Decision recommendation method and device, computer equipment and storage medium

Abstract

The invention discloses a decision recommendation method and device, computer equipment and a storage medium, belongs to the technical field of artificial intelligence, and is applied to a car insurance risk control decision scene. According to the method, real-time fusion and dynamic feature alignment of the multi-source heterogeneous data are realized by constructing the federal learning data center, and the bottleneck of traditional data islands and scattered storage is broken. A pre-trained dynamic deep learning model is utilized, a space-time attention mechanism and an incremental updating engine are introduced into the system, space-time association of sensor data and business events is effectively captured, and cross-scene generalization ability improvement and quick response to market fluctuation are achieved. The explainable decision network splits complex model output into clear feature contribution degrees, and by matching with key influence factor extraction and natural language decision correction, the transparency of decisions and the understanding efficiency of business personnel are greatly improved. The real-time recommendation system based on reinforcement learning guarantees the timeliness of abnormal early warning and the accuracy of decision execution.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a decision recommendation method and device, a computer device and a storage medium. BACKGROUND

[0002] In the field of modern enterprise management and data-driven decision-making, accurate analysis of abnormal indicators and recommendation of business decisions are of vital importance, directly related to the operational efficiency and market competitiveness of enterprises. This field is not only the core pillar of enterprise digital transformation, but also the key to improving resource allocation efficiency and reducing operational risks. However, although many current solutions have made some progress in data analysis and decision support, there are some deep-seated limitations, such as the difficulty of dealing with data dispersion and heterogeneity in complex environments, and the lack of sufficient adaptability and timely response mechanisms when facing dynamic business scenarios, resulting in poor decision-making results and even trust crises. For example, in the decision-making of the property insurance field, vehicle data in the car insurance business is scattered in heterogeneous systems such as the Internet of Vehicles and repair shops, making it difficult for traditional methods to integrate. When the market fluctuates and policies change, it is impossible to adjust strategies in a timely manner, resulting in unreasonable pricing and inaccurate risk assessment, which makes customers question the fairness of the decision-making process and seriously affects business development and industry reputation.

[0003] By analyzing the challenges in this field, it can be found that data fragmentation is the primary obstacle. Because data within an enterprise and across organizations is often distributed in different systems and regions, it is isolated and difficult to integrate, which directly leads to the inability of analysis models to obtain comprehensive information during training and application. This problem further exacerbates the lack of adaptability of models in different scenarios, as the lack of a unified data foundation makes it difficult for models to capture the dynamic characteristics of business environments, thereby affecting the accuracy and timeliness of decisions.

[0004] Therefore, how to build a decision support framework that can effectively integrate scattered data resources and dynamically adapt to complex business scenarios has become a key problem in improving the effectiveness of abnormal indicator analysis and business decision recommendation. SUMMARY

[0005] The purpose of the embodiments of the present application is to propose a decision recommendation method, device, computer device and storage medium to solve the technical problem of how to build a decision support framework that can effectively integrate scattered data resources and dynamically adapt to complex business scenarios.

[0006] To solve the above technical problems, the embodiments of the present application provide a decision recommendation method, which adopts the following technical solutions:

[0007] A decision recommendation method, comprising:

[0008] The data stream from the multi-source heterogeneous system is acquired, and real-time feature alignment is performed on the data stream through a preset distributed architecture to generate dynamic interaction features;

[0009] The dynamic deep learning model is used to process the dynamic interaction features, and the spatio-temporal correlation between the dynamic interaction features is captured.

[0010] Based on the dynamic interaction features and the spatio-temporal correlation, the pre-trained anomaly recognition model is used for data anomaly recognition to obtain a data anomaly recognition result.

[0011] The pre-constructed explainable decision network is used to decompose the feature contribution degree of the data anomaly recognition result to generate a feature contribution degree report.

[0012] The key influencing factors causing the data anomaly are determined according to the feature contribution degree report.

[0013] The preset decision is corrected according to the key influencing factors, a decision basis report is generated, and the corrected decision and the decision basis report are output.

[0014] To solve the above technical problems, the embodiment of the application also provides a decision recommendation device, which adopts the technical scheme as follows:

[0015] A decision recommendation device comprises:

[0016] A feature alignment module is configured to acquire data stream from a multi-source heterogeneous system, and perform real-time feature alignment on the data stream through a preset distributed architecture to generate dynamic interaction features.

[0017] A spatio-temporal correlation module is configured to use a pre-trained dynamic deep learning model to process the dynamic interaction features, and capture the spatio-temporal correlation between the dynamic interaction features.

[0018] An anomaly recognition module is configured to use a pre-trained anomaly recognition model to perform data anomaly recognition based on the dynamic interaction features and the spatio-temporal correlation, and obtain a data anomaly recognition result.

[0019] A contribution decomposition module is configured to use a pre-constructed explainable decision network to decompose the feature contribution degree of the data anomaly recognition result to generate a feature contribution degree report.

[0020] A key factor module is configured to determine key influencing factors causing the data anomaly according to the feature contribution degree report.

[0021] A decision recommendation module is configured to correct a preset decision according to the key influencing factors, generate a decision basis report, and output the corrected decision and the decision basis report.

[0022] To solve the above technical problems, the embodiment of the present application also provides a computer device which adopts the technical scheme as follows:

[0023] A computer device comprises a memory and a processor, the memory stores computer readable instructions, and the processor executes the computer readable instructions to realize the steps of the decision recommendation method according to any one of the above.

[0024] To solve the above technical problems, the embodiment of the present application also provides a computer readable storage medium which adopts the technical scheme as follows:

[0025] A computer readable storage medium stores computer readable instructions, and the computer readable instructions are executed by a processor to realize the steps of the decision recommendation method according to any one of the above.

