Insurance fraud risk real-time assessment method based on dynamic Bayesian network

Abstract

The invention discloses an insurance fraud risk real-time evaluation method based on a dynamic Bayesian network. The method comprises the following steps: S1, collecting time sequence behavior data of an insurance user; s2, performing preprocessing to generate a structured observation sequence; s3, dividing into continuous time slices according to time, and constructing a time sequence sample; s4, constructing a dynamic Bayesian network; s5, calculating a posterior probability of a state variable in each time slice, and generating a sample weight; s6, executing an improved increment EM algorithm based on the sample weight, and optimizing dynamic Bayesian network parameters; s7, monitoring posterior probability entropy change of the state variable, and performing structure adjustment; and S8, the reasoning process is executed again, and a risk scoring report is generated. According to the method, the dynamic Bayesian network and the incremental EM algorithm are fused, the insurance fraud risk time sequence evaluation model is constructed, and the method has the advantages of being high in real-time performance, self-adaptive in structure, interpretable in result and the like.

Description

Technical Field

[0001] The present invention relates to the field of financial risk control, and in particular to a real-time insurance fraud risk assessment method based on a dynamic Bayesian network. Background Art

[0002] With the rapid development of financial technology, the insurance business is gradually moving towards a digital and intelligent operation model. A large amount of user behavior data, transaction data, and device data are collected and recorded in real time, providing data support for the insurance risk control system. However, insurance fraud has also become more covert, intelligent, and complex. The traditional detection method that relies on rule matching and static models has gradually exposed its shortcomings in recognition accuracy and response speed. New risk identification methods are urgently needed to conduct more accurate and dynamic modeling and analysis of the evolution of user behavior.

[0003] Currently, mainstream insurance fraud detection methods mostly use static supervised learning or unsupervised learning models, such as support vector machines, random forests, extreme gradient boosting trees, and deep neural networks. These methods usually structurally assume that samples are independent of each other and focus more on behavioral characteristics at the current moment. They lack the ability to model user behavior that changes over time. Therefore, such methods find it difficult to effectively capture the evolution of potential fraudulent behavior on the timeline, such as increased behavior frequency, sudden changes in operation paths, frequent equipment replacement, etc., which may be important clues that may be precursors to fraud.

[0004] To overcome these shortcomings, recent research has begun to introduce time series models and graphical models for fraud detection, typically using hidden Markov models and dynamic Bayesian networks. Dynamic Bayesian networks, by structurally introducing state dependencies between time steps, can effectively characterize the temporal evolution of a system. They are particularly well-suited for modeling the conditional relationships between state transitions in user behavior and observed variables. Consequently, dynamic Bayesian networks possess the ability to infer potential user behavior from sequence data and are gradually being introduced into the fields of financial risk control and anomaly detection. However, in practical deployment, traditional DBN modeling approaches still face numerous challenges.

[0005] First, most current DBN-based fraud detection solutions rely on offline batch training for parameter estimation, usually using the standard EM algorithm to iteratively solve the maximum likelihood estimate of the model on the complete dataset. This method cannot update model parameters in real time when facing dynamic streaming data, causing the model to lag behind data changes and making it difficult to capture the subtle evolution of user behavior in a timely manner. When user behavior mutates (such as identity switching, abnormal geographic location login, etc.), the model often cannot complete self-adjustment in a short period of time, seriously affecting the timeliness of risk response. In addition, the EM algorithm assumes that all samples contribute equally to model learning. In insurance fraud scenarios, the observed data is highly heterogeneous and uncertain, with a large number of missing values, disguised behaviors, and behavioral pseudo-heterogeneity, which can easily lead to the parameter estimation process being misled by erroneous samples, thereby reducing the model stability and generalization ability.

[0006] Secondly, the existing DBN model structure is generally a static design. Once the state space size and transfer structure are determined, they remain fixed after training is completed. In reality, insurance fraud behavior patterns are often dynamically expanded. Fraudulent behavior is no longer limited to patterns under traditional rules, but continues to evolve through new paths, equipment and means. The static state space is difficult to express this evolutionary trend, and it is very easy to have problems such as insufficient state expression granularity and incoherent transfer paths, making it difficult for the model to adapt to the ever-changing behavioral strategies in long-term operation. In addition, in traditional models, structural adjustments require expert knowledge for reconstruction or manual intervention, and lack of data-driven adaptive mechanisms.

[0007] Therefore, although some current studies have attempted to introduce probabilistic graphical models, time series learning and incremental update mechanisms into insurance risk control, there is still a lack of a complete solution with streaming modeling capabilities, dynamic structure scalability, real-time parameter updates and interpretable risk results. Traditional DBN and EM algorithm frameworks have failed to fully solve the practical bottlenecks in the real-time identification of insurance fraud scenarios, especially when faced with actual business needs such as rapid changes in observation data, dynamic evolution of behavioral strategies, and system processing of highly concurrent streaming data. The existing technology still has obvious deficiencies in timeliness, scalability and robustness. Summary of the Invention

[0008] One purpose of the present invention is to propose a real-time insurance fraud risk assessment method based on a dynamic Bayesian network. The present invention integrates a dynamic Bayesian network with an improved incremental EM algorithm to construct a time-series risk assessment model for insurance user behavior data, realizes state evolution modeling, online parameter optimization and structural adaptive expansion, and improves the recognition accuracy and response efficiency of highly concealed and slowly evolving fraudulent behaviors.

