Intelligent loan routing method, system, program product and related device

Abstract

The invention relates to an intelligent loan routing method and system, a program product and a related device. The method comprises the following steps: receiving a loan request initiated by a service system, and obtaining an initial payment channel set according to the loan request; performing state prediction on each payment channel through a channel state prediction model to obtain a state prediction value of each payment channel in the initial payment channel set; selecting an available payment channel set from the initial payment channel set by using a Pareto optimization model; according to a preset loan payment channel table, whether a payment channel capable of independently processing the loan amount exists in the available payment channel set or not is judged, and if yes, the payment channel is selected as a first loan execution scheme; otherwise, splitting the loan amount to obtain a plurality of sub-amounts, and screening out payment channels corresponding to the plurality of sub-amounts as a second loan execution scheme; and according to the first loan execution scheme or the second loan execution scheme, assembling a loan message for loan processing.

Description

Technical Field

[0001] This invention relates to the field of computer application technology, and more specifically, to an intelligent loan routing method, system, program product, and related devices. Background Technology

[0002] With the rapid development of fintech, lending business models are constantly evolving, product forms are becoming increasingly diversified, and the number of cooperating funding sources and payment channels is growing rapidly. Against this backdrop, loan disbursement, as a core link in the lending process, directly impacts customer experience, capital turnover efficiency, and the ability to control costs, affecting both customer satisfaction and the scale of business expansion.

[0003] However, the current loan routing mechanism still mainly relies on preset static rules (such as fixed priority, threshold splitting or simple round-robin), which has obvious bottlenecks. For example, it cannot predict the status fluctuations of payment channels in advance, the routing strategy is rigid, the channel resource utilization rate is low, and the overall cost remains high. Summary of the Invention

[0004] This invention provides an intelligent loan routing method, system, program product, and related device. By dynamically sensing the future state of payment channels through a channel status prediction model, payment channels can be flexibly selected, thereby improving the utilization rate of payment channels and reducing costs.

[0005] One embodiment of the present invention is an intelligent loan routing method, which includes:

[0006] Receive a loan request initiated by the business system, wherein the loan request includes at least the product number and the loan amount;

[0007] Based on the product number, query the pre-set product payment channel configuration table to obtain the initial payment channel set associated with the product number;

[0008] The status prediction model is used to predict the status of each payment channel in the initial payment channel set, so as to obtain the status prediction value of each payment channel in the initial payment channel set.

[0009] Based on the predicted status values ​​of each payment channel and the estimated transaction fees of each payment channel obtained from the pre-set payment channel fee information table, a Pareto optimization model is used to select an available payment channel set from the initial payment channel set.

[0010] Based on the pre-set loan payment channel table, determine whether there is a payment channel in the set of available payment channels that can handle the loan amount independently. If there is, select that payment channel as the first loan execution plan.

[0011] If it is determined that there is no payment channel that can handle the loan amount alone, the loan amount is split into multiple sub-amounts according to the product number and the preset product splitting configuration table, and the payment channels corresponding to the multiple sub-amounts are selected from the set of available payment channels as the second loan execution plan.

[0012] According to the first loan disbursement execution plan or the second loan disbursement execution plan, the loan disbursement message is assembled and the loan is processed.

[0013] In some implementations, a separate channel state prediction model is configured for each payment channel, and the channel state prediction model is based on a lightweight long short-term memory network (LSTM).

[0014] In some implementations, for each payment channel in the initial payment channel set, state prediction using a corresponding channel state prediction model includes:

[0015] The payment channel is retrieved from a time-series database for multiple time-series characteristic indicators within a predetermined time window in the past. These time-series characteristic indicators include statistical characteristics, trend characteristics, and periodic characteristics.

[0016] The time-series characteristic indicators are input into the channel status prediction model of the payment channel to obtain the status prediction value of the payment channel within the future prediction window. The status prediction value includes at least the prediction success rate, the prediction average delay, and the prediction risk score.

[0017] In some embodiments, the time series database includes a time series characteristic index data table, which is formed in the following manner:

[0018] By using scheduled tasks, historical loan disbursement data from various payment channels are aggregated at fixed time granularities to form a basic window data table;

[0019] Based on the aforementioned basic window data table, time-series characteristic indicator data for each payment channel are calculated according to the channel coding dimension. The time-series characteristic indicator data includes the average success rate over the past 1 hour, the success rate trend over the past 6 hours, the maximum average delay over the past 1 hour, the minimum proportion of risk events over the past 6 hours, the average standard deviation of delay over the past 1 hour, the variance of success rate over the past 30 minutes, the standard deviation of the number of risk events over the past 1 hour, the average success rate of the same hour over the past 7 days, the average average delay of the same window over the past 4 weeks, the current window online status, the average instantaneous success rate monitored in real time over the past 30 minutes, the number of risk events in the previous window, the ratio of the average success rate to the average delay over the past 1 hour, and risk warnings.

[0020] The time-series characteristic index data is stored in the time-series characteristic index data table according to the channel code.

[0021] In some implementations, the intelligent loan routing method further includes: recording the loan execution result and using the result to update the time series database for subsequent retraining of the channel status prediction model.

[0022] In some implementations, selecting the set of available payment channels from the initial set of payment channels using a Pareto optimization model includes:

[0023] The multiple target values ​​of each payment channel in the initial payment channel set are unified into a minimization direction, wherein the multiple target values ​​include at least the predicted success rate, the predicted risk score, the predicted average delay, and the estimated transaction fee;

[0024] By applying the Pareto dominance relation, all non-dominated solutions are selected from the initial set of payment channels to form the Pareto front set;

[0025] Based on the product number, query the product payment channel configuration table to obtain the weight coefficients corresponding to the predicted success rate, predicted risk score, predicted average delay, and estimated handling fee.

[0026] The payment channels in the Pareto frontier set are weighted and scored based on the weighting coefficients, and then sorted from high to low scores to form the set of available payment channels.

[0027] In some implementations, the intelligent loan routing method further includes: filtering each payment channel in the initial payment channel set according to preset business bottom line rules before processing with Pareto dominance relationship, wherein the business bottom line rules include a predicted risk score threshold and a predicted average delay threshold, thereby eliminating payment channels that exceed the predicted risk score threshold or the predicted average delay threshold.

[0028] In some implementations, when it is determined that there is no payment channel capable of processing the loan amount independently, splitting the loan amount includes:

[0029] Based on the rules in the product splitting configuration table, generate multiple loan amount splitting schemes that meet the constraints;

[0030] Construct a bipartite graph with the splitting scheme and the payment channel as two nodes, where the weight of the edge is the cost incurred by the sub-amount in the splitting scheme on the payment channel;

[0031] The Hungarian matching algorithm is applied to calculate the minimum matching cost under each splitting scheme, and the splitting scheme with the lowest cost and its channel matching result are selected as the second loan disbursement execution scheme.

[0032] In some implementations, generating multiple loan amount splitting schemes that satisfy the constraints includes:

[0033] The loan amount is discretized by a fixed step size to generate a discrete amount candidate set;

[0034] A restricted greedy algorithm and a backtracking pruning algorithm are used to search for all combinations of amounts that satisfy the constraints of minimum split amount and maximum number of splits from the discrete amount candidate set, forming a set of splitting schemes.

[0035] In some implementations, the intelligent loan routing method further includes: after obtaining the initial payment channel set, querying a preset loan payment channel table in conjunction with the entrusted payment field in the loan request, and filtering out payment channels that support entrusted payment and are in an enabled state, so as to modify the initial payment channel set.

