Resource scheduling method and device, electronic equipment, medium and program product

Abstract

This application provides a resource scheduling method, apparatus, electronic device, medium, and program product, relating to the application of large-scale models in fintech scenarios, and applicable to the fields of big data technology and artificial intelligence technology. The method includes: acquiring multi-source heterogeneous data, wherein the multi-source heterogeneous data includes at least customer-side, environment-side, and system-side data; constructing a multimodal feature representation based on the multi-source heterogeneous data; inputting the multimodal feature representation into a customer behavior prediction model to perform multi-task prediction, obtaining prediction state information; obtaining resource allocation strategy information based on the prediction state information and resource scheduling objectives; and performing resource scheduling based on the resource allocation strategy information.

Description

Technical Field

[0001] This application relates to the application of large models in financial technology scenarios, and further to the fields of big data technology and artificial intelligence technology, and more specifically to a resource scheduling method, device, equipment, medium and program product. Background Technology

[0002] With the rapid evolution of financial markets, banks are gradually introducing quantitative modeling and data-driven methods into asset and liability management to improve their responsiveness to customer behavior and market changes.

[0003] However, existing technologies still have significant shortcomings in the following aspects: First, customer financial behavior modeling mostly relies on static statistical methods or rule-driven models, making it difficult to capture the dynamic impact of complex market environment changes on customer preferences and capital flows; Second, adjustments to the maturity structure of assets and liabilities are usually based on macroeconomic indicators or empirical rules, lacking the ability to analyze and predict the linkage of micro-level customer behavior patterns; Third, the impact of market changes on the asset and liability sides of banks is asymmetrical, and existing systems lack a real-time identification and modeling mechanism for such asymmetrical shocks, resulting in a lag in asset-liability matching adjustments. Summary of the Invention

[0004] In view of the above problems, this application provides resource scheduling methods, apparatus, equipment, media and program products.

[0005] According to a first aspect of this application, a resource scheduling method is provided, the method comprising: acquiring multi-source heterogeneous data, the multi-source heterogeneous data including at least customer-side, environment-side, and system-side data; constructing a multimodal feature representation based on the multi-source heterogeneous data; inputting the multimodal feature representation into a customer behavior prediction model to perform multi-task prediction to obtain prediction state information; obtaining resource allocation strategy information based on the prediction state information and a resource scheduling objective; and performing resource scheduling based on the resource allocation strategy information.

[0006] According to an embodiment of this application, the step of inputting the multimodal feature representation into a customer behavior prediction model to perform multi-task prediction and obtain prediction state information includes: performing multi-task prediction based on the multimodal features using multiple task sub-models in the customer behavior prediction model, wherein: based on the first task sub-model, the customer's deposit loss probability is predicted, and a loss risk level is output; based on the second task sub-model, the customer's loan repayment time is predicted, and risk function information is output; based on the third task sub-model, the customer's financial migration is predicted, and a migration probability distribution is output; and the loss risk level, the risk function information, and the migration probability distribution are fused to obtain the prediction state information.

[0007] According to embodiments of this application, the training process of the customer behavior prediction model includes at least one of the following: constructing target training samples based on a financial knowledge graph, and training the customer behavior prediction model based on the target training samples, wherein the target training samples include at least causal relationship information between customer behavior and market events; constructing supervised training samples based on cue word engineering, and training the customer behavior prediction model based on the supervised training samples, wherein the supervised training samples include at least financial scenario labels; and training the customer behavior prediction model based on parameter fine-tuning of a large model, wherein parameter fine-tuning includes updating the matrix parameters with preset rank constraints in the attention mechanism of the large model.

[0008] According to embodiments of this application, the training process of the customer behavior prediction model includes at least one of the following: constructing a counterfactual intervention scenario, calculating the average treatment effect of the counterfactual intervention scenario on customer behavior, and training the customer behavior prediction model based on the average treatment effect; training the customer behavior prediction model based on instrumental variables, wherein the instrumental variables are used to control for confounding factors including at least market sentiment.

[0009] According to an embodiment of this application, obtaining resource allocation strategy information based on the predicted state information and the resource scheduling objective includes: mapping the predicted state information into a state space vector; defining the action space of the reinforcement learning model; constructing a multi-objective reward function based on the resource scheduling objective; and generating the resource allocation strategy information using the reinforcement learning model based on the state space vector, the action space, and the multi-objective reward function.

[0010] According to embodiments of this application, the constraints of the action space include hard constraints, which are used to limit target resource allocation operations that can meet preset regulatory indicator requirements; and / or the multi-objective reward function includes soft constraints, which are used to penalize resource allocation operations that do not meet the target optimization requirements, or to reward resource allocation operations that exceed the target optimization requirements.

[0011] According to an embodiment of this application, the method further includes: performing interpretability analysis on the resource allocation strategy information to obtain explanatory information of the resource allocation strategy information; determining the risk level of the resource allocation strategy information based on market volatility; and iteratively optimizing the resource allocation strategy information based on the explanatory information and the risk level.

