Event processing path prediction method, system and device based on time sequence feature fusion

Abstract

The application discloses a matter processing path prediction method, system and device based on time sequence feature fusion, which comprises the following steps: first, a time observation window is delimited, five categories of time sequence data, i.e., state evolution, interactive feedback, carrier attribute, external environment and event object subject behavior, are extracted, and the time sequence data is preprocessed; then, a fusion time sequence representation vector is generated through a time sequence encoder, a cross-factor attention and a gating mechanism. The whole-process processing path is constructed into a path atlas in the form of a directed acyclic graph, and a dynamic prediction atlas jump parameter is relied on an event comprehensive representation. The candidate paths are screened in combination with the fusion representation, a path value network and a Monte Carlo tree search, and the optimal path is determined through feasibility verification, sensitivity analysis and stress testing. In the execution process, the data is monitored in real time, and the model parameters are automatically re-planned and iterated in an abnormal situation. The application can mine the dynamic correlation of multiple factors in an event, realize global path optimization of the whole process of the event, and improve the rationality and adaptability of path planning.

Description

Technical Field

[0001] This application belongs to the field of data processing and process optimization technology, specifically relating to a processing path prediction method, system, device and medium based on time series feature fusion. Background Technology

[0002] In complex event handling processes, once an event enters multiple stages, various processing paths are available (including but not limited to: directly initiating standard procedures → adjudication → execution; initiating procedures → negotiation during procedures → execution; applying for a urging order → execution; preliminary preservation of assets → initiating procedures later → execution; initiating procedures → adjudication → bankruptcy filing; arbitration → execution; handling multiple parties separately → consolidated execution, etc.). Different paths differ significantly in terms of processing efficiency, resource consumption, and final outcomes. Current path selection faces the following challenges: Event states have dynamic evolutionary characteristics. The responsiveness, willingness to cooperate, and state attributes of related event objects change continuously over time. Static features cannot capture the impact of these evolutionary trends on path selection. There are complex interactions among different time-series factors (for example, the combination of a continuous increase in the cumulative delay value and a continuous decrease in the response rate may indicate that the subject will adopt a confrontational strategy, while a sudden change in attributes may indicate a shift in behavior). Individual analysis of each factor cannot reveal these interaction patterns. The processing path is a multi-stage sequential decision-making process. The choices made in the preceding stages will change the available paths and expected results in the subsequent stages. Therefore, it is necessary to perform global joint optimization of the path rather than making greedy decisions stage by stage. Historical case data may contain selection biases (processors tend to choose specific paths for specific types of cases), and learning directly from historical data may lead to systematic biases in path recommendations.

[0003] Therefore, there is an urgent need for a method that can integrate multi-dimensional temporal features, capture the dynamic interaction between factors, and perform global optimization prediction of multi-stage paths. Summary of the Invention

[0004] The technical problem to be solved by this application is to provide a processing path prediction method, system, device and medium based on time series feature fusion, so as to solve the problems in the prior art that cannot capture the dynamic evolution trend of event objects and the interaction relationship between factors due to the use of only static features or single factor analysis, cannot achieve global path optimization due to the adoption of stage-by-stage greedy decision-making, and cause systematic deviation in path recommendation due to historical data selection bias.

[0005] To address the aforementioned technical problems, this application provides the following technical solution: Firstly, this application provides an event processing path prediction method based on temporal feature fusion, including: The time series of a multi-factor group of events to be processed is obtained, and the time series of the multi-factor group is preprocessed to obtain the preprocessed time series of the multi-factor group. The time series of the multi-factor group includes: delayed evolution time series, response change time series, state attribute change time series, environmental dynamic time series and event object subject behavior pattern time series. The temporal evolution representations of each dimension are extracted based on the temporal encoder within each factor group, and the dynamic interaction relationships between different factor groups are captured through a cross-factor attention mechanism to generate a fused temporal representation vector. The processing path is modeled to form a directed acyclic graph. Nodes in the directed acyclic graph represent processing stage states, and directed edges represent legal state transitions. The graph is also accompanied by parameterized predictions of transition duration, transition resource consumption, and transition success probability based on the representation parameters of the events to be processed. Based on the fused temporal representation vector, and combined with the path value network and Monte Carlo tree search algorithm, candidate processing paths are searched in the directed acyclic graph to form a candidate path set; Feasibility verification and sensitivity analysis are performed on each candidate path in the candidate path set to obtain the globally optimal path.

[0006] Furthermore, the method also includes: The process is executed according to the global optimal path, and the data is monitored and processed in real time. If the processed data reaches the processing path re-evaluation trigger condition, the processing path is re-evaluated. The path re-evaluation trigger condition includes: the cumulative deviation exceeds the threshold, the transfer step fails, and the change in object representation exceeds the threshold. When any of the conditions is met, the temporal feature fusion step and subsequent steps are re-executed from the current state.

[0007] Furthermore, the preprocessing includes: missing value imputation, sequence normalization, and trend feature enhancement.