[0036] Another embodiment of the present invention is an intelligent loan routing system, which executes the intelligent loan routing method described in any embodiment of the present invention, wherein the intelligent loan routing system includes:

[0037] The request receiving and preprocessing module is used to: receive a loan request initiated by the business system, wherein the loan request includes at least a product number and a loan amount; and based on the product number, query a pre-set product payment channel configuration table to obtain an initial payment channel set associated with the product number.

[0038] The channel status prediction module is used to predict the status of each payment channel in the initial payment channel set through the channel status prediction model, and obtain the status prediction value of each payment channel in the initial payment channel set.

[0039] The multi-objective optimization processing module is used to select a set of available payment channels from the initial set of payment channels based on the predicted state values ​​of each payment channel and the estimated handling fees of each payment channel obtained from the preset payment channel fee information table, using the Pareto optimization model.

[0040] The amount splitting and optimal matching module is used to: determine whether there is a payment channel in the set of available payment channels that can handle the loan amount independently, based on a preset loan payment channel table; if there is, select the payment channel as the first loan execution plan; if it is determined that there is no payment channel that can handle the loan amount independently, split the loan amount into multiple sub-amounts according to the product number and a preset product splitting configuration table, and select the payment channels corresponding to the multiple sub-amounts from the set of available payment channels as the second loan execution plan.

[0041] The loan disbursement triggering execution module is used to assemble a loan disbursement message and perform loan disbursement processing according to the first loan disbursement execution plan or the second loan disbursement execution plan.

[0042] Furthermore, embodiments of the present invention relate to a computer program product comprising a computer program that, when executed by a processor, implements the steps of the intelligent loan routing method described in any embodiment of the present invention. Correspondingly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, characterized in that, when executed by a processor, the computer program implements the steps of the intelligent loan routing method described in any embodiment of the present invention.

[0043] In addition, embodiments of the present invention also provide a computer device comprising:

[0044] A memory on which computer programs are stored;

[0045] A processor, when executing the computer program, implements the steps of the intelligent loan routing method according to any embodiment of the present invention.

[0046] The following beneficial effects can be achieved through the embodiments of the present invention:

[0047] The embodiments of the present invention use a channel status prediction model to dynamically perceive the future status of payment channels, avoid high-risk channels in advance, significantly improve the success rate of the first loan disbursement, and reduce failed retries and customer complaints.

[0048] Furthermore, the embodiments of the present invention are based on Pareto multi-objective optimization and the Hungarian algorithm to split and combine the loan amount into different payment channels, dynamically balancing cost, timeliness and risk, and automatically generating the globally optimal loan solution, taking into account both business efficiency and risk control requirements.

[0049] The intelligent loan routing system of this invention is flexible and efficient, supports complex business scenarios, and automatically handles complex business scenarios such as large amounts through dynamic amount splitting and multi-channel parallel scheduling. It supports strategy iteration at the minute level and flexibly adapts to rapid business changes and channel expansion.

[0050] Various aspects, features, and advantages of embodiments of the present invention will be specifically described below in conjunction with the accompanying drawings. These aspects, features, and advantages will become clearer from the following detailed description taken in conjunction with the accompanying drawings. Attached Figure Description

[0051] Figure 1 This is a basic block diagram of an exemplary implementation of an intelligent loan routing system.

[0052] Figure 2 This is a basic flowchart of an exemplary implementation of an intelligent loan routing method.

[0053] Figure 3 This is a basic block diagram of a computer device according to an exemplary embodiment. Detailed Implementation

[0054] The exemplary embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the present invention can be implemented in various forms and is not limited to the specific embodiments described herein or shown in the accompanying drawings.

[0055] The terms “comprising,” “including,” and “having,” etc., are used herein to indicate the presence of certain features, steps, operations, elements, and / or components, but do not exclude the presence or addition of other features, steps, operations, elements, components, or combinations thereof. Unless explicitly defined by context, the terms “first,” “second,” and similar terms do not indicate any priority or order, but are used only to distinguish different objects in the description.

[0056] Figure 1 An exemplary block diagram of the intelligent loan routing system of the present invention is shown. In an exemplary embodiment, the intelligent loan routing system 1000 executes the intelligent loan routing method described in any embodiment of the present invention, wherein the intelligent loan routing system 1000 includes a request receiving and preprocessing module 1001, a channel status prediction module 1002, a multi-objective optimization processing module 1003, an amount splitting and optimal matching module 1004, and a loan triggering execution module.

[0057] In some implementations, the database 3000 is pre-configured with a loan payment channel table, a payment channel fee information table, a product payment channel configuration table, and a product splitting configuration table. In some implementations, the system of the present invention is configured with a visual page, such as a web page, so that the relevant configuration tables can be added, deleted, modified, and queried through the visual page. It can also be integrated with relevant OA approval processes to automatically realize the automatic modification and process management of relevant information.

[0058] In some implementations, the loan disbursement payment channel table includes fields such as channel code (unique), channel name, channel status ("1" indicates enabled / "0" indicates disabled), support for real-time loan disbursement ("1" indicates yes / "0" indicates no), maximum single transaction amount, daily cumulative limit, and support for entrusted payment ("1" indicates yes / "0" indicates no). The payment channel fee information table includes fields such as channel code (unique), fee collection type ("01" indicates fixed rate / "02" indicates tiered rate / "03" indicates fixed fee per transaction, etc.), additional fees, fee settlement cycle (T+0 real-time settlement / T+1 next day settlement / T+30 monthly settlement, etc.), and specific parameters for fee calculation (this field is in JSON format). The structure of the specific parameters for fee calculation can vary depending on the fee collection type. The product payment channel configuration table includes fields such as: product number (unique), product name, payment channel set (multiple payment channel codes are separated by commas; if all channels are supported, use ALL), success rate weight, risk score weight, time consumption weight, and fee weight. The fields in the product splitting configuration table include: product number (unique), minimum splitting amount (if the loan amount is greater than this amount, the loan amount can be split and disbursed), and maximum number of splitting transactions (maximum number of sub-orders for splitting and disbursing loans).

[0059] In an exemplary embodiment, the request receiving and preprocessing module 1001 is configured to: receive a loan request initiated by the business system 2000, the loan request including at least a product number and a loan amount; based on the product number, query a pre-set product payment channel configuration table in the database 3000 to obtain an initial payment channel set associated with the product number. The channel status prediction module 1002 is configured to: perform status prediction on each payment channel in the initial payment channel set using a channel status prediction model to obtain the status prediction value of each payment channel in the initial payment channel set. The multi-objective optimization processing module 1003 is configured to: based on the status prediction value of each payment channel and the estimated handling fee of each payment channel obtained from the pre-set payment channel fee information table in the database 3000, use a Pareto optimization model to select an available payment channel set from the initial payment channel set. The amount splitting and optimal matching module 1004 is configured to: determine whether there is a payment channel in the set of available payment channels that can handle the loan amount independently, based on the pre-set loan payment channel table in the database 3000; if so, select that payment channel as the first loan execution plan; if it is determined that there is no payment channel that can handle the loan amount independently, split the loan amount into multiple sub-amounts according to the product number and the pre-set product splitting configuration table, and select the payment channels corresponding to the multiple sub-amounts from the set of available payment channels as the second loan execution plan. The loan triggering execution module 1005 is configured to assemble a loan message and perform loan processing according to the first loan execution plan or the second loan execution plan.

[0060] The following is combined Figure 2 An exemplary intelligent loan routing method performed by the intelligent loan routing system is described.