[0009] According to an embodiment of the present invention, a method for real-time insurance fraud risk assessment based on a dynamic Bayesian network includes the following steps:

[0010] S1. Collect the time series behavior data of insurance users and construct an observation sequence by sorting by timestamp;

[0011] S2. Preprocessing the observation sequence to generate a structured observation sequence;

[0012] S3. Setting the length and step size of the sliding time window, dividing the structured observation sequence into continuous time segments according to time, and constructing a time series sample;

[0013] S4. Define the temporal graph structure of the dynamic Bayesian network and construct the dynamic Bayesian network;

[0014] S5. Input the time series samples into a dynamic Bayesian network, calculate the posterior probability of the state variable in each time segment, and generate sample weights based on the uncertainty index of the posterior probability;

[0015] S6. Execute the improved incremental EM algorithm based on the sample weights to perform weighted updates on the state transition probability and observation conditional probability, and optimize the dynamic Bayesian network parameters;

[0016] S7, monitoring the change in the posterior probability entropy of the state variable. If the entropy change rate in the continuous time segment exceeds the preset threshold, the structural adjustment mechanism is triggered, and a new state variable node is inserted into the dynamic Bayesian network to reconstruct the state transition relationship and the corresponding observation conditional probability;

[0017] S8. Based on the optimized dynamic Bayesian network, re-execute the inference process, integrate the posterior probabilities of the state variables of the time segments, and generate a risk score report.

[0018] Optionally, the time series behavior data includes insurance policy application information, claim records, payment behavior, device usage information and geographic location information.

[0019] Optionally, the preprocessing includes time alignment, missing value filling, normalization and outlier removal.

[0020] Optionally, the S3 specifically includes:

[0021] S31. Set the sliding time window length and sliding step size, divide the structured observation sequence into windows according to the timestamp, and generate multiple continuous time segments. Each window corresponds to an observation segment of length L.

[0022] S32. For each sliding window, construct the corresponding time series sample:

[0023]

[0024] in, is the i-th time series sample, Indicates that at time step t i The structured observation sequence of L represents the length of the sliding time window. Indicates that at time step t i+L-1 A structured observation sequence of

[0025] S33. Calculate the number of all sliding time windows based on the total sequence length, window length, and step size, and construct a time series sample set:

[0026]

[0027] in, represents a set of time series samples, represents the Nth time series sample, N represents the number of sliding time windows, Y represents the total length of the time series sample, and S represents the sliding step size;

[0028] S34. Associating a corresponding timestamp sequence label with each time series sample;

[0029] S35. Input the constructed time series sample set into the dynamic Bayesian network as the observation sequence input under continuous time steps.

[0030] Optionally, the timing graph structure is represented as a directed graph, including a node set and an edge set, the node set including state variables and corresponding observation variables, and the edge set including time transfer edges between state variables and generation edges from state variables to observation variables.

[0031] Optionally, the S5 specifically includes:

[0032] S51. Input each set of time series samples into the dynamic Bayesian network, and use the initial state probability, state transition probability and observation conditional probability to perform the forward reasoning process and calculate the forward probability distribution of the state variable for each time step:

[0033]

[0034] Among them, α t (i) means that at time step t, the state is i and the observation sequence is O t1:t The forward probability, α t-1 (j) represents the forward probability of state j at time step t-1, A j,i =P(S t =i|S t-1 =j) represents the transition probability from state j to state i, B i (O t )=P(O t ∣S t =i) means generating observation variable O in state it The conditional probability of K represents the size of the state space, O t represents the observed variable at time step t;

[0035] S52, the posterior probability of the state variable at each time step Perform normalization to obtain the standardized posterior probability:

[0036]

[0037] in, Represents the observation sequence from time step t1 to t Under the condition, the state variable S t is the posterior probability of i, α t (i) means that at time step t, the state is i and the observation sequence is The forward probability of represents the sum of the forward probabilities of all states at time step t, represents the observation sequence from time step t1 to t;

[0038] S53. Calculate the uncertainty index based on the standardized posterior probability of each time step, and use the Shannon entropy function to define the uncertainty measure of the state variable:

[0039]

[0040] Among them, H(S t ) represents the state variable S at time step t t The uncertainty entropy value of Represents the observation sequence from time step t1 to t Under the condition, the state variable S t is the posterior probability of i;

[0041] S54. Generate sample weights corresponding to time series samples based on uncertainty entropy values:

[0042]

[0043] Among them, w t Represents the sample weight corresponding to the time series sample at time step t, H(S t ) represents the state variable S at time step t t The uncertainty entropy value of ∈ is represented by a positive constant to prevent the denominator from being zero.

[0044] Optionally, the S6 specifically includes:

[0045] S61. After receiving a new observation sample, the dynamic Bayesian network structure is kept unchanged, and the observation sequence and the corresponding sample weight sequence in the current sliding time window are extracted to perform incremental parameter update;

[0046] S62. In the expectation step of the improved incremental EM algorithm, the weighted joint posterior probability of state transition is calculated by combining the current dynamic Bayesian network parameter set and sample weights:

[0047] Q t (i,j)=w t ·P(S t-1 =i,S t =j|O 1:T ,Θ);

[0048] Among them, Q t (i, j) represents the weighted joint posterior probability of the state transitioning from i to j at time step t, w t represents the sample weight at time step t, P(S t-1 =i,S t =j|O 1:T ,Θ) means that under the parameter set Θ, given the observation sequence O 1:T The joint probability of state from i to j, O 1:T represents the observation sequence from time step 1 to time step T, Θ is the current set of dynamic Bayesian network parameters;

[0049] S63. In the maximization step, the state transition probability is updated using a weighted incremental update method based on the weighted joint posterior probability:

[0050]

[0051] in, represents the transition probability of state transition from i to j in the tth iteration, η represents the learning rate, and its value range is (0,1). represents the transition probability of state transition from i to j in the t-1th iteration, K represents the total number of state variables, Q t (i, j) represents the weighted joint posterior probability of the state transitioning from i to j at time step t, Q t (i,k) represents the weighted joint posterior probability of the state transitioning from i to k at time step t;

[0052] S64. Based on the current observation sequence and sample weights, update the observation conditional probability and perform weighted update on the observation variables in each state:

[0053]

[0054] in, represents the conditional probability that the observation value is o in state j in the tth iteration, η represents the learning rate, and its value range is (0,1). represents the conditional probability that the observation value is o in state j in the t-1th iteration, P(S t =j|O 1:T ,Θ) represents the posterior probability of state j at time step t, Indicator function indicating whether the observation value is equal to o, the value is 1 or 0, T represents the number of time steps;

[0055] S65. Write the updated state transition probability and observation condition probability back to the dynamic Bayesian network parameter set to form a new round of optimized network parameters.

[0056] Optionally, the S7 specifically includes:

[0057] S71. At each time step after completing the parameter update, obtaining the uncertainty entropy value of the state variable based on the current state posterior probability of the dynamic Bayesian network;

[0058] S72. Set the entropy monitoring sliding window length to L h , calculate the average rate of change of entropy:

[0059]

[0060] Among them, ΔH represents the average rate of change of entropy value in the sliding window, H t represents the uncertainty entropy value of time step t, H t-1 represents the uncertainty entropy value of time step t-1, L h represents the length of the entropy monitoring sliding window, |H t -H t-1 | represents the change amplitude of uncertainty entropy value in adjacent time steps;

[0061] S73. Setting an entropy change rate threshold. If the average change rate of the entropy value is greater than the entropy change rate threshold, triggering a structural adjustment mechanism.

[0062] S74, perform the structure expansion operation, add new state variables to the state set, and change the original state space from {S1, S2, ..., S K} is expanded to {S1,S2,...,S K ,S new}, and update the state transition probability and observation conditional probability;

[0063] S75. Expand the dimension of the state transition probability. After adding the new state, initialize the state transition probability of the newly added rows and columns:

[0064]

[0065] Among them, a new,j represents the probability of the new state transitioning to state j, a i,new represents the probability of state i transitioning to a new state, K represents the size of the state space, and K+1 represents the size of the expanded state space. It means "any";

[0066] S76. Expand the observation conditional probability and initialize the observation conditional probability for the newly added state.

[0067] Optionally, the S8 specifically includes:

[0068] S81. After the dynamic Bayesian network completes structural adjustment and parameter optimization, the forward reasoning algorithm is re-executed based on the updated model parameter set to calculate the posterior probability of the state variables at each time step, reflecting the evolution of the insurance user's fraud risk during the time series process;

[0069] S82. Combine the state posterior probabilities corresponding to each time step in chronological order to form a fraud probability trajectory:

[0070]

[0071] Among them, r t represents the maximum fraud probability value at time step t, represents the subset of states related to fraud, K ′ Indicates the size of the state space after structural adjustment, P(S t =i|O 1:T ,Θ ′ ) represents the corresponding state posterior probability, Θ ′ represents the updated set of dynamic Bayesian network parameters;

[0072] S83, combining the state transition path and the corresponding state transition probability in the dynamic Bayesian network to extract the state change chain P trans ={(S t ,S t+1 ,a ij )}, describing the transfer pattern of user risk change trajectory;

[0073] S84. For each time step observation variable vector, calculate the conditional probability relationship with the high-risk state variable and select the key observation variable set:

[0074]

[0075] in, represents the set of key observation variables, x j (t) represents the j-th observed variable value at time step t, represents the partial derivative of the posterior probability of the fraud state with respect to the observed variable, δ represents the sensitivity threshold, S fraud represents a high-risk state variable, P(S t =S fraud ∣O t ) represents the posterior probability of the fraud state;

[0076] S85. Calculate the risk level score based on the maximum fraud probability value:

[0077]

[0078] Among them, S risk represents the risk level score, r t represents the maximum fraud probability value at time step t, r j represents the maximum fraud probability value at time step j;

[0079] S86. Combine the above-mentioned fraud probability trajectory, state change chain, key observation variable set and final risk level score to generate a risk scoring report.

[0080] Optionally, the improved incremental EM algorithm gradually updates the dynamic Bayesian network parameters by introducing a sample weight mechanism based on state uncertainty and coupling it with a structure expansion mechanism based on the entropy change rate.

[0081] The beneficial effects of the present invention are:

[0082] First, the present invention proposes a real-time insurance fraud risk assessment method based on a dynamic Bayesian network. By constructing a time dependency relationship between state variables and observation variables, the temporal evolution process of insurance user behavior data is systematically modeled, breaking through the technical limitation of existing static models that cannot capture dynamic changes in behavior. Compared with traditional fraud detection methods that rely on batch learning, the present invention uses a dynamic Bayesian network to infer user status, and on this basis introduces an improved incremental EM algorithm driven by sample uncertainty to realize online incremental optimization and update of the model state transition probability and observation conditional probability, thereby significantly improving the real-time and adaptability of the model, and meeting the high concurrency and fast-changing behavior data processing scenarios in insurance business.