[0026] Compared with the prior art, the embodiment of the present application has the following beneficial effects:

[0027] The present application discloses a decision recommendation method and device, a computer device and a storage medium, which belong to the field of artificial intelligence and are applied to the vehicle insurance risk control decision scene. The present application realizes real-time fusion and dynamic feature alignment of multi-source heterogeneous data by constructing a federated learning data hub, greatly improves the integrity and consistency of data, and breaks the bottleneck of traditional data silos and scattered storage. Using a pre-trained dynamic deep learning model, the system introduces a spatiotemporal attention mechanism and an incremental update engine, effectively capturing the spatiotemporal correlation of sensor data and business events, and realizing the improvement of cross-scene generalization ability and the rapid response to market fluctuations. The interpretable decision network disassembles the complex model output into clear feature contribution, cooperates with key influence factor extraction and natural language decision correction, and greatly improves the transparency of the decision and the understanding efficiency of the business personnel. The real-time recommendation system based on reinforcement learning realizes millisecond-level processing of massive sensor data, guarantees the timeliness of the abnormal warning and the accuracy of the decision execution. The present application effectively breaks the data silos, improves the dynamic adaptability of the model and the accuracy of the decision. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the schemes in the present application, the drawings needed in the description of the embodiments of the present application will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0029] Figure 1 An exemplary system architecture diagram in which the present application can be applied is shown;

[0030] Figure 2A flowchart of one embodiment of the decision recommendation method according to this application is shown;

[0031] Figure 3 A schematic diagram of one embodiment of the decision recommendation device according to this application is shown;

[0032] Figure 4 A schematic diagram of the structure of one embodiment of a computer device according to this application is shown. Detailed Implementation

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0034] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0035] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0036] like Figure 1 As shown, system architecture 100 may include terminal device 101, network 102, and server 103. Terminal device 101 may be a laptop 1011, tablet 1012, or mobile phone 1013. Network 102 is used as a medium to provide a communication link between terminal device 101 and server 103. Network 102 may include various connection types, such as wired, wireless communication links, or fiber optic cables.

[0037] Users can use terminal device 101 to interact with server 103 via network 102 to receive or send messages, etc. Various communication client applications can be installed on terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0038] The terminal device 101 can be various electronic devices with a display screen and supporting web browsing, in addition to the notebook computer 1011, the tablet computer 1012 or the mobile phone 1013, the terminal device 101 can also be an electronic book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 player (Moving Picture Experts Group Audio Layer IV), a laptop computer and a desktop computer, etc.

[0039] The server 103 can be a server providing various services, for example, a background server providing support for a page displayed on the terminal device 101.

[0040] It should be noted that the decision recommendation method provided in the embodiments of the present application is generally executed by a server / terminal device, and accordingly, the decision recommendation apparatus is generally arranged in the server / terminal device.

[0041] It should be understood that, Figure 1 The number of terminal devices, networks and servers in the system is only illustrative, and the system can have any number of terminal devices, networks and servers according to the needs of implementation.

[0042] With reference to Figure 2 , a flow chart of one embodiment of the decision recommendation method according to the present application is shown. The decision recommendation method includes the following steps:

[0043] S201, acquiring data streams from multiple source heterogeneous systems, and performing real-time feature alignment on the data streams through a preset distributed architecture to generate dynamic interaction features;

[0044] Specifically, the system first accesses the data from the CRM, ERP, SCM, IoT device, financial system and other multiple source heterogeneous systems to the real-time processing platform through a unified data access layer, and the commonly used technologies include the distributed stream processing framework based on Flink or SparkStreaming. Since there are significant differences in the time stamp, dimension and granularity of different data sources, the system needs to design a sliding time window mechanism (such as a window size of 15 minutes and a sliding interval of 30 seconds), and align different time axis events through the Watermark mechanism and out-of-order data buffering when data is ingested. For different source data feature dimensions, the distributed feature factory is used for feature standardization and coding (such as One-Hot, Embedding vectorization), and real-time derivation of interaction features, for example, cross combination of the average volatility of sensor data and the ERP inventory turnover rate.

[0045] To protect data security and privacy, a federated feature derivation mechanism can also be introduced during feature alignment. Through homomorphic encryption, cross-organizational data is aggregated to avoid the leakage of raw data. The final dynamic interaction features are buffered and pushed through message queues such as Kafka to ensure that deep learning models can subscribe to and process these aligned feature streams in real time.

[0046] S202, using a pre-trained dynamic deep learning model to process the dynamic interaction features and capture the spatio-temporal correlation between the dynamic interaction features;

[0047] Specifically, the system inputs the real-time aligned dynamic interaction features into a pre-trained deep learning model. This model is usually extended based on the Transformer architecture, embedding residual attention modules and spatio-temporal convolution units to capture complex correlations in both time and space dimensions. The input end uses multi-head self-attention mechanisms to weight different feature channels to learn the interaction patterns between features, and combines position encoding to handle the time series characteristics of the feature stream. For spatial correlations, such as device sensor deployment locations and supply chain node geographic distribution, graph convolution (GCN) or spatio-temporal graph attention (ST-GAT) can be used to model the spatial topology structure and embed it into the Transformer for joint spatio-temporal modeling. In addition, model parameters are updated regularly through online incremental learning strategies. When the system detects that the market volatility index or external environmental parameters reach a certain threshold, it triggers Monte Carlo sampling to generate multiple candidate parameter sets and selects the optimal model parameter set through KL divergence evaluation. The entire model uses optimized TensorRT or ONNX Runtime for acceleration during the inference phase to meet the high throughput requirements of real-time processing.

[0048] A dynamic deep learning model is an architecture that combines deep neural network structures with time and spatial dimension modeling techniques when processing time series stream data and high-dimensional interaction features. It is usually based on Transformer, LSTM, TCN (Temporal Convolutional Network), etc. as the basic model architecture. By introducing residual attention mechanisms, spatio-temporal convolution blocks, and graph neural network embeddings, it dynamically updates weight parameters to adapt to environmental changes. It not only encodes the dependency relationships between past and current data, but also learns new patterns in an incremental manner during runtime.

[0049] For example, in the Internet of Vehicles scenario, the dynamic deep learning model can receive vehicle OBD sensor data, road condition monitoring information and weather forecast features in real time, capture the coupling relationship between the speed of the vehicle at different road nodes and the environmental temperature through the spatio-temporal attention module, and trigger incremental updates when the hourly market volatility index changes, so that the model quickly adapts to sudden traffic control or extreme weather, thereby showing stronger dynamic response capability and cross-scene adaptability in anomaly detection and prediction tasks.

