Methods and systems for identifying bank customers who have lost credit.

Abstract

This application provides a method and system for identifying bank credit churn customers, relating to the field of artificial intelligence technology. It involves acquiring a set of post-loan behavior data and a set of credit contract attribute data for credit customers; dividing the post-loan behavior data set into behavioral evolution stages to obtain fund flow stage division results and channel access stage division results; constructing a contract state transition chain for the credit contract attribute data set to obtain the credit customer contract state transition chain; inputting the fund flow stage division results, channel access stage division results, and contract state transition chain into a churn tendency joint identification model for churn tendency identification processing to obtain the credit customer churn tendency identification result; and based on this credit customer churn tendency identification result, identifying a list of credit customers requiring intervention from the credit customer group. This invention improves the accuracy and efficiency of bank credit churn identification and intervention.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a method and system for identifying bank credit loss customers. Background Technology

[0002] In bank lending operations, customer churn directly leads to deterioration of credit asset quality and reduction in interest income, posing a core risk issue that banks urgently need to address. After obtaining a loan, a customer's repayment behavior, fund usage habits, and frequency of interaction with the bank all change over time. These changes often indicate a loosening or even termination of the credit relationship between the customer and the bank. In practice, banks typically only realize churn after a customer has shown significant delinquency or has proactively closed their account. By this time, it is often too late to recover the customer, and the bank not only loses future interest income but also incurs additional costs associated with disposing of non-performing assets. To detect churn risk in advance, existing technologies often employ analysis methods based on single-dimensional data. For example, they may use only the customer's repayment records at a specific point in time or rely solely on fixed attributes such as historical loan amounts to construct simple scoring rules, applying a uniform standard to assess the probability of churn for all credit customers.

[0003] However, individuals within the credit customer group vary greatly in terms of capital turnover patterns, frequency of interaction with banking channels, and the types and terms of their credit contracts. These differences exhibit non-linear dynamic changes over time, making it difficult to effectively distinguish between customers truly at risk of churn and those exhibiting only short-term fluctuations, relying solely on static data or a single, simple rule. Furthermore, existing technologies often mix information from different sources and of different natures without distinguishing their inherent temporal evolution patterns and contract status change logic. This results in limited model capabilities for capturing complex churn signals, leading to high false positive and false negative rates in early warning results. Consequently, banks struggle to develop targeted customer retention strategies, hindering the initiative and accuracy of customer relationship maintenance. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of this application, embodiments of this application provide a method for identifying bank credit churn customers, the method comprising: Obtain a set of post-loan behavior data and a set of credit contract attribute data for credit customers. The set of post-loan behavior data includes a sequence of account fund flow records and a sequence of channel access behavior records. The set of credit contract attribute data includes a sequence of credit contract status records. The post-loan behavior data set is divided into behavior evolution stages to obtain the fund flow stage division results and the channel access stage division results. The credit contract attribute data set is processed to construct a contract state transition chain, thereby obtaining the contract state transition chain of the credit customer. The results of the fund flow stage segmentation, the results of the channel access stage segmentation, and the contract state transition chain input churn tendency joint identification model are used to perform churn tendency identification processing to obtain the churn tendency identification result of the credit customer. Based on the churn tendency identification results, a list of credit customers to be intervened in is drawn from the credit customer group.

[0005] In another aspect, embodiments of this application also provide a bank credit loss customer identification system, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0006] Based on the above, this invention simultaneously collects and separately organizes a post-loan behavior data set containing account fund flow record sequences and channel access behavior record sequences, as well as a credit contract attribute data set containing credit contract validity status record sequences. This integrates the dynamic behavioral information of credit customers at the fund usage and channel interaction levels, as well as the status change information at the contract validity level, into a unified analysis process. Furthermore, it performs behavioral evolution stage segmentation processing on the post-loan behavior data set to extract the stage-specific evolution characteristics of fund flow and channel access, and performs contract status migration chain construction processing on the credit contract attribute data set to extract the temporal migration path of the contract validity status. This allows for the separate analysis of the stage-specific leaps in customer behavior and the temporal migration of contract status. Instead of a mixed and crude processing approach, this method extracts structured data and integrates the results of fund flow stage segmentation, channel access stage segmentation, and contract state transition chain as complementary dimensions into a churn tendency joint identification model for comprehensive judgment. This allows the model to simultaneously utilize behavioral evolution stage signals and contract state transition signals for cross-verification and collaborative decision-making within the same identification framework. Ultimately, based on the churn tendency identification results, the model accurately identifies the list of credit customers requiring intervention. This enables banks to concentrate customer retention resources on high-risk customer groups confirmed through multi-dimensional collaborative identification, significantly improving the comprehensive collaborative optimization level of bank credit churn customer identification in terms of dynamic behavioral perception depth, contract state evolution modeling accuracy, and intervention decision resource allocation efficiency. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the execution flow of the bank credit loss customer identification method provided in the embodiments of this application.

[0008] Figure 2This is a schematic diagram of a scenario for post-loan information collection and credit status migration analysis provided in the embodiments of this application.

[0009] Figure 3 This is a schematic diagram of a multi-source information fusion and loss tendency identification scenario provided in the embodiments of this application.

[0010] Figure 4 This is a schematic diagram of a closed-loop scenario for the hierarchical distribution and follow-up of the customer list to be intervened, provided in an embodiment of this application. Detailed Implementation

[0011] This embodiment uses a single application scenario throughout: identifying the tendency of bank loan customers to churn in the post-loan stage. In this embodiment, all loan customers have signed a data authorization agreement, authorizing the bank to collect, store, and use post-loan data related to credit risk management and customer service during the loan term. The authorized scope includes account fund flow records, channel access behavior records, loan contract status records, images of business premises, images of business documents, voice calls, text interactions, business processing videos, interface screenshots, location trajectories, address books, publicly available third-party information, surveillance videos of business premises, and publicly available text and images on social media. For information involving personal identity, field-level anonymization and differential privacy methods are used before the data enters the subsequent processing flow, ensuring that no single data record can be directly attributed to a specific natural person. The processing complies with the requirements of relevant laws and regulations regarding minimum data necessity, purpose limitation, and security protection.

[0012] The present application will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a method for identifying lost bank credit customers according to an embodiment of this application. The following is a detailed description of this method for identifying lost bank credit customers.

[0013] Step S110: Obtain the post-loan behavior data set and the credit contract attribute data set of the credit customer. The post-loan behavior data set includes the account fund flow record sequence and the channel access behavior record sequence. The credit contract attribute data set includes the credit contract existence status record sequence.

[0014] This embodiment uses the bank's post-loan churn risk identification for micro and small enterprise (MSE) loan customers as a unified application scenario. The loan customers are MSE owners with outstanding loan contracts with the bank. The post-loan behavior data set comes from the bank's core accounting system and electronic channel log system. Each record in the account fund flow record sequence includes a fund flow direction identifier D1, a fund flow amount value M1, a fund flow occurrence timestamp T1, a fund flow counterparty type identifier C1, and a fund flow channel identifier P1. Each record in the channel access behavior record sequence includes a channel type identifier Q1, a channel access start timestamp T2, a channel access duration value L1, and a channel access page path identifier R1. The loan contract attribute data set comes from the bank's loan management system. Each record in the loan contract status record sequence includes a contract status identifier S1, a contract status record timestamp T3, and contract status record content text X1.

[0015] The data collection process follows the data authorization agreement signed with credit customers, which clearly stipulates that post-loan behavioral data and credit contract attribute data are used for post-loan risk assessment and early warning management. The scope of data collection is limited to necessary information related to the performance of credit contracts.

[0016] Specifically, data fields involving personal identification information are masked and hashed for anonymization. Counterparty identification information in account fund flow records is tokenized, retaining only the counterparty type identifier. After data collection, extreme value truncation is performed on the fund flow amount value M1 in the account fund flow record sequence; outliers exceeding a set reasonable range are truncated and replaced according to the range boundary values. Missing values ​​are imputed for the channel access duration value L1, using the median of records of the same channel type.

[0017] Step S120: Perform behavioral evolution stage segmentation processing on the post-loan behavior data set to obtain the fund flow stage segmentation result and the channel access stage segmentation result.

[0018] Stage boundary identification based on adjacent record difference detection is performed on the account fund flow record sequence and the channel access behavior record sequence respectively. The fund flow stage segmentation result contains multiple fund flow stage units, each carrying the stage start and end record positions and the set of records within the stage. The channel access stage segmentation result contains multiple channel access stage units, each carrying the stage start and end access record positions and the set of access records within the stage.

[0019] Step S121: Extract the account fund flow record sequence from the post-loan behavior data set, perform fund flow direction change capture processing on the account fund flow record sequence in chronological order to obtain a fund flow direction change sequence, and perform continuous same-direction segment merging processing on the fund flow direction change sequence to obtain multiple fund flow same-direction segment units, each fund flow same-direction segment unit carrying the segment start record position and segment end record position.

[0020] After extracting the account fund flow record sequence, all account fund flow records are arranged in ascending order according to the fund flow occurrence timestamp T1, resulting in an ordered account fund flow record sequence. For each pair of time-adjacent records in the ordered account fund flow record sequence, the fund flow direction identifier D1 of the two records is compared. If the fund flow direction identifier D1 of the two adjacent records is the same, a direction-invariant marker V11 is generated at that adjacent position. If the fund flow direction identifier D1 of the two adjacent records is different, a direction-change marker V12 is generated at that adjacent position. All the direction markers at all adjacent positions are arranged in chronological order to form a fund flow direction change sequence B11. For each direction change marker V12 appearing in the fund flow direction change sequence B11, a segment division point is set between the adjacent records corresponding to the direction change marker V12. All records between adjacent segment division points constitute a fund flow in the same direction segment unit F11. Each fund flow in the same direction segment unit F11 carries the segment start record position F11s and the segment end record position F11e. The segment start record position F11s is the sequence number of the earliest record in the ordered account fund flow record sequence within that segment unit. The segment termination record position F11e is the sequence number of the latest record in the ordered account fund flow record sequence within that segment unit.

[0021] Step S122: Perform capital flow amplitude comparison processing on adjacent capital flow segment units in the same direction to obtain a capital flow amplitude difference sequence between adjacent segments. Perform capital flow stage boundary identification processing based on the capital flow amplitude difference sequence between adjacent segments to obtain a set of capital flow stage boundary positions.

[0022] For each segment unit F11 with the same direction of capital flow, extract all capital flow amount values ​​M1 within the segment, and calculate the ratio of the standard deviation to the mean of the capital flow amount values ​​M1 within the segment. This ratio is used as the capital flow amplitude A11 of that segment. For each pair of adjacent segments, the amplitude difference value K11 = A11 after - A11 before. The amplitude difference values ​​K11 of all adjacent segment pairs are arranged in chronological order to form the capital flow amplitude difference sequence E11 between adjacent segments. When performing boundary identification processing on the capital flow amplitude difference sequence E11 between adjacent segments, the absolute value of each amplitude difference value K11 in the amplitude difference sequence E11 is compared with the preset amplitude difference threshold Y11. The preset amplitude difference threshold Y11 is dynamically determined based on the distribution characteristics of all amplitude difference values ​​K11, Y11 = (Q3 - Q1) × c1, where Q3 is the upper quartile of the amplitude difference value K11 distribution, Q1 is the lower quartile of the amplitude difference value K11 distribution, and c1 is a preset coefficient. The adjacent segment positions corresponding to amplitude difference values ​​K11 whose absolute values ​​are greater than the preset amplitude difference threshold Y11 are marked as boundary positions of the capital flow stage. All marked boundary positions are arranged in chronological order to form the capital flow stage boundary position set G11.

[0023] Step S123: Perform stage segmentation processing on the account fund flow record sequence based on the set of fund flow stage boundary positions to obtain the fund flow stage division result. The fund flow stage division result includes multiple fund flow stage units and the stage start and end record positions corresponding to each fund flow stage unit.

[0024] Based on the sequence number of each boundary position in the set G11 of fund flow stage boundary positions within the ordered account fund flow record sequence, the ordered account fund flow record sequence is divided into multiple consecutive subsequences. All records between two adjacent boundary positions constitute a fund flow stage unit H11. Each fund flow stage unit H11 carries a stage start record position H11s and a stage end record position H11e. The stage start record position H11s is the sequence number of the earliest record in the ordered account fund flow record sequence within that stage unit. The stage end record position H11e is the sequence number of the latest record in the ordered account fund flow record sequence within that stage unit. All fund flow stage units H11 arranged in chronological order constitute the fund flow stage division result.

[0025] Step S124: Extract the channel access behavior record sequence from the post-loan behavior data set, perform access channel switching capture processing on the channel access behavior record sequence in chronological order between adjacent access records to obtain an access channel switching sequence, and perform continuous same-channel segment merging processing on the access channel switching sequence to obtain multiple channel access same-channel segment units, each channel access same-channel segment unit carrying the segment start access record position and segment end access record position.