[0061] In an exemplary implementation, the intelligent loan routing method begins with process S100. In S100, a loan request initiated by the business system is received, the loan request including at least a product number and a loan amount. In S200, based on the product number, a preset product payment channel configuration table is queried to obtain an initial set of payment channels associated with that product number. In S300, a channel status prediction model is used to predict the status of each payment channel in the initial payment channel set, obtaining a status prediction value for each payment channel in the initial payment channel set. In S400, based on the status prediction values ​​of each payment channel and the estimated handling fees of each payment channel obtained from a preset payment channel fee information table, a Pareto optimization model is used to select a set of available payment channels from the initial payment channel set. In S500, a preset loan payment channel table is used to determine whether there exists a payment channel in the set of available payment channels capable of processing the loan amount independently. If the determination is yes in S500, then in S600, that payment channel is selected as the first loan execution plan. If the determination in S500 is negative, then in S700, the loan amount is split into multiple sub-amounts according to the product number and a preset product splitting configuration table, and payment channels corresponding to the multiple sub-amounts are selected from the set of available payment channels as the second loan disbursement execution plan. In S800, a loan disbursement message is assembled and the loan is processed according to the first or second loan disbursement execution plan.

[0062] In some implementations, a separate channel state prediction model is configured for each payment channel, and the channel state prediction model is based on a lightweight long short-term memory (LSTM) network.

[0063] In some implementations, in S300, for each payment channel in the initial payment channel set, the state prediction using a corresponding channel state prediction model includes: obtaining multiple time-series characteristic indicators of the payment channel within a predetermined time window from a time-series database, the time-series characteristic indicators including statistical characteristics, trend characteristics, and periodic characteristics; inputting the time-series characteristic indicators into the channel state prediction model of the payment channel to obtain the state prediction value of the payment channel within a future prediction window, the state prediction value including at least the prediction success rate, the prediction average delay, and the prediction risk score.

[0064] In some implementations, the time-series database includes a time-series characteristic indicator data table, which is formed by: aggregating historical loan disbursement data from various payment channels at a fixed time granularity through a scheduled task to form a basic window data table; and calculating time-series characteristic indicator data for each payment channel according to the channel coding dimension based on the basic window data table. The time-series characteristic indicator data includes the average success rate over the past 1 hour (success_rate_1h_mean), the success rate trend over the past 6 hours (success_rate_6h_trend), the maximum average delay over the past 1 hour (avg_delay_1h_max), the minimum risk event ratio over the past 6 hours (risk_ratio_6h_min), the average standard deviation of delay over the past 1 hour (delay_std_1h_mean), and the variance of the success rate over the past 30 minutes (succ). The time-series characteristic index data includes: risk_rate_30m_var, standard deviation of risk events in the past hour (risk_count_1h_std), average success rate of the same hour in the past 7 days (success_rate_same_hour_7d), average delay of the same window in the past 4 weeks (avg_delay_same_weekday_4w), online status of the current window (online_status_current), average instantaneous success rate of real-time monitoring in the past 30 minutes (instant_success_30m), number of risk events in the previous window (risk_events_last_window), ratio of average success rate to average delay in the past hour (success_delay_ratio_1h), and risk warning (risk_trend_alert). This time-series characteristic index data is stored in a time-series characteristic index data table according to channel encoding. In some implementations, the time-series database and database 3000 are the same database, meaning the basic window data table, the time-series characteristic index data table, and the aforementioned preset configuration table reside in the same database. Optionally, the time-series database may be a separate database, different from database 300.

[0065] In some implementations, the system of this invention connects payment data from all payment channel databases using JDBC (Java Database Connectivity) and dynamic data source technology, and processes and extracts data according to the smallest set unit. For example, the system of this invention uses a scheduled task to extract data with a time granularity of 5 minutes as the smallest aggregation unit (e.g., 00:00-00:05, 00:05-00:10…23:55-24:00), generating 24×12=288 time windows per day, i.e., 288 corresponding data entries. The extracted data forms the basic window data table, whose fields include: channel code, start time (e.g., 20251111083000, YYYYMMDDHHMMSS), end time (e.g., 20251111083500, YYYYMMDDHHMMSS), total transaction volume, number of successful transactions, success rate, average latency, total handling fee, total transaction volume, average transaction amount, successful loan amount, average successful loan amount, number of risk events, percentage of risk events, and instantaneous success rate. Specifically, "average latency" refers to the total time taken for all transactions within the window divided by the total transaction volume (unit: seconds); "number of risk events" refers to the number of risky transactions, such as: abnormal card status, anti-money laundering triggers, etc.; "instantaneous success rate within the window" is calculated by taking 10 statistics every 30 seconds within the time window, for a total of 10 statistics within a 5-minute window, and then calculating the average of the 10 statistics.

[0066] In some implementations, the data in the base window data table is processed according to the [channel code] dimension and the following rules to form a time-series feature indicator data table. Its fields include: channel number, start time (e.g., 20251111083000), end time (e.g., 20251111083500), and the following indicator fields (14-dimensional time-series features):

[0067] success_rate_1h_mean: The average success rate over the past hour (12 five-minute windows), reflecting the overall success rate of the channel in the past hour;

[0068] success_rate_6h_trend: Calculates the difference in success rate between the last window and the first window over the past 6 hours (72 5-minute windows), and then divides it by 72. It reflects the trend of success rate over the past 6 hours.

[0069] avg_delay_1h_max: Takes the maximum value of the average delay of 12 five-minute windows over the past hour, which can be used as an indicator of channel service pressure;

[0070] risk_ratio_6h_min: The minimum percentage of risk events over the past 6 hours (72 five-minute windows). It represents the lowest level of risk event percentage in the past 6 hours and can be used to determine the historical lower limit of channel risk.

[0071] delay_std_1h_mean: The mean of the standard deviations of delay over the past 1 hour and 12 5-minute windows, used to measure the average volatility of channel delay over the past hour;

[0072] success_rate_30m_var: The variance of the success rate over the past 30 minutes (6 five-minute windows), which reflects the degree of fluctuation in the success rate over the past 30 minutes (the larger the variance, the more unstable the success rate).

[0073] risk_count_1h_std: The standard deviation of the number of risk events over the past hour and 12 five-minute windows, reflecting the fluctuation of the number of risk events in the past hour;

[0074] success_rate_same_hour_7d: The average success rate of the same hour period (e.g., 08:30-09:30) every day over the past 7 days, used to capture historical patterns of the channel during the same period;

[0075] avg_delay_same_weekday_4w: The average delay for the same weekday and the same window over the past four weeks, reflecting the weekly cycle pattern;

[0076] online_status_current: The online status of the current 5-minute window, 1 indicates online and 0 indicates offline. It directly reflects whether the channel is currently available.

[0077] instant_success_30m: The average instantaneous success rate monitored in real time over the past 30 minutes (6 five-minute windows), which reflects the channel's ultra-short-term real-time performance;

[0078] risk_events_last_window: The number of risk events in the previous 5-minute window. This is the latest risk signal and can be used to detect risk anomalies in a timely manner.

[0079] success_delay_ratio_1h: The ratio of the average success rate to the average delay over the past hour, which comprehensively reflects the "success rate / delay" performance (the higher the ratio, the higher the cost-effectiveness of the channel).

[0080] risk_trend_alert: If the risk ratio trend over the past 2 hours is greater than 0.1 and the number of risk events in the previous window is greater than 5, it is marked as 1 (triggering an alert), otherwise it is 0. It is used to quickly identify an upward trend in risk.