[0050] S203, based on the dynamic interaction features and the spatio-temporal correlation, using a pre-trained anomaly recognition model to perform data anomaly recognition, and obtaining a data anomaly recognition result;

[0051] Specifically, the system embeds the features processed by the spatio-temporal deep processing into the anomaly recognition model, usually using a structure based on an autoencoder (AutoEncoder), a variational autoencoder (VAE) or a deep one-class classification (Deep One-Class Classification) to learn the distribution of the feature space. During the pre-training process, the reconstruction distribution of the normal data pattern has been established. When the real-time input features deviate significantly, the model will output an anomaly score. In order to improve the adaptability of anomaly recognition to complex spatio-temporal scenarios, an anomaly detection method based on adversarial training (such as GANomaly) can be combined to train the discriminator in the generative adversarial network to improve the sensitivity to subtle anomalies. The inference stage of the model reduces the delay through a batch fine-tuning strategy, and combines the sliding window anomaly distribution statistics, such as using the CUSUM control chart and time series residual detection method, to aggregate and analyze the anomaly trend in the continuous time period. The final output data anomaly recognition result not only includes the anomaly label, but also can be accompanied by a confidence score and positioning information, such as the device node, time period and feature dimension where the anomaly occurs.

[0052] S204, using a pre-constructed explainable decision network to decompose the feature contribution degree of the data anomaly recognition result, and generating a feature contribution degree report;

[0053] Specifically, in view of the black-box characteristics of the anomaly identification model, the system introduces an explainability algorithm module, such as a SHAP (SHapley Additive exPlanations) based explanation engine, to perform feature level contribution decomposition on a single prediction result. The engine will sample a large number of perturbed feature inputs in the background and calculate the marginal contribution of each feature to the change in abnormal score under different perturbation combinations. Finally, the feature contribution matrix is output in the form of Shapley value. For real-time streaming data, the explanation process needs to be optimized, and weighted sampling or local proxy model (LIME) can be used to reduce the dimension of high-dimensional feature space and speed up the process. The feature contribution report not only records the numerical contribution, but also converts the complex contribution information into business understandable text description through natural language generation (NLG) module, such as "sensor temperature fluctuation accounts for 35% of the abnormal score, and inventory turnover rate accounts for 20%".

[0054] S205, determining the key influencing factors causing data anomaly according to the feature contribution report;

[0055] Specifically, the system sorts and aggregates the feature contribution matrix output by the SHAP algorithm to filter out a number of features with the highest contribution as key influencing factors. In this process, a contribution threshold can be set, such as a feature set with cumulative contribution exceeding 70%, or a feature with contribution to abnormal score exceeding a certain percentage. At the same time, in order to avoid misjudgment caused by accidental high contribution of a single feature, the system can introduce feature importance stability evaluation, such as cross-validation of contribution distribution in different time periods or different sliding windows, and use statistical methods (such as Bootst rap sampling and confidence interval calculation) to improve the reliability of the screening results. For data in different business domains, the system will also conduct semantic association analysis on the features in combination with the domain knowledge graph to identify features with more causal relationship with abnormal events, rather than just high correlation. The final key influencing factors are output in the form of a structured list, with metadata such as feature dimension, time interval, contribution ratio, confidence score, etc. for direct reference by the decision modification link.

[0056] S206, modifying the preset decision according to the key influencing factors, generating a decision basis report, and outputting the modified decision and the decision basis report.

[0057] Specifically, the system will call the preset business rule library and reinforcement learning strategy module according to the identified key influencing factors to automatically correct the current operating decision. The rule library can store multi-dimensional decision logic such as "if the inventory turnover rate is abnormal, adjust the purchase frequency", "if the equipment temperature fluctuation is too high, preventive maintenance is advanced", and the reinforcement learning module searches for the optimal solution in the strategy space through algorithms such as PPO or DDPG, and optimizes the current decision scheme combined with real-time environmental feedback. During the correction process, the system will perform Monte Carlo simulation on the risk, cost and return of different strategies, and select the scheme with the optimal expected value or the lowest risk as the final output. The corrected decision and its reasoning path will be written into the decision basis report, which includes the key features that trigger the correction, the original decision logic, the specific parameters after correction and the expected effect, etc. The report is presented in both structured data and natural language forms, making it easy for business personnel to review and audit the archive. Finally, the corrected decision and report will be pushed to the enterprise execution system (such as ERP, MES) through API or message bus, realizing the automatic closed-loop decision recommendation process.

[0058] Further, the step of acquiring data streams from multiple source heterogeneous systems and performing real-time feature alignment on the data streams through a preset distributed architecture to generate dynamic interaction features comprises:

[0059] Accessing multiple data sources through a distributed feature factory, setting a sliding time window for processing the data streams in each data source;

[0060] Performing time axis deviation correction on the data streams to generate aligned data streams;

[0061] Combining cross-institutional features for the aligned data streams in each data source to generate dynamic interaction features, and storing the dynamic interaction features in a data hub;

[0062] Detecting abnormal fluctuations in the data streams;

[0063] If abnormal fluctuations are detected in the data streams, triggering a feature update mechanism to regenerate dynamic interaction features.

[0064] In this embodiment, the distributed feature factory converges multi-source heterogeneous data such as ERP, CRM, IoT sensors, financial systems, etc. in real time through a high-throughput data access layer, and configures an independent sliding time window (such as a 15-minute window, a 30-second sliding window) for each data stream for segmented processing. The system uses timestamp re-alignment and Watermark delay processing technology to automatically correct the time axis deviation between different sources, ensuring that multi-dimensional features of the same business event can be aligned under the same time reference. Subsequently, homomorphic encryption and feature derivation algorithms are used to combine cross-institutional features of the aligned data stream, generating multi-dimensional dynamic interaction features, such as multiplying real-time sensor temperature fluctuation rate and financial cost ratio to form a new feature. Dynamic interaction features are written into a distributed data hub (such as HDFS or Federated Data Lake) for continuous calling by downstream models. The system also monitors statistical fluctuations and adapts threshold detection to determine whether there are abnormal fluctuations in the data stream in real time. Once the threshold is triggered, the feature update mechanism is started to recalculate and publish new interaction features to adapt to changes in external environment or business status.

[0065] Through the above steps, the system realizes high-precision real-time alignment of multi-source data and dynamic feature update, significantly improving the freshness of data features and the accuracy of decision input.

[0066] Further, the step of using a pre-trained dynamic deep learning model to process the dynamic interaction features to capture the spatio-temporal correlation between the dynamic interaction features includes:

[0067] Using the embedded attention mechanism of the dynamic deep learning model to generate time feature weights and spatial feature weights of the dynamic interaction features respectively;

[0068] Correlating the time feature weights and spatial feature weights to construct spatio-temporal correlation weights;

[0069] Using the spatio-temporal correlation weights to weight process the dynamic interaction features to obtain weighted dynamic interaction features;

[0070] Performing local convolution operation on the weighted dynamic interaction features and analyzing the local convolution results to obtain the spatio-temporal correlation between the dynamic interaction features.