[0026] After extracting the channel access behavior record sequence, all channel access behavior records are arranged in ascending order according to the channel access start timestamp T2, resulting in an ordered channel access behavior record sequence. For each pair of time-adjacent access records in the ordered channel access behavior record sequence, their channel type identifiers Q1 are compared. If the channel type identifiers Q1 of two adjacent access records are the same, a channel invariant marker W11 is generated at that adjacent position. If the channel type identifiers Q1 of two adjacent access records are different, a channel switching marker W12 is generated at that adjacent position. All channel markers at all adjacent positions are arranged in chronological order to form an access channel switching sequence B12. For each channel switching marker W12 appearing in the access channel switching sequence B12, a segment splitting point is set between the adjacent access records corresponding to that channel switching marker W12. All access records between adjacent segment splitting points constitute a channel access co-channel segment unit F12. Each channel access co-channel segment unit F12 carries the segment start access record position F12s and the segment end access record position F12e. The segment start access record position F12s is the sequence number of the earliest access record in the ordered channel access behavior record sequence within this segment unit. The segment end access record position F12e is the sequence number of the latest access record in the ordered channel access behavior record sequence within this segment unit.

[0027] Step S125: Compare the access frequencies of adjacent channel access segments to obtain a sequence of access frequency differences between adjacent segments. Based on the sequence of access frequency differences between adjacent segments, perform channel access stage boundary identification processing to obtain a set of channel access stage boundary locations.

[0028] For each channel access segment unit F12, extract the number of access records N11 and the segment duration U11. The segment duration U11 is the difference between the timestamp corresponding to the segment termination access record position F12e and the timestamp corresponding to the segment start access record position F12s. Access frequency A12 = N11 ÷ U11. If U11 = 0, then access frequency A12 = N11 ÷ u0, where u0 is a preset minimum time unit. For each pair of adjacent segments, the frequency difference value K12 = after A12 - before A12. The frequency difference values ​​K12 of all adjacent segment pairs are arranged in chronological order to form the access frequency difference sequence E12 between adjacent segments. When performing boundary identification processing on the access frequency difference sequence E12 between adjacent segments, the absolute value of each frequency difference value K12 in the frequency difference sequence E12 is compared with the preset frequency difference threshold Y12. The preset frequency difference threshold Y12 is dynamically determined based on the distribution characteristics of all frequency difference values ​​K12. Y12 = (Q4 - Q5) × c2, where Q4 is the upper quartile of the frequency difference value K12 distribution, Q5 is the lower quartile of the frequency difference value K12 distribution, and c2 is a preset coefficient. The adjacent segment positions corresponding to frequency difference values ​​K12 with an absolute value greater than the preset frequency difference threshold Y12 are marked as channel access stage boundary positions. All marked boundary positions are arranged in chronological order to form the channel access stage boundary position set G12.

[0029] Step S126: Perform stage segmentation processing on the channel access behavior record sequence based on the set of channel access stage boundary positions to obtain the channel access stage division result. The channel access stage division result includes multiple channel access stage units and the stage start and end access record positions corresponding to each channel access stage unit.

[0030] Based on the sequence number of each boundary position in the set G12 of channel access phase boundary positions within the ordered channel access behavior record sequence, the ordered channel access behavior record sequence is divided into multiple consecutive subsequences. All access records between two adjacent boundary positions constitute a channel access phase unit H12. Each channel access phase unit H12 carries the phase start access record position H12s and the phase end access record position H12e. The phase start access record position H12s is the sequence number of the earliest access record in this phase unit within the ordered channel access behavior record sequence. The phase end access record position H12e is the sequence number of the latest access record in this phase unit within the ordered channel access behavior record sequence. All channel access phase units H12 are arranged in chronological order to form the channel access phase division result.

[0031] Step S130: Perform contract state transition chain construction processing on the credit contract attribute data set to obtain the contract state transition chain of the credit customer.

[0032] The system performs state change capture, state dwelling fragment merging, state transition relationship extraction, state migration chain construction, state dwelling duration association, state record content association, and attribute merging processing on the sequence of credit contract existence state records to construct a contract state migration chain carrying state dwelling duration attribute and state record content attribute.

[0033] Step S131: Extract the sequence of credit contract status records from the credit contract attribute data set, and perform contract status change capture processing on the sequence of credit contract status records in chronological order between adjacent status records to obtain a contract status change sequence.

[0034] After extracting the sequence of credit contract status records, all credit contract status records are arranged in ascending order according to the contract status record timestamp T3 to obtain an ordered sequence of credit contract status records. For each pair of time-adjacent status records in the ordered sequence, their contract status identifiers S1 are compared. If the contract status identifiers S1 of two adjacent status records are the same, a status-invariant marker Z11 is generated at that adjacent position. If the contract status identifiers S1 of two adjacent status records are different, a status-change marker Z12 is generated at that adjacent position. All the status markers at all adjacent positions are arranged in chronological order to form a contract status change sequence B13.

[0035] Step S132: Perform adjacent identical state merging processing on the contract state change sequence to obtain multiple contract state dwelling segment units. Each contract state dwelling segment unit carries the segment start state record position and the segment end state record position.

[0036] For each state change marker Z12 appearing in the contract state change sequence B13, a segmentation point is set between the adjacent state records corresponding to that state change marker Z12. All state records between adjacent segmentation points constitute a contract state residence segment unit F13. Each contract state residence segment unit F13 carries the segment start state record position F13s and the segment end state record position F13e. The segment start state record position F13s is the sequence number of the earliest state record in this segment unit in the ordered sequence of credit contract existence state records. The segment end state record position F13e is the sequence number of the latest state record in this segment unit in the ordered sequence of credit contract existence state records.

[0037] Step S133: Extract the contract state identifier corresponding to each contract state dwelling segment unit, arrange the contract state identifiers in chronological order to obtain a contract state identifier sequence, perform adjacent state transition relationship extraction processing on the contract state identifier sequence to obtain a contract state transition relationship set, the contract state transition relationship set containing the transition relationship between every two adjacent contract state identifiers.

[0038] For each contract state residency segment unit F13, extract the contract state identifier S1 that corresponds to all state records within the segment. Arrange all contract state identifiers S1 in chronological order according to the segment unit to form a contract state identifier sequence Y13. For each pair of adjacent contract state identifiers S1 in the contract state identifier sequence Y13, extract the transition relationship between them. The transition relationship includes the state identifier S1a before the transition and the state identifier S1b after the transition. The transition relationships of all adjacent contract state identifier pairs constitute the contract state transition relationship set J13.

[0039] Step S134: Construct a contract state transition chain based on the contract state transition relationship set. The contract state transition chain includes a state node set and a state transition edge set. Each state node in the state node set corresponds to a contract state identifier, and each state transition edge in the state transition edge set corresponds to a transition relationship.

[0040] Deduplicatively process all distinct contract state identifiers S1 appearing in the contract state transition relationship set J13 to obtain the complete set of state identifiers. Create a state node O11 for each contract state identifier S1 in the complete set of state identifiers. State node O11 contains a node identifier and the corresponding contract state identifier S1. Create a state transition edge O12 for each transition relationship in the contract state transition relationship set J13. State transition edge O12 contains an edge identifier, the state identifier S1a before the transition, the state identifier S1b after the transition, and the number of times the transition relationship occurs. All state nodes O11 constitute the state node set O13. All state transition edges O12 constitute the state transition edge set O14. State node set O13 and state transition edge set O14 together constitute the contract state transition chain O15.

[0041] Step S135: Extract the state dwell time corresponding to each contract state dwell time segment unit, associate the state dwell time with the corresponding state node in the contract state transition chain, and obtain the contract state transition chain carrying the state dwell time attribute.

[0042] For each contract state dwell segment unit F13, the state dwell duration U13 = timestamp corresponding to F13e - timestamp corresponding to F13s. The state dwell duration U13 is associated as an attribute with the state node O11 in the contract state transition chain O15 that corresponds to the contract state identifier S1 of that segment unit. Each state node O11 corresponds to a dwell duration list, used to store all state dwell durations U13 corresponding to that state node O11. After the association is completed, the contract state transition chain O15 carrying the state dwell duration attribute is obtained.

[0043] Step S136: Extract the state record content corresponding to each contract state resident fragment unit, associate the state record content with the corresponding state node in the contract state migration chain to obtain the contract state migration chain carrying the state record content attribute, and perform attribute merging processing with the contract state migration chain carrying the state record content attribute to obtain the contract state migration chain of the credit customer.

[0044] For each contract state dwelling segment unit F13, extract the contract state record content text X1 of all state records within the segment. Associate the list of contract state record content text X1 as an attribute with the state node O11 in the contract state migration chain O15 corresponding to the contract state identifier S1 of that segment unit. Each state node O11 corresponds to a list of record content. After association, a contract state migration chain O15 carrying the state record content attribute is obtained. Perform attribute merging processing on the contract state migration chain O15 carrying the state dwelling duration attribute and the contract state migration chain O15 carrying the state record content attribute. The attribute merging process aligns according to the node identifier of state node O11, merging the dwelling duration attribute and the record content attribute onto the same state node O11. After merging, each state node O11 simultaneously carries the contract state identifier S1, the dwelling duration list, and the record content list. Finally, the contract state migration chain O15 for the credit customer is obtained.

[0045] Step S140: Perform churn tendency identification processing on the results of the fund flow stage division, the results of the channel access stage division, and the contract state transition chain input churn tendency joint identification model to obtain the churn tendency identification result of the credit customer.

[0046] In this embodiment, the churn tendency joint identification model adopts a Transformer encoder architecture. This model includes an input embedding layer, six identical encoder layers, and a classification output layer. Each encoder layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer, with each sublayer followed by a normalization layer and residual connections. The input embedding layer maps the graph structure representations of the fund flow stage representation set, the channel access stage representation set, and the contract state transition chain into a sequence of embedding vectors of a unified dimension. The multi-head self-attention sublayer calculates the attention weights between any two positions in the embedding vector sequence, and the feedforward neural network sublayer performs a non-linear transformation on the attention output. The classification output layer pools the embedding vector sequence output from the last layer of the encoder to obtain a joint encoded representation, which is then used to generate a churn tendency category probability distribution through a fully connected layer and a normalized exponential function. The model training data comes from anonymized post-loan behavior data and contract status data of historically settled loan customers, with the training label indicating whether early settlement or credit transfer occurred. The model was trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. The mean absolute percentage error was used as the criterion for early stopping.

[0047] Step S141: Extract the start and end record positions of each capital flow stage unit and the set of capital flow records within the stage from the capital flow stage division results, and perform record density distribution capture processing on the set of capital flow records within each stage to obtain the capital flow stage density distribution characterization.

[0048] For each fund flow stage unit H11, extract the stage start record position H11s, the stage end record position H11e, and the fund flow record set I11 within the stage. Sort all records in the fund flow record set I11 in ascending order by the fund flow occurrence timestamp T1. Divide the time interval between the timestamp corresponding to the stage start record position H11s and the timestamp corresponding to the stage end record position H11e into a predetermined number of time buckets. Count the number of records falling into each time bucket to obtain a histogram of record quantity distribution. Divide the number of records in each bucket in the histogram by the time span of that bucket to obtain the record density value for that bucket. The record density values ​​of all buckets constitute the fund flow stage density distribution representation P11.

[0049] Step S142: Perform capital flow direction ratio capture processing on the capital flow record set within each capital flow stage unit to obtain the capital flow stage flow direction ratio representation. Construct a capital flow stage representation set based on the capital flow stage density distribution representation and the capital flow stage flow direction ratio representation.

[0050] For each fund flow stage unit H11, within the set of fund flow records I11, count the frequency of each fund flow direction identifier D1. Divide the frequency of each fund flow direction identifier D1 by the total number of records in the set of fund flow records I11 within the stage to obtain the flow proportion value of that fund flow direction identifier D1. The flow proportion values ​​corresponding to all different fund flow direction identifiers D1 constitute the fund flow stage flow proportion representation P12. Concatenate the fund flow stage density distribution representation P11 and the fund flow stage flow proportion representation P12 along the feature dimension to obtain the fund flow stage representation vector R11 of the fund flow stage unit H11. The fund flow stage representation vectors R11 of all fund flow stage units H11 are arranged in chronological order to form the fund flow stage representation set R12.

[0051] Step S143: Extract the start and end access record positions and the set of channel access records within each channel access stage unit from the channel access stage division results, and perform access channel distribution capture processing on the set of channel access records within each stage to obtain the channel distribution representation of the channel access stage.

[0052] For each channel access stage unit H12, extract the stage start access record position H12s, the stage end access record position H12e, and the channel access record set I12 within the stage. For all records in the channel access record set I12 within the stage, count the number of times each channel type identifier Q1 appears. Divide the number of times each channel type identifier Q1 appears by the total number of records in the channel access record set I12 within the stage to obtain the channel distribution percentage value for that channel type identifier Q1. The channel distribution percentage values ​​corresponding to all different channel type identifiers Q1 constitute the channel distribution representation P13 of the channel access stage.

[0053] Step S144: Perform access time interval capture processing on the channel access record set within each channel access stage unit to obtain the channel access stage time interval representation, and construct a channel access stage representation set based on the channel distribution representation and the channel access stage time interval representation.

