances risk interception and business profit within a safe, compliant, and cost-controllable feasible domain. This significantly improves the system's decision-making intelligence level and business closed-loop implementation capability.

[0084] It is also important to know that, after obtaining the target object processing strategy in this embodiment, the object risk assessment system first automatically analyzes the decision calculation process of the internal fusion engine and the specific contribution of each heterogeneous sub-model to the final assessment result, and then generates and outputs a structured interpretable report to clearly point out the key risk factors affecting the final processing strategy and their data sources, thereby meeting industry compliance requirements and the transparency requirements of business communication. At the same time, the system will also continuously monitor real-time dynamic risk events related to the target object in subsequent business execution cycles. Once a specific risk event is triggered, the system will immediately extract the event feature information and use it as a feedback signal to start a reassessment, thereby obtaining the online updated underlying sub-model parameters and the iteratively optimized fusion decision strategy. This completely breaks the limitations of traditional single-assessment models and realizes full-link closed-loop optimization from front-end dynamic business response to back-end model optimization.

[0085] As can be seen, in the above scheme, the target scene context features can be extracted and target multi-source heterogeneous data can be collected by receiving object processing requests, breaking the limitation of the single data dimension of traditional assessment schemes. Subsequently, the target heterogeneous sub-model set is dynamically matched based on the target scene context features, effectively solving the defect of decoupling between traditional static fixed models and complex and variable scenarios from the underlying architecture level. Next, the target heterogeneous sub-model set is used to perform submodal feature mapping and model confidence evaluation respectively. After obtaining dynamic fusion weights by quantizing the joint target scene context features, weighted feature fusion is performed on the multi-dimensional initial feature tensor to obtain a comprehensive risk representation vector. This enables the object risk assessment system to adaptively strengthen high reliability risk features and suppress interference noise from non-matching scenarios, greatly improving the three-dimensionality and accuracy of the underlying risk characterization. Finally, the target object processing strategy is generated based on the comprehensive risk representation vector, opening up a closed loop from real-time risk perception to dynamic strategy issuance, improving the adaptive decision-making efficiency and accuracy of the object risk assessment system in multi-mode business scenarios.

[0086] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0087] In one embodiment, an object risk assessment device is provided, which corresponds one-to-one with the object risk assessment method in the above embodiments. For example... Figure 6 As shown, the object risk assessment device includes a request feature extraction module, a data acquisition module, a model matching module, a feature mapping module, a confidence assessment module, a quantification mapping module, a feature fusion module, and a risk assessment module. Detailed descriptions of each functional module are as follows: The request feature extraction module is used to receive object processing requests and extract features from the object processing requests to obtain target scene context features; The data acquisition module is used to perform multimodal data acquisition based on object processing requests to obtain target multi-source heterogeneous data; The model matching module is used to perform model matching based on the context features of the target scene to obtain a set of target heterogeneous sub-models; The feature mapping module is used to perform modal feature mapping on multi-source heterogeneous data of the target based on the set of target heterogeneous sub-models, and obtain multi-dimensional initial feature tensors. The confidence assessment module is used to assess the model confidence of target multi-source heterogeneous data based on the target heterogeneous sub-model set, and obtain multi-dimensional output confidence. The quantization mapping module is used to perform joint quantization mapping on the multi-dimensional output confidence based on the context features of the target scene to obtain dynamic fusion weights; The feature fusion module is used to perform weighted feature fusion on multi-dimensional initial feature tensors based on dynamic fusion weights to obtain a comprehensive risk representation vector. The risk assessment module is used to conduct object risk assessment based on the comprehensive risk representation vector and obtain the target object processing strategy.

[0088] This invention provides an object risk assessment device. First, it receives object processing requests, extracts target scene context features, and collects multi-source heterogeneous target data, breaking the limitation of traditional assessment schemes with their single data dimension. Then, it dynamically matches a set of heterogeneous target sub-models based on the target scene context features, effectively solving the problem of decoupling traditional static fixed models from complex and variable scenarios at the underlying architecture level. Next, it uses the target heterogeneous sub-model set to perform submodal feature mapping and model confidence assessment. After obtaining dynamic fusion weights through quantization mapping of the joint target scene context features, it performs weighted feature fusion on the multi-dimensional initial feature tensors to obtain a comprehensive risk representation vector. This enables the object risk assessment system to adaptively strengthen high-reliability risk features and suppress interference noise from mismatched scenarios, greatly improving the three-dimensionality and accuracy of the underlying risk characterization. Finally, it generates target object processing strategies based on this comprehensive risk representation vector, establishing a closed-loop process from real-time risk perception to dynamic strategy issuance, improving the adaptive decision-making efficiency and accuracy of the object risk assessment system in multi-modal business scenarios.