[0071] In the embodiment, the pre-trained dynamic deep learning model adopts an improved Transformer framework as the basic structure. First, the input dynamic interaction feature sequence is separated in time and space dimensions through an embedding attention module. The generation of time feature weights relies on a multi-head self-attention mechanism to learn the weight of the feature dependence on time. The space feature weights are encoded by introducing a graph neural network (GCN) or a spatio-temporal graph attention module to encode the topological structure and geographical adjacency relationship between business nodes or sensors. After obtaining the time and space feature weights, the model fuses and calculates the two types of weights through a feature interaction unit to form a unified spatio-temporal correlation weight matrix. Then, the original dynamic interaction features are weighted using the matrix to obtain weighted dynamic interaction features that better reflect multi-dimensional correlations. To further mine local patterns, the model performs local convolution operations (such as TCN or Depthwise Convolution) on the weighted feature map to capture abnormal local features within a short time sequence and parse the convolution results into spatio-temporal correlation representations. Finally, a set of feature vectors representing the spatio-temporal interaction relationship between multi-source data is output to provide input for anomaly identification and decision recommendation.

[0072] Through the above steps, the model can deeply capture the spatio-temporal dependence of dynamic interaction features, significantly improving the richness of feature representation and the accuracy of downstream identification and prediction.

[0073] Further, the step of associating the time feature weights and the space feature weights to construct the spatio-temporal correlation weight includes:

[0074] Using the spatio-temporal separation layer of the dynamic deep learning model to separate the dynamic interaction features to obtain time features and space features;

[0075] Using data fusion technology, an initial feature combination matrix is constructed based on the time features and the space features to obtain a preliminary spatio-temporal characteristic description;

[0076] According to the preliminary spatio-temporal characteristic description, a correlation analysis method is used to fuse the time feature weights and the space feature weights to determine a comprehensive weight distribution map;

[0077] The spatio-temporal correlation weight is extracted from the comprehensive weight distribution map.

[0078] In this embodiment, the dynamic deep learning model first splits the input dynamic interaction features through a space-time separation layer, such as a multi-head attention-based feature separation module, extracting the feature sequence evolving over time as a time feature vector and the features reflecting the differences between different data sources and different spatial nodes as a spatial feature vector. Subsequently, the system maps the time features and spatial features into the same high-dimensional feature space using data fusion techniques such as feature crossing, feature tensor splicing, and normalization processing, constructing an initial feature combination matrix that describes the potential interaction between time variation patterns and spatial distribution patterns. Then, the model performs deep fusion of time feature weights and spatial feature weights through correlation analysis methods such as mutual information-based feature selection, covariance matrix analysis, or graph convolution-based weight aggregation, considering the feature contribution of both time and space dimensions to form a comprehensive weight distribution map. Finally, the system extracts the spatio-temporal correlation weights from the comprehensive weight distribution map to guide the weighted and convolutional processing of dynamic interaction features, achieving high-precision characterization of spatio-temporal patterns in complex multi-source data.

[0079] In the property insurance field, assume that an insurance company needs to analyze real-time driving data (speed, number of sudden stops, tire pressure changes) uploaded by OBD devices in vehicles from all over the country and accident-prone road segments, weather conditions, and road construction information in different regions when conducting vehicle insurance risk control. The system first extracts the driving state sequence of a vehicle every day in the past week as a time feature vector W t , and extracts the geographical location, road grade, and surrounding accident rate information of the area where the vehicle belongs as a spatial feature vector W s . During fusion, the system first performs dimension alignment and normalization processing on both, concatenating the time feature weight vector W t and the spatial feature weight vector W s into a feature pair [W t , W s ], and then calculating their interaction effects through a data fusion network (such as a multi-layer perception MLP or an attention gate unit), such as multiplying the weight of each time slice by the corresponding spatial risk factor and then performing weighted summation to obtain a time-space interaction matrix M ts . For example, the sudden stop feature at night (high time weight) will be given a higher comprehensive weight on "rainy and waterlogged road segments" (high spatial weight), while the weight will be lower on "dry road segments on highways" (low spatial weight). Through this element-by-element weighting and nonlinear fusion, a comprehensive weight distribution map is finally generated, allowing the model to clearly capture the strong correlation regions between time and space features.

[0080] Subsequently, the model combines the driving behavior patterns in the time dimension and the risk distribution in the space dimension into a feature matrix using feature splicing and normalization. Then, through the correlation analysis method, the time feature of "increased frequency of sudden braking at night" is combined with the spatial feature of "distribution of waterlogged road sections in a certain city" to obtain a comprehensive weight distribution map. Finally, the spatio-temporal correlation weight is extracted, enabling the model to accurately identify higher accident risks at specific times and locations, thereby providing a refined basis for pricing and risk control.

[0081] Through the above steps, the spatio-temporal correlation weight can be accurately constructed, the depth and expression ability of feature fusion are improved, and more comprehensive spatio-temporal relationship inputs are provided for model processing.

[0082] Further, the step of using a pre-constructed explainable decision network to perform feature contribution degree decomposition on the data anomaly identification result to generate a feature contribution degree report includes:

[0083] In the data anomaly identification result, the feature contribution degree of each dynamic interaction feature is calculated respectively;

[0084] The feature contribution degree decomposition is performed on the feature contribution degree of each dynamic interaction feature using the explainable decision network to determine the contribution degree of a single data stream feature;

[0085] The contribution degrees of all single data stream features are summarized to obtain a feature contribution degree report.

[0086] In this embodiment, the explainable decision network decomposes the contribution of each dynamic interaction feature based on the intermediate representation result of the deep neural network. First, the system calculates the marginal contribution degree of each feature node in the anomaly identification result. The SHAP (Shapley Additive Explanations) method or integrated gradient method can be used to decompose the overall anomaly score output by the complex model into the independent contribution value of each feature. Subsequently, the explainable decision network further decomposes the contribution degree of each feature, and restores the contribution value of the original single data stream feature (such as "brake frequency" and "road humidity") from the high-level feature (such as "interaction term of brake frequency and road humidity"). The system establishes a feature traceability graph in this process, and through the hierarchical mapping relationship, the feature interaction is decomposed into the basic input features, so that the positive or negative influence degree of each feature on the final anomaly score is clear at a glance. Finally, the contribution degrees of all single data stream features are summarized and sorted to form a feature contribution degree report. The report not only contains the contribution percentage of each feature, but also generates a natural language description, indicating the possible causes of the anomaly and the key features to focus on, facilitating decision makers to quickly understand and adjust the strategy.