[0008] Furthermore, the step of extracting temporal evolution representations of each dimension based on the temporal encoder within each factor group, and capturing the dynamic interaction relationships between different factor groups through a cross-factor attention mechanism to generate a fused temporal representation vector includes: The temporal features of each factor group are encoded using a Transformer encoder with causal masking to extract the temporal evolution representation within the factor. The representations of each factor group are combined into a factor sequence, and the dynamic interaction weights between factor groups are calculated through a cross-factor attention mechanism. Then, the weights are adaptively weighted through a gated fusion layer to generate the final fused temporal representation vector.

[0009] Furthermore, the processing path is modeled to form a directed acyclic graph (DAG), where nodes represent processing stage states, directed edges represent valid state transitions, and the graph includes parameterized predictions of transition duration, transition resource consumption, and transition success probability based on the events to be processed. The processing path is modeled to form a directed acyclic graph; Define multiple legal transition paths from the initial state to the final state achieved through disposal; For each directed edge in the directed acyclic graph, a transition attribute prediction network is established, and the transition duration, transition cost, and transition success probability are used as variables that are dynamically parameterized by the comprehensive characterization of the object to be transferred.

[0010] Further, the step of searching for the globally optimal processing path in the directed acyclic graph based on the fused temporal representation vector and combining the path value network with the Monte Carlo tree search algorithm includes: Construct a path value function and generate a path value network based on the path value function; Based on the path value network, evaluate the remaining expected value of each state node in the directed acyclic graph; Based on the remaining expected value, the Monte Carlo tree search algorithm is used to explore paths in the directed acyclic graph. In the selection phase, the success probability of transition is used as a priori guidance, and in the simulation phase, the output of the path value network is used to replace random simulation. Candidate paths are selected based on search efficiency and path value, and a candidate path set is formed.

[0011] Furthermore, the step of performing feasibility verification and sensitivity analysis on each candidate path in the candidate path set to obtain the globally optimal path includes: Each candidate path in the candidate path is verified for legal and logical feasibility, and paths that do not meet the transfer conditions are eliminated; Calculate the sensitivity of each candidate path to changes in each time series factor, and obtain the stability of the candidate path based on the sensitivity. The resilience index of the candidate path is calculated by conducting stress tests under adverse stress scenarios. By combining the path value, stability, and resilience metrics of the candidate paths, the globally optimal path is obtained.

[0012] Secondly, this application also provides an event processing path prediction system based on temporal feature fusion, comprising: The data acquisition module is used to acquire the time series of a multi-factor group of events to be processed, and to preprocess the time series of the multi-factor group to obtain the preprocessed time series of the multi-factor group. The time series of the multi-factor group includes: delayed evolution time series, response change time series, state attribute change time series, environmental dynamic time series, and event object subject behavior pattern time series. The feature extraction and fusion module is used to extract temporal evolution representations of each dimension based on the temporal encoder within each factor group, and to capture the dynamic interaction relationships between different factor groups through a cross-factor attention mechanism to generate a fused temporal representation vector. The processing path construction module is used to model the processing path and form a directed acyclic graph. The nodes in the directed acyclic graph represent the processing stage state, the directed edges represent the legal state transitions, and are accompanied by the transition duration, transition resource consumption and transition success probability predicted by the representation parameterization of the event to be processed. The filtering module is used to search for candidate processing paths in the directed acyclic graph based on the fused temporal representation vector and in combination with the path value network and Monte Carlo tree search algorithm, to form a candidate path set; The analysis module is used to perform feasibility verification and sensitivity analysis on each candidate path in the candidate path set to obtain the globally optimal path.

[0013] Thirdly, this application also provides a computer electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the event processing path prediction method based on temporal feature fusion as described in any one of the above.

[0014] Fourthly, this application also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the event processing path prediction method based on temporal feature fusion as described in any one of the above.

[0015] This application provides an event processing path prediction method, system, device, and medium based on temporal feature fusion, the advantages of which are: This application extracts the time series of events across five dimensions: delay evolution, response change, state attribute variation, environmental dynamics, and event subject behavior patterns, and performs trend enhancement. It automatically captures dynamic interactions between factors using intra-factor temporal coding and cross-factor attention mechanisms to generate a fused temporal representation. Simultaneously, it models the processing path as a directed acyclic graph, dynamically parameterizing the transition time, resource consumption, and success probability of each edge based on the event representation. Furthermore, it combines Monte Carlo tree search and path value network to achieve globally optimal path search, overcoming the limitations of static features and single-factor analysis, as well as the defects of stage-by-stage greedy decision-making. Further, it reduces the impact of historical data selection bias through feasibility verification, gradient-based sensitivity analysis, and multi-scenario stress testing. During execution, it triggers path reassessment and online model updates based on actual deviation monitoring, transition failures, or significant changes in event representation, forming a closed loop of prediction-execution-observation-correction. This significantly improves the dynamic adaptability, global optimization capability, and system robustness of complex transaction processing path prediction. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating an event processing path prediction method based on temporal feature fusion in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of an event processing path prediction system based on temporal feature fusion in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer electronic device according to an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Please see Figure 1This application provides an event processing path prediction method based on temporal feature fusion. This method relies on multi-dimensional temporal data mining, feature fusion, graph structure modeling, and intelligent search algorithms to achieve automated, global, and dynamic planning of event processing paths. It can adapt to state changes throughout the entire event lifecycle, improving the rationality and reliability of path selection. The method includes at least the following execution steps: S10. Obtain the time series of the multi-factor group of the event to be processed, and preprocess the time series of the multi-factor group to obtain the preprocessed time series of the multi-factor group, wherein the time series of the multi-factor group includes: delayed evolution time series, response change time series, state attribute change time series, environmental dynamic time series and event object subject behavior pattern time series.