[0081] It should be understood that although the present invention has been described above or below with specific numerical values, the present invention is not limited thereto, and these numerical values ​​can be appropriately modified according to actual applications.

[0082] In some implementations, the channel status prediction model of the present invention is based on a lightweight LSTM and adopts a "one channel, one model" architecture. Each payment channel independently trains a dedicated LSTM prediction model to capture the unique temporal characteristics and state change patterns of that channel.

[0083] The channel status prediction model based on lightweight LSTM of the present invention will be described below by way of example.

[0084] The channel status prediction model of this invention consists of five core modules that are connected in sequence according to the data flow: input layer, LSTM layer, regularization layer (Dropout), fully connected layer (shared semantic representation layer), and output layer (multi-task head).

[0085] Input layer

[0086] Input data: Each sample contains historical statistical features from the past 12 five-minute windows (i.e., 1 hour).

[0087] Feature dimensions: A total of 14 time-series features, derived from the time-series feature index data table, including: continuous features (such as success_rate_1h_mean, delay_std_1h_mean, avg_delay_1h_max, etc.) and binary classification features (such as online_status_current, risk_trend_alert).

[0088] Input tensor shape: (number of samples, time step = 12, feature dimension = 14).

[0089] LSTM layer

[0090] Hierarchical structure: It adopts a stack of two layers of lightweight LSTM units. This design effectively controls model complexity and computational overhead while ensuring model prediction accuracy. The number of hidden units (Hidd en Size) in each layer is set to 64, aiming to balance model expressiveness and computational efficiency.

[0091] The internal structure of an LSTM unit includes an input gate (responsible for regulating the flow of new input information into the cell state and determining which new information should be added to the cell state), a forget gate (controlling the degree of forgetting of information in the cell state and determining which old information should be discarded from the cell state, thereby achieving dynamic information updates), and an output gate (based on the current cell state and input information, determining the output content of the LSTM unit, i.e., which information should be passed to the next layer or the output layer).

[0092] Information Flow and Memory Mechanism: At each time step, the LSTM unit dynamically adjusts the cell state through the synergistic action of the input gate, forget gate, and output gate. The input gate allows new information to enter, the forget gate discards irrelevant information, and the output gate controls the output of information, collectively achieving effective memory and updating of historical information.

[0093] Output processing: The output of the LSTM layer uses the hidden state of the last time step as a compressed representation of the temporal semantics. This design captures the most important features of the entire sequence while reducing the output dimensionality, facilitating processing by subsequent fully connected layers. In some variants, pooling (such as average pooling or max pooling) of the entire sequence output can also be considered to further extract sequence-level features, but in this model, to maintain simplicity and computational efficiency, the hidden state of the last time step is chosen.

[0094] Regularization layer (Dropout)

[0095] Adding a Dropout layer after the LSTM layer primarily prevents overfitting during training by randomly discarding the outputs of some neurons, thus significantly improving the model's generalization ability. This mechanism simulates the effect of ensemble learning, enhancing the model's adaptability to unknown data through diverse combinations of sub-networks.

[0096] Fully connected layer (shared semantic representation layer)

[0097] Hierarchical Structure: A two-layer fully connected network (64→32D→16D) is employed, with each layer followed by a ReLU activation function to achieve non-linear feature transformation. The first layer compresses the 64-dimensional vector output by Dropout to 32-dimensional vectors, and the second layer further abstracts it into 16-dimensional shared semantic features. Finally, the 16-dimensional vector serves as the unified input basis for the three prediction tasks, realizing knowledge transfer and feature recombination between tasks.

[0098] Weight constraints: L2 regularization (coefficient 1e-4) is used to prevent overfitting, and Layer Normalization (LN) is used to improve the convergence speed.

[0099] High-level abstract features are extracted and used as a shared semantic foundation for the three prediction tasks, enabling knowledge transfer between tasks.

[0100] Multi-task Headers

[0101] As shown in Table 1 below, the three independent branches predict the three core state indicators of the payment channel within the next 5-minute window, achieving knowledge sharing and collaborative optimization between tasks. Each branch designs a dedicated network structure and loss function for different prediction objectives, and is ultimately jointly trained using a weighted loss function.

[0102] Table 1

[0103] Output head Number of neurons Activation function Output meaning loss function Success rate prediction head 1 Sigmoid Transaction success rate ∈ [0,1] Binary Cross-Entropy Average delay prediction head 1 Linear (no activation) Average latency (seconds, real value) MAE or Huber Loss Risk score prediction head 1 Sigmoid Risk probability ∈ [0,1] Binary Cross-Entropy

[0104] In some implementations, for the channel status prediction model, all training samples are strictly screened according to channel code (e.g., only data with channel_code='HCFC001' is used); the input features are all aggregated statistics of the channel over the past hour, directly reflecting its current operating status and historical trends; the outputs are prediction success rate, average prediction latency, and prediction risk score, defined as follows:

[0105] Predicted success rate (success_rate): Used to predict the stability of the transaction link and trigger backup route switching;

[0106] Predicted average latency (avg_delay): aids in monitoring and user experience optimization;

[0107] Predictive risk score: Identify abnormal fluctuations in advance (such as fraud attacks or system failures).

[0108] Therefore, the triples (success_rate, avg_delay, risk_score) output by the model constitute a comprehensive profile of the health status of the corresponding channel in the next 5 minutes.

[0109] In some implementations, during model training, samples from specified channels are selected from the time-series feature index data table, for example, channel_code='HCFC001'. The samples are divided into training, validation, and test sets according to a fixed time window (e.g., 5-minute granularity). Each window contains a set of input features (e.g., statistics from the past hour) and corresponding target variables (e.g., success rate, latency, risk score in the next window). The samples are strictly divided in time order and divided into training, validation, and test sets in a 7:2:1 ratio to ensure that the training set → validation set → test set progresses sequentially in time.

[0110] In some implementations, sample data is preprocessed uniformly based on training set statistics to ensure distribution consistency. The 14-dimensional features are differentiated into continuous numerical features (e.g., fields such as success_rate_1h_mean, avg_delay_1h_max, delay_std_1h_mean, and success_delay_ratio_1h in the time-series feature index data table, which are Z-score standardized) and binary features (e.g., fields such as online_status_current and risk_trend_alert in the time-series feature index data table, which retain their original 0 / 1 values ​​and are not standardized). Missing values ​​are then processed accordingly; for example, continuous feature fields are filled with the historical global mean across all channels, and binary feature fields are filled with zeros.

[0111] Configure the model training as follows.

[0112] The loss function for model training, as shown in Equation (1), is achieved through weighted multi-task loss, with the loss function weights w1:w2:w3 = 1:1:1.

[0113]

[0114] The parameters involved in the formula are explained in Table 2:

[0115] Table 2:

[0116]

[0117] Set the Dropout Rate to 0.25.

[0118] Other parameters are configured according to Table 3 below.

[0119] Table 3:

[0120] Configuration items default value Adjustable range Adjustment basis Batch Size 64 32-128 Single channel sample size Initial learning rate 1e-3 5e-4–2e-3 Convergence speed and stability Optimizer AdamW – Required, supports valid L2 regular expressions Loss weights 1:1:1 Adjusted by business importance Prioritizing success rate / delay / risk Early Stop Patience Value 10 5-15 Data noise level and periodicity Maximum learning rate decay times 4 3-5 To prevent indefinite decay, stop losses in time.

[0121] For each payment channel, according to the above configuration, train the model using the training set, validate it using the validation set, and test it using the test set.