[0087] Through the above steps, the system can clearly present the influence degree of each feature on the abnormal result, improve the result interpretation transparency, and assist business personnel in quickly positioning the problem root cause.

[0088] Further, the step of determining the key influencing factors causing the data anomaly according to the feature contribution report specifically includes:

[0089] In the feature contribution report, it is judged in turn whether the contribution degree of each data stream feature exceeds the preset contribution degree threshold value;

[0090] The data stream features exceeding the threshold value are obtained to obtain a threshold value exceeding feature set;

[0091] The related features between each threshold value exceeding feature in the threshold value exceeding feature set are judged in turn;

[0092] The threshold value exceeding features with strong related features are determined as key influencing factors.

[0093] In the present embodiment, the system will first traverse all feature items in the feature contribution report, compare the contribution degree value of each feature with the preset threshold value, filter out threshold value exceeding features having greater influence on the abnormal identification result, and form a preliminary high contribution feature set. Subsequently, the system performs two-by-two correlation calculation on the features in the set, and the correlation analysis can be based on Pearson correlation coefficient, mutual information score or graph structure similarity, so as to identify the strong correlation between the features. For example, in the car insurance scene in the property insurance field, the two features of “nighttime emergency braking frequency” and “rainy day slippery road accident density” both exceed the threshold value, and there is a significant positive correlation between them. The system determines this feature combination as a key influencing factor. In addition, in order to avoid accidental interference, the system also performs stability test, and statistically verifies the correlation of the related features in multiple time windows, so as to ensure that the selected key influencing factors have persistence and representativeness. Through the above multi-level filtering and correlation analysis, a key influencing factor list is finally formed.

[0094] Through the above steps, the system can extract the key factors that truly drive the anomaly from the complex features, and significantly improve the pertinence and efficiency of problem positioning and decision making.

[0095] Further, the step of modifying the preset decision according to the key influencing factors, generating a decision basis report, and outputting the modified decision and the decision basis report specifically includes:

[0096] The key influencing factors are introduced into the pre-trained natural language model to generate a natural language description of the predicted decision;

[0097] The semantic deviation between the natural language description of the predicted decision and the natural language description of the preset decision is calculated to obtain a decision semantic deviation;

[0098] correcting the preset decision based on the semantic deviation of the decision, to generate a corrected decision;

[0099] associating the corrected decision with the dynamic interaction features, to generate a decision basis report, and outputting the corrected decision and the decision basis report.

[0100] In this embodiment, the system first inputs the key influencing factors identified through the aforementioned steps into a pre-trained natural language generation model (such as GPT or BERT derived models based on Transformer architecture), which combines industry knowledge base and historical decision data to generate a predictive decision description for the current abnormal situation. This natural language description not only covers the potential causes of the anomaly, but also proposes corresponding countermeasures and recommended solutions. Subsequently, the system compares the predictive decision description with the enterprise's preset standard decision description through semantic comparison techniques (such as cosine similarity calculation based on sentence vectors or semantic matching networks), quantifying the semantic deviation between the two. This deviation reflects the adaptability and effectiveness of the preset decision under the current situation. According to the deviation result, the system uses a dynamic decision correction algorithm to automatically adjust the preset decision, including adjusting parameters, optimizing strategies or supplementing conditions, to generate a more accurate corrected decision. Finally, the system associates the corrected decision with relevant dynamic interaction features to form a comprehensive decision basis report containing abnormal cause analysis, correction measures and feature basis, ensuring the transparency and traceability of the decision, and synchronously outputs the corrected decision and the decision basis report.

[0101] Taking car insurance claim risk control as an example, the system detects that there have been many sudden braking accidents in a certain area caused by slippery roads in rainy days, and the key influencing factors include "high wet road slip index in rainy days" and "abnormal frequency of sudden braking of vehicles at night". The natural language model generates a predictive decision description based on these factors: "suggest increasing the risk coefficient of vehicles in that area during rainy days and strengthening the warning measures for night driving". The system compares this description with the original preset "uniform discount strategy" and finds that there is a large deviation between them, indicating that the original decision cannot effectively respond to the current risk. Therefore, the system automatically corrects the decision to "suspend the discount in rainy days in that area and strengthen the risk warning notification", and generates a decision basis report detailing the relevant wet road data and sudden braking abnormal features, providing clear operational guidance for the risk control team.

[0102] Through the above steps, a closed loop from abnormal features to natural language decision suggestions is realized, improving the intelligence and explainability of the decision, and enhancing the understanding and execution ability of business personnel.

[0103] In the specific embodiments of the present application, a compliance risk control system is also provided, in which differential privacy technology and blockchain technology are introduced. The differential privacy technology is integrated into the training and updating process of the dynamic deep learning model, and by injecting Laplace noise with privacy protection properties (ε=0.3) in the gradient updating stage, the risk of customer identity re-identification is effectively reduced, ensuring privacy security during data use, especially in cross-institutional federated learning scenarios. At the same time, an intelligent contract record based on blockchain technology is automatically generated for each key decision node in the decision recommendation system. These records include decision input data, model parameters, feature contribution, and final decision results, ensuring that all decision chains are transparent and tamper-proof. The blockchain record chain supports multi-level auditing and tracing, facilitating efficient inspection and compliance verification by regulatory authorities, significantly reducing compliance audit preparation time, enhancing the credibility and compliance of the system, and promoting precise and real-time business decisions in compliance with laws and regulations.