[0054] For each channel access phase unit H12, within the set of channel access records I12, all records are sorted in ascending order by the channel access start timestamp T2. For each pair of time-adjacent records, the access time interval value = after T2 - before T2. All access time interval values ​​constitute an access time interval list. The values ​​in the access time interval list are statistically binned, and the interval values ​​are mapped to the corresponding bins according to preset bin boundaries. The number of interval values ​​falling into each bin is counted, and the number of interval values ​​in each bin is divided by the total number of access time intervals in the list to obtain the interval distribution ratio value for that bin. The interval distribution ratio values ​​of all preset bins constitute the channel access phase time interval representation P14. The channel access phase channel distribution representation P13 and the channel access phase time interval representation P14 are concatenated along the feature dimension to obtain the channel access phase representation vector R13 for the channel access phase unit H12. The channel access phase representation vectors R13 of all channel access phase units H12 are arranged in chronological order to form the channel access phase representation set R14.

[0055] Step S145: Input the set of fund flow stage representations, the set of channel access stage representations, and the contract state transition chain into the churn tendency joint identification model for joint encoding processing of stage representations and state transition chains to obtain joint encoded representations. Based on the joint encoded representations, perform churn tendency identification processing to obtain the churn tendency identification result of the credit customer.

[0056] The set of fund flow stage representations R12, the set of channel access stage representations R14, and the contract state transition chain O15 are input into the joint identification model for churn tendency. The contract state transition chain O15 is first processed by a graph convolutional layer, mapping the contract state identifier S1, dwell time list, and record content list of each state node O11 to a state node embedding vector. All state node embedding vectors are arranged according to the order of their appearance in the contract state identifier sequence Y13, resulting in the contract state transition chain representation R15. Each fund flow stage representation vector R11 in the set of fund flow stage representations R12, each channel access stage representation vector R13 in the set of channel access stage representations R14, and each state node embedding vector in the contract state transition chain representation R15 are linearly mapped to the same dimension and then concatenated sequentially to form an embedding vector sequence. This embedding vector sequence is then input into six encoder layers for multi-head self-attention calculation and feedforward neural network nonlinear transformation. The embedding vector sequence output from the last encoder layer is average-pooled to obtain the joint encoding representation R16. The joint encoding representation R16 is input into the classification output layer, which contains a fully connected layer and a normalized exponential function, outputting a probability distribution of churn tendency categories. The churn tendency category corresponding to the maximum probability in the probability distribution is taken as the churn tendency identification result for credit customers.

[0057] Step S150: Based on the churn tendency identification results, a list of credit customers to be intervened in is drawn from the credit customer group.

[0058] The churn tendency identification results for each credit customer in the credit customer group are analyzed to obtain a churn tendency intensity distribution. Credit customer identifiers whose churn tendency intensity meets preset churn tendency intensity conditions are extracted from the churn tendency intensity distribution to obtain a candidate credit customer identifier set. Churn trigger analysis is performed on the churn tendency identification results corresponding to each credit customer identifier in the candidate credit customer identifier set to obtain a churn trigger description for each credit customer identifier. Churn stage analysis is performed on the churn tendency identification results corresponding to each credit customer identifier in the candidate credit customer identifier set to obtain a churn stage description for each credit customer identifier. The candidate credit customer identifier set, the churn trigger descriptions for each credit customer identifier, and the churn stage descriptions for each credit customer identifier are grouped by customer dimension to obtain a credit customer churn profile set. Intervention urgency stratification is performed on each credit customer churn profile in the credit customer churn profile set to obtain an intervention urgency description for each credit customer churn profile. The set of credit customer churn profiles is sorted based on the intervention urgency description corresponding to each credit customer churn profile. The sorted set of credit customer churn profiles is then extracted from the sorted set of credit customer churn profiles. Credit customer identifiers whose intervention urgency descriptions meet the preset intervention urgency conditions are then extracted to obtain the list of credit customers to be intervened.

[0059] Step S151: Perform churn tendency intensity analysis on the churn tendency identification results to obtain a churn tendency intensity distribution. Extract credit customer identifiers that meet the preset churn tendency intensity conditions from the churn tendency intensity distribution to obtain a set of candidate credit customer identifiers.

[0060] For each credit customer's churn tendency identification result, the maximum probability value output by the normalized exponential function is taken as the churn tendency intensity value. All credit customer churn tendency intensity values ​​are then arranged by credit customer identifier to obtain a churn tendency intensity distribution. A preset churn tendency intensity condition is that the churn tendency intensity value is greater than a preset intensity threshold. Credit customer identifiers with churn tendency intensity values ​​greater than the preset intensity threshold are selected from the churn tendency intensity distribution to form a candidate credit customer identifier set.

[0061] Step S152: Perform churn cause analysis on the churn tendency identification result corresponding to each credit customer identifier in the candidate credit customer identifier set to obtain a churn cause description corresponding to each credit customer identifier.

[0062] For each credit customer identifier in the candidate credit customer identifier set, extract the joint encoding representation R16 from the last encoder layer output before the classification output layer in the churn tendency identification result for that credit customer. Calculate the attention weight for each dimension of the joint encoding representation R16 according to the feature dimensions, sort them from largest to smallest, and select the top-ranked dimensions. Based on the mapping relationship between dimensions and original features recorded during training, map the top-ranked dimensions back to their original feature names to obtain the churn cause description. The churn cause description contains several original feature names and their corresponding attention weight values.

[0063] Step S153: Perform churn stage parsing processing on the churn tendency identification result corresponding to each credit customer identifier in the candidate credit customer identifier set to obtain the churn stage description corresponding to each credit customer identifier.

[0064] For each credit customer identifier in the candidate credit customer identifier set, extract the last fund flow stage unit H11 from the fund flow stage segmentation result of that credit customer and the last state node O11 in the contract state transition chain O15. Compare the timestamp corresponding to the stage start record position H11s of the last fund flow stage unit H11 with the dwell time list of the last state node O11 to determine the churn stage description. The churn stage description is defined as early churn, mid-churn, or late churn. The churn stage description is determined based on the combination relationship between the average fund flow amount M1 within the last fund flow stage unit H11 and the maximum dwell time in the dwell time list corresponding to the last state node O11.

[0065] Step S154: Perform customer-dimensional grouping processing on the candidate credit customer identifier set, the description of the churn cause corresponding to each credit customer identifier, and the description of the churn stage corresponding to each credit customer identifier to obtain a credit customer churn profile set.

[0066] For each credit customer identifier in the candidate credit customer identifier set, a triple is formed by combining the credit customer identifier, the description of the churn cause corresponding to the credit customer identifier, and the description of the churn stage corresponding to the credit customer identifier. All triples of credit customer identifiers constitute the credit customer churn profile set. Each element in the credit customer churn profile set is a credit customer churn profile.

[0067] Step S155: Perform intervention urgency stratification processing on each credit customer churn profile in the credit customer churn profile set to obtain an intervention urgency description corresponding to each credit customer churn profile.

[0068] For each credit customer churn profile in the credit customer churn profile set, extract the original feature name with the highest attention weight from the churn stage description and churn cause description. Based on the combination of the churn stage description and the original feature name with the highest attention weight, the intervention urgency is divided into several levels. The intervention urgency is described as high urgency, medium urgency, or low urgency. The specific stratification rules are as follows: for profiles where the churn stage description is late churn stage and the original feature name with the highest attention weight belongs to the cash flow related feature, the intervention urgency is described as high urgency; for profiles where the churn stage description is mid-churn stage and the original feature name with the highest attention weight belongs to the channel access related feature, the intervention urgency is described as medium urgency; and for the remaining profiles, the intervention urgency is described as low urgency.

[0069] Step S156: Sort the set of credit customer churn profiles based on the intervention urgency description corresponding to each credit customer churn profile to obtain a sorted set of credit customer churn profiles. Extract credit customer identifiers whose intervention urgency descriptions meet the preset intervention urgency conditions from the sorted set of credit customer churn profiles to obtain the list of credit customers to be intervened.

[0070] For each profile in the credit customer churn profile set, the intervention urgency description is mapped to a ranking priority value. High urgency corresponds to a ranking priority value of 3, medium urgency to a ranking priority value of 2, and low urgency to a ranking priority value of 1. The credit customer churn profile set is sorted from highest to lowest ranking priority value. Profiles with the same ranking priority value are then sorted from highest to lowest churn tendency value. The preset intervention urgency condition is that the intervention urgency description is either high or medium urgency. From the sorted credit customer churn profile set, the credit customer identifiers corresponding to the profiles with the intervention urgency description of high or medium urgency are extracted to form a list of credit customers to be intervened in.

[0071] Step S210: Obtain the set of image data of the business premises and the set of image data of business documents corresponding to the credit customer. The set of image data of the business premises includes multiple frames of real-scene image units of the business premises, and the set of image data of business documents includes multiple frames of scanned image units of business documents.

[0072] In this embodiment, the business premises image data set is collected through compliant business premises monitoring equipment authorized by the credit customer. Each frame of the business premises real-scene image unit is a two-dimensional digital image matrix, containing the number of pixels in the height direction H1, the number of pixels in the width direction W1, and the number of color channels C1. The business document image data set is obtained through scanned copies of business documents uploaded by the credit customer through the bank's designated channels. Each frame of the business document scanned image unit is a two-dimensional digital image matrix, containing the number of pixels in the height direction H2, the number of pixels in the width direction W2, and the number of color channels C2. All business premises real-scene image units are normalized in size, uniformly scaled to a fixed size of H1×W1×C1. All business document scanned image units are normalized in size, uniformly scaled to a fixed size of H2×W2×C2. Image areas involving personal identification information or sensitive business data are blurred.

[0073] Step S220: Perform visual state extraction processing on the image data set of the business premises to obtain a visual state representation set of the business premises. The visual state representation set of the business premises includes the representation of the goods display status, the representation of the personnel activity status, and the representation of the equipment operation status within the business premises.

[0074] Each frame of the real-world image of the business premises from the image dataset is input into the visual feature extraction network. The visual feature extraction network employs a residual convolutional neural network architecture, containing 16 residual blocks and a global average pooling layer. After each frame of the real-world image of the business premises undergoes layer-by-layer convolution and pooling operations in the residual convolutional neural network, a visual feature vector is output at the global average pooling layer. The dimension of this visual feature vector is D1. The visual feature vectors corresponding to all frames in the business premises image dataset are arranged in chronological order to obtain the visual state representation set of the business premises. The visual state representation set of the business premises is spatially divided into goods display area, personnel activity area, and equipment operation area. Feature sub-vectors of the corresponding areas are extracted from the visual feature vectors according to the spatial division to obtain the representations of the goods display state, personnel activity state, and equipment operation state within the business premises, respectively.

[0075] Step S230: Perform business document layout parsing processing on the business document image data set to obtain a business document layout structure representation set, which includes business document field layout representation, business document field content representation, and business document field association representation.

[0076] Each frame of a scanned image of a business invoice from the business invoice image dataset is input into a layout resolution network. The layout resolution network uses a deformable convolutional network to extract layout region features, and then generates candidate field region boxes through a region proposal network. Each candidate field region box corresponds to a business invoice field. Optical character recognition (OCR) is performed on each business invoice field, converting image pixels into a field text sequence. Character embedding mapping is applied to the field text sequence to obtain a field content vector. The spatial coordinates of all field region boxes in the scanned image unit are encoded to obtain a field layout vector. The spatial relative position relationship and textual semantic relationship between any two field region boxes are fused and encoded to obtain an inter-field association vector. The field layout vector, field content vector, and inter-field association vector are concatenated along the feature dimension to obtain the layout structure representation of that frame of the scanned image unit of the business invoice. The layout structure representations of all frames constitute the business invoice layout structure representation set.

[0077] Step S240: Input the visual state representation set of the business premises and the layout structure representation set of the business documents into the loss tendency joint identification model to perform joint encoding processing of business visual state and business document layout, and obtain the joint encoding representation of business visual documents.

[0078] The visual state representation set of the business premises and the layout structure representation set of business documents are respectively input into the two input branches of the churn tendency joint identification model. The visual state representation set of the business premises is mapped by a multilayer perceptron within the input branch to obtain an intermediate representation of the visual state of the business premises. The layout structure representation set of business documents is also mapped by a multilayer perceptron within the input branch to obtain an intermediate representation of the layout structure of business documents. The intermediate representations of the visual state of the business premises and the intermediate representations of the layout structure of business documents are concatenated along the feature dimension to obtain a concatenated representation of the visual and structural representations of the business documents. This concatenated representation of the visual and structural representations of the business documents is input into the joint encoding branch, and after processing by a multilayer perceptron and a self-attention layer, a joint encoded representation of the visual and structural representations of the business documents is obtained.

[0079] Step S250: The joint coding representation of the business visual documents and the churn tendency identification result of the credit customers are fused together to obtain the churn tendency identification result of the integrated business visual document information. Based on the churn tendency identification result of the integrated business visual document information, a list of credit customers to be intervened in is delineated from the credit customer group.

[0080] The joint coding representation from the operational visual documents and the joint coding representation from the churn tendency identification results are concatenated along the feature dimension to obtain a fused representation. This fused representation is then input into a fusion classification layer, which contains a fully connected layer and a normalized exponential function, outputting a probability distribution of the fused churn tendency categories. The churn tendency category corresponding to the maximum probability in the probability distribution is taken as the churn tendency identification result from the fused operational visual document information. The churn tendency identification result from the fused operational visual document information replaces the original churn tendency identification result. Following steps S150 to S156, a list of credit customers requiring intervention is drawn from the credit customer group.