[0122] Each channel model is stored independently, for example (model_HCFC001.pth), and the mapping relationship between channel codes and numbers is maintained in the channel model table. The fields of the channel model table include: id, channel number, model name, release time, and availability ("1" indicates availability / "0" indicates unavailability).

[0123] In some implementations, in S300, the production environment matches the model name in the channel model table with the channel code, and dynamically loads the corresponding channel status prediction model by obtaining the name, thereby realizing the prediction function of the channel.

[0124] In some implementations, the 14-dimensional features of a payment channel for the most recent 12 5-minute windows are obtained in real time from the time-series feature index data table; the features are standardized according to the training phase rules and missing values ​​are filled; finally, the corresponding channel model is input and the ternary prediction result is output.

[0125] In some implementations, the model's performance is evaluated daily using three key metrics: success rate (AUC), average latency (MAE), and risk score (Brier Score). If any metric significantly deteriorates (e.g., AUC drops by >5%), the retraining process for the channel model is automatically triggered, forming a closed-loop intelligent operation and maintenance system of "prediction → execution → observation → learning".

[0126] In some implementations, communication with the channel model is achieved through an API interface. By inputting the channel code, the API outputs three dimensions of information corresponding to that channel: success rate, average latency, and risk score, facilitating access by other systems or modules. For example:

[0127] Input message: {"channel_code":"HCFC001"}

[0128] Output message: {"success_rate":0.982,"avg_delay_sec":1.45,"risk_score":0.03}

[0129] In some implementations, the intelligent loan routing method further includes: recording the loan execution result and using the result to update the time series database for subsequent retraining of the channel status prediction model.

[0130] In some implementations, in step S400, selecting the set of available payment channels from the initial payment channel set using the Pareto optimization model includes: unifying multiple objective values ​​of each payment channel in the initial payment channel set into a minimization direction, wherein the multiple objective values ​​include at least the prediction success rate, prediction risk score, prediction average delay, and estimated transaction fee; applying Pareto dominance to filter out all non-dominated solutions from the initial payment channel set to form a Pareto front set; querying the product payment channel configuration table based on the product number to obtain the weight coefficients corresponding to the prediction success rate, prediction risk score, prediction average delay, and estimated transaction fee; weighting the payment channels in the Pareto front set based on the weight coefficients, and sorting them from high to low scores to form the set of available payment channels.

[0131] In some implementations, the intelligent loan routing method further includes: filtering each payment channel in the initial payment channel set according to preset business bottom line rules before processing with Pareto dominance relationship, wherein the business bottom line rules include a predicted risk score threshold and a predicted average delay threshold, thereby eliminating payment channels that exceed the predicted risk score threshold or the predicted average delay threshold.

[0132] For example, the system of the present invention introduces the Pareto optimization function to automatically identify the set of high-quality candidates that are “irreplaceable” (i.e., Pareto front) from multiple payment channels without any subjective weight preset, thereby screening out the payment channels that can be used for lending, i.e., the set of available payment channels.

[0133] The specific implementation is as follows:

[0134] (1) Data acquisition: For candidate lending channels (i.e., the initial set of payment channels), collect four target values: predicted success rate (the higher the better), predicted risk score (the lower the better), predicted time (unit: seconds, the lower the better), and estimated handling fee (the lower the better).

[0135] (2) Unify the optimization direction. Pareto dominance requires all target directions to be consistent, so the "success rate" is negative and uniformly transformed into a minimization problem of "the smaller the better".

[0136] (3) Pareto payment judgment function and extraction of Pareto front. Here, we define whether the first channel is dominated by the second channel, and then traverse all channels and retain the channel that is not dominated by any other channel.

[0137] (4) Built-in pre-filtering rules apply business bottom-line rules for further filtering before or after Pareto calculation to ensure feasibility. For example, rules such as risk score > 0.5 (predicted risk score threshold) and prediction time > 30 seconds (predicted average delay threshold) are set to hard-kill candidate payment channels that exceed the threshold.

[0138] (5) Result set sorting: For the selected channel set that meets the conditions, match the four weight coefficients of the corresponding product configuration in the product payment channel configuration table according to the product number input by the business system, and calculate the score according to the weighted summation method (as shown in formula (2)). Arrange them in order from high to low score.

[0139]

[0140] Among them, w j f is the weight of the j-th objective. ijIt is the value of the i-th channel on the j-th target.

[0141] In some implementations, in S700, when it is determined that there is no payment channel that can handle the loan amount independently, splitting the loan amount includes: generating multiple loan amount splitting schemes that meet the constraints according to the rules in the product splitting configuration table; constructing a bipartite graph with the splitting scheme and the payment channel as two nodes, where the weight of the edge is the cost incurred by the sub-amount in the splitting scheme on the payment channel; applying the Hungarian matching algorithm to calculate the minimum matching cost under each splitting scheme, and selecting the splitting scheme with the lowest cost and its channel matching result as the second loan execution scheme.

[0142] In some implementations, generating multiple loan amount splitting schemes that satisfy the constraints includes: discretizing the loan amount by a fixed step size to generate a discrete amount candidate set; and using a restricted greedy algorithm and a backtracking pruning algorithm to search for all amount combinations that satisfy the minimum splitting amount and the maximum number of splitting transactions constraints from the discrete amount candidate set to form a splitting scheme set.

[0143] For example, processing S500 to S700 involves "multi-objective optimization processing", which, based on the selected set of available payment channels A, selects payment channels that can meet the requirements of a single loan according to the priority of the available payment channels in the set of available payment channels and the configuration information in the loan payment channel table corresponding to the payment channel code value. It then directly returns the loan channel and loan amount information as the first loan execution plan, for example: {{"channel_code":"HCFC001","amt":3000}}.

[0144] When a single loan cannot be disbursed, the system matches the product number submitted by the business system with the minimum split amount and maximum number of split transactions in the product payment split configuration table. Payment channels with loan amounts exceeding the minimum split amount are then selected from the available payment channel set to form a second payment channel set. Dynamic splitting is performed according to the maximum number of split transactions requirement. The split amounts (sub-amounts) are matched with payment channels in the second payment channel set (these channels must meet requirements such as not exceeding single-transaction limits). The corresponding loan disbursement fee (channel cost) is calculated, and finally, the second loan disbursement execution plan (multiple sub-amounts and their respective payment channels) with the lowest cost is selected.

[0145] As can be understood from the above, when there is no payment channel that satisfies a single loan disbursement, the system of this invention, given the total loan amount, finds a set of splitting schemes through steps such as discretization, combination generation, and matching optimization to minimize the overall loan disbursement cost. The specific implementation is as follows:

[0146] (1) Discretizing the amount space: Specifically, the system of the present invention discretizes the amount range with a fixed step size, dividing it according to a step size of 1000 to form a discrete amount candidate set. For example, if the loan amount is 15000, the discrete points are {0, 1000, 2000, ... 15000}.

[0147] (2) Generate representative splitting schemes. Specifically, use a restricted greedy algorithm to prioritize larger amounts (e.g., traverse discrete points from high to low) to quickly approach the target. Then, combine it with a backtracking pruning algorithm. If the current part sum exceeds the total amount, backtrack immediately; if the remaining amount cannot be achieved by combining the remaining candidate amounts, prune (constraints such as the maximum number of splits or the minimum split amount can be set for further pruning). Finally, output all splitting schemes that meet the constraints (each splitting scheme is a list of discrete amounts, such as [8000, 7000], [5000, 10000], [5000, 5000, 5000], etc.).