[0020] It should be noted that the "event object subject" refers to the core entity of the pending matter targeted by the technical solution of this application, that is, the object of the entire prediction and processing flow. Specifically, it can be a matter entity with delayed performance characteristics that requires the selection of the optimal processing path, and its state is dynamically characterized by multi-dimensional temporal factors (such as delay evolution, response changes, state attribute changes, environmental dynamics, and subject behavior patterns). By extracting the temporal characteristics of this subject and fusing them to generate a comprehensive representation, the system can predict the most suitable processing path for it in a personalized manner, and dynamically adjust the strategy according to the changes in the subject's state during execution.

[0021] Specifically, this application first obtains time series data categorized into five major factor groups, comprehensively covering all dimensions of dynamic information about the event. These include: delayed evolution time series, response change time series, state attribute change time series, environmental dynamics time series, and event subject behavior pattern time series. These five types of time series data fully record the event development trend, feedback status, entity state changes, external environmental fluctuations, and related event subject behavior changes, providing raw data support for subsequent feature analysis and path planning. Details are as follows: For the One pending event, within the time window Inside (of which) For the current decision-making moment, The preset number of historical observation steps, (For the preset time sampling interval), extract the time series in the following dimensions: Factor group F1 – Delayed evolution time series (time series for short) (denoted as F1) The sequence includes: a cumulative delay base sequence (the unprocessed delay base balance at each time step); a cumulative delay add-on sequence (the cumulative delay add-on value at each time step); an increasing delay days sequence (the delay days corresponding to each time step); a new delay amount sequence (the incremental delay amount added at each time step); and a delay amount as a percentage of the initial baseline sequence. denoted as the number of feature sub-dimensions of factor group F1.

[0022] Factor group F2 – Response change time series (denoted as F2) This includes: a response connection rate sequence (the success rate of contact reaching the target within each time step); a commitment fulfillment rate sequence (the actual percentage of commitments fulfilled by the target within each time step); a response latency sequence (the average latency from contact to response within each time step); a proactive contact frequency sequence (the number of times the target proactively contacts the initiator within each time step); a communication attitude rating sequence (the rating of the target's communication attitude by the processing personnel within each time step); and a clear refusal intensity sequence (the intensity rating of the target's clear refusal to fulfill commitments within each time step). is the number of feature sub-dimensions of factor group F2.

[0023] Factor group F3 – Time series of state attribute changes (denoted as) This includes: a sequence of changes in fund account balances (the change in account balance at each time step); a sequence of events indicating whether a property transfer, mortgage establishment, or cancellation event has occurred at each time step; a sequence of events indicating whether a vehicle transfer or new registration event has occurred at each time step; a sequence of events indicating whether a change in equity, legal representative, or registered capital has occurred at each time step; a sequence of signals indicating new obligations (indicators of whether new loan records or new enforcement information have appeared at each time step); and a sequence of changes in regular income (the trend value of salary or operating income changes at each time step). is the number of feature sub-dimensions of factor group F3.

[0024] Factor group F4 – Environmental dynamics time series (denoted as F4) This includes: a sequence of backlogged events by jurisdiction (changes in the number of unresolved events by the agency at each time step); a sequence of average processing times for similar events (average processing time for similar events at each time step); a sequence of changes in regional achievement rates (statistical values ​​of achievement rates for each region at each time step); a sequence of completion rates (completion rates for each region at each time step); and a sequence of rule and policy change signals (indicators of whether new policies or rules affecting the processing flow have been introduced at each time step). is the number of feature sub-dimensions of factor group F4.

[0025] Factor group F5 – Event object subject behavior pattern time sequence (denoted as This includes: a residence change signal sequence (whether a change of residence address is detected at each time step); a communication method change sequence (whether a change or deactivation of a communication number occurs at each time step); a social network activity sequence (changes in activity on public social media platforms at each time step); an abnormal consumption behavior sequence (whether abnormal consumption behavior that is inconsistent with the ability to perform obligations occurs at each time step); and a related party behavior change sequence (signals of behavioral changes by guarantors or closely related parties at each time step). is the number of feature sub-dimensions of factor group F5.

[0026] Secondly, the following preprocessing was performed on the original time series of each factor group: Missing value imputation: For missing time steps caused by inconsistent sampling frequency or missing data sources, a combination of forward imputation and linear interpolation is used. ; in, and These are the most recent forward and backward effective observation time steps, respectively. These are the linear interpolation coefficients.

[0027] Sequence normalization: Sliding window Z-score standardization is applied to the time series for each factor dimension. ; in, and They are respectively based on time steps Within the local window centered The mean and standard deviation of the dimensional features. This is a preset numerical stability constant; in one embodiment, Values .