[0083] Secondly, the present invention constructs a sample weight mechanism by utilizing the Shannon entropy of the state posterior distribution, thereby solving the problem in the traditional EM algorithm that all sample update weights are equal and easily disturbed by abnormal data, and enhancing the model's recognition ability of high-confidence behaviors and robustness to abnormal samples. At the same time, the present invention designs a structural self-evolution mechanism with the state entropy change rate as a trigger condition. When it is monitored that the model structure is insufficient to express the behavior evolution, it automatically adds state nodes, dynamically adjusts the state transition structure and observation condition parameters, and realizes adaptive expansion of the model at the structural level, avoiding the expression bottleneck of the static graph structure when facing the evolution of complex fraud strategies, and improving the structural adaptability and recognition accuracy of the system in long-term operation.

[0084] Furthermore, the present invention is no longer limited to fraud judgment at a single moment. Instead, it generates a complete fraud probability trajectory by integrating the posterior probabilities of states at each time step. Combined with state transition paths, sensitivity analysis of key observed variables, and final risk level output, it constructs an interpretable risk score report, supporting insurance risk control systems in tracking and reviewing suspicious behavior throughout the entire process, meeting regulatory requirements for model auditability and interpretability. This holistic approach achieves comprehensive optimization with clear modeling structure, efficient parameter updates, scalable structure, and interpretable results, providing enhanced real-time decision support capabilities and business deployment value in a dynamic data environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0086] Figure 1 This is a flow chart of a real-time insurance fraud risk assessment method based on dynamic Bayesian networks proposed by the present invention;

[0087] Figure 2 This is a schematic diagram of the weighted update mechanism of the improved incremental EM algorithm for the real-time insurance fraud risk assessment method based on dynamic Bayesian networks proposed in the present invention, based on the state transition probability and observation conditional probability;

[0088] Figure 3 This is a schematic diagram of the model structure expansion mechanism based on the state entropy change rate triggering of a real-time insurance fraud risk assessment method based on a dynamic Bayesian network proposed in the present invention. DETAILED DESCRIPTION

[0089] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0090] refer to Figure 1-3A real-time insurance fraud risk assessment method based on a dynamic Bayesian network includes the following steps:

[0091] S1. Collect the time series behavior data of insurance users and construct an observation sequence by sorting by timestamp;

[0092] S2. Preprocessing the observation sequence to generate a structured observation sequence;

[0093] S3. Setting the length and step size of the sliding time window, dividing the structured observation sequence into continuous time segments according to time, and constructing a time series sample;

[0094] S4. Define the temporal graph structure of the dynamic Bayesian network and construct the dynamic Bayesian network;

[0095] S5. Input the time series samples into a dynamic Bayesian network, calculate the posterior probability of the state variable in each time segment, and generate sample weights based on the uncertainty index of the posterior probability;

[0096] S6. Execute the improved incremental EM algorithm based on the sample weights to perform weighted updates on the state transition probability and observation conditional probability, and optimize the dynamic Bayesian network parameters;

[0097] S7, monitoring the change in the posterior probability entropy of the state variable. If the entropy change rate in the continuous time segment exceeds the preset threshold, the structural adjustment mechanism is triggered, and a new state variable node is inserted into the dynamic Bayesian network to reconstruct the state transition relationship and the corresponding observation conditional probability;

[0098] S8. Based on the optimized dynamic Bayesian network, re-execute the inference process, integrate the posterior probabilities of the state variables of the time segments, and generate a risk score report.

[0099] The present invention constructs a dynamic Bayesian network model to model the temporal behavior data of insurance users. It combines the incremental EM optimization guided by sample weights with the adaptive expansion mechanism driven by structural entropy to achieve dynamic identification and risk scoring of insurance fraud behaviors. It has the advantages of real-time updating capability, structural scalability and strong score interpretability, which significantly improves the intelligence and accuracy of the fraud detection system.

[0100] In this embodiment, the time series behavior data includes insurance policy application information, claim settlement records, payment behavior, device usage information and geographic location information.

[0101] By introducing multi-source heterogeneous time-series behavioral data including policy applications, claims records, payment behaviors, equipment usage information and geographic location information, this invention effectively enriches the observation dimension of the model, provides a more comprehensive behavioral basis for fraud behavior modeling, and improves the recognition coverage and model discrimination ability of the dynamic Bayesian network for complex fraud behaviors.

[0102] In this embodiment, the preprocessing includes time alignment, missing value filling, normalization and outlier elimination.

[0103] The present invention adopts multiple preprocessing steps such as time alignment, missing value filling, normalization and anomaly elimination, which effectively improves the quality and consistency of time series data, reduces abnormal interference in observed variables, helps to enhance the stability and accuracy of subsequent modeling processes, and improves the reliability and robustness of fraud identification.

[0104] In this embodiment, S3 specifically includes:

[0105] S31. Set the sliding time window length and sliding step size, divide the structured observation sequence into windows according to the timestamp, and generate multiple continuous time segments. Each window corresponds to an observation segment of length L.

[0106] S32. For each sliding window, construct the corresponding time series sample:

[0107]

[0108] in, is the i-th time series sample, O ti Indicates that at time step t i The structured observation sequence of L represents the length of the sliding time window. Indicates that at time step t i+L-1 A structured observation sequence of

[0109] S33. Calculate the number of all sliding time windows based on the total sequence length, window length, and step size, and construct a time series sample set:

[0110]

[0111] in, represents a set of time series samples, represents the Nth time series sample, N represents the number of sliding time windows, Y represents the total length of the time series sample, and S represents the sliding step size;

[0112] S34. Associating a corresponding timestamp sequence label with each time series sample;

[0113] S35. Input the constructed time series sample set into the dynamic Bayesian network as the observation sequence input under continuous time steps.