[0089] Specific limitations regarding the object risk assessment device can be found in the limitations of the object risk assessment method described above, and will not be repeated here. Each module in the aforementioned object risk assessment device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0090] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side object risk assessment method.

[0091] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 8 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements the functions or steps of a client-side object risk assessment method. In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Receive object processing requests and extract features from the object processing requests to obtain target scene context features; Based on object-oriented processing requests, multimodal data acquisition is performed to obtain target multi-source heterogeneous data; Based on the contextual features of the target scene, model matching is performed to obtain a set of heterogeneous sub-models for the target. Based on the target heterogeneous sub-model set, modal feature mapping is performed on the target multi-source heterogeneous data to obtain a multi-dimensional initial feature tensor. Based on the target heterogeneous sub-model set, the model confidence of the target multi-source heterogeneous data is evaluated to obtain multi-dimensional output confidence. Based on the contextual features of the target scene, the multi-dimensional output confidence is jointly quantized and mapped to obtain dynamic fusion weights; Based on dynamic fusion weights, a weighted feature fusion is performed on the multi-dimensional initial feature tensor to obtain a comprehensive risk representation vector. Based on the comprehensive risk representation vector, object risk assessment is performed to obtain the target object processing strategy.

[0092] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Receive object processing requests and extract features from the object processing requests to obtain target scene context features; Based on object-oriented processing requests, multimodal data acquisition is performed to obtain target multi-source heterogeneous data; Based on the contextual features of the target scene, model matching is performed to obtain a set of heterogeneous sub-models for the target. Based on the target heterogeneous sub-model set, modal feature mapping is performed on the target multi-source heterogeneous data to obtain a multi-dimensional initial feature tensor. Based on the target heterogeneous sub-model set, the model confidence of the target multi-source heterogeneous data is evaluated to obtain multi-dimensional output confidence. Based on the contextual features of the target scene, the multi-dimensional output confidence is jointly quantized and mapped to obtain dynamic fusion weights; Based on dynamic fusion weights, a weighted feature fusion is performed on the multi-dimensional initial feature tensor to obtain a comprehensive risk representation vector. Based on the comprehensive risk representation vector, object risk assessment is performed to obtain the target object processing strategy.

[0093] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0094] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0095] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0096] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0097] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for assessing object risk, characterized in that, The method includes: Receive an object processing request and extract features from the object processing request to obtain target scene context features; Based on the object processing request, multimodal data acquisition is performed to obtain target multi-source heterogeneous data; Based on the target scene context features, model matching is performed to obtain a set of target heterogeneous sub-models; Based on the target heterogeneous sub-model set, modal feature mapping is performed on the target multi-source heterogeneous data to obtain a multi-dimensional initial feature tensor. Based on the target heterogeneous sub-model set, the model confidence of the target multi-source heterogeneous data is evaluated to obtain multi-dimensional output confidence. Based on the target scene context features, the multidimensional output confidence is jointly quantized and mapped to obtain dynamic fusion weights; Based on the dynamic fusion weights, the multi-dimensional initial feature tensor is weighted and fused to obtain a comprehensive risk representation vector. Based on the comprehensive risk representation vector, an object risk assessment is performed to obtain the target object processing strategy.

2. The object risk assessment method as described in claim 1, characterized in that, The target heterogeneous sub-model set includes graph relation sub-models and time-series behavior sub-models; The step of performing modal feature mapping on the target multi-source heterogeneous data based on the target heterogeneous sub-model set to obtain a multi-dimensional initial feature tensor includes: Based on the graph relation sub-model, a random walk of heterogeneous nodes is performed on the target multi-source heterogeneous data to obtain a sequence of node association and interaction paths. Based on the graph relation sub-model, graph features are aggregated on the node association interaction path sequence to obtain the association risk feature tensor; Based on the aforementioned temporal behavior sub-model, dynamic temporal features are extracted from the target multi-source heterogeneous data to obtain the behavior offset feature tensor. The associated risk feature tensor and the behavior offset feature tensor are mapped into a vector space to obtain a unified dimension feature. The unified dimensional features are integrated into a multimodal tensor to obtain the initial multidimensional feature tensor.