[0028] Enhanced trend features: First and second differences are additionally calculated for each time series as trend and acceleration features. ; ; Ultimately, the enhanced time series representation of each factor group is a concatenation of the original normalized sequence and the differencing sequence: ; The output of step S10 is a set of multidimensional enhanced temporal features for all events: ; The multidimensional enhanced temporal feature set is used as the input for step S20.

[0029] S20. Extract the temporal evolution representation of each dimension based on the temporal encoder within each factor group, and capture the dynamic interaction relationship between different factor groups through a cross-factor attention mechanism to generate a fused temporal representation vector.

[0030] Specifically, for each factor group, a dedicated temporal encoder is invoked to extract the temporal evolution representation under the corresponding dimension, accurately capturing the inherent patterns of change in a single data set over time. Based on this, a cross-factor attention mechanism is introduced to deeply explore the dynamic relationships and interaction logic between different factor groups, quantifying the degree of mutual influence among factors. Combining factor representations and interaction weights, feature fusion is completed, generating a fused temporal representation vector that fully represents the overall state of the event to be processed. This vector condenses all temporal information and serves as a unified feature input for subsequent path modeling, value assessment, and path search.

[0031] In one embodiment of this application, step S20 includes: S21. Encode the temporal features of each factor group using a Transformer encoder with causal masking to extract the temporal evolution representation within the factor.

[0032] S22. The representations of each factor group are combined into a factor sequence, and the dynamic interaction weights between factor groups are calculated through a cross-factor attention mechanism. Then, the weights are adaptively weighted through a gated fusion layer to generate the final fused temporal representation vector.

[0033] Specifically, first, intra-factor temporal coding is performed: For each factor group First, the temporal evolution representation of the factor is extracted using an intra-factor temporal encoder. A temporal Transformer encoder with positional encoding is employed. Location coding: ; in, For time step index, For dimensional indexing, Hide dimensions for the preset model; In one embodiment, The value is 128.

[0034] Input projection: ; in, For factor groups The input projection matrix.

[0035] Multi-head self-attention computation (the first) layer): ; ; in, The first Layer query, key-value projection matrix; : Dimensions for each attention head The preset number of attention heads; in one embodiment, The value is 8; It is a causal mask matrix that ensures that each time step only focuses on the current and previous time steps.

[0036] go through After layer encoding, the hidden state at the last time step is taken as the factor set. Temporal representation vector: ; In one embodiment, The value is 3.

[0037] Secondly, cross-factor temporal attention fusion is performed. The temporal representation vectors of the five factor groups are combined into a factor sequence, and the dynamic interaction relationships between factors are captured through a cross-factor attention mechanism. ; Cross-factor attention computation: ; ; in, Let be the cross-factor attention weight matrix, and its i-th Each element represents a factor group. For factor groups The intensity of information contribution.

[0038] Next, gating fusion is performed, and the gating fusion mechanism is as follows: To adaptively control the contribution weights of each factor group to the final representation, a gated fusion layer is introduced: ; ; in, This is the gated weight matrix; This is the gated bias vector; For factor groups The gating weight value, with a range of ; This is the Sigmoid activation function.

[0039] Finally, a joint representation of temporal and static features is performed: The final comprehensive case representation is formed by combining temporal representations with static event characteristics (including but not limited to the initial scale of the event, the code for the type of protection, the code for the type of subject, and the code for the region to which the event belongs). ; in, The static feature vector of the case (after embedding and normalization). This is a comprehensive representation dimension.

[0040] The output of step S20 is the set of fused temporal representation vectors for all cases. This serves as the input for step S30.

[0041] S30. Model the processing path to form a directed acyclic graph. Nodes in the directed acyclic graph represent processing stage states, and directed edges represent legal state transitions. The graph also includes parameterized predictions of transition duration, transition resource consumption, and transition success probability based on the events to be processed.

[0042] Specifically, a structured model is constructed for the entire processing chain of the event to be processed, and the complete processing flow is abstracted into a directed acyclic graph. Each node in the directed acyclic graph corresponds to a different processing stage state in the event progression; the directed edges between nodes represent the legal and compliant state transition relationships between different processing stages.

[0043] Meanwhile, three types of core transition attribute parameters are configured for each directed edge: transition duration, transition resource consumption, and transition success probability. These parameters are not fixed constants, but rather dynamically parameterized predictions are made through a dedicated prediction network based on the comprehensive representation of the events to be processed, so that the path attributes can be adapted to the personalized characteristics of different events.

[0044] In one embodiment of this application, step S30 includes: S31. Model the processing path to form a directed acyclic graph.

[0045] S32. Define multiple legal transition paths from the initial state to the final state achieved through disposal.

[0046] S33. For each directed edge in the directed acyclic graph, establish a transfer attribute prediction network, and use the transfer duration, transfer cost and transfer success probability as variables that are dynamically parameterized by the comprehensive characterization of the object to be transferred.

[0047] Specifically, the event processing path is modeled as a directed acyclic graph. , where: node set Each node This represents a processing stage state, including but not limited to: Initial state (the point at which an event enters the processing decision); Status of asset preservation application; Status of the urging order application; : Standard procedure filing status (ordinary procedure); : Standard procedure filing status (simplified procedure); Arbitration application status; Special procedure application status; : The negotiation / settlement status in the program; Status of the first-instance judgment; Status of appeal in the second instance; Status of obtaining effective legal documents; : Execution application status; Asset control status; Auction status; Asset recovery status; : Execution and reconciliation status; Additional responsible party status; Bankruptcy filing status; : Temporarily terminated status (this process is temporarily suspended); : Resume execution status; : Processing completed (effect achieved).