[0148] (3) Construct a bipartite graph to model the matching relationship between the above splitting scheme and the optional payment channels as a bipartite graph.

[0149] The structure of the bipartite graph is as follows:

[0150] Left node: Splitting schemes ([8000,7000], [5000,10000], [5000,5000,5000], etc.)

[0151] Right-hand node: Select payment channel (Channel No. 1, Channel No. 2)

[0152] Edge weight: The overall cost (e.g., transaction fees) of a certain splitting scheme on a certain channel.

[0153] (4) Transform the bipartite graph into a cost matrix, where each element represents the cost of the corresponding splitting scheme-channel combination. Then, run the Kuhn-Munkres algorithm to output the minimum matching cost under this splitting scheme.

[0154] (5) Global optimal selection: Specifically, iterate through all splitting schemes, calculate the minimum matching cost for each scheme, and select the scheme with the minimum matching cost as the final loan disbursement plan, i.e., the second loan disbursement execution scheme (which includes a set of payment channel number and splitting amount), for example: {{"channel_code":"HCFC001","amt":7000},{"channel_code":"HCFC002","amt":8000}}

[0155] In some implementations, in step S800, based on either the first or second loan execution plan, the selected payment channel and corresponding loan amount are obtained. An API interface is then invoked to assemble the loan message required by the corresponding channel for loan processing. Further, in some implementations, the relevant payment information for the loan processing is recorded in the relevant table of the corresponding payment channel, facilitating the timely task to obtain the latest time window information in real time and promoting model prediction.

[0156] In some implementations, the intelligent loan routing method further includes: after obtaining the initial payment channel set, querying a preset loan payment channel table in conjunction with the entrusted payment field in the loan request, and filtering out payment channels that support entrusted payment and are in an enabled state, so as to modify the initial payment channel set.

[0157] The above provides a detailed description of the intelligent loan routing system and method according to exemplary embodiments of the present invention. To facilitate understanding, a specific example is provided below to illustrate the invention.

[0158] First, the business system initiates a loan application through a loan disbursement request, which transmits the following parameters:

[0159]

[0160] After compliance and parameter validity checks, the system retrieves a set of candidate payment channels from the product payment channel configuration table based on the product number (product_code). Assuming the matched candidate payment channel set is ["HCFC001","HCFC002","ALL"] (ALL represents support for all channels, but further judgment based on other conditions is needed during actual filtering), it also matches the loan disbursement payment channel table with "whether it is entrusted (is_entrusted)" (assuming 1 indicates entrusted payment) to further narrow down the scope by filtering out payment channels that support entrusted payment and are in an enabled state. Here, it is assumed that the final initial payment channel set is ["HCFC001","HCFC002"], which serves as the initial channels available for loan disbursement.

[0161] Then, for each channel code in the selected initial payment channel set, the API interface of the corresponding channel status prediction model is called (e.g., inputting {"channel_code":"HCFC001"} and {"channel_code":"HCFC002"}). This allows the corresponding effective channel status prediction model to be retrieved from the channel model table based on the channel code, and the feature data of each channel over the past hour is directly obtained from the time-series feature index data table. The feature data of each channel is then input into the corresponding channel status prediction model. Each model outputs three dimensions of information for the corresponding channel in the next 5-minute window: prediction success rate, prediction average delay, and prediction risk score. For example, the following might be returned:

[0162] {"channel_code":"HCFC001","success_rate":0.98,"avg_delay_sec":1.2,"risk_score":0.02}

[0163] {"channel_code":"HCFC002","success_rate":0.95,"avg_delay_sec":1.8,"risk_score":0.04}

[0164] Furthermore, the acquired data is encapsulated into a List format, which serves as a selection set for loan disbursement and payment channels.

[0165] [{"channel_code":"HCFC001","success_rate":0.98,"avg_delay_sec":1.2,"risk_score":0.02}

[0166] ,{"channel_code":"HCFC002","success_rate":0.95,"avg_delay_sec":1.8,"risk_score":0.04} ]

[0168] Next, based on the prediction results of the available payment channels above, for candidate lending channels ["HCFC001", "HCFC002"] in the initial payment channel set, four target values ​​are collected: prediction success rate (the higher the better), prediction risk score (the lower the better), prediction time (here, the average prediction delay is assumed to be used as a reference for prediction time, in seconds, the lower the better), and estimated handling fee (which is obtained from the payment channel fee information table; assuming the handling fee for HCFC001 is 5 yuan and the handling fee for HCFC002 is 3 yuan). These are summarized in Table 4 below.

[0169] Table 4:

[0170]

[0171] With a unified optimization approach, this system takes the success rate as negative, thus transforming all fields into a "smaller is better" minimization problem, and obtaining the latest target value, as shown in Table 5 below.

[0172] Table 5:

[0173]

[0174] The Pareto payment judgment function iterates through all candidate lending channels, retaining individuals that are not dominated by any other channel. After judgment, HCFC001 and HCFC002 are not dominated by each other, so the Pareto front is ["HCFC001","HCFC002"].

[0175] Assuming filtering is performed using rules such as risk score > 0.5 and prediction time > 30 seconds, since neither HCFC001 nor HCFC002 meets the filtering conditions, the candidate loan channel set remains ["HC FC001", "HCFC002"].

[0176] Based on the product number entered into the business system, the system matches the corresponding product payment channel configuration table with four weighted coefficients (assuming a success rate weight of 0.4, a risk score weight of 0.3, a time consumption weight of 0.2, and a transaction fee weight of 0.1), and calculates the score using a weighted summation method.

[0177] HCFC001 score: 0.4×0.98+0.3×(1-0.02)+0.2×(1 / 1.2)+0.1×(1 / 5)=0.392+0.294+0.167+0.02=0.873

[0178] HCFC002 score: 0.4×0.95+0.3×(1-0.04)+0.2×(1 / 1.8)+0.1×(1 / 3)=0.38+0.288+0.111+0.033=0.812

[0179] Arranged in descending order of scores, the set of available payment channels is ["HCFC001","HCFC002"].

[0180] Next, based on the channels (["HCFC001","HCFC002"]) in the set of available payment channels above, and combined with the configuration information in the loan disbursement payment channel table, it is determined whether there is a channel that can meet the requirements for a single loan disbursement.

[0181] Assuming the maximum single amount for HCFC001 is 20,000 yuan, which is greater than the loan amount of 15,000 yuan, a single loan can be disbursed directly, and the loan channel and loan amount information (i.e., the first loan execution plan) will be returned: {"channel_code":"HCFC001",3"amt":15000}.

[0182] Assuming that the maximum single loan amount for HCFC001 and HCFC002 is 10,000 yuan, which cannot meet the single loan disbursement requirement, then proceed to the following steps:

[0183] Based on the product number entered by the business system, match the minimum split amount (assuming it is 5,000 yuan) and the maximum number of split transactions (assuming it is 3 transactions) in the product payment split configuration table. Filter out the payment channels from the available payment channel set that have a loan amount of 15,000 yuan entered by the business system that is greater than the minimum split amount, forming a second payment channel set, assuming it is ["HCFC001", "HCFC002"].

[0184] The loan amount is split into multiple items. First, the amount range is discretized by a fixed step size of 1000, forming a discrete amount candidate set: {0,1000,2000,…,15000}.

[0185] Then, using a restricted greedy algorithm and a backtracking pruning algorithm, all order splitting schemes that satisfy the constraints are generated, such as: [5000,10000], [6000,9000], [7000,8000], [5000,5000,5000], etc. (constraints such as a maximum of 3 splitting orders and a minimum single order amount of 5000 yuan must be met).