[0104] In the above embodiments, the present application discloses a decision recommendation method, which belongs to the field of artificial intelligence and is applied to the field of car insurance risk control decision-making. By constructing a federated learning data hub, the present application realizes real-time fusion and dynamic feature alignment of multi-source heterogeneous data, greatly improving the integrity and consistency of data and breaking through the bottleneck of traditional data silos and decentralized storage. Using a pre-trained dynamic deep learning model, the system introduces a spatio-temporal attention mechanism and an incremental updating engine, effectively capturing the spatio-temporal correlation of sensor data and business events, and achieving improved generalization ability and rapid response to market fluctuations across scenarios. The explainable decision network breaks down complex model outputs into clear feature contributions, combined with key influence factor extraction and natural language decision correction, greatly improving the transparency of decisions and the efficiency of business personnel understanding. The real-time recommendation system based on reinforcement learning achieves millisecond-level processing of massive sensor data, ensuring the timeliness of anomaly early warning and the accuracy of decision execution. The present application effectively breaks down data silos, improves the dynamic adaptability of the model and the accuracy of the decision.

[0105] In the present embodiment, the electronic device (e.g. Figure 1 The server shown in the figure can receive instructions or obtain data through wired or wireless connection. It should be noted that the above-mentioned wireless connection can include but is not limited to 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other now known or future developed wireless connection methods.

[0106] It should be emphasized that, in order to further ensure the privacy and security of the above-mentioned data flow information, the above-mentioned data flow information can also be stored in a node of a blockchain.

[0107] The blockchain referred to in the present application is a new application mode of distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm and other computer technologies. The blockchain is essentially a decentralized database, which is a series of data blocks associated using cryptographic methods, each data block containing information of a batch of network transactions, used to verify the validity (anti-fake) of the information and generate the next block. The blockchain can include a blockchain underlying platform, a platform product service layer, and an application service layer.

[0108] The embodiments of the present application can acquire and process related data based on artificial intelligence technology. Artificial intelligence (AI) is the use of digital computers or computer-controlled machines to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0109] Artificial intelligence basic technologies generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics, etc. Artificial intelligence software technology mainly includes computer vision technology, robot technology, biometric technology, speech processing technology, natural language processing technology, and machine learning / deep learning, etc.

[0110] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by computer readable instructions instructing related hardware, and the computer readable instructions can be stored in a computer readable storage medium. The program can include the processes of the above-mentioned embodiments when executed, wherein the storage medium can be a non-volatile storage medium such as a magnetic disc, an optical disc, a read-only memory (ROM), or a random access memory (RAM).

[0111] It should be understood that although each step in the flowchart of the accompanying drawings is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and they can be executed in other orders. Moreover, at least part of the steps in the flowchart of the accompanying drawings can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence is not necessarily sequential, but can be executed alternately or alternately with at least part of other steps or sub-steps or stages of other steps.

[0112] Further referring to Figure 3 , as an implementation of the method shown in the above Figure 2 , the present application provides an embodiment of a decision recommendation device, which corresponds to the method embodiment shown in Figure 2 , and the device can be applied in various electronic devices.

[0113] As shown in Figure 3 , the decision recommendation device 300 comprises:

[0114] a feature alignment module 301 configured to acquire data streams from multiple source heterogeneous systems and perform real-time feature alignment on the data streams through a preset distributed architecture to generate dynamic interaction features;

[0115] a space-time correlation module 302 configured to process the dynamic interaction features by using a pre-trained dynamic deep learning model to capture space-time correlation between the dynamic interaction features;

[0116] an anomaly identification module 303 configured to perform data anomaly identification by using a pre-trained anomaly identification model based on the dynamic interaction features and the space-time correlation to obtain a data anomaly identification result;

[0117] a contribution decomposition module 304 configured to perform feature contribution degree decomposition on the data anomaly identification result by using a pre-constructed explainable decision network to generate a feature contribution degree report;

[0118] a key factor module 305 configured to determine key influencing factors of data anomaly according to the feature contribution degree report;

[0119] a decision recommendation module 306 configured to correct a preset decision according to the key influencing factors, generate a decision basis report, and output the corrected decision and the decision basis report.

[0120] Further, the feature alignment module 301 specifically comprises:

[0121] a sliding processing unit configured to access multiple data sources through a distributed feature factory, and set a sliding time window for processing data streams in each data source;

[0122] a bias correction unit configured to perform time axis bias correction on the data streams to generate aligned data streams;

[0123] a cross-institution combination unit configured to perform cross-institution feature combination on the aligned data streams in each data source to generate dynamic interaction features, and store the dynamic interaction features in a data hub;

[0124] a fluctuation detection unit configured to detect abnormal fluctuations of the data streams;

[0125] The feature updating unit is configured to trigger a feature updating mechanism to regenerate the dynamic interaction feature if an abnormal fluctuation is detected in the data stream.

[0126] Further, the spatio-temporal correlation module 302 specifically comprises:

[0127] The attention weight unit is configured to generate a time feature weight and a space feature weight of the dynamic interaction feature respectively using an embedded attention mechanism of the dynamic deep learning model.

[0128] The weight correlation unit is configured to correlate the time feature weight and the space feature weight to construct a spatio-temporal correlation weight.

[0129] The weighted processing unit is configured to perform weighted processing on the dynamic interaction feature using the spatio-temporal correlation weight to obtain a weighted dynamic interaction feature.

[0130] The local convolution unit is configured to perform a local convolution operation on the weighted dynamic interaction feature and analyze the local convolution result to obtain the spatio-temporal correlation between the dynamic interaction features.

[0131] Further, the weight correlation unit specifically comprises:

[0132] The feature separation sub-unit is configured to perform feature separation on the dynamic interaction feature using a spatio-temporal separation layer of the dynamic deep learning model to obtain a time feature and a space feature.

[0133] The feature fusion sub-unit is configured to construct an initial feature combination matrix by the time feature and the space feature using a data fusion technology to obtain a preliminary spatio-temporal characteristic description.

[0134] The weight fusion sub-unit is configured to fuse the time feature weight and the space feature weight according to the preliminary spatio-temporal characteristic description using a correlation analysis method to determine a comprehensive weight distribution map.

[0135] The correlation weight acquisition sub-unit is configured to extract the spatio-temporal correlation weight from the comprehensive weight distribution map.

[0136] Further, the contribution decomposition module 304 specifically comprises:

[0137] The contribution degree calculation unit is configured to calculate a feature contribution degree of each dynamic interaction feature in the data anomaly identification result.

[0138] The contribution degree decomposition unit is configured to perform feature contribution degree decomposition on the feature contribution degree of each dynamic interaction feature using an explainable decision network to determine a contribution degree of a single data stream feature.

[0139] The contribution degree summary unit is configured to summarize the contribution degrees of all single data stream features to obtain a feature contribution degree report.