[0048] edge set Each directed edge Indicates from state to state Legal transfer, subject to transfer conditions And predicted transfer attributes: ; in, : Set of conditions for transfer (including but not limited to preconditions, time limits, and evidentiary conditions); : Estimated transfer time (from state) Entering the state The expected time (in days) is a positive real number. : Estimated resource consumption for the transfer (expected resource expenditure for this transfer step), with a range of non-negative real numbers; : Successful transfer probability (the probability that this transfer step will be completed successfully), range is .

[0049] Understandably, a complete event handling path Defined from the initial state Completed A directed path: ; in, This represents the number of intermediate nodes in the path. For path number The intermediate state nodes traversed by the step.

[0050] The global attribute of a path is defined as follows: Total path duration: ; Total resource consumption of the path: ; Path success probability (the joint probability that all transition steps are successful): ; Expected outcome of the path: ; in, The expected outcome when the path is successfully completed is determined by the processing method before the path's terminal state.

[0051] It should be noted that the transition attributes of each edge in the path graph are not fixed constants, but rather functions of event characteristics. For each edge... Its transfer properties are characterized by a comprehensive representation of events. Parametric prediction: ; in, For the edge Attribute prediction network, These are the corresponding learnable parameters.

[0052] The output of step S30 is a parameterized path graph structure. and the parameter set of each edge attribute prediction network This serves as the input for step S40.

[0053] S40. Based on the fused temporal representation vector and combined with the path value network and Monte Carlo tree search algorithm, candidate processing paths are searched in the directed acyclic graph to form a candidate path set.

[0054] Specifically, in this step, the fused temporal representation vector output by S20 is used as the basic input. Combined with the pre-trained path value network and Monte Carlo tree search algorithm, a full-link path retrieval is carried out in the directed acyclic graph space constructed by S30. All legal transition links are traversed, and multiple processing paths with excellent comprehensive performance are selected. Finally, a candidate path set is formed.

[0055] In one embodiment of this application, step S40 includes: S41. Construct a path value function and generate a path value network based on the path value function.

[0056] S42. Based on the path value network, evaluate the remaining expected value of each state node in the directed acyclic graph.

[0057] S43. Based on the remaining expected value, the Monte Carlo tree search algorithm is used to explore paths in the directed acyclic graph. In the selection phase, the success probability of transition is used as a priori guidance. In the simulation phase, the output of the path value network is used to replace random simulation. Candidate paths are selected based on search efficiency and path value, and a candidate path set is formed.

[0058] Specifically, in this step, the path value function is first defined. Define path Regarding the event The comprehensive value function: ; in, :event In the path The expected results were achieved; :event In the path The expected total resource consumption; :event In the path Total expected duration (in years); :event The initial scale base, with data sourced from the event ledger system; Annualized opportunity cost rate; : Preset path value dimension weights, with a value range of positive real numbers, and the data source is the business target configuration table.

[0059] Secondly, construct a path value network: Building a path value network It is used to quickly assess the expected value of a path starting from the current state: ; in: State Node The embedding vector encodes the semantic and positional information of the state; Learnable parameters of the value network; From state The expected residual value upon arrival at the point of completion; It should be noted that the path value network architecture is a multilayer perceptron: ; The path value network is trained using historical complete path data, and its loss function is: ; in, From state The actual cumulative value to the end of the path (calculated by backtracking the actual achieved results).

[0060] Next, the Monte Carlo Tree Search (MCTS) algorithm is used to search for the optimal path in the path graph: Search tree node: Each search tree node corresponds to a state in the path graph. Record the number of visits and cumulative value .

[0061] Selection phase: From the current state Starting from the upper confidence boundary (UCB) strategy, select the next transfer step: ; in: The preset exploration coefficient controls the balance between exploration and utilization; its value range is positive real numbers. In one embodiment, The value is 1.5.

[0062] :event From current state to state The success probability of the transition (calculated by the edge attribute prediction network in step S30) is used as a priori probability to guide the search direction. Expansion phase: When an unvisited node is reached, it is added to the search tree.

[0063] Simulation phase: Starting from the newly expanded nodes, utilizing the value network Quickly estimate the value of the remaining path (an alternative to stochastic simulation in traditional MCTS):

[0064] Backtracking phase: The simulated values ​​are backtracked along the search path to update the statistics of each node. ; ; Iteration termination: Execute the preset... The search terminates after several iterations; in one embodiment, the search is terminated after a certain number of iterations. The value is 800.

[0065] Optimal path extraction: Starting from the root node, select the child node with the most visits at each step until the terminal state is reached, forming a recommended path: ; Path diversity preservation: Besides the optimal path, retain the top-ranked paths by access frequency. The alternative paths are selected to form a candidate path set: ; The paths are sorted from highest to lowest value estimate; in one embodiment, the paths are sorted from highest to lowest value estimate. The value is 3.