[0081] Step S310: Obtain the post-loan business call voice data set and the post-loan business text interaction data set corresponding to the credit customer. The post-loan business call voice data set contains multiple post-loan business call voice segment units, and the post-loan business text interaction data set contains multiple post-loan business text interaction message units.

[0082] In this embodiment, the post-loan business call voice data set is collected through the bank's compliant outbound telephone system. Each post-loan business call voice segment unit includes a voice waveform sampling sequence and a call timestamp. The post-loan business text interaction data set is collected through the bank's official online customer service system. Each post-loan business text interaction message unit includes the text message content and a message sending timestamp. The voice waveform sampling sequence undergoes pre-emphasis and frame segmentation processing, with each frame containing a fixed number of sampling points. The text message content undergoes word segmentation and illegal character filtering processing. Text content involving personal identification information is replaced.

[0083] Step S320: Perform call prosody parsing on the post-loan business call voice data set to obtain a call voice prosody representation set, which includes call voice speech rate evolution representation, call voice pause distribution representation, and call voice intonation change representation.

[0084] The speech waveform sampling sequence of each post-loan business call segment is input into the prosodic analysis network. The prosodic analysis network uses a bidirectional long short-term memory network to extract speech frame-level features. The frame-level features are arranged in chronological order to obtain the speech frame-level feature sequence. A first-order difference operation is performed on the speech frame-level feature sequence to obtain the speech rate variation feature sequence. The speech rate variation feature sequence is segmented and statistically analyzed in chronological order. The mean speech rate and standard deviation of each segment are concatenated to form a speech rate evolution representation of the call. Continuous frames with energy values ​​below a preset energy threshold in the speech frame-level feature sequence are merged, and the duration and position of each pause are statistically analyzed. The histogram of pause duration distribution and the histogram of pause position distribution are concatenated to form a pause distribution representation of the call. The fundamental frequency trajectory in the speech frame-level feature sequence is extracted, and trend fitting is performed on the fundamental frequency trajectory. The fitting slope and fitting residual are concatenated to form a pitch variation representation of the call. The prosodic representation of a segment is obtained by concatenating the representations of speech rate evolution, pause distribution, and intonation changes along the feature dimensions. The prosodic representations of all segments constitute the set of prosodic representations of spoken speech.

[0085] Step S330: Perform text interaction semantic parsing on the post-loan business text interaction data set to obtain a text interaction semantic representation set, which includes text interaction topic evolution representation, text interaction expression tendency representation, and text interaction response interval representation.

[0086] The text message content of each post-loan business text interaction message unit in the post-loan business text interaction dataset is input into a text semantic coding network. The text semantic coding network uses a pre-trained text encoder to map the text message content into text semantic vectors. All text message content is sorted in ascending order by message sending timestamp, resulting in a text semantic vector sequence. A sliding window topic clustering is performed on the text semantic vector sequence; the cluster center vector of the text semantic vectors within each window constitutes the topic representation for that window. The topic representations of all windows, arranged chronologically, constitute the text interaction topic evolution representation. Sentiment polarity classification is performed on the text semantic vector sequence; the category distribution vector of each message constitutes the expression tendency representation for that message. The expression tendency representations of all messages constitute the text interaction expression tendency representation. For every two adjacent messages in the text semantic vector sequence, the difference in message sending timestamps is calculated to obtain the response interval value. The distribution histogram of all response interval values ​​constitutes the text interaction response interval representation. The text interaction topic evolution representation, text interaction expression tendency representation, and text interaction response interval representation are concatenated along the feature dimensions to obtain a set of text interaction semantic representations.

[0087] Step S340: Input the call speech prosody representation set and the text interaction semantic representation set into the churn tendency joint identification model for call prosody and text semantic joint encoding processing to obtain the call text joint encoding representation.

[0088] The call speech prosodic representation set and the text interaction semantic representation set are respectively input into the two input branches of the churn tendency joint recognition model. The call speech prosodic representation set is mapped by a recurrent neural network within the input branch to obtain an intermediate call speech prosodic representation. The text interaction semantic representation set is also mapped by a recurrent neural network within the input branch to obtain an intermediate text interaction semantic representation. The intermediate call speech prosodic representation and the intermediate text interaction semantic representation are concatenated along the feature dimension to obtain the concatenated call text representation. The concatenated call text representation is input into the joint encoding branch, and after processing by a multilayer perceptron and a self-attention layer, the joint encoded call text representation is obtained.

[0089] Step S350: The joint encoding representation of the call text and the churn tendency identification result of the credit customer are fused together to obtain the churn tendency identification result of the fused call text information. Based on the churn tendency identification result of the fused call text information, a list of credit customers to be intervened in is delineated from the credit customer group.

[0090] The joint encoded representation of the call text and the joint encoded representation of the churn tendency identification result are concatenated along the feature dimension to obtain a fused representation. This fused representation is input into a fusion classification layer, which contains a fully connected layer and a normalized exponential function, and outputs a probability distribution of the fused churn tendency categories. The churn tendency category corresponding to the maximum probability in the probability distribution is taken as the churn tendency identification result of the fused call text information. The churn tendency identification result of the fused call text information replaces the original churn tendency identification result, and following steps S150 to S156, a list of credit customers to be intervened with is drawn from the credit customer group.

[0091] Step S410: Obtain the post-loan business processing video data set and the post-loan business processing interface screenshot data set corresponding to the credit customer. The post-loan business processing video data set contains multiple post-loan business processing video clip units, and the post-loan business processing interface screenshot data set contains multiple post-loan business processing interface screenshot units.

[0092] In this embodiment, the post-loan business processing video data set is collected through screen recording on the bank's compliant business processing terminal. Each post-loan business processing video segment unit includes a continuous frame image sequence and a frame timestamp. The post-loan business processing interface screenshot data set is collected through timed screenshots on the bank's compliant business processing terminal. Each post-loan business processing interface screenshot unit is a two-dimensional digital image matrix. Frame extraction processing is performed on the continuous frame image sequence, with a fixed extraction interval. All frame images are normalized in size. Interface areas involving personal identification information or sensitive account information are blurred.

[0093] Step S420: Perform business processing operation trajectory parsing processing on the post-loan business processing video data set to obtain a business processing operation trajectory representation set, which includes business processing operation path representation, business processing operation pause representation, and business processing operation repetition representation.

[0094] The continuous frame image sequence of each post-loan business processing video segment is input into the operation trajectory parsing network. The operation trajectory parsing network uses a 3D convolutional neural network to extract spatiotemporal features, and then uses an object detection branch to locate the cursor position in each frame, obtaining a cursor position coordinate sequence. The cursor position coordinate sequence is arranged chronologically to obtain the operation path coordinate sequence. This operation path coordinate sequence is input into a path encoding network to obtain the business processing operation path representation. Time periods where the cursor position coordinates remain unchanged across multiple consecutive frames in the cursor position coordinate sequence are merged, and the duration and position of each pause are calculated to obtain the business processing operation pause representation. Repeated subsequences are detected in the operation path coordinate sequence, and the frequency and type of repeated subsequences are counted to obtain the business processing operation repetition representation. The business processing operation path representation, business processing pause representation, and business processing repetition representation are concatenated along the feature dimension to obtain the operation trajectory representation for that segment. The operation trajectory representations of all segments constitute the business processing operation trajectory representation set.

[0095] Step S430: Perform visual state parsing processing on the set of screenshots of the post-loan business processing interface to obtain a set of visual state representations of the screenshots. The set of visual state representations of the screenshots includes the layout state representation of the screenshots, the content state representation of the screenshots, and the switching state representation of the screenshots.

[0096] Each frame of the post-loan business processing interface screenshot is input into the interface visual state analysis network. The interface visual state analysis network uses a convolutional neural network to extract the interface screenshot feature map. The interface screenshot feature map is divided into spatial grids, and the mean feature value within each grid constitutes the layout feature of that grid. The layout features of all grids are arranged according to their spatial position to form the interface screenshot layout state representation. Global average pooling is performed on the interface screenshot feature map to obtain the interface screenshot content state representation. For two temporally adjacent interface screenshot units, the degree of difference between their feature maps is calculated, and the difference values ​​are arranged in chronological order to form the interface screenshot switching state representation. The interface screenshot layout state representation, interface screenshot content state representation, and interface screenshot switching state representation are concatenated along the feature dimension to obtain the interface screenshot visual state representation set.

[0097] Step S440: Input the business processing operation trajectory representation set and the interface screenshot visual state representation set into the loss tendency joint identification model for joint encoding processing of operation trajectory and interface state to obtain the joint encoding representation of operation interface.

[0098] The set of business processing operation trajectory representations and the set of interface screenshot visual state representations are respectively input into the two input branches of the churn tendency joint identification model. The set of business processing operation trajectory representations is mapped by a recurrent neural network within the input branch to obtain an intermediate representation of the business processing operation trajectory. The set of interface screenshot visual state representations is also mapped by a recurrent neural network within the input branch to obtain an intermediate representation of the interface screenshot visual state. The intermediate representations of the business processing operation trajectory and the intermediate representations of the interface screenshot visual state are concatenated along the feature dimension to obtain the concatenated representation of the operation interface. This concatenated representation of the operation interface is input into the joint encoding branch, and after processing by a multilayer perceptron and a self-attention layer, the joint encoded representation of the operation interface is obtained.

[0099] Step S450: The joint encoding representation of the operation interface and the churn tendency identification result of the credit customer are fused together to obtain the churn tendency identification result of the fused operation interface information. Based on the churn tendency identification result of the fused operation interface information, a list of credit customers to be intervened is delineated from the credit customer group.

[0100] The joint encoded representation of the user interface and the joint encoded representation of the churn tendency identification result are concatenated along the feature dimension to obtain a fused representation. This fused representation is input into a fusion classification layer, which contains a fully connected layer and a normalized exponential function, and outputs a probability distribution of the fused churn tendency categories. The churn tendency category corresponding to the maximum probability in the probability distribution is taken as the churn tendency identification result of the fused user interface information. The churn tendency identification result of the fused user interface information replaces the original churn tendency identification result, and following steps S150 to S156, a list of credit customers to be intervened with is drawn from the credit customer group.

[0101] Step S510: Obtain the post-loan location trajectory data set and the post-loan address book data set corresponding to the credit customer. The post-loan location trajectory data set contains multiple location trajectory point units, and the post-loan address book data set contains multiple address book record units.

[0102] In this embodiment, the post-loan location trajectory data set is collected through a location service data interface authorized by the loan customer. Each location trajectory point unit includes longitude coordinates, latitude coordinates, and a location timestamp. The post-loan address book data set is collected through the address information change records of the loan customer in the bank's system. Each address book record unit includes an address type identifier, an address change type identifier, and an address change timestamp. The location trajectory point units are processed by regional rasterization, mapping the longitude and latitude coordinates to geographic raster numbers of a preset size. The detailed address text in the address book record units is hierarchically processed, retaining only administrative division information at the district / county level or above.

[0103] Step S520: Perform location trajectory dwelling parsing processing on the post-loan location trajectory data set to obtain a location trajectory dwelling representation set, which includes location trajectory dwelling area representation, location trajectory dwelling duration representation, and location trajectory migration path representation.

[0104] The location trajectory points in the post-loan location trajectory dataset are sorted in ascending order by location timestamp. For two adjacent location trajectory points, the raster distance between their geographic raster numbers is calculated. Consecutive location trajectory points with a raster distance less than a preset raster distance threshold and a time interval less than a preset time interval threshold are merged into a single dwell segment. For each dwell segment, the geographic raster number that appears most frequently within the segment is used as the dwell area representation. For each dwell segment, the difference between the last location timestamp and the first location timestamp within the segment is used as the dwell duration representation. For two adjacent dwell segments, the dwell area representations of the preceding and following dwell segments are extracted to form migration path pairs. All migration path pairs constitute the location trajectory migration path representation. The location trajectory dwell area representation, location trajectory dwell duration representation, and location trajectory migration path representation are concatenated along the feature dimension to obtain the location trajectory dwell representation set.

[0105] Step S530: Perform address book change parsing processing on the post-loan address book data set to obtain an address book change representation set, which includes address book new record representation, address book deleted record representation, and address book modified record representation.

[0106] The address book records in the post-loan address book dataset are sorted in ascending order by address change timestamp. These records are then categorized into three types based on the address change type: new, deleted, and modified. For new records, the distribution of address type identifiers is statistically analyzed, and this distribution vector represents the new record. For deleted records, the distribution of address type identifiers is statistically analyzed, and this distribution vector represents the deleted record. For modified records, the changes in address type identifiers before and after the modification are statistically analyzed, and this change type distribution vector represents the modified record. The new, deleted, and modified record representations are concatenated along the feature dimension to obtain the address book change representation set.

[0107] Step S540: Input the location trajectory dwelling representation set and the address book change representation set into the loss tendency joint identification model to perform joint encoding processing of location trajectory and address book change to obtain the location address joint encoding representation.