[0186] The matching relationship between each order splitting scheme and the optional payment channels is modeled as a bipartite graph, a cost matrix is ​​constructed, and then the Hungarian algorithm is run to output the minimum matching cost under each order splitting scheme.

[0187] Iterate through all the above order splitting schemes, calculate the minimum matching cost for each scheme, and select the scheme with the minimum cost as the final loan disbursement plan, i.e., make the optimal choice. Assume that the optimal final loan disbursement scheme (i.e., the second loan disbursement execution scheme) is:

[0188] [{"channel_code":"HCFC001","amt":7000},{"channel_code":"HCFC002","amt":8000}]

[0189] Based on the final selected lending channel and corresponding lending amount, the lending message required by the corresponding channel is assembled for lending processing. For example, for HCFC001 and HCFC002, messages are assembled according to their specified message formats and sent to the corresponding channels for lending. Simultaneously, relevant payment information is recorded in the relevant tables of the corresponding payment channel database, so that the system can use scheduled tasks to obtain the latest time window information in real time and perform feature processing, thereby promoting more accurate predictions by the prediction model.

[0190] It should be understood that the steps, processes, and operations described in the above embodiments can be implemented by computer software programs. Therefore, the embodiments of the present invention also relate to a computer program product, which includes a computer program that, when executed by a processor, implements the steps, processes, or operations of the intelligent loan routing method described in any embodiment of the present invention. In some embodiments, program code for executing the computer program of the embodiments of the present invention can be written in any combination of one or more programming languages. For example, the computer program can be written using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on a personal computing device, partially on a personal computing device, partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the personal computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0191] It should also be understood that the computer program product may be stored on a computer-readable storage medium, such as hard disks, floppy disks, magnetic tapes, optical disks, solid-state drives, flash memory, etc.

[0192] Furthermore, it can be understood that the intelligent loan routing method of any embodiment of the present invention can be executed by a computer device. Accordingly, embodiments of the present invention also relate to a computer device, comprising: a memory storing a computer program thereon; and a processor that, when executing the computer program, implements the steps (processing, operations, etc.) of the intelligent loan routing method described in any embodiment of the present disclosure.

[0193] Figure 3 An exemplary structure of a computer device is shown. Figure 3As shown, the computer device 100 may include at least a processor 101 and a memory 102, wherein the memory 102 may be or include a storage medium storing a computer program (or computer-readable instructions). The processor 101 executes the computer program to perform some or all of the operations in the intelligent loan routing method described in any embodiment of this disclosure, including but not limited to: receiving a loan request initiated by a business system, the loan request including at least a product number and a loan amount; querying a preset product payment channel configuration table based on the product number to obtain an initial payment channel set associated with the product number; performing a status prediction on each payment channel in the initial payment channel set using a channel status prediction model to obtain a status prediction value for each payment channel in the initial payment channel set; and based on each payment channel... The predicted status value and the estimated handling fees of each payment channel obtained from the preset payment channel fee information table are used to select an available payment channel set from the initial payment channel set using a Pareto optimization model. Based on the preset loan disbursement payment channel table, it is determined whether there is a payment channel in the available payment channel set that can handle the loan amount independently. If so, this payment channel is selected as the first loan disbursement execution plan. If it is determined that there is no payment channel that can handle the loan amount independently, the loan amount is split into multiple sub-amounts according to the product number and a preset product splitting configuration table. Payment channels corresponding to these sub-amounts are then selected from the available payment channel set as the second loan disbursement execution plan. A loan disbursement message is assembled and processed according to either the first or second loan disbursement execution plan.

[0194] In some embodiments, the computer device 100 may further include a data storage device 103, a display 104, a speaker 105, and a communication module 106. The processor 101, memory 102, and data storage device 103 communicate with each other via a bus and interact with peripheral devices, such as the display 104 and communication module 106, via the bus and I / O module 107.

[0195] Specifically, data storage device 103 can store application programs and their various configuration files and data. Computer programs are stored on memory 102. Processor 101 executes the computer programs to perform various processes, operations, or steps of the intelligent routing loan method. Display 104 can be used to display various data, images, etc.

[0196] In optional embodiments, the communication module 106 may be omitted where communication with the other systems, devices, or equipment is not required. In some embodiments, the computer device 100 may further include a speaker 105 for voice output.

[0197] In some implementations, processor 101 may include any suitable semiconductor-based electronic processing unit, chip, microchip, or integrated circuit (IC). Memory 102 is a programmable memory, which may include any suitable electronic storage device configured to store instructions and be reprogrammable. For example, programmable memory may include an erasable programmable read-only memory (EPROM) device. Data storage device 103 may be a persistent storage device, which may include any suitable electronic storage configured to retain stored information as power cycles. For example, data storage device 103 may include a hard disk drive, solid-state drive (SSD), flash memory drive, hybrid drive, etc., or any combination thereof.

[0198] Therefore, processor 101 can control the output of information on I / O module 107 by storing information in memory and / or executing programs / instructions stored in memory. For example, several aspects of the methods described herein can be executed by processor 101 according to programs / instructions stored in memory (e.g., memory 102 and / or data storage device 103).

[0199] Furthermore, the processor 101 can electronically communicate with the I / O module 107 and / or the communication module 106 to receive or send relevant instructions and information. The I / O module 107 may include any suitable mechanical or virtual user interface configured to enable a user to interact with the computer device 100 or to allow the user to perform one or more functions of the computer device 100 itself, such as a graphical user interface (GUI) on a screen or other display. In some instances, the user interface may include a voice interface capable of speech recognition, through which an operator can provide voice commands to the processor.

[0200] The communication module 106 may include any suitable devices and / or structures configured to facilitate information exchange between the computer device 100 and external electronic devices (e.g., business system servers). The communication module 106 may include devices configured to send and / or receive wireless or wired information with other devices. For example, the communication module 106 may include antennas, transceivers, connectors for wired data reception and / or transmission, data exchange equipment, etc., or any combination thereof. The communication module 106 may also include ancillary components, such as filtering circuits, encryption / decryption circuits, and / or integrated circuit (IC) chips for processing signals (e.g., (Chip). In some implementations, the communication module 106 may include a WiFi device configured to connect to a local wireless network.

[0201] In some embodiments, computer device 100 may include a smartphone, wearable computer, portable / mobile electronic device, tablet computer, smartwatch, personal digital assistant (PDA), personal computer (PC), desktop computer, laptop computer, server, etc. Computer device 100 may include or have installed one or more application programs (APPs), one of which is configured to perform the smart routing loan disbursement method described herein.

[0202] Although not shown, it should be understood that the computer device 100 also includes a power supply component, which may include any suitable device and / or structure configured to provide an electrical interface between the computer device 100 and a power source. The power source may include any suitable source of electrical energy, such as a battery, socket, capacitor, fuel cell, etc., or any combination thereof. Additionally or optionally, the power source may be included within the power supply component. For example, a battery or battery pack may be included within the computer device 100. In some embodiments, the battery may be rechargeable, for example, charged via a cable or an interface provided by the power supply component. In some embodiments, the power supply component may share features with or repeat features of the communication module 106. For example, a USB or micro-USB cable connector may be included in the computer device 100, so that power supply or data communication is performed by the same component.

[0203] Those skilled in the art should understand that the above disclosure is merely illustrative of embodiments of the present invention, and the scope of patent protection claimed in this application is not limited thereto. Various modifications, alterations, substitutions, and other changes can be made to the embodiments disclosed herein without departing from the spirit of the invention, and such changes are within the scope covered by the claims of this application.