[0066] The output of step S40 is a set of candidate paths for each event. This serves as the input for step S50.

[0067] S50. Perform feasibility verification and sensitivity analysis on each candidate path in the candidate path set to obtain the global optimal path.

[0068] Specifically, in this step, for each candidate path in the candidate path set, the feasibility of implementation, the sensitivity analysis of time series characteristics, and the stress test of multiple scenarios are carried out in sequence. The path value, operational stability, and risk resistance resilience are comprehensively evaluated, and the globally optimal path that is suitable for the current event is selected from the candidate paths.

[0069] In one embodiment of this application, step S50 includes: S51. Perform legal and logical feasibility verification on each candidate path in the candidate paths, and eliminate paths that do not meet the transfer conditions.

[0070] S52. Calculate the sensitivity of each candidate path to changes in each time series factor, and obtain the stability of the candidate path based on the sensitivity.

[0071] S53. Perform stress tests by setting adverse stress scenarios and calculate the resilience index of the candidate path.

[0072] S54. By combining the path value, stability, and resilience indicators of the candidate paths, the globally optimal path is obtained.

[0073] Specifically, the first step is to conduct a legal and logical feasibility verification. For path Each transfer step in Verify its transfer conditions Whether in the event The current state satisfies: ; in, For the event The current state vector (including but not limited to statute of limitations status, evidence possession status, jurisdiction determination status, preservation conditions met status, and enforcement basis obtained status) is sourced from the event status management system. Paths that do not meet the feasibility conditions are removed from the candidate set.

[0074] Secondly, a sensitivity analysis of time-series features is conducted: Assess the sensitivity of the recommended path to changes in various time-series factors and identify key risk factors: For factor groups The sensitivity of the path value to the time series representation of this factor group is calculated. ; in, It represents the average of the absolute values ​​of the components of the gradient vector.

[0075] The path stability index is defined as a weighted sum of the sensitivities of each factor: ; in, The preset factor sensitivity weights reflect the degree of uncertainty of future changes in each factor. The value range is positive real numbers, and the data source is historical factor volatility statistics.

[0076] Scenario stress test: The recommended path was stress-tested under multiple adverse scenarios: definition Each stress scenario (including but not limited to: a significant decrease in the assets of the event subject, a significant decline in the efficiency of the handling agency, the subject adopting extreme confrontational strategies, jurisdictional disputes arising during the handling process, and key evidence being questioned) requires further analysis of each scenario. Modify the corresponding time-series feature inputs and recalculate the path value: ; in, To be in stressful situations The following is a revised comprehensive representation of the event.

[0077] Calculate the worst-case value of the path: ; Define path resilience metrics: ; The range is The closer the value is to 1, the smaller the value loss of the path under adverse scenarios.

[0078] Final Path Scoring and Selection: The final score is calculated by combining path value, stability, and resilience. ; in, A preset stability preference index, with a value range of positive real numbers; in one embodiment, The value is 0.3; A preset resilience preference index, with a value range of positive real numbers; in one embodiment, The value is 0.5; The path with the highest final score will be selected as the recommended path: ; The output of step S50 is the verified optimal processing path for each event. .

[0079] In one embodiment of this application, the method further includes: S60. Execute according to the global optimal path and monitor and process the data in real time during the execution process. If the processed data reaches the processing path re-evaluation trigger condition, the processing path is re-evaluated. The path re-evaluation trigger condition includes: the cumulative deviation exceeds the threshold, the transfer step fails to execute, and the change in object representation exceeds the threshold. If any condition is met, the temporal feature fusion step and subsequent steps are re-executed from the current state.

[0080] Specifically, firstly, the execution of the processing path is monitored in real time: For paths that have already been started Collect actual transfer attributes after each path node transfer is completed: ; And compare it with the predicted value to calculate the cumulative deviation: ; Secondly, the path reassessment trigger conditions: Path re-evaluation (re-execution of steps S20 to S50 from the currently reached nodes) is triggered when any of the following conditions are met: Condition R1: The cumulative deviation exceeds the preset path deviation threshold. : ; In one embodiment, The value is 0.3.

[0081] Condition R2: A certain transition step failed ( Therefore, it is necessary to replan the subsequent path from the current state.

[0082] Condition R3: Time-series feature monitoring detects significant change events (including but not limited to large changes in the target's assets, the emergence of new negative records by the target, and the issuance of new rules by regulatory agencies that affect the event), leading to a comprehensive characterization of the case. Significant changes occurred: ; In one embodiment, The value is 0.2.

[0083] Finally, the model parameters are updated online: Based on actual data from path execution, the following model parameters are continuously updated: Edge attribute prediction network update: ; in, The loss function for edge attribute prediction (the deviation between the predicted transfer time, cost, and success probability and the actual values). The learning rate is a preset value for the edge attribute model; in one embodiment, The value is 0.001.

[0084] Value network update: ; in, The loss is calculated based on the actual path value of newly completed cases. The preset value network learning rate; in one embodiment, The value is 0.0005.

[0085] Cross-factor fusion model update: Updated at a preset fusion model cycle Incremental training is performed on the multi-factor fusion model from step S20, using the newly added path result data as a supervision signal: ; In one embodiment, The value is 30 days. The value is 0.0002.