[0114] By setting a sliding time window and step size, the present invention continuously segments the structured observation sequence and constructs a time series sample set, thereby retaining the temporal sequence information of the behavioral data, providing continuous input for the dynamic Bayesian network, enhancing the model's ability to model behavioral evolution trends, and improving the accuracy of temporal identification of fraudulent behavior.

[0115] By constructing a time series graph structure containing state variables and observation variables, the present invention clarifies the time transfer relationship between states and the generation relationship from states to observations, enabling the dynamic Bayesian network to have the ability to model the causal relationship between the evolution of behavioral states and observation characteristics, thereby improving the expressiveness and semantic rationality of the model structure.

[0116] In this embodiment, the timing graph structure is represented as a directed graph, including a node set and an edge set, the node set includes state variables and corresponding observation variables, and the edge set includes time transfer edges between state variables and generation edges from state variables to observation variables.

[0117] By constructing a time series graph structure containing state variables and observation variables, the present invention clarifies the time transfer relationship between states and the generation relationship from states to observations, enabling the dynamic Bayesian network to have the ability to model the causal relationship between the evolution of behavioral states and observation characteristics, thereby improving the expressiveness and semantic rationality of the model structure.

[0118] In this embodiment, the S5 specifically includes:

[0119] S51. Input each set of time series samples into the dynamic Bayesian network, and use the initial state probability, state transition probability and observation conditional probability to perform the forward reasoning process and calculate the forward probability distribution of the state variable for each time step:

[0120]

[0121] Among them, α t (i) means that at time step t, the state is i and the observation sequence is O t1:t The forward probability, α t-1 (j) represents the forward probability of state j at time step t-1, A j,i =P(S t =i|S t-1 =j) represents the transition probability from state j to state i, B i (O t )=P(O t ∣S t =i) means generating observation variable O in state i t The conditional probability of K represents the size of the state space, O t represents the observed variable at time step t;

[0122] S52, the posterior probability P(S) of the state variable at each time step t =i|O t1:t ) is normalized to obtain the standardized posterior probability:

[0123]

[0124] Among them, P(S t =i|O t1:t ) represents the observation sequence O from time step t1 to t t1:t Under the condition, the state variable S t is the posterior probability of i, α t (i) means that at time step t, the state is i and the observation sequence is O t1:t The forward probability of represents the sum of the forward probabilities of all states at time step t, O t1:t represents the observation sequence from time step t1 to t;

[0125] S53. Calculate the uncertainty index based on the standardized posterior probability of each time step, and use the Shannon entropy function to define the uncertainty measure of the state variable:

[0126]

[0127] Among them, H(S t ) represents the state variable S at time step t t The uncertainty entropy value of Represents the observation sequence from time step t1 to t Under the condition, the state variable S t is the posterior probability of i;

[0128] S54. Generate sample weights corresponding to time series samples based on uncertainty entropy values:

[0129]

[0130] Among them, w t Represents the sample weight corresponding to the time series sample at time step t, H(S t ) represents the state variable S at time step t t The uncertainty entropy value of ∈ is represented by a positive constant to prevent the denominator from being zero.

[0131] The present invention uses a forward reasoning algorithm to calculate the posterior probability of state variables, and combines it with the Shannon entropy function to evaluate state uncertainty and generate sample weights, so that the model can focus on behavioral segments with high information content and strong confidence during the update process, thereby improving the efficiency and accuracy of subsequent incremental learning and enhancing the model's sensitivity to changes in fraud risks.

[0132] In this embodiment, S6 specifically includes:

[0133] S61. After receiving a new observation sample, the dynamic Bayesian network structure is kept unchanged, and the observation sequence and the corresponding sample weight sequence in the current sliding time window are extracted to perform incremental parameter update;

[0134] S62. In the expectation step of the improved incremental EM algorithm, the weighted joint posterior probability of state transition is calculated by combining the current dynamic Bayesian network parameter set and sample weights:

[0135] Q t (i,j)=w t ·P(S t-1 =i,S t =j|O 1:T ,Θ);

[0136] Among them, Q t (i, j) represents the weighted joint posterior probability of the state transitioning from i to j at time step t, w t represents the sample weight at time step t, P(S t-1 =i,S t =j|O 1:T ,Θ) means that under the parameter set Θ, given the observation sequence O 1:T The joint probability of state from i to j, O 1:T represents the observation sequence from time step 1 to time step T, Θ is the current set of dynamic Bayesian network parameters;

[0137] S63. In the maximization step, the state transition probability is updated using a weighted incremental update method based on the weighted joint posterior probability:

[0138]

[0139] in, represents the transition probability of state transition from i to j in the tth iteration, η represents the learning rate, and its value range is (0,1). represents the transition probability of state transition from i to j in the t-1th iteration, K represents the total number of state variables, Q t (i, j) represents the weighted joint posterior probability of the state transitioning from i to j at time step t, Q t (i,k) represents the weighted joint posterior probability of the state transitioning from i to k at time step t;

[0140] S64. Based on the current observation sequence and sample weights, update the observation conditional probability and perform weighted update on the observation variables in each state:

[0141]

[0142] in, represents the conditional probability that the observation value is o in state j in the tth iteration, η represents the learning rate, and its value range is (0,1). represents the conditional probability that the observation value is o in state j in the t-1th iteration, P(S t =j|O 1:T ,Θ) represents the posterior probability of state j at time step t, Indicator function indicating whether the observation value is equal to o, the value is 1 or 0, T represents the number of time steps;

[0143] S65. Write the updated state transition probability and observation condition probability back to the dynamic Bayesian network parameter set to form a new round of optimized network parameters.