[0140] Further, the key factor module 305 specifically comprises:

[0141] A threshold judgment unit, configured to judge, in the feature contribution report, whether the contribution degree of each data stream feature exceeds a preset contribution degree threshold in sequence;

[0142] A super-threshold feature unit, configured to obtain data stream features exceeding the threshold to obtain a super-threshold feature set;

[0143] A feature correlation unit, configured to judge, in sequence, the correlation features between each super-threshold feature in the super-threshold feature set;

[0144] A key factor unit, configured to determine the super-threshold feature pair with strong correlation features as a key impact factor.

[0145] Further, the decision recommendation module 306 specifically comprises:

[0146] A natural language description unit, configured to import the key impact factor into a pre-trained natural language model to generate a natural language description of the predicted decision;

[0147] A semantic deviation unit, configured to calculate the semantic deviation between the natural language description of the predicted decision and the natural language description of the preset decision to obtain a decision semantic deviation;

[0148] A decision correction unit, configured to correct the preset decision based on the decision semantic deviation to generate a corrected decision;

[0149] A decision output unit, configured to associate the corrected decision and the dynamic interaction feature to generate a decision basis report and output the corrected decision and the decision basis report.

[0150] In the above embodiment, the application discloses a decision recommendation device, which belongs to the field of artificial intelligence and is applied to the vehicle insurance risk control decision scene. The application realizes real-time fusion and dynamic feature alignment of multi-source heterogeneous data by constructing a federal learning data hub, greatly improves the integrity and consistency of data, and breaks the bottleneck of traditional data silos and scattered storage. Using a pre-trained dynamic deep learning model, the system introduces a spatiotemporal attention mechanism and an incremental update engine, effectively capturing the spatiotemporal association of sensor data and business events, and realizing the improvement of cross-scene generalization ability and rapid response to market fluctuations. The interpretable decision network disassembles the complex model output into clear feature contribution, cooperates with key impact factor extraction and natural language decision correction, greatly improves the transparency of the decision and the understanding efficiency of the business personnel. The real-time recommendation system based on reinforcement learning realizes the millisecond-level processing of massive sensor data, guarantees the timeliness of the abnormal warning and the accuracy of the decision execution. The application effectively breaks the data silos, improves the dynamic adaptability of the model and the accuracy of the decision.

[0151] To solve the above technical problems, the embodiments of the present application further provide a computer device. For details, please refer to Figure 4 , Figure 4 The basic structure block diagram of the computer device of the present embodiment is shown in the figure.

[0152] The computer device 4 comprises a memory 41, a processor 42 and a network interface 43 which are connected to each other through a system bus. It should be noted that only the computer device 4 with the memory 41, the processor 42 and the network interface 43 is shown in the figure, but it should be understood that all the components shown are not required to be implemented, and more or less components can be alternatively implemented. Among them, the computer device herein is a device capable of automatically performing numerical calculation and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0153] The computer device can be a desktop computer, a notebook computer, a palm computer, a cloud server and other computing devices. The computer device can interact with the user through a keyboard, a mouse, a remote controller, a touchpad or a voice control device.

[0154] The memory 41 includes at least one type of readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory 41 can be an internal storage unit of the computer device 4, such as a hard disk or a memory of the computer device 4. In other embodiments, the memory 41 can also be an external storage device of the computer device 4, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the computer device 4. Of course, the memory 41 can also include both the internal storage unit and the external storage device of the computer device 4. In this embodiment, the memory 41 is generally used to store an operating system and various application software installed on the computer device 4, such as computer readable instructions of the decision recommendation method, etc. In addition, the memory 41 can also be used to temporarily store various data that have been output or will be output.

[0155] The processor 42 can be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip in some embodiments. The processor 42 is generally used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to run computer readable instructions or process data stored in the memory 41, such as computer readable instructions of the decision recommendation method.

[0156] The network interface 43 can include a wireless network interface or a wired network interface, and is generally used to establish a communication connection between the computer device 4 and other electronic devices.

[0157] The present application also provides an embodiment, i.e., to provide a computer device, which includes a memory and a processor, the memory stores computer readable instructions, and the processor executes the computer readable instructions to implement the steps of the decision recommendation method as described above, i.e., to implement:

[0158] A decision recommendation method includes:

[0159] Obtaining data streams from multiple source heterogeneous systems, and performing real-time feature alignment on the data streams through a preset distributed architecture to generate dynamic interaction features;

[0160] The pre-trained dynamic deep learning model is used to process the dynamic interaction features, and the spatio-temporal correlation between the dynamic interaction features is captured.

[0161] Based on the dynamic interaction features and the spatio-temporal correlation, a pre-trained anomaly recognition model is used to perform data anomaly recognition, and a data anomaly recognition result is obtained.

[0162] The pre-constructed explainable decision network is used to perform feature contribution degree decomposition on the data anomaly recognition result, and a feature contribution degree report is generated.

[0163] The key influencing factors causing the data anomaly are determined according to the feature contribution degree report.

[0164] The preset decision is corrected according to the key influencing factors, a decision basis report is generated, and the corrected decision and the decision basis report are output.

[0165] The application also provides another embodiment, that is, to provide a computer readable storage medium, the computer readable storage medium stores computer readable instructions, the computer readable instructions can be executed by at least one processor, so that the at least one processor executes the steps of the decision recommendation method as described above, that is, to realize:

[0166] A decision recommendation method comprises:

[0167] Obtaining data streams from multiple source heterogeneous systems, and performing real-time feature alignment on the data streams through a pre-set distributed architecture to generate dynamic interaction features;

[0168] The pre-trained dynamic deep learning model is used to process the dynamic interaction features, and the spatio-temporal correlation between the dynamic interaction features is captured.

[0169] Based on the dynamic interaction features and the spatio-temporal correlation, a pre-trained anomaly recognition model is used to perform data anomaly recognition, and a data anomaly recognition result is obtained.

[0170] The pre-constructed explainable decision network is used to perform feature contribution degree decomposition on the data anomaly recognition result, and a feature contribution degree report is generated.

[0171] The key influencing factors causing the data anomaly are determined according to the feature contribution degree report.

[0172] The preset decision is corrected according to the key influencing factors, a decision basis report is generated, and the corrected decision and the decision basis report are output.

[0173] Those skilled in the art can clearly understand that the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, it can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes a plurality of instructions for making a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device) execute the methods described in various embodiments of the present application.