[0086] This application provides an event processing path prediction method based on temporal feature fusion, the advantages of which are: This application extracts the time series of events across five dimensions: delay evolution, response change, state attribute variation, environmental dynamics, and event subject behavior patterns, and performs trend enhancement. It automatically captures dynamic interactions between factors using intra-factor temporal coding and cross-factor attention mechanisms to generate a fused temporal representation. Simultaneously, it models the processing path as a directed acyclic graph, dynamically parameterizing the transition time, resource consumption, and success probability of each edge based on the event representation. Furthermore, it combines Monte Carlo tree search and path value network to achieve globally optimal path search, overcoming the limitations of static features and single-factor analysis, as well as the defects of stage-by-stage greedy decision-making. Further, it reduces the impact of historical data selection bias through feasibility verification, gradient-based sensitivity analysis, and multi-scenario stress testing. During execution, it triggers path reassessment and online model updates based on actual deviation monitoring, transition failures, or significant changes in event representation, forming a closed loop of prediction-execution-observation-correction. This significantly improves the dynamic adaptability, global optimization capability, and system robustness of complex transaction processing path prediction.

[0087] Please see Figure 2 This application also provides an event processing path prediction system 200 based on temporal feature fusion, comprising: The data acquisition module 201 is used to acquire the time series of a multi-factor group of the event to be processed, and to preprocess the time series of the multi-factor group to obtain the preprocessed time series of the multi-factor group. The time series of the multi-factor group includes: delayed evolution time series, response change time series, state attribute change time series, environmental dynamic time series and event object subject behavior pattern time series. The feature extraction and fusion module 202 is used to extract the temporal evolution representation of each dimension according to the temporal encoder in each factor group, and capture the dynamic interaction relationship between different factor groups through the cross-factor attention mechanism to generate a fused temporal representation vector. The processing path construction module 203 is used to model the processing path and form a directed acyclic graph. The nodes in the directed acyclic graph represent the processing stage state, the directed edges represent the legal state transitions, and are accompanied by parameterized predictions of the transition duration, transition resource consumption, and transition success probability by the event to be processed. The filtering module 204 is used to search for candidate processing paths in the directed acyclic graph based on the fused temporal representation vector and in combination with the path value network and Monte Carlo tree search algorithm, to form a candidate path set; The analysis module 205 is used to perform feasibility verification and sensitivity analysis on each candidate path in the candidate path set to obtain the globally optimal path.

[0088] Please see Figure 3This application also provides a computer electronic device 300, including a memory 503 and a processor 302. The memory 303 stores a computer program, and the processor 302 executes the computer program to implement the steps of the event processing path prediction method based on temporal feature fusion described above.

[0089] Specifically, the computer electronic device 300 includes: a transceiver 301, a bus interface, and a processor 302. The processor 302 is used to acquire the time series of a multi-factor group of events to be processed, and to preprocess the time series of the multi-factor group to obtain a preprocessed time series of the multi-factor group. The time series of the multi-factor group includes: a delayed evolution time series, a response change time series, a state attribute change time series, an environmental dynamic time series, and an event object subject behavior pattern time series. The processor extracts the temporal evolution representation of each dimension based on the time encoder within each factor group and captures it through a cross-factor attention mechanism. The dynamic interaction relationships between different factor groups generate a fused temporal representation vector; the processing path is modeled to form a directed acyclic graph (DAG), where nodes represent processing stage states, directed edges represent legal state transitions, and are accompanied by parameterized predictions of transition duration, transition resource consumption, and transition success probability based on the representation parameters of the events to be processed; candidate processing paths are searched in the DAG based on the fused temporal representation vector, combined with a path value network and a Monte Carlo tree search algorithm, forming a candidate path set; feasibility verification and sensitivity analysis are performed on each candidate path in the candidate path set to obtain the globally optimal path.

[0090] In this embodiment of the application, the computer electronic device 300 further includes a memory 303. Figure 3 In this context, the bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits of one or more processors 302 (represented by processor 302) and various circuits of memory 303 (represented by memory 303). The bus architecture can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 301 can be multiple elements, including transmitters and receivers, providing a unit for communicating with various other devices over a transmission medium. The processor 302 is responsible for managing the bus architecture and general processing, and the memory 303 can store data used by the processor 302 during operation.

[0091] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the event processing path prediction method based on temporal feature fusion described above.

[0092] In this embodiment, the computer-readable storage medium can be a non-volatile storage medium or a volatile storage medium. For example, the computer storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0093] In all examples shown and described herein, any specific values ​​should be interpreted as merely exemplary and not as limitations; therefore, other examples of exemplary embodiments may have different values.