[0144] The present invention introduces an improved incremental EM algorithm based on sample weights to perform weighted updates on state transition probabilities and observation conditional probabilities, thereby achieving rapid adjustment of model parameters without retraining the entire model, effectively improving the model's adaptability and real-time performance to new behavior patterns, and at the same time improving the stability and anti-interference ability of model parameter estimation.

[0145] In this embodiment, the S7 specifically includes:

[0146] S71. At each time step after completing the parameter update, obtaining the uncertainty entropy value of the state variable based on the current state posterior probability of the dynamic Bayesian network;

[0147] S72. Set the entropy monitoring sliding window length to L h , calculate the average rate of change of entropy:

[0148]

[0149] Among them, ΔH represents the average rate of change of entropy value in the sliding window, H t represents the uncertainty entropy value of time step t, H t-1 represents the uncertainty entropy value of time step t-1, L h represents the length of the entropy monitoring sliding window, |H t -H t-1 | represents the change amplitude of uncertainty entropy value in adjacent time steps;

[0150] S73. Setting an entropy change rate threshold. If the average change rate of the entropy value is greater than the entropy change rate threshold, triggering a structural adjustment mechanism.

[0151] S74, perform the structure expansion operation, add new state variables to the state set, and change the original state space from {S1, S2, ..., S K} is expanded to {S1,S2,...,S K ,S new}, and update the state transition probability and observation conditional probability;

[0152] S75. Expand the dimension of the state transition probability. After adding the new state, initialize the state transition probability of the newly added rows and columns:

[0153]

[0154] Among them, a new,j represents the probability of the new state transitioning to state j, a i,new represents the probability of state i transitioning to a new state, K represents the size of the state space, and K+1 represents the size of the expanded state space. It means "any";

[0155] S76. Expand the observation conditional probability and initialize the observation conditional probability for the newly added state.

[0156] The present invention triggers a structural expansion mechanism based on the posterior entropy change rate, dynamically inserts new state nodes, and synchronously expands the state transition and observation probability structures, thereby realizing the adaptive evolution capability of the model structure, effectively solving the problem that static structures are difficult to cope with complex behaviors, and enhancing the scalability and expressiveness of the model under long-term operation.

[0157] In this embodiment, the S8 specifically includes:

[0158] S81. After the dynamic Bayesian network completes structural adjustment and parameter optimization, the forward reasoning algorithm is re-executed based on the updated model parameter set to calculate the posterior probability of the state variables at each time step, reflecting the evolution of the insurance user's fraud risk during the time series process;

[0159] S82. Combine the state posterior probabilities corresponding to each time step in chronological order to form a fraud probability trajectory:

[0160]

[0161] Among them, r t represents the maximum fraud probability value at time step t, represents the subset of states related to fraud, K ′ Indicates the size of the state space after structural adjustment, P(S t =i|O 1:T ,Θ ′ ) represents the corresponding state posterior probability, Θ ′ represents the updated set of dynamic Bayesian network parameters;

[0162] S83, combining the state transition path and the corresponding state transition probability in the dynamic Bayesian network to extract the state change chain P trans ={(S t ,S t+1 ,a ij )}, describing the transfer pattern of user risk change trajectory;

[0163] S84. For each time step observation variable vector, calculate the conditional probability relationship with the high-risk state variable and select the key observation variable set:

[0164]

[0165] in, represents the set of key observation variables, x j (t) represents the j-th observed variable value at time step t, represents the partial derivative of the posterior probability of the fraud state with respect to the observed variable, δ represents the sensitivity threshold, S fraud represents a high-risk state variable, P(S t =S fraud ∣O t ) represents the posterior probability of the fraud state;

[0166] S85. Calculate the risk level score based on the maximum fraud probability value:

[0167]

[0168] Among them, S risk represents the risk level score, r t represents the maximum fraud probability value at time step t, r j represents the maximum fraud probability value at time step j;

[0169] S86. Combine the above-mentioned fraud probability trajectory, state change chain, key observation variable set and final risk level score to generate a risk scoring report.

[0170] The present invention generates a risk scoring report by integrating state posterior probability, state transition path, sensitivity of key observation variables and risk level score, providing a traceable and explainable fraud risk output form for insurance risk control, supporting the dual-path application of manual review and automatic interception, and improving the business availability and compliance of fraud identification results.

[0171] In this embodiment, the improved incremental EM algorithm gradually updates the dynamic Bayesian network parameters by introducing a sample weight mechanism based on state uncertainty and coupling it with a structure expansion mechanism based on the entropy change rate.

[0172] By coupling the sample weight mechanism based on state uncertainty and the structural expansion mechanism based on the entropy change rate into the incremental EM algorithm, the present invention enables the model to have both high-confidence learning ability and structural adaptive adjustment ability, forming a parameter-structure dual-dimensional collaborative optimization path, and significantly improving the intelligence, self-evolution ability and practical value of the dynamic Bayesian network in fraud detection.

[0173] Example 1:

[0174] In order to verify the feasibility of the present invention in implementation, the present invention was applied to the online auto insurance claims system of a property insurance company in Nantong, Jiangsu Province. In recent years, the insurance company has continued to promote the digital claims process, and has connected automated processing modules to mobile insurance, online reporting, rapid damage assessment, intelligent review and other links. However, with the convenience of online claims processing, the company's risk pressure in claims fraud has continued to rise in recent years, especially from the second half of 2022 to the beginning of 2023. A large number of claims records with concentrated reporting behaviors, abnormal geographical locations, and frequent changes in equipment identification appeared in Nantong. The burden of manual review was heavy, and traditional rules and static scoring models were seriously insufficient to identify "gang fraud" that is highly concealed and slowly evolving.