[0174] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in a distributed computing environment in which tasks are performed by remote processing devices connected by a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0175] It should be noted that the non-company software tools or components appearing in various embodiments of the present application are only illustrative and do not represent actual use.

[0176] Obviously, the above-described embodiments are only part of the embodiments of the present application, and are not all the embodiments. The preferred embodiments of the present application are given in the drawings, but do not limit the patent scope of the present application. The present application can be realized in many different forms, and on the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing specific embodiments, or make equivalent replacements to some technical features. Any equivalent structure made by using the contents of the specification and drawings, directly or indirectly applied to other related technical fields, is also within the scope of the patent protection of the present application.

Claims

1. A decision recommendation method characterized by, The method comprises the following steps: acquiring data streams from a plurality of heterogeneous systems, and performing real-time feature alignment on the data streams through a preset distributed architecture to generate dynamic interaction features; processing the dynamic interaction features using a pre-trained dynamic deep learning model to capture the spatio-temporal correlation between the dynamic interaction features; based on the dynamic interaction features and the spatio-temporal correlation, performing data anomaly recognition using a pre-trained anomaly recognition model to obtain a data anomaly recognition result; using a pre-constructed explainable decision network to perform feature contribution degree decomposition on the data anomaly recognition result to generate a feature contribution degree report; determining key influencing factors of data anomaly according to the feature contribution degree report; correcting a preset decision according to the key influencing factors to generate a decision basis report, and outputting the corrected decision and the decision basis report.

2. The decision recommendation method of claim 1, wherein, The step of acquiring data streams from a plurality of heterogeneous systems and performing real-time feature alignment on the data streams through a preset distributed architecture to generate dynamic interaction features comprises the following steps: accessing a plurality of data sources through a distributed feature factory, setting a sliding time window for processing data streams in each data source; performing time axis deviation correction on the data streams to generate aligned data streams; combining cross-institutional features of the aligned data streams in each data source to generate dynamic interaction features, and storing the dynamic interaction features in a data hub; detecting abnormal fluctuations in the data streams; if abnormal fluctuations are detected in the data streams, triggering a feature update mechanism to regenerate the dynamic interaction features.

3. The decision recommendation method of claim 1, wherein, The step of processing the dynamic interaction features using a pre-trained dynamic deep learning model to capture the spatio-temporal correlation between the dynamic interaction features comprises the following steps: using an embedded attention mechanism of the dynamic deep learning model to generate time feature weights and space feature weights of the dynamic interaction features respectively; associating the time feature weights and the space feature weights to construct spatio-temporal correlation weights; using the spatio-temporal correlation weights to perform weighted processing on the dynamic interaction features to obtain weighted dynamic interaction features; performing local convolution operation on the weighted dynamic interaction features and analyzing the local convolution result to obtain the spatio-temporal correlation between the dynamic interaction features.

4. The decision recommendation method of claim 3, wherein, The step of associating the time feature weights and the space feature weights to construct spatio-temporal correlation weights comprises the following steps: using a spatio-temporal separation layer of the dynamic deep learning model to separate features of the dynamic interaction features to obtain time features and space features; using data fusion technology, constructing an initial feature combination matrix based on the time features and the space features to obtain a preliminary spatio-temporal characteristic description; based on the preliminary spatio-temporal characteristic description, using correlation analysis method to perform weight fusion on the time feature weights and the space feature weights to determine a comprehensive weight distribution map; extracting the spatio-temporal correlation weights from the comprehensive weight distribution map.

5. The decision recommendation method of claim 1, wherein, The step of using a pre-constructed explainable decision network to perform feature contribution degree decomposition on the data anomaly recognition result to generate a feature contribution degree report comprises the following steps: In the data anomaly identification result, the feature contribution degrees of the dynamic interaction features are calculated respectively; The feature contribution degree decomposition is performed on the feature contribution degrees of the dynamic interaction features using the interpretable decision network to determine the contribution degrees of single data stream features; The contribution degrees of all single data stream features are summarized to obtain the feature contribution degree report.

6. The decision recommendation method of claim 5, wherein, The step of determining the key influencing factors causing the data anomaly according to the feature contribution degree report specifically includes: In the feature contribution degree report, whether the contribution degree of each data stream feature exceeds a preset contribution degree threshold is judged in sequence; The data stream features exceeding the threshold are obtained to obtain a threshold-exceeding feature set; The correlation features between each threshold-exceeding feature in the threshold-exceeding feature set are judged in sequence; The threshold-exceeding feature pair with strong correlation features is determined as the key influencing factor.

7. The decision recommendation method of claim 1, wherein, The step of correcting the preset decision according to the key influencing factor, generating a decision basis report, and outputting the corrected decision and the decision basis report specifically includes: The key influencing factor is introduced into a pre-trained natural language model to generate a natural language description of the predicted decision; The semantic deviation between the natural language description of the predicted decision and a natural language description of the preset decision is calculated to obtain a decision semantic deviation; The preset decision is corrected based on the decision semantic deviation to generate the corrected decision; The corrected decision and the dynamic interaction features are associated to generate the decision basis report, and the corrected decision and the decision basis report are output.

8. A decision recommendation apparatus characterized by comprising: It includes: A feature alignment module is configured to obtain data streams from multiple source heterogeneous systems and perform real-time feature alignment on the data streams through a preset distributed architecture to generate dynamic interaction features; A space-time correlation module is configured to process the dynamic interaction features using a pre-trained dynamic deep learning model to capture the space-time correlation between the dynamic interaction features; An anomaly identification module is configured to perform data anomaly identification based on the dynamic interaction features and the space-time correlation using a pre-trained anomaly identification model to obtain a data anomaly identification result; A contribution decomposition module is configured to perform feature contribution degree decomposition on the data anomaly identification result using a pre-built interpretable decision network to generate a feature contribution degree report; A key factor module is configured to determine key influencing factors causing the data anomaly according to the feature contribution degree report; A decision recommendation module is configured to correct a preset decision according to the key influencing factors, generate a decision basis report, and output the corrected decision and the decision basis report.

9. A computer device, comprising: The computer readable storage medium stores computer readable instructions, and the computer readable instructions are executed by the processor to implement the steps of the decision recommendation method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer readable instructions, and the computer readable instructions are executed by the processor to implement the steps of the decision recommendation method according to any one of claims 1 to 7.

Images