[0108] The location trajectory dwelling representation set and the address book change representation set are respectively input into the two input branches of the churn tendency joint identification model. The location trajectory dwelling representation set is mapped by a recurrent neural network within the input branch to obtain an intermediate location trajectory dwelling representation. The address book change representation set is also mapped by a recurrent neural network within the input branch to obtain an intermediate address book change representation. The intermediate location trajectory dwelling representation and the intermediate address book change representation are concatenated along the feature dimension to obtain a concatenated location-address representation. This concatenated location-address representation is input into the joint encoding branch, and after processing by a multilayer perceptron and a self-attention layer, a joint location-address encoded representation is obtained.

[0109] Step S550: The location address joint encoding representation and the credit customer churn tendency identification result are fused together to obtain the churn tendency identification result with fused location address information. Based on the churn tendency identification result with fused location address information, a list of credit customers to be intervened in is delineated from the credit customer group.

[0110] The joint encoding representation of location address and the joint encoding representation of churn tendency identification result are concatenated along the feature dimension to obtain a fused representation. This fused representation is input into a fusion classification layer, which contains a fully connected layer and a normalized exponential function, and outputs a probability distribution of the fused churn tendency categories. The churn tendency category corresponding to the maximum probability in the probability distribution is taken as the churn tendency identification result of the fused location address information. The churn tendency identification result of the fused location address information replaces the original churn tendency identification result, and following steps S150 to S156, a list of credit customers to be intervened with is drawn from the credit customer group.

[0111] For example, step S610: Obtain the post-loan third-party public information data set corresponding to the credit customer, wherein the post-loan third-party public information data set includes third-party public text information units and third-party public image information units.

[0112] In this embodiment, the post-loan third-party publicly available information dataset is collected through a compliant third-party publicly available information platform. The third-party publicly available text information unit includes the publicly available text title, the publicly available text body, and the publicly available text publication timestamp. The third-party publicly available image information unit includes the publicly available image description text, the publicly available image publication timestamp, and the publicly available image source identifier. The publicly available text body undergoes word segmentation and sensitive information filtering. The publicly available images undergo source legality verification and image content compliance filtering.

[0113] Step S620: Perform public text information parsing processing on the third-party public text information units in the post-loan third-party public information data set to obtain a public text information representation set. The public text information representation set includes public text information topic representation, public text information release time representation, and public text information release source representation.

[0114] The title and body of each publicly available third-party text information unit are input into a text topic encoding network. The text topic encoding network uses a pre-trained text encoder to map the title and body of the publicly available text into title vectors and body vectors, respectively. These vectors are then concatenated along the feature dimensions to obtain the topic representation of the publicly available text. The publication timestamp of the publicly available text is mapped into a time feature vector to obtain the publication time representation. The publication source identifier of the publicly available text information unit is mapped into a source feature vector to obtain the publication source representation. The topic representation, publication time representation, and publication source representation of the publicly available text information are then concatenated along the feature dimensions to obtain the overall representation of the publicly available text information. All the publicly available text information representations constitute the publicly available text information representation set.

[0115] Step S630: Perform public image information parsing processing on the third-party public image information units in the post-loan third-party public information data set to obtain a public image information representation set. The public image information representation set includes public image information scene representation, public image information associated object representation, and public image information release time representation.

[0116] Each publicly available third-party image information unit's publicly available image is input into an image scene encoding network. The image scene encoding network uses a convolutional neural network to extract image scene feature maps. Global average pooling is applied to the image scene feature maps to obtain the publicly available image information scene representation. Object detection is performed on the image scene feature maps to identify the main objects in the image, and the category distribution vector of the main objects is used as the associated object representation of the publicly available image information. The publicly available image release timestamp is mapped to a time feature vector to obtain the publicly available image information release time representation. The publicly available image information scene representation, the publicly available image information associated object representation, and the publicly available image information release time representation are concatenated along the feature dimension to obtain the representation of the entire publicly available image information unit. All publicly available image information representations constitute the publicly available image information representation set.

[0117] Step S640: Input the publicly available text information representation set and the publicly available image information representation set into the loss tendency joint identification model for joint encoding processing of publicly available text and publicly available images to obtain the joint encoded representation of publicly available information.

[0118] The publicly available text information representation set and the publicly available image information representation set are respectively input into the two input branches of the churn tendency joint identification model. The publicly available text information representation set is mapped through an attention pooling layer within the input branch to obtain an intermediate publicly available text information representation. The publicly available image information representation set is also mapped through an attention pooling layer within the input branch to obtain an intermediate publicly available image information representation. The intermediate publicly available text information representation and the intermediate publicly available image information representation are concatenated along the feature dimension to obtain the concatenated publicly available information representation. This concatenated publicly available information representation is input into the joint encoding branch, and after processing by a multilayer perceptron and a self-attention layer, the jointly encoded publicly available information representation is obtained.

[0119] Step S650: The joint encoding representation of the public information and the churn tendency identification result of the credit customer are fused together to obtain the churn tendency identification result of the third-party public information. Based on the churn tendency identification result of the third-party public information, a list of credit customers to be intervened is delineated from the credit customer group.

[0120] The joint encoded representation of publicly available information and the joint encoded representation of churn tendency identification results are concatenated along the feature dimension to obtain a fused representation. This fused representation is input into a fusion classification layer, which contains a fully connected layer and a normalized exponential function, and outputs a probability distribution of churn tendency categories. The churn tendency category corresponding to the maximum probability in the probability distribution is taken as the churn tendency identification result fused with third-party publicly available information. The original churn tendency identification result is replaced with the churn tendency identification result fused with third-party publicly available information. Following steps S150 to S156, a list of credit customers to be intervened with is drawn from the credit customer group.

[0121] Step S710: Obtain the set of post-loan business premises monitoring video data corresponding to the credit customer, wherein the set of post-loan business premises monitoring video data includes multiple business premises monitoring video clip units.

[0122] In this embodiment, the post-loan business premises monitoring video data set is collected through compliant business premises monitoring equipment authorized by the loan customer. Each business premises monitoring video segment unit contains a continuous frame image sequence and a frame timestamp. Frame extraction processing is performed on the continuous frame image sequence. Size normalization processing is performed on all frame images. Areas involving personally identifiable information are blurred.

[0123] Step S720: Perform monitoring image parsing on the post-loan business premises monitoring video data set to obtain a business premises monitoring image representation set, which includes representations of changes in the display of goods in the business premises, changes in the entry and exit of personnel in the business premises, and changes in the operation of equipment in the business premises.

[0124] The continuous frame sequence of each segment of the surveillance video footage from the business premises is input into the surveillance video parsing network. The network employs a 3D convolutional neural network to extract spatiotemporal features. These features are then divided into goods display areas based on spatial location, and the spatiotemporal feature change trajectories within these areas are extracted to obtain a representation of changes in goods display. Similarly, the network divides personnel activity areas based on spatial location, and the spatiotemporal feature change trajectories within these areas are extracted to obtain a representation of changes in personnel entering and exiting the business premises. Finally, the network divides equipment operation areas based on spatial location, and the spatiotemporal feature change trajectories within these areas are extracted to obtain a representation of changes in equipment operation. These representations are then concatenated along the feature dimensions to obtain a set of surveillance video representations of the business premises.

[0125] Step S730: Obtain the post-loan operating document image data set corresponding to the credit customer, wherein the post-loan operating document image data set contains multiple frames of operating document image units.

[0126] In this embodiment, the post-loan operating document image data set is obtained through scanned copies of operating documents uploaded by loan customers through designated bank channels. Each frame of the operating document image unit is a two-dimensional digital image matrix. Size normalization is performed on all frames of operating document image units. Image regions involving sensitive operating data are blurred.

[0127] Step S740: Perform business document layout parsing processing on the post-loan business document image data set to obtain a business document layout representation set, which includes business document field distribution representation, business document field content representation, and business document field association representation.

[0128] Each frame of the business document image is input into the layout resolution network. The layout resolution network uses a deformable convolutional network to extract layout region features, and then generates candidate field region boxes through a region proposal network. Each candidate field region box corresponds to a business document field. Optical character recognition is performed on each business document field to obtain a field text sequence. Character embedding mapping is performed on the field text sequence to obtain a field content vector. The spatial coordinates of all field region boxes in the business document image unit are encoded to obtain a field distribution vector. The spatial relative position relationship and textual semantic relationship between any two field region boxes are fused and encoded to obtain a field association vector. The field distribution vector, field content vector, and field association vector are concatenated along the feature dimension to obtain the layout representation of the business document image unit of that frame. The layout representations of all frames constitute the business document layout representation set.

[0129] Step S750: Input the set of representations of the monitoring screen of the business premises and the set of representations of the business document layout into the multimodal large model for joint understanding of business vision to obtain joint understanding representation of business vision.

[0130] The sets of representations of the business premises monitoring screen and the sets of representations of business documents are input into the two modal input branches of the multimodal large model, respectively. The set of representations of the business premises monitoring screen is mapped by a video modal encoder to obtain an intermediate representation of the monitoring screen. The set of representations of the business documents is mapped by an image modal encoder to obtain an intermediate representation of the document layout. The intermediate representations of the monitoring screen and the intermediate representations of the document layout are concatenated along the feature dimension to obtain a joint visual representation of the business premises. This joint visual representation of the business premises is input into a cross-modal attention layer, which calculates the attention weights between the intermediate representations of the monitoring screen and the intermediate representations of the document layout, and outputs the joint visual representation of the business premises.

[0131] Step S760: Perform stage division and operational visual comparison processing on the operational visual joint understanding representation and the results of the capital flow stage division to obtain the operational visual comparison representation of the capital flow stage.

[0132] The timestamps corresponding to the start and end records (H11s and H11e) of each cash flow stage unit H11 in the joint understanding and representation of business operations and the segmentation of cash flow stages are time-aligned. Segments in the joint understanding and representation of business operations whose timestamps fall within the time range of each cash flow stage unit H11 are compared with the cash flow stage unit H11. The compared representations are then stitched together along the feature dimensions to obtain the business visual comparison representation of the cash flow stage for that cash flow stage unit H11. The business visual comparison representations of all cash flow stage units H11 constitute the set of business visual comparison representations of the cash flow stages.

[0133] Step S770: Perform stage division and operational visual comparison processing on the operational visual joint understanding representation and the channel access stage division results to obtain the operational visual comparison representation of the channel access stage.

[0134] The timestamps corresponding to the start and end access record positions H12s and H12e of each channel access stage unit H12 in the joint understanding and representation of the operational visuals are time-aligned. Segments in the joint understanding and representation of the operational visuals that fall within the time range of each channel access stage unit H12 are compared with the channel access stage unit H12. The compared representations are then concatenated along the feature dimensions to obtain the operational visual comparison representation of the channel access stage for that channel access stage unit H12. The operational visual comparison representations of all channel access stage units H12 constitute the set of operational visual comparison representations for the channel access stages.

[0135] Step S780: Input the operational visual comparison representation of the capital flow stage and the operational visual comparison representation of the channel access stage into the churn tendency joint identification model for operational visual fusion churn identification processing to obtain the churn tendency identification result of the fused operational visuals, and delineate the list of credit customers to be intervened from the credit customer group based on the churn tendency identification result of the fused operational visuals.

[0136] The sets of visual representations for the cash flow stage and the channel access stage are respectively input into the two input branches of the joint identification model for churn tendency. The set of visual representations for the cash flow stage is mapped by a recurrent neural network within the input branch to obtain an intermediate visual representation for the cash flow stage. The set of visual representations for the channel access stage is also mapped by a recurrent neural network within the input branch to obtain an intermediate visual representation for the channel access stage. The intermediate visual representations for the cash flow stage and the intermediate visual representations for the channel access stage are concatenated along the feature dimension to obtain a fused visual representation. This fused visual representation is input into the joint encoding branch and processed by a multilayer perceptron and a self-attention layer to obtain a fused visual joint encoding representation. This fused visual joint encoding representation is input into the classification output layer, which outputs the probability distribution of churn tendency categories. The churn tendency category corresponding to the maximum probability in the probability distribution is taken as the churn tendency identification result of the fused visual representation. Following steps S150 to S156, a list of credit customers to be intervened with is selected from the credit customer group.

[0137] Step S810: Obtain the post-loan business call voice data set corresponding to the credit customer, wherein the post-loan business call voice data set contains multiple business call voice segment units.

[0138] In this embodiment, the post-loan business call voice data set is collected through the bank's compliant outbound telephone system. Each business call voice segment unit includes a voice waveform sampling sequence and a call timestamp. The voice waveform sampling sequence undergoes pre-emphasis and frame segmentation. Voice content involving personal identification information is filtered.

[0139] Step S820: Perform call voice content transcription processing on the post-loan business call voice data set to obtain a call voice transcription text set.

[0140] The speech waveform sampling sequence of each business call voice segment is input into the speech recognition network. The speech recognition network adopts an end-to-end speech recognition architecture, including an acoustic encoder and a text decoder. The acoustic encoder maps the speech waveform sampling sequence into an acoustic feature sequence. The text decoder decodes the acoustic feature sequence frame by frame into a text character sequence. The text character sequences are concatenated to form a complete call speech transcription text. The call speech transcription texts of all segments constitute the call speech transcription text set.

[0141] Step S830: Perform call text semantic parsing processing on the call speech-to-text set to obtain a call text semantic representation set, which includes call topic evolution representation, call request type representation and call emotional tendency representation.