3. The object risk assessment method as described in claim 1, characterized in that, The step of performing joint quantization mapping on the multidimensional output confidence based on the target scene context features to obtain dynamic fusion weights includes: Based on the target scene context features and the pre-acquired model historical performance data, a context tensor mapping is performed to obtain the scene perception feature matrix. Based on the scene-aware feature matrix, attention feature interaction is performed on the multidimensional output confidence to obtain the confidence residual feature; The confidence residual features are assigned nonlinear weights to obtain the dynamic fusion weights.

4. The object risk assessment method as described in claim 1, characterized in that, The step of performing weighted feature fusion on the multi-dimensional initial feature tensor based on the dynamic fusion weights to obtain a comprehensive risk representation vector includes: Based on the dynamic fusion weights, the multi-dimensional initial feature tensor is weighted feature projection to obtain a multi-dimensional weighted feature tensor. A nonlinear cross-mapping is performed on the multidimensional weighted feature tensor to obtain a dimension-reduced fusion feature matrix; The reduced-dimensionality fusion feature matrix is ​​vectorized and smoothed to obtain the comprehensive risk representation vector.

5. The object risk assessment method as described in claim 1, characterized in that, The step of performing model matching based on the target scene context features to obtain a target heterogeneous sub-model set includes: Based on the contextual features of the target scene, multidimensional semantic clustering analysis is performed to obtain multimodal scene semantic clusters; Based on the multimodal scene semantic cluster and the preset model performance benchmark topology, gated routing calculation is performed to obtain the activation probability of candidate sub-model routes. Based on the candidate sub-model routing activation probability and the preset activation threshold, feature mask filtering is performed to obtain the target activation mask; Based on the target activation mask, model extraction is performed on the preset polymorphic risk assessment model library to obtain the target heterogeneous sub-model set.

6. The object risk assessment method as described in any one of claims 1-5, characterized in that, The step of performing multimodal data acquisition based on the object processing request to obtain target multi-source heterogeneous data includes: Based on the object processing request, multi-source data concurrent capture is performed to obtain heterogeneous raw data streams; The heterogeneous raw data stream is subjected to time-domain interpolation to obtain a time-aligned data stream; Based on a preset feature space transformation matrix, the time-aligned data stream is dimensionally aligned and mapped to obtain the target multi-source heterogeneous data.

7. The object risk assessment method according to any one of claims 1-5, characterized in that, The process of conducting object risk assessment based on the comprehensive risk representation vector to obtain the target object processing strategy includes: Based on the comprehensive risk representation vector, a nonlinear probability mapping is performed to obtain the object risk quantification index. Based on the pre-acquired resource cost constraints and the preset dynamic business scheduling strategy, the boundary constraints of the object risk quantification index are calculated to obtain the feasible domain for strategy optimization. Based on the feasible region of the strategy optimization, dynamic strategy optimization is performed to obtain the target object processing strategy.

8. A device for assessing the risk of an object, characterized in that, include: The request feature extraction module is used to receive object processing requests and extract features from the object processing requests to obtain target scene context features; The data acquisition module is used to perform multimodal data acquisition based on the object processing request to obtain target multi-source heterogeneous data; The model matching module is used to perform model matching based on the target scene context features to obtain a set of target heterogeneous sub-models; The feature mapping module is used to perform modal feature mapping on the target multi-source heterogeneous data based on the target heterogeneous sub-model set to obtain a multi-dimensional initial feature tensor; The confidence assessment module is used to assess the model confidence of the target multi-source heterogeneous data based on the target heterogeneous sub-model set, and obtain multi-dimensional output confidence. The quantization mapping module is used to perform joint quantization mapping on the multidimensional output confidence based on the target scene context features to obtain dynamic fusion weights; The feature fusion module is used to perform weighted feature fusion on the multi-dimensional initial feature tensor based on the dynamic fusion weights to obtain a comprehensive risk representation vector; The risk assessment module is used to perform object risk assessment based on the comprehensive risk characterization vector to obtain the target object processing strategy.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the object risk assessment method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the object risk assessment method as described in any one of claims 1 to 7.

Images