[0175] In this actual scenario, the real-time insurance fraud risk assessment method based on dynamic Bayesian network proposed in the present invention is deployed in the online claims risk control system, and is connected to multi-dimensional behavioral data streams including policy application records, claim timestamps, payment methods, device identification numbers, and user geographic location change trajectories. The system monitors user behavior at a minute-level collection frequency, processes continuous behavioral data in a sliding time window manner, and constructs a dynamic Bayesian network model to infer user status. Compared with the original static scoring model based on XGBoost, the dynamic model adopted by the present invention can identify the evolution path of user status over time, and provide timely warnings for certain suspicious behaviors that "slowly migrate from normal status to high risk."

[0176] During the deployment process, the system conducted an offline replay of 9,138 claims records from the past three months as a backtest set, and used subsequent real manual review results as control labels to compare the recognition capabilities of the model of the present invention with those of the traditional model. During the experiment, the model of the present invention introduced a sample weight mechanism based on posterior probability entropy, which made it possible to pay more attention to key time windows with high information density and strong reasoning credibility during the training phase, and effectively filtered out some noise samples caused by device switching. At the same time, when the model detected that the entropy of the state sequence of some behavioral users fluctuated violently, the structure expansion mechanism was triggered, new state nodes were dynamically inserted, and the state transfer path was optimized, so that the model has a stronger representation capability for complex and diverse behavioral structures.

[0177] In addition, the synergistic effect of the incremental EM algorithm and the structural expansion mechanism is recorded during the model operation. When the system continues to receive new data, there is no need to retrain the entire model. Only the parameter matrix needs to be updated to adapt to the new risk situation.

[0178] Table 1 Performance comparison of the proposed method and the traditional method on the backtest set

[0179]

[0180] In terms of fraud identification accuracy, the model of the present invention reached 91.6%, an improvement of 13.5% over traditional methods, fully demonstrating the advantages of dynamic modeling in identifying behavioral sequence evolution. Secondly, the false alarm rate dropped significantly from 12.3% of the original method to 4.7%, reducing the manual review costs and compliance interference caused by a large number of misjudgments, and improving the practicality and trustworthiness of the model.

[0181] In terms of response efficiency, the model of the present invention shortens the average risk response time to 4.8 minutes through real-time sliding window input and incremental parameter update mechanism, compared with 28.6 minutes of the traditional model, effectively meeting the business demand for "real-time intervention" in insurance fraud identification. In addition, in terms of the early warning rate of suspicious samples, the method of the present invention reached 82.4%, which is almost more than twice the 39.7% of the traditional method, indicating that the model can capture the status mutation of potential risk users earlier, and gain valuable risk disposal time for the risk control department.

[0182] Regarding model update strategies, this invention utilizes a collaborative optimization process combining an incremental EM algorithm with a state structure expansion mechanism, enabling real-time online model updates and avoiding the lag inherent in traditional methods that rely on weekly batch retraining. This adaptive optimization approach not only enhances long-term model adaptability but also significantly reduces operational costs, making it suitable for long-term deployment and operation in dynamic, highly concurrent insurance business systems.

[0183] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A real-time insurance fraud risk assessment method based on dynamic Bayesian networks, characterized in that: The steps include: S1. Collect the time series behavior data of insurance users and construct an observation sequence by sorting by timestamp; S2. Preprocessing the observation sequence to generate a structured observation sequence; S3. Setting the length and step size of the sliding time window, dividing the structured observation sequence into continuous time segments according to time, and constructing a time series sample; S4. Define the temporal graph structure of the dynamic Bayesian network and construct the dynamic Bayesian network; S5. Input the time series samples into a dynamic Bayesian network, calculate the posterior probability of the state variable in each time segment, and generate sample weights based on the uncertainty index of the posterior probability; S6. Execute the improved incremental EM algorithm based on the sample weights to perform weighted updates on the state transition probability and observation conditional probability, and optimize the dynamic Bayesian network parameters; S7, monitoring the change in the posterior probability entropy of the state variable. If the entropy change rate in the continuous time segment exceeds the preset threshold, the structural adjustment mechanism is triggered, and a new state variable node is inserted into the dynamic Bayesian network to reconstruct the state transition relationship and the corresponding observation conditional probability; S8. Based on the optimized dynamic Bayesian network, re-execute the inference process, integrate the posterior probabilities of the state variables of the time segments, and generate a risk score report.

2. The real-time insurance fraud risk assessment method based on dynamic Bayesian network according to claim 1 is characterized in that: The time series behavior data includes insurance policy application information, claim records, payment behavior, device usage information and geographic location information.

3. The real-time insurance fraud risk assessment method based on dynamic Bayesian network according to claim 1 is characterized in that: The preprocessing includes time alignment, missing value filling, normalization and outlier removal.

4. The real-time insurance fraud risk assessment method based on dynamic Bayesian network according to claim 1 is characterized in that: The S3 specifically includes: S31. Set the sliding time window length and sliding step size, divide the structured observation sequence into windows according to the timestamp, and generate multiple continuous time segments. Each window corresponds to an observation segment of length L. S32. For each sliding window, construct a corresponding time series sample; S33. Calculate the number of all sliding time windows according to the total sequence length, window length, and step size, and construct a time series sample set. S34. Associating a corresponding timestamp sequence label with each time series sample; S35. Input the constructed time series sample set into the dynamic Bayesian network as the observation sequence input under continuous time steps.