[0142] Each segment of the call speech-to-text dataset is input into a text semantic encoding network. The network uses a pre-trained text encoder to map each segment into a text semantic vector. All segments are sorted in ascending order by call timestamp, resulting in a sequence of text semantic vectors. A sliding window topic clustering method is applied to this sequence, with the cluster center vector within each window constituting the topic representation for that window. The topic representations of all windows are arranged chronologically to form the call topic evolution representation. The sequence is then categorized by appeal type, with the appeal type distribution vector for each segment constituting the call appeal type representation. Finally, the sequence is categorized by sentiment polarity, with the sentiment polarity distribution vector for each segment constituting the call sentiment tendency representation. The call topic evolution representation, appeal type representation, and call sentiment tendency representation are concatenated along the feature dimensions to obtain a set of call text semantic representations.

[0143] Step S840: Perform call prosody parsing on the post-loan business call voice data set to obtain a call voice prosody representation set, which includes call speed variation representation, call pause distribution representation, and call tone variation representation.

[0144] The speech waveform sampling sequence of each business call speech segment is input into the prosodic analysis network. The prosodic analysis network uses a bidirectional long short-term memory network to extract speech frame-level features. First-order difference operations are performed on the speech frame-level feature sequence to obtain the speech rate variation feature sequence. The speech rate variation feature sequence is segmented and statistically analyzed in chronological order, and the mean speech rate and standard deviation of each segment are concatenated to form a speech rate variation representation. Consecutive frames in the speech frame-level feature sequence with energy values ​​below a preset energy threshold are merged, and the duration and position of each pause are statistically analyzed. The histogram of pause duration distribution and the histogram of pause position distribution are concatenated to form a speech pause distribution representation. The fundamental frequency trajectory in the speech frame-level feature sequence is extracted, and trend fitting is performed on the fundamental frequency trajectory. The fitting slope and fitting residual are concatenated to form a speech tone variation representation. The speech rate variation representation, the speech pause distribution representation, and the speech tone variation representation are concatenated along the feature dimension to obtain the prosodic representation of that segment. The prosodic representations of all segments constitute the set of speech prosodic representations.

[0145] Step S850: Input the semantic representation set of the call text and the prosodic representation set of the call speech into a multimodal large model for joint understanding of the call speech, and obtain the joint understanding representation of the call speech.

[0146] The semantic representation set of the call text and the prosodic representation set of the call speech are respectively input into the two modal input branches of the multimodal large model. The semantic representation set of the call text is mapped by the text modality encoder to obtain the intermediate semantic representation of the call text. The prosodic representation set of the call speech is mapped by the speech modality encoder to obtain the intermediate prosodic representation of the call speech. The intermediate semantic representation of the call text and the intermediate prosodic representation of the call speech are concatenated along the feature dimension to obtain the joint concatenated representation of the call speech. The joint concatenated representation of the call speech is input into the cross-modal attention layer, which calculates the attention weights between the intermediate semantic representation of the call text and the intermediate prosodic representation of the call speech, and outputs the joint understanding representation of the call speech.

[0147] Step S860: Perform state transition and call voice comparison processing on the joint understanding representation of the call voice and the contract state transition chain to obtain the contract state transition call voice comparison representation.

[0148] The joint unification representation of the call audio is time-aligned with the dwell time list of each state node O11 in the contract state transition chain O15. Segments in the joint unification representation of the call audio whose timestamps fall within the dwell time range of each state node O11 are compared with state node O11. The compared representations are then concatenated along the feature dimension to obtain the contract state transition call audio comparison representation for that state node O11. The contract state transition call audio comparison representations of all state nodes O11 constitute the contract state transition call audio comparison representation set.

[0149] Step S870: Input the contract state transition call voice comparison representation into the churn tendency joint identification model for call voice fusion churn identification processing to obtain the churn tendency identification result of the fused call voice, and delineate the list of credit customers to be intervened from the credit customer group based on the churn tendency identification result of the fused call voice.

[0150] The contract state transition call voice reference set is input into the churn tendency joint identification model. The contract state transition call voice reference set is processed through a graph attention layer within the model to aggregate information between nodes, resulting in a fused joint coding representation of the call voice. This fused joint coding representation is input into a classification output layer, which outputs a probability distribution of churn tendency categories in the fused call voice. The churn tendency category corresponding to the maximum probability in the probability distribution is taken as the churn tendency identification result of the fused call voice. Following steps S150 to S156, a list of credit customers to be intervened with is selected from the credit customer group.

[0151] Step S910: Obtain the post-loan business processing interface operation video data set corresponding to the credit customer, wherein the post-loan business processing interface operation video data set contains multiple business processing interface operation video segment units.

[0152] In this embodiment, the video data set of post-loan business processing interface operations is collected through screen recording of the bank's compliant business processing terminal. Each video segment of the business processing interface operation unit includes a continuous frame image sequence and a frame timestamp. Frame extraction processing is performed on the continuous frame image sequence. Size normalization processing is performed on all frame images. Interface areas involving personal identification information are blurred.

[0153] Step S920: Perform interface operation trajectory parsing processing on the video data set of the post-loan business processing interface to obtain an interface operation trajectory representation set, which includes interface operation path representation, interface operation pause representation, and interface operation rollback representation.

[0154] The continuous frame image sequence of each business processing interface operation video segment is input into the operation trajectory parsing network. The operation trajectory parsing network uses a 3D convolutional neural network to extract spatiotemporal features, and then uses an object detection branch to locate the cursor position in each frame, obtaining a cursor position coordinate sequence. The cursor position coordinate sequence is arranged in chronological order to obtain the interface operation path coordinate sequence. The interface operation path coordinate sequence is input into a path encoding network to obtain the interface operation path representation. Time periods in the cursor position coordinate sequence where the cursor position coordinates remain unchanged across multiple consecutive frames are merged, and the duration and position of each pause are counted to obtain the interface operation pause representation. Backtracking actions are detected in the interface operation path coordinate sequence, and the frequency and type of backtracking actions are counted to obtain the interface operation backtrack representation. The interface operation path representation, interface operation pause representation, and interface operation backtrack representation are concatenated along the feature dimension to obtain the interface operation trajectory representation for that segment. The interface operation trajectory representations of all segments constitute the interface operation trajectory representation set.

[0155] Step S930: Obtain a set of screenshots of the post-loan business processing interface corresponding to the credit customer, wherein the set of screenshots of the post-loan business processing interface includes multiple frames of business processing interface screenshots.

[0156] In this embodiment, the screenshot data set of the post-loan business processing interface is collected through timed screenshots taken from the bank's compliant business processing terminal. Each frame of the business processing interface screenshot unit is a two-dimensional digital image matrix. The size of all frame screenshot units is normalized. Interface areas involving personal identification information are blurred.

[0157] Step S940: Perform interface visual state parsing processing on the set of screenshots of the post-loan business processing interface to obtain a set of interface visual state representations. The set of interface visual state representations includes interface layout state representation, interface text state representation, and interface jump state representation.

[0158] Each frame of the business processing interface screenshot is input into the interface visual state analysis network. The interface visual state analysis network uses a convolutional neural network to extract the interface screenshot feature map. The interface screenshot feature map is divided into spatial grids, and the mean feature value within each grid constitutes the layout feature of that grid. The layout features of all grids are arranged spatially to form the interface layout state representation. Optical character recognition is performed on the interface screenshot feature map, converting image pixels into a sequence of interface text. Character embedding mapping is then performed on the interface text sequence to obtain the interface text state representation. For two temporally adjacent interface screenshot units, the degree of difference between their feature maps is calculated, and the difference values ​​are arranged chronologically to form the interface transition state representation. The interface layout state representation, interface text state representation, and interface transition state representation are concatenated along the feature dimension to obtain the interface visual state representation set.

[0159] Step S950: Input the interface operation trajectory representation set and the interface visual state representation set into the multimodal large model for joint understanding processing of the business processing interface to obtain the joint understanding representation of the business processing interface.

[0160] The interface operation trajectory representation set and the interface visual state representation set are respectively input into the two modal input branches of the multimodal large model. The interface operation trajectory representation set is mapped by a trajectory modality encoder to obtain an intermediate representation of the interface operation trajectory. The interface visual state representation set is mapped by a visual modality encoder to obtain an intermediate representation of the interface visual state. The intermediate representations of the interface operation trajectory and the intermediate representations of the interface visual state are concatenated along the feature dimension to obtain a joint concatenated representation of the business processing interface. This joint concatenated representation of the business processing interface is input into a cross-modal attention layer. The cross-modal attention layer calculates the attention weights between the intermediate representations of the interface operation trajectory and the intermediate representations of the interface visual state, and outputs the joint understanding representation of the business processing interface.

[0161] Step S960: Perform a comparison between the business processing interface and the fund flow stage division result to obtain a comparison representation of the business processing interface for the fund flow stage.

[0162] The timestamps corresponding to the start record position H11s and the end record position H11e of each fund flow stage unit H11 in the joint deconstruction representation of the business processing interface and the fund flow stage segmentation results are time-aligned. Segments in the joint deconstruction representation of the business processing interface whose timestamps fall within the time range of each fund flow stage unit H11 are compared with the fund flow stage unit H11. The compared representations are then concatenated along the feature dimensions to obtain the business processing interface comparison representation for that fund flow stage unit H11. The business processing interface comparison representations for all fund flow stage units H11 constitute the set of business processing interface comparison representations for all fund flow stage units H11.

[0163] Step S970: Perform a comparison between the business processing interface and the channel access stage division result by combining the joint understanding and representation of the business processing interface to obtain a comparison representation of the business processing interface in the channel access stage.

[0164] The timestamps corresponding to the start access record position H12s and the end access record position H12e of each channel access stage unit H12 in the joint unification representation of the business processing interface and the channel access stage segmentation results are time-aligned. The segments in the joint unification representation of the business processing interface whose timestamps fall within the time range of each channel access stage unit H12 are compared with the channel access stage unit H12. The compared representations are then concatenated along the feature dimensions to obtain the channel access stage business processing interface comparison representation for that channel access stage unit H12. The channel access stage business processing interface comparison representations of all channel access stage units H12 constitute the channel access stage business processing interface comparison representation set.

[0165] Step S980: Input the business processing interface comparison representation of the fund flow stage and the business processing interface comparison representation of the channel access stage into the churn tendency joint identification model to perform business processing interface fusion churn identification processing, obtain the churn tendency identification result of the fused business processing interface, and delineate the list of credit customers to be intervened from the credit customer group based on the churn tendency identification result of the fused business processing interface.

[0166] The sets of interface representations for the business processing stage during the fund flow stage and the set of interface representations for the business processing stage during the channel access stage are respectively input into the two input branches of the churn tendency joint identification model. The set of interface representations for the business processing stage during the fund flow stage is mapped by a recurrent neural network within the input branch to obtain an intermediate representation of the business processing interface during the fund flow stage. The set of interface representations for the business processing stage during the channel access stage is also mapped by a recurrent neural network within the input branch to obtain an intermediate representation of the business processing interface during the channel access stage. The intermediate representations of the business processing interface during the fund flow stage and the intermediate representations of the business processing interface during the channel access stage are concatenated along the feature dimension to obtain a fused and concatenated representation of the business processing interface. This fused and concatenated representation is input into the joint encoding branch, and after processing by a multilayer perceptron and a self-attention layer, a joint encoded representation of the fused business processing interface is obtained. This joint encoded representation is input into the classification output layer, which outputs the probability distribution of the churn tendency categories of the fused business processing interface. The churn tendency category corresponding to the maximum probability in the probability distribution is taken as the churn tendency identification result of the fused business processing interface. Following steps S150 to S156, a list of credit customers to be intervened with is selected from the credit customer group.

[0167] Step S1010: Obtain the post-loan location trajectory data set corresponding to the credit customer, wherein the post-loan location trajectory data set contains multiple location trajectory point units.

[0168] In this embodiment, the post-loan location trajectory data set is collected through a location service data interface authorized by the loan customer. Each location trajectory point unit contains longitude coordinates, latitude coordinates, and a location timestamp. The location trajectory point units are processed by regional rasterization, mapping the longitude and latitude coordinates to geographic raster numbers of a preset size.

[0169] Step S1020: Perform location trajectory residence area parsing processing on the post-loan location trajectory data set to obtain a location trajectory residence area representation set, which includes residence area distribution representation, residence area duration representation and residence area migration representation.

[0170] The location trajectory points in the post-loan location trajectory dataset are sorted in ascending order by location timestamp. For two adjacent location trajectory points, the raster distance between their geographic raster numbers is calculated. Consecutive location trajectory points with a raster distance less than a preset threshold and a time interval less than a preset time interval threshold are merged into a single dwell segment. For each dwell segment, the geographic raster number with the highest frequency within the segment is used as the dwell area identifier. The dwell area identifiers of all dwell segments are statistically analyzed to obtain a dwell area distribution representation. For each dwell segment, the difference between the last location timestamp and the first location timestamp within the segment is calculated as the dwell duration value; the dwell duration values ​​corresponding to all dwell segments constitute a dwell area duration representation. For two adjacent dwell segments, the dwell area identifiers of the preceding and following segments are extracted to form migration pairs; all migration pairs constitute a dwell area migration representation. The dwell area distribution representation, dwell area duration representation, and dwell area migration representation are concatenated along the feature dimension to obtain a set of location trajectory dwell area representations.