[0094] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0095] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that, in alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0096] In addition, the functional modules or units in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0097] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a terminal device (which may be a smartphone, personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0098] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An event processing path prediction method based on temporal feature fusion, characterized in that, include: The time series of a multi-factor group of events to be processed is obtained, and the time series of the multi-factor group is preprocessed to obtain the preprocessed time series of the multi-factor group. The time series of the multi-factor group includes: delayed evolution time series, response change time series, state attribute change time series, environmental dynamic time series and event object subject behavior pattern time series. The temporal evolution representations of each dimension are extracted based on the temporal encoder within each factor group, and the dynamic interaction relationships between different factor groups are captured through a cross-factor attention mechanism to generate a fused temporal representation vector. The processing path is modeled to form a directed acyclic graph. Nodes in the directed acyclic graph represent processing stage states, and directed edges represent legal state transitions. The graph is also accompanied by parameterized predictions of transition duration, transition resource consumption, and transition success probability based on the representation parameters of the events to be processed. Based on the fused temporal representation vector, and combined with the path value network and Monte Carlo tree search algorithm, candidate processing paths are searched in the directed acyclic graph to form a candidate path set; Feasibility verification and sensitivity analysis are performed on each candidate path in the candidate path set to obtain the globally optimal path.

2. The method according to claim 1, characterized in that, The method further includes: The process is executed according to the global optimal path, and the data is monitored and processed in real time. If the processed data reaches the processing path re-evaluation trigger condition, the processing path is re-evaluated. The path re-evaluation trigger condition includes: the cumulative deviation exceeds the threshold, the transfer step fails, and the change in object representation exceeds the threshold. When any of the conditions is met, the temporal feature fusion step and subsequent steps are re-executed from the current state.

3. The method according to claim 1, characterized in that, The preprocessing includes: missing value imputation, sequence normalization, and trend feature enhancement.

4. The method according to claim 1, characterized in that, The step involves extracting temporal evolution representations of each dimension based on the temporal encoder within each factor group, and capturing the dynamic interaction relationships between different factor groups through a cross-factor attention mechanism to generate a fused temporal representation vector, including: The temporal features of each factor group are encoded using a Transformer encoder with causal masking to extract the temporal evolution representation within the factor. The representations of each factor group are combined into a factor sequence, and the dynamic interaction weights between factor groups are calculated through a cross-factor attention mechanism. Then, the weights are adaptively weighted through a gated fusion layer to generate the final fused temporal representation vector.

5. The method according to claim 1, characterized in that, The processing path is modeled to form a directed acyclic graph (DAG), where nodes represent processing stage states, directed edges represent valid state transitions, and the graph includes parameterized predictions of transition duration, transition resource consumption, and transition success probability based on the events to be processed. The processing path is modeled to form a directed acyclic graph; Define multiple legal transition paths from the initial state to the final state achieved through disposal; For each directed edge in the directed acyclic graph, a transition attribute prediction network is established, and the transition duration, transition cost, and transition success probability are used as variables that are dynamically parameterized by the comprehensive characterization of the object to be transferred.

6. The method according to claim 1, characterized in that, The step of searching for the globally optimal processing path in the directed acyclic graph based on the fused temporal representation vector and combining the path value network with the Monte Carlo tree search algorithm includes: Construct a path value function and generate a path value network based on the path value function; Based on the path value network, evaluate the remaining expected value of each state node in the directed acyclic graph; Based on the remaining expected value, the Monte Carlo tree search algorithm is used to explore paths in the directed acyclic graph. In the selection phase, the success probability of transition is used as a priori guidance, and in the simulation phase, the output of the path value network is used to replace random simulation. Candidate paths are selected based on search efficiency and path value, and a candidate path set is formed.

7. The method according to claim 1, characterized in that, The process of performing feasibility verification and sensitivity analysis on each candidate path in the candidate path set to obtain the globally optimal path includes: Each candidate path in the candidate path is verified for legal and logical feasibility, and paths that do not meet the transfer conditions are eliminated; Calculate the sensitivity of each candidate path to changes in each time series factor, and obtain the stability of the candidate path based on the sensitivity. The resilience index of the candidate path is calculated by conducting stress tests under adverse stress scenarios. By combining the path value, stability, and resilience metrics of the candidate paths, the globally optimal path is obtained.

8. An event processing path prediction system based on temporal feature fusion, characterized in that, include: The data acquisition module is used to acquire the time series of a multi-factor group of events to be processed, and to preprocess the time series of the multi-factor group to obtain the preprocessed time series of the multi-factor group. The time series of the multi-factor group includes: delayed evolution time series, response change time series, state attribute change time series, environmental dynamic time series, and event object subject behavior pattern time series. The feature extraction and fusion module is used to extract temporal evolution representations of each dimension based on the temporal encoder within each factor group, and to capture the dynamic interaction relationships between different factor groups through a cross-factor attention mechanism to generate a fused temporal representation vector. The processing path construction module is used to model the processing path and form a directed acyclic graph. The nodes in the directed acyclic graph represent the processing stage state, the directed edges represent the legal state transitions, and are accompanied by the transition duration, transition resource consumption and transition success probability predicted by the representation parameterization of the event to be processed. The filtering module is used to search for candidate processing paths in the directed acyclic graph based on the fused temporal representation vector and in combination with the path value network and Monte Carlo tree search algorithm, to form a candidate path set; The analysis module is used to perform feasibility verification and sensitivity analysis on each candidate path in the candidate path set to obtain the globally optimal path.

9. A computer electronic device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the event processing path prediction method based on temporal feature fusion as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the event processing path prediction method based on temporal feature fusion as described in any one of claims 1-7.

Images