5. The real-time insurance fraud risk assessment method based on dynamic Bayesian network according to claim 1 is characterized in that: The timing graph structure is represented as a directed graph, including a node set and an edge set. The node set includes state variables and corresponding observation variables. The edge set includes time transfer edges between state variables and generation edges from state variables to observation variables.

6. The method for real-time insurance fraud risk assessment based on dynamic Bayesian network according to claim 1, characterized in that: The S5 specifically includes: S51. Input each set of time series samples into the dynamic Bayesian network, and use the initial state probability, state transition probability and observation conditional probability to perform the forward reasoning process and calculate the forward probability distribution of the state variable for each time step: Among them, α t (i) means that at time step t, the state is i and the observation sequence is O t1:t The forward probability, α t-1 (j) represents the forward probability of state j at time step t-1, A j,i =P(S t =i|S t-1 =j) represents the transition probability from state j to state i, B i (O t )=P(O t ∣S t =i) means generating observation variable O in state i t The conditional probability of K represents the size of the state space, O t represents the observed variable at time step t; S52, normalizing the posterior probability of the state variable at each time step to obtain a standardized posterior probability; S53. Calculate uncertainty indicators based on the standardized posterior probability of each time step, and use the Shannon entropy function to define the uncertainty measure of the state variable; S54. Generate sample weights corresponding to time series samples based on uncertainty entropy values: Among them, w t Represents the sample weight corresponding to the time series sample at time step t, H(S t ) represents the state variable S at time step t t The uncertainty entropy value of ∈ is represented by a positive constant to prevent the denominator from being zero.

7. The method for real-time insurance fraud risk assessment based on dynamic Bayesian network according to claim 1, characterized in that: The S6 specifically includes: S61. After receiving a new observation sample, the dynamic Bayesian network structure is kept unchanged, and the observation sequence and the corresponding sample weight sequence in the current sliding time window are extracted to perform incremental parameter update; S62. In the expectation step of the improved incremental EM algorithm, the weighted joint posterior probability of the state transition is calculated by combining the current dynamic Bayesian network parameter set and the sample weights; S63. In the maximization step, the state transition probability is updated using a weighted incremental update method based on the weighted joint posterior probability: in, represents the transition probability of state transition from i to j in the tth iteration, η represents the learning rate, and its value range is (0,1). represents the transition probability of state transition from i to j in the t-1th iteration, K represents the total number of state variables, Q t (i, j) represents the weighted joint posterior probability of the state transitioning from i to j at time step t, Q t (i,k) represents the weighted joint posterior probability of the state transitioning from i to k at time step t; S64. Based on the current observation sequence and sample weights, update the observation conditional probability and perform weighted update on the observation variables in each state: in, represents the conditional probability that the observation value is o in state j in the tth iteration, η represents the learning rate, and its value range is (0,1). represents the conditional probability that the observation value is o in state j in the t-1th iteration, P(S t =j|O 1:T ,Θ) represents the posterior probability of state j at time step t, Indicator function indicating whether the observation value is equal to o, the value is 1 or 0, T represents the number of time steps; S65. Write the updated state transition probability and observation condition probability back to the dynamic Bayesian network parameter set to form a new round of optimized network parameters.

8. The method for real-time insurance fraud risk assessment based on dynamic Bayesian network according to claim 1, characterized in that: The S7 specifically includes: S71. At each time step after completing the parameter update, obtaining the uncertainty entropy value of the state variable based on the current state posterior probability of the dynamic Bayesian network; S72. Set the entropy monitoring sliding window length to L h , calculate the average rate of change of entropy value; S73. Setting an entropy change rate threshold. If the average change rate of the entropy value is greater than the entropy change rate threshold, triggering a structural adjustment mechanism. S74, perform the structure expansion operation, add new state variables to the state set, and change the original state space from {S1, S2, ..., S K } is expanded to {S1,S2,...,S K ,S new }, and update the state transition probability and observation conditional probability; S75. Expand the dimension of the state transition probability, add the new state, and initialize the state transition probability of the newly added rows and columns; S76. Expand the observation conditional probability and initialize the observation conditional probability for the newly added state.

9. The method for real-time insurance fraud risk assessment based on dynamic Bayesian network according to claim 1, characterized in that: The S8 specifically includes: S81. After the dynamic Bayesian network completes structural adjustment and parameter optimization, the forward reasoning algorithm is re-executed based on the updated model parameter set to calculate the posterior probability of the state variables at each time step, reflecting the evolution of the insurance user's fraud risk during the time series process; S82. Combining the state posterior probabilities corresponding to each time step in chronological order to form a fraud probability trajectory; S83. Combining the state transition path and the corresponding state transition probability in the dynamic Bayesian network, extract the state change chain and describe the transition pattern of the user risk change trajectory; S84. For each time step observation variable vector, calculate the conditional probability relationship with the high-risk state variable and screen the key observation variable set; S85. Calculate the risk level score based on the maximum fraud probability value: Among them, S risk represents the risk level score, r t represents the maximum fraud probability value at time step t, r j represents the maximum fraud probability value at time step j; S86. Combine the above-mentioned fraud probability trajectory, state change chain, key observation variable set and final risk level score to generate a risk scoring report.

10. The real-time insurance fraud risk assessment method based on dynamic Bayesian network according to claim 1 is characterized in that: The improved incremental EM algorithm gradually updates the dynamic Bayesian network parameters by introducing a sample weight mechanism based on state uncertainty and coupling it with a structure expansion mechanism based on the entropy change rate.

Images