[0171] Step S1030: Obtain the post-loan address book data set corresponding to the credit customer, wherein the post-loan address book data set contains multiple address book record units.

[0172] In this embodiment, the post-loan address book dataset is collected from the address information change records of loan customers in the banking system. Each address book record unit includes an address type identifier, an address change type identifier, and an address change timestamp. The detailed address text in the address book record unit is processed hierarchically, retaining only administrative division information at the district / county level or above.

[0173] Step S1040: Perform address book record change parsing processing on the post-loan address book data set to obtain an address book record change representation set, which includes address book new record representation, address book deleted record representation, and address book modified record representation.

[0174] The address book records in the post-loan address book dataset are sorted in ascending order by address change timestamp. These records are then categorized into three types based on the address change type: new, deleted, and modified. For new records, the distribution of address type identifiers is statistically analyzed, and the distribution vector represents the new record. For deleted records, the distribution of address type identifiers is statistically analyzed, and the distribution vector represents the deleted record. For modified records, the changes in address type identifiers before and after the modification are statistically analyzed, and the change type distribution vector represents the modified record. The new, deleted, and modified record representations are concatenated along the feature dimension to obtain the address book record change representation set.

[0175] Step S1050: Input the location trajectory dwell area representation set and the address book record change representation set into the multimodal large model for joint understanding processing of location trajectory and address book changes to obtain the joint understanding representation of location and address.

[0176] The location trajectory dwelling region representation set and the address book record change representation set are respectively input into the two modal input branches of the multimodal large model. The location trajectory dwelling region representation set is mapped by the location modality encoder to obtain the intermediate representation of the location trajectory dwelling region. The address book record change representation set is mapped by the address modality encoder to obtain the intermediate representation of the address book record change. The intermediate representation of the location trajectory dwelling region and the intermediate representation of the address book record change are concatenated along the feature dimension to obtain the joint location-address representation. The joint location-address representation is input into the cross-modal attention layer, which calculates the attention weights between the intermediate representation of the location trajectory dwelling region and the intermediate representation of the address book record change, and outputs the joint location-address representation.

[0177] Step S1060: Perform state transition and location address comparison processing on the joint deconstruction representation of the location address and the contract state transition chain to obtain the contract state transition location address comparison representation.

[0178] The joint unification representation of location addresses is time-aligned with the residence duration list of each state node O11 in the contract state transition chain O15. Segments of the joint unification representation whose timestamps fall within the residence duration range of each state node O11 are compared with state node O11. The compared representations are then concatenated along the feature dimension to obtain the contract state transition location address mapping representation for that state node O11. The contract state transition location address mapping representations of all state nodes O11 constitute the contract state transition location address mapping representation set.

[0179] Step S1070: Input the contract state migration location address comparison characterization into the churn tendency joint identification model for location address fusion churn identification processing to obtain the churn tendency identification result of the fused location address, and delineate the list of credit customers to be intervened from the credit customer group based on the churn tendency identification result of the fused location address.

[0180] The contract state transition location address mapping representation set is input into the churn tendency joint identification model. This representation set undergoes node-to-node information aggregation via a graph attention layer within the model, yielding a fused joint encoding representation of the location address. This fused joint encoding representation is then input into a classification output layer, which outputs a probability distribution of churn tendency categories for the fused location addresses. The churn tendency category corresponding to the maximum probability in the probability distribution is taken as the churn tendency identification result for the fused location address. Following steps S150 to S156, a list of credit customers requiring intervention is drawn from the credit customer group.

[0181] Step S1110: Obtain the post-loan social media public text data set corresponding to the credit customer, wherein the post-loan social media public text data set contains multiple social media public text units.

[0182] In this embodiment, the post-loan social media public text data set is collected through a compliant public social media platform. Each social media public text unit includes the public text content and a timestamp of the public text publication. The public text content undergoes word segmentation and sensitive information filtering.

[0183] Step S1120: Perform public text theme parsing processing on the post-loan social media public text data set to obtain a public text theme representation set, which includes public text business theme representation, public text recruitment theme representation and public text relocation theme representation.

[0184] The public text content of each social media post is input into a text topic classification network. This network uses a pre-trained text encoder to map the public text content into semantic vectors. These semantic vectors are then input into a topic classification layer, which contains fully connected layers and a normalized exponential function, outputting probabilities for business-related topics, recruitment-related topics, and relocation-related topics. These probabilities are then concatenated to form a public text topic representation. The public text topic representations of all social media posts constitute the public text topic representation set.

[0185] Step S1130: Obtain the post-loan social media public image data set corresponding to the credit customer, wherein the post-loan social media public image data set contains multiple frames of social media public image units.

[0186] In this embodiment, the post-loan social media publicly available image data set is collected through a compliant and publicly available social media platform, and each frame of the social media publicly available image unit is a two-dimensional digital image matrix. All frames of social media publicly available image units are subjected to size normalization processing. Areas involving personally identifiable information are blurred.

[0187] Step S1140: Perform public image scene parsing processing on the post-loan social media public image data set to obtain a public image scene representation set, which includes public image business scene representation, public image personnel scene representation and public image venue scene representation.

[0188] Each frame of publicly shared social media image units is input into an image scene classification network. This network uses a convolutional neural network to extract image scene feature maps. Global average pooling is then applied to these feature maps to obtain image scene vectors. These vectors are then input into a scene classification layer, which contains fully connected layers and a normalized exponential function. The layer outputs probabilities for business scenes, personnel scenes, and venue scenes. These probabilities are concatenated to form a publicly shared image scene representation. The publicly shared image scene representations of all frames of publicly shared social media image units constitute the publicly shared image scene representation set.

[0189] Step S1150: Input the publicly available text topic representation set and the publicly available image scene representation set into a multimodal large model for joint understanding processing of publicly available social media information to obtain a joint understanding representation of publicly available social media information.

[0190] The publicly available text topic representations and publicly available image scene representations are input into the two modal input branches of the multimodal large model, respectively. The publicly available text topic representations are mapped by a text modality encoder to obtain intermediate publicly available text topic representations. The publicly available image scene representations are mapped by an image modality encoder to obtain intermediate publicly available image scene representations. These intermediate publicly available text topic representations and intermediate publicly available image scene representations are concatenated along the feature dimension to obtain a joint concatenated representation of social media public information. This joint concatenated representation of social media public information is then input into a cross-modal attention layer. The cross-modal attention layer calculates the attention weights between the intermediate publicly available text topic representations and the intermediate publicly available image scene representations, outputting a joint understanding representation of social media public information.

[0191] Step S1160: Perform a comparison process between the social media public information joint understanding representation and the fund flow stage segmentation result to obtain a comparison representation of social media public information for fund flow stages.

[0192] The timestamps corresponding to the start record position H11s and the end record position H11e of each fund flow stage unit H11 in the joint understanding representation of social media public information and the fund flow stage segmentation results are time-aligned. Segments in the joint understanding representation of social media public information whose timestamps fall within the time range of each fund flow stage unit H11 are compared with fund flow stage unit H11. The compared representations are then concatenated along the feature dimension to obtain the fund flow stage social media public information comparison representation for that fund flow stage unit H11. All fund flow stage social media public information comparison representations for all fund flow stage units H11 constitute the fund flow stage social media public information comparison representation set.

[0193] Step S1170: Perform state transition and social media public information comparison processing on the joint understanding representation of the social media public information and the contract state transition chain to obtain the contract state transition social media public information comparison representation.

[0194] The joint understanding representation of public social media information is time-aligned with the residence duration list of each state node O11 in the contract state transition chain O15. Segments of the joint understanding representation of public social media information whose timestamps fall within the residence duration range of each state node O11 are compared with state node O11. The compared representations are then concatenated along the feature dimension to obtain the contract state transition public social media information comparison representation for that state node O11. The contract state transition public social media information comparison representations of all state nodes O11 constitute the contract state transition public social media information comparison representation set.

[0195] Step S1180: Input the social media public information comparison representation of the fund flow stage and the social media public information comparison representation of the contract state transition into the churn tendency joint identification model to perform social media public information fusion churn identification processing, obtain the churn tendency identification result of fused social media public information, and delineate the list of credit customers to be intervened from the credit customer group based on the churn tendency identification result of fused social media public information.

[0196] The sets of social media public information comparison representations for the fund flow stage and the sets of social media public information comparison representations for the contract state transition are respectively input into the two input branches of the churn tendency joint identification model. The set of social media public information comparison representations for the fund flow stage is mapped by a recurrent neural network within the input branch to obtain an intermediate representation of social media public information for the fund flow stage. The set of social media public information comparison representations for the contract state transition is processed by a graph attention layer within the input branch to aggregate information between nodes, resulting in an intermediate representation of social media public information for the contract state transition. The intermediate representations of social media public information for the fund flow stage and the intermediate representations of social media public information for the contract state transition are concatenated along the feature dimension to obtain a fused and concatenated representation of social media public information. This fused and concatenated representation of social media public information is input into the joint encoding branch, and after processing by a multilayer perceptron and a self-attention layer, a joint encoded representation of fused social media public information is obtained. This joint encoded representation of fused social media public information is input into the classification output layer, which outputs the probability distribution of churn tendency categories of fused social media public information. The churn tendency category corresponding to the maximum probability in the probability distribution is taken as the churn tendency identification result of fused social media public information. Following steps S150 to S156, a list of credit customers to be intervened in is selected from the credit customer group.

[0197] In practical application scenarios, such as Figure 2 As shown, taking a loan customer operating a community convenience store as an example, during the post-loan management period, real-time monitoring footage of the customer's business premises and images of sales receipts from the POS system were continuously collected and compared with the account's cash flow data recorded by the bank. The cash flow data showed that the customer's account balance had been declining for three consecutive months, and had repeatedly fallen below the monthly loan repayment amount at key points. Combined with the customer's existing loan contract status records, the post-loan management operation formed a credit status migration trajectory from normal to watch status, and then further from watch status to overdue status. Based on this status migration trajectory, the back-end analysis operation calculated the customer's current churn intensity and identified the direct cause of churn (e.g., revenue diversion due to the opening of new stores nearby) and the churn stage (e.g., early churn stage). Subsequently, the above information on churn intensity, churn cause, and churn stage was packaged into a churn identification output file. This output file was distributed to specific workstations in the bank's credit management department through a list distribution process. At the same time, based on the customer's contract expiration date and repayment calendar, the system automatically marks the execution time point for the next post-loan monitoring, ensuring that risk tracking of the customer is continuous and uninterrupted.

[0198] like Figure 3As shown, when determining the churn tendency of the aforementioned convenience store customers through multi-source analysis, not only were their cash flow and credit status considered, but also customer flow footage from the store's surveillance video, recordings of their post-loan business calls with bank customer service, location trajectories when logging into the mobile banking app, and their interface operation paths when conducting business through online banking channels were also incorporated. For example, surveillance video showed recent sparse customer traffic, voice analysis revealed that the customer spoke rapidly and paused frequently, location trajectories indicated that the customer had recently been frequently appearing in cross-city areas different from the loan approval location, and interface operation trajectories showed that the customer repeatedly stopped at the "early settlement" page but did not ultimately submit the request. These four types of information were input in parallel into the multimodal large model joint understanding and processing stage. This stage, through a matrix-gridized feature space, cross-analyzed the video footage, speech prosody, spatiotemporal migration paths, and interface operation habits, ultimately identifying a high churn tendency from the customer's overall performance. The results of identifying customers with high churn potential are presented as a sorted list of customers. The convenience store customer's identifier is highlighted separately in the list and extracted into a folder for the list of customers to be intervened in for credit purposes. The collection and evaluation of multi-source information for this customer is carried out continuously at fixed intervals to ensure that the latest business dynamics and interactive behaviors are included in the analysis in a timely manner.

[0199] like Figure 4 As shown, after the list of credit customers to be intervened with is generated, the bank's credit management department executes specific follow-up actions based on this list. Taking this convenience store customer as an example, the list clearly lists the customer's identifier, the reason for churn (revenue diversion), the stage of churn (early churn), and the urgency of intervention (high priority) in tabular form. Based on the difference in the urgency of intervention, the customer is classified into a high-priority group and assigned to the corresponding credit management workstation. The management personnel receive the task through their office terminal. In the terminal interface, the management personnel contact the customer's registered mobile phone number according to the schedule of the follow-up plan calendar using the outbound calling function. During the call, the management personnel record in detail the customer's feedback on business difficulties and willingness to resolve them in the chat interaction window within the interface. After the call ends, the management personnel enter the follow-up results into the system and dynamically update the status of the customer list.

[0200] Figure 2 The following is a schematic diagram of the hardware structure of a bank credit churn customer identification system 100 provided in this application embodiment for implementing the above-described bank credit churn customer identification method, as shown in the figure. Figure 2 As shown, the bank credit loss customer identification system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0201] In one possible design, the bank credit churn customer identification system 100 can be a single server or a group of servers. The server group can be centralized or distributed (e.g., the bank credit churn customer identification system 100 can be a distributed system). In some embodiments, the bank credit churn customer identification system 100 can be local or remote. For example, the bank credit churn customer identification system 100 can access information and / or data stored in machine-readable storage medium 120 via a network. As another example, the bank credit churn customer identification system 100 can directly connect to machine-readable storage medium 120 to access stored information and / or data. In some embodiments, the bank credit churn customer identification system 100 can be implemented on a server. By way of example only, the server can include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, internal cloud, multi-tiered cloud, etc., or any combination thereof.