Claims

1. A smart loan routing method, characterized in that, include: Receive a loan request initiated by the business system, wherein the loan request includes at least the product number and the loan amount; Based on the product number, query the pre-set product payment channel configuration table to obtain the initial payment channel set associated with the product number; The status prediction of each payment channel in the initial payment channel set is performed by the channel status prediction model to obtain the status prediction value of each payment channel in the initial payment channel set. Based on the predicted status values ​​of each payment channel and the estimated transaction fees of each payment channel obtained from the pre-set payment channel fee information table, a Pareto optimization model is used to select an available payment channel set from the initial payment channel set. Based on the pre-set loan payment channel table, determine whether there is a payment channel in the set of available payment channels that can handle the loan amount independently. If there is, select that payment channel as the first loan execution plan. If it is determined that there is no payment channel that can handle the loan amount alone, the loan amount is split into multiple sub-amounts according to the product number and the preset product splitting configuration table, and the payment channels corresponding to the multiple sub-amounts are selected from the set of available payment channels as the second loan execution plan. According to the first loan disbursement execution plan or the second loan disbursement execution plan, the loan disbursement message is assembled and the loan is processed.

2. The intelligent loan routing method according to claim 1, characterized in that, A separate channel state prediction model is configured for each payment channel, and the channel state prediction model is based on a lightweight long short-term memory network (LSTM).

3. The intelligent loan routing method according to claim 2, characterized in that, For each payment channel in the initial payment channel set, the corresponding channel state prediction model is used to predict the state, including: The payment channel is retrieved from a time-series database for multiple time-series characteristic indicators within a predetermined time window in the past. These time-series characteristic indicators include statistical characteristics, trend characteristics, and periodic characteristics. The time-series characteristic indicators are input into the channel status prediction model of the payment channel to obtain the status prediction value of the payment channel within the future prediction window. The status prediction value includes at least the prediction success rate, the prediction average delay, and the prediction risk score.

4. The intelligent loan routing method according to claim 3, characterized in that, The time-series database includes a time-series feature index data table. The time-series characteristic index data table is formed in the following manner: By using scheduled tasks, historical loan disbursement data from various payment channels are aggregated at fixed time granularities to form a basic window data table; Based on the aforementioned basic window data table, time-series characteristic indicator data for each payment channel are calculated according to the channel coding dimension. The time-series characteristic indicator data includes the average success rate over the past 1 hour, the success rate trend over the past 6 hours, the maximum average delay over the past 1 hour, the minimum proportion of risk events over the past 6 hours, the average standard deviation of delay over the past 1 hour, the variance of success rate over the past 30 minutes, the standard deviation of the number of risk events over the past 1 hour, the average success rate of the same hour over the past 7 days, the average average delay of the same window over the past 4 weeks, the current window online status, the average instantaneous success rate monitored in real time over the past 30 minutes, the number of risk events in the previous window, the ratio of the average success rate to the average delay over the past 1 hour, and risk warnings. The time-series characteristic index data is stored in the time-series characteristic index data table according to the channel code.

5. The intelligent loan routing method according to claim 3, characterized in that, Also includes: Record the loan disbursement results and use these results to update the time series database for subsequent retraining of the channel status prediction model.

6. The intelligent loan routing method according to claim 3, characterized in that, The step of selecting the set of available payment channels from the initial set of payment channels using the Pareto optimization model includes: The multiple target values ​​of each payment channel in the initial payment channel set are unified into a minimization direction, wherein the multiple target values ​​include at least the predicted success rate, the predicted risk score, the predicted average delay, and the estimated transaction fee; By applying the Pareto dominance relation, all non-dominated solutions are selected from the initial set of payment channels to form the Pareto front set; Based on the product number, query the product payment channel configuration table to obtain the weight coefficients corresponding to the predicted success rate, predicted risk score, predicted average delay, and estimated handling fee. The payment channels in the Pareto frontier set are weighted and scored based on the weighting coefficients, and then sorted from high to low scores to form the set of available payment channels.

7. The intelligent loan routing method according to claim 6, characterized in that, Also includes: Before applying the Pareto dominance relationship, each payment channel in the initial payment channel set is filtered according to the preset business bottom line rules. The business bottom line rules include a predicted risk score threshold and a predicted average delay threshold, thereby eliminating payment channels that exceed the predicted risk score threshold or the predicted average delay threshold.

8. The intelligent loan routing method according to claim 1, characterized in that, When it is determined that there is no payment channel capable of processing the loan amount independently, the loan amount is split into the following steps: Based on the rules in the product splitting configuration table, generate multiple loan amount splitting schemes that meet the constraints; Construct a bipartite graph with the splitting scheme and the payment channel as two nodes, where the weight of the edge is the cost incurred by the sub-amount in the splitting scheme on the payment channel; The Hungarian matching algorithm is applied to calculate the minimum matching cost under each splitting scheme, and the splitting scheme with the lowest cost and its channel matching result are selected as the second loan disbursement execution scheme.

9. The intelligent loan routing method according to claim 8, characterized in that, The generation of multiple loan amount splitting schemes that satisfy the constraints includes: The loan amount is discretized by a fixed step size to generate a discrete amount candidate set; A restricted greedy algorithm and a backtracking pruning algorithm are used to search for all combinations of amounts that satisfy the constraints of minimum split amount and maximum number of splits from the discrete amount candidate set, forming a set of splitting schemes.

10. The intelligent loan routing method according to claim 1, characterized in that, Also includes: After obtaining the initial payment channel set, the system queries the pre-set payment channel table based on the entrusted payment field in the loan request, and filters out payment channels that support entrusted payment and are in the enabled state, so as to modify the initial payment channel set.

11. An intelligent loan routing system, which executes the intelligent loan routing method according to any one of claims 1 to 10, wherein, The intelligent loan routing system includes: The request receiving and preprocessing module is used to: receive a loan request initiated by the business system, wherein the loan request includes at least a product number and a loan amount; and based on the product number, query a pre-set product payment channel configuration table to obtain an initial payment channel set associated with the product number. The channel status prediction module is used to predict the status of each payment channel in the initial payment channel set through the channel status prediction model, and obtain the status prediction value of each payment channel in the initial payment channel set. The multi-objective optimization processing module is used to select a set of available payment channels from the initial set of payment channels based on the predicted state values ​​of each payment channel and the estimated handling fees of each payment channel obtained from the preset payment channel fee information table, using the Pareto optimization model. The amount splitting and optimal matching module is used to: determine whether there is a payment channel in the set of available payment channels that can handle the loan amount independently, based on a preset loan payment channel table; if there is, select the payment channel as the first loan execution plan; if it is determined that there is no payment channel that can handle the loan amount independently, split the loan amount into multiple sub-amounts according to the product number and a preset product splitting configuration table, and select the payment channels corresponding to the multiple sub-amounts from the set of available payment channels as the second loan execution plan. The loan disbursement triggering execution module is used to assemble a loan disbursement message and perform loan disbursement processing according to the first loan disbursement execution plan or the second loan disbursement execution plan.

12. A computer program product comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the intelligent loan routing method according to any one of claims 1 to 10.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the intelligent loan routing method according to any one of claims 1 to 10.

14. A computer device, characterized in that, include: A memory on which computer programs are stored; A processor that, when executing the computer program, implements the steps of the intelligent loan routing method according to any one of claims 1 to 10.

Images