[0202] Machine-readable storage medium 120 may store data and / or instructions. In some embodiments, machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, machine-readable storage medium 120 may store data and / or instructions used by the bank credit loss customer identification system 100 to perform or use in order to complete the exemplary methods described in this application.

[0203] In a specific implementation, one or more processors 110 execute computer-executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the bank credit loss customer identification method as described in the above method embodiment. The processor 110, the machine-readable storage medium 120, and the communication unit 140 are connected through the bus 130. The processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0204] The specific implementation process of processor 110 can be found in the various method embodiments executed by the above-mentioned bank credit loss customer identification system 100. The implementation principle and technical effect are similar, and will not be repeated here.

[0205] Furthermore, this application embodiment also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned method for identifying bank credit loss customers is implemented.

[0206] It should be noted that, in order to simplify the description disclosed in this application and thus help to understand one or more embodiments of the invention, the foregoing description of the embodiments of this application may sometimes combine multiple features into one embodiment, drawing or description thereof.

Claims

1. A method for identifying bank customers experiencing credit churn, characterized in that, The method includes: Obtain a set of post-loan behavior data and a set of credit contract attribute data for credit customers. The set of post-loan behavior data includes a sequence of account fund flow records and a sequence of channel access behavior records. The set of credit contract attribute data includes a sequence of credit contract status records. The post-loan behavior data set is divided into behavior evolution stages to obtain the fund flow stage division results and the channel access stage division results. The credit contract attribute data set is processed to construct a contract state transition chain, thereby obtaining the contract state transition chain of the credit customer. The results of the fund flow stage segmentation, the results of the channel access stage segmentation, and the contract state transition chain input churn tendency joint identification model are used to perform churn tendency identification processing to obtain the churn tendency identification results of the credit customer. Based on the churn tendency identification results, a list of credit customers to be intervened in is drawn from the credit customer group.

2. The method for identifying bank credit loss customers according to claim 1, characterized in that, The process of dividing the post-loan behavior data set into behavioral evolution stages yields results in fund flow stage division and channel access stage division, including: The account fund flow record sequence is extracted from the post-loan behavior data set. The changes in the direction of fund flow between adjacent records are captured in the order of time to obtain the fund flow direction change sequence. The fund flow direction change sequence is then merged into consecutive segments in the same direction to obtain multiple fund flow segment units. Each fund flow segment unit carries the segment start record position and the segment end record position. The adjacent segments of the same flow of funds are compared to obtain a sequence of differences in the amplitude of the flow of funds between adjacent segments. Based on the sequence of differences in the amplitude of the flow of funds between adjacent segments, the boundary of the flow of funds stage is identified to obtain a set of boundary positions of the flow of funds stage. The fund flow record sequence of the account is segmented into stages based on the set of boundary positions of the fund flow stages to obtain the fund flow stage division result. The fund flow stage division result includes multiple fund flow stage units and the start and end record positions of each fund flow stage unit. Extract the channel access behavior record sequence from the post-loan behavior data set, perform access channel switching capture processing between adjacent access records in the channel access behavior record sequence according to the time sequence to obtain the access channel switching sequence, and perform continuous same-channel segment merging processing on the access channel switching sequence to obtain multiple channel access same-channel segment units. Each channel access same-channel segment unit carries the segment start access record position and segment end access record position. The access frequency of adjacent channel access segments is compared to obtain a sequence of access frequency differences between adjacent segments. Based on the sequence of access frequency differences between adjacent segments, channel access stage boundary identification is performed to obtain a set of channel access stage boundary locations. Based on the set of boundary positions of the channel access stage, the channel access behavior record sequence is segmented into stages to obtain the channel access stage division result. The channel access stage division result includes multiple channel access stage units and the start and end access record positions of each channel access stage unit.

3. The method for identifying bank credit loss customers according to claim 1, characterized in that, The step of constructing a contract state transition chain on the credit contract attribute data set to obtain the contract state transition chain of the credit customer includes: Extract the sequence of credit contract status records from the credit contract attribute data set, and perform contract status change capture processing on the sequence of credit contract status records in chronological order between adjacent status records to obtain a contract status change sequence. The contract state change sequence is merged with adjacent identical states to obtain multiple contract state resident segment units. Each contract state resident segment unit carries the segment start state record position and the segment end state record position. Extract the contract state identifier corresponding to each contract state dwelling segment unit, arrange the contract state identifiers in chronological order to obtain a contract state identifier sequence, perform adjacent state transition relationship extraction processing on the contract state identifier sequence to obtain a contract state transition relationship set, the contract state transition relationship set contains the transition relationship between every two adjacent contract state identifiers; A contract state transition chain is constructed based on the set of contract state transition relationships. The contract state transition chain includes a set of state nodes and a set of state transition edges. Each state node in the set of state nodes corresponds to a contract state identifier, and each state transition edge in the set of state transition edges corresponds to a transition relationship. Extract the state dwell time corresponding to each contract state dwell time segment unit, and associate the state dwell time with the corresponding state node in the contract state transition chain to obtain the contract state transition chain carrying the state dwell time attribute. Extract the state record content corresponding to each contract state residency segment unit, associate the state record content with the corresponding state node in the contract state migration chain to obtain a contract state migration chain carrying state record content attributes, and perform attribute merging processing with the contract state migration chain carrying state record content attributes to obtain the contract state migration chain of the credit customer.

4. The method for identifying bank credit loss customers according to claim 1, characterized in that, The step of performing churn tendency identification processing on the joint identification model of the fund flow stage segmentation result, the channel access stage segmentation result, and the contract state transition chain input churn tendency, to obtain the churn tendency identification result of the credit customer, includes: Extract the start and end record positions of each capital flow stage unit and the set of capital flow records within each stage from the capital flow stage division results. Perform record density distribution capture processing on the set of capital flow records within each stage to obtain the capital flow stage density distribution characterization. For each fund flow stage unit, the fund flow record set within the stage is processed to capture the proportion of fund flow direction, and a fund flow stage proportion representation is obtained. Based on the fund flow stage density distribution representation and the fund flow stage proportion representation, a fund flow stage representation set is constructed. Extract the start and end access record positions and the set of channel access records within each channel access stage unit from the channel access stage division results, and perform access channel distribution capture processing on the set of channel access records within each stage to obtain the channel distribution representation of the channel access stage. For each channel access stage unit, the channel access record set within the stage is processed by capturing the access time interval to obtain the channel access stage time interval representation. Based on the channel access stage channel distribution representation and the channel access stage time interval representation, a channel access stage representation set is constructed. The set of fund flow stage representations, the set of channel access stage representations, and the contract state transition chain are input into the joint identification model for churn tendency identification. The stage representations and state transition chains are jointly encoded to obtain a joint encoded representation. Based on the joint encoded representation, churn tendency identification is performed to obtain the churn tendency identification result of the credit customer.

5. The method for identifying bank credit loss customers according to claim 1, characterized in that, The step of identifying a list of credit customers to be intervened with from the credit customer group based on the churn tendency identification results includes: The churn tendency identification results are processed by churn tendency intensity analysis to obtain churn tendency intensity distribution. Credit customer identifiers whose churn tendency intensity meets the preset churn tendency intensity conditions are extracted from the churn tendency intensity distribution to obtain a set of candidate credit customer identifiers. The churn tendency identification result corresponding to each credit customer identifier in the candidate credit customer identifier set is subjected to churn cause analysis processing to obtain the churn cause description corresponding to each credit customer identifier; The churn tendency identification result corresponding to each credit customer identifier in the candidate credit customer identifier set is subjected to churn stage parsing processing to obtain the churn stage description corresponding to each credit customer identifier; The candidate credit customer identifier set, the description of the churn cause corresponding to each credit customer identifier, and the description of the churn stage corresponding to each credit customer identifier are grouped according to the customer dimension to obtain a credit customer churn profile set. The intervention urgency level of each credit customer churn profile in the credit customer churn profile set is stratified to obtain an intervention urgency level description for each credit customer churn profile. The set of credit customer churn profiles is sorted based on the intervention urgency description corresponding to each credit customer churn profile to obtain a sorted set of credit customer churn profiles. Credit customer identifiers whose intervention urgency descriptions meet the preset intervention urgency conditions are extracted from the sorted set of credit customer churn profiles to obtain the list of credit customers to be intervened.

6. The method for identifying bank credit loss customers according to claim 1, characterized in that, The method further includes: Obtain a set of image data of the business premises and a set of image data of business documents corresponding to the credit customers. The set of image data of the business premises includes multiple frames of real-scene image units of the business premises, and the set of image data of business documents includes multiple frames of scanned image units of business documents. The visual state extraction processing of the business premises image data set is performed to obtain a visual state representation set of the business premises, which includes the representation of the goods display state, the representation of the personnel activity state, and the representation of the equipment operation state in the business premises. The business document image data set is subjected to business document layout parsing processing to obtain a business document layout structure representation set, which includes business document field layout representation, business document field content representation and business document field association representation; The visual state representation set of the business premises and the layout structure representation set of the business documents are input into the loss tendency joint identification model for joint coding of business visual state and business document layout to obtain the joint coding representation of business visual documents. The joint coding representation of the operational visual documents and the churn tendency identification results of the credit customers are fused together to obtain the churn tendency identification results of the integrated operational visual document information. Based on the churn tendency identification results of the integrated operational visual document information, a list of credit customers to be intervened in is delineated from the credit customer group.

7. The method for identifying bank credit loss customers according to claim 1, characterized in that, The method further includes: Obtain a set of post-loan business call voice data and a set of post-loan business text interaction data corresponding to the credit customer. The set of post-loan business call voice data includes multiple post-loan business call voice segment units, and the set of post-loan business text interaction data includes multiple post-loan business text interaction message units. The post-loan business call voice data set is subjected to call voice prosody parsing to obtain a call voice prosody representation set, which includes call voice speech rate evolution representation, call voice pause distribution representation, and call voice tone change representation. The post-loan business text interaction data set is subjected to text interaction semantic parsing processing to obtain a text interaction semantic representation set, which includes text interaction topic evolution representation, text interaction expression tendency representation and text interaction response interval representation. The call speech prosody representation set and the text interaction semantic representation set are input into the call prosody and text semantic joint encoding model to obtain the call text joint encoding representation. The joint encoded representation of the call text is fused with the churn tendency identification result of the credit customer to obtain the churn tendency identification result of the fused call text information. Based on the churn tendency identification result of the fused call text information, a list of credit customers to be intervened in is delineated from the credit customer group.

8. The method for identifying bank credit loss customers according to claim 1, characterized in that, The method further includes: Obtain a set of video data and a set of screenshots of post-loan business processing interfaces corresponding to the credit customer. The set of video data includes multiple video segments of post-loan business processing, and the set of screenshots of post-loan business processing interfaces includes multiple frames of screenshots of post-loan business processing interfaces. The post-loan business processing video data set is subjected to business processing operation trajectory parsing processing to obtain a business processing operation trajectory representation set, which includes business processing operation path representation, business processing operation pause representation and business processing operation repetition representation. The set of screenshots of the post-loan business processing interface is subjected to interface screenshot visual state parsing processing to obtain a set of interface screenshot visual state representations. The set of interface screenshot visual state representations includes interface screenshot layout state representation, interface screenshot content state representation, and interface screenshot switching state representation. The set of business processing operation trajectory representations and the set of interface screenshot visual state representations are input into the loss tendency joint identification model for joint encoding of operation trajectory and interface state to obtain joint encoding representation of operation interface. The joint encoding representation of the operation interface and the churn tendency identification result of the credit customer are fused together to obtain the churn tendency identification result of the fused operation interface information. Based on the churn tendency identification result of the fused operation interface information, a list of credit customers to be intervened is delineated from the credit customer group.

9. The method for identifying bank credit loss customers according to claim 1, characterized in that, The method further includes: Obtain a set of post-loan location trajectory data and a set of post-loan address book data corresponding to the credit customer. The set of post-loan location trajectory data contains multiple location trajectory point units, and the set of post-loan address book data contains multiple address book record units. The post-loan location trajectory data set is subjected to location trajectory dwelling parsing processing to obtain a location trajectory dwelling representation set, which includes location trajectory dwelling area representation, location trajectory dwelling duration representation, and location trajectory migration path representation. The post-loan address book data set is subjected to address book change parsing processing to obtain an address book change representation set, which includes address book new record representation, address book deleted record representation and address book modified record representation; The location trajectory dwelling character set and the address book change character set are input into the loss tendency joint identification model for joint encoding of location trajectory and address book change to obtain the location address joint encoding character. The location address joint encoding representation and the credit customer churn tendency identification result are fused together to obtain the churn tendency identification result with fused location address information. Based on the churn tendency identification result with fused location address information, a list of credit customers to be intervened in is delineated from the credit customer group.

10. A bank customer loss identification system, characterized in that, The device includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the bank credit loss customer identification method according to any one of claims 1-9.

Images