[0012] According to an embodiment of this application, the method further includes at least one of the following: acquiring incremental customer behavior data based on a preset time interval, incrementally training the customer behavior prediction model using the incremental customer behavior data and an elastic weight consolidation algorithm; deploying an original resource allocation strategy and an updated resource allocation strategy based on the resource allocation strategy information in parallel, and performing resource scheduling based on the updated resource allocation strategy in response to the updated resource allocation strategy's improvement in operational indicators relative to the original resource allocation strategy within a preset evaluation period reaching a preset threshold.

[0013] A second aspect of this application provides a resource scheduling apparatus, comprising: a data acquisition module, configured to: acquire multi-source heterogeneous data, the multi-source heterogeneous data including at least customer-side, environment-side, and system-side data, and construct a multimodal feature representation based on the multi-source heterogeneous data; a multi-task prediction module, configured to: input the multimodal feature representation into a customer behavior prediction model to perform multi-task prediction and obtain prediction status information; a resource allocation strategy acquisition module, configured to: obtain resource allocation strategy information based on the prediction status information and resource scheduling objectives; and a scheduling module, configured to: perform resource scheduling based on the resource allocation strategy information.

[0014] According to embodiments of this application, the multi-task prediction module can also be used to perform multi-task prediction based on the multimodal features and using multiple task sub-models in the customer behavior prediction model. Specifically, it predicts the probability of customer deposit churn based on the first task sub-model and outputs a churn risk level; it predicts the loan repayment time of customer based on the second task sub-model and outputs risk function information; it predicts the financial migration of customer based on the third task sub-model and outputs a migration probability distribution; and it fuses the churn risk level, the risk function information, and the migration probability distribution to obtain the prediction state information.

[0015] According to an embodiment of this application, the resource allocation strategy acquisition module can also be used to map the predicted state information into a state space vector; define the action space of the reinforcement learning model; construct a multi-objective reward function based on the resource scheduling objective; and generate the resource allocation strategy information using the reinforcement learning model based on the state space vector, the action space, and the multi-objective reward function.

[0016] According to embodiments of this application, the resource allocation strategy acquisition module can also be used to perform interpretability analysis on the resource allocation strategy information to obtain explanatory information of the resource allocation strategy information; determine the risk level of the resource allocation strategy information based on market volatility; and iteratively optimize the resource allocation strategy information based on the explanatory information and the risk level.

[0017] According to embodiments of this application, the scheduling module can also be used to deploy the original resource configuration strategy and the updated resource configuration strategy based on the resource configuration strategy information in parallel. In response to the updated resource configuration strategy's improvement in operational indicators relative to the original resource configuration strategy within a preset evaluation period reaching a preset threshold, resource scheduling is performed based on the updated resource configuration strategy.

[0018] According to embodiments of this application, the resource scheduling device may further include a training module. The training module can be used to construct target training samples based on a financial knowledge graph, and train the customer behavior prediction model based on the target training samples, wherein the target training samples at least include causal relationship information between customer behavior and market events; construct supervised training samples based on cue word engineering, and train the customer behavior prediction model based on the supervised training samples, wherein the supervised training samples at least include financial scenario labels; and train the customer behavior prediction model by fine-tuning the parameters of a large model, wherein the parameter fine-tuning includes updating the matrix parameters with preset rank constraints in the attention mechanism of the large model.

[0019] According to embodiments of this application, the training module can also be used to construct counterfactual intervention scenarios, calculate the average treatment effect of the counterfactual intervention scenarios on customer behavior, train the customer behavior prediction model based on the average treatment effect, and train the customer behavior prediction model based on instrumental variables, wherein the instrumental variables are used to control for confounding factors including at least market sentiment.

[0020] According to an embodiment of this application, the training module can also be used to acquire incremental customer behavior data based on a preset time interval, and use the incremental customer behavior data and the elastic weight consolidation algorithm to incrementally train the customer behavior prediction model.

[0021] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0022] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0023] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0024] According to embodiments of this application, by introducing multi-source heterogeneous data fusion and a multi-task customer behavior prediction model, the asymmetric propagation path of market shocks on the asset and liability sides can be identified, and optimized scheduling strategies can be generated in a timely manner. This improves system computing performance while enhancing the real-time and intelligent nature of resource scheduling. Simultaneously, the prediction model can capture customer behavior change trends in advance, enabling resource scheduling to shift from passive response to proactive adjustment, thereby significantly shortening resource scheduling response time and improving the system's real-time adaptability. Furthermore, dynamic prediction and precise scheduling based on customer behavior can provide service resources that better meet individual needs, thereby enhancing customer trust and satisfaction with the system's services. Attached Figure Description

[0025] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0026] Figure 1 The illustrations depict application scenarios of resource scheduling methods, apparatuses, devices, media, and program products according to embodiments of this application.

[0027] Figure 2 A flowchart illustrating a resource scheduling method according to an embodiment of this application is shown schematically;

[0028] Figure 3 A flowchart illustrating a method for performing multi-task prediction according to some exemplary embodiments of this application is shown schematically;

[0029] Figure 4 A flowchart illustrating a method for obtaining resource allocation strategy information according to some exemplary embodiments of this application is shown schematically;

[0030] Figure 5 A schematic diagram illustrating the structure of a resource scheduling apparatus according to an embodiment of this application is shown; and

[0031] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a resource scheduling method according to an embodiment of this application. Detailed Implementation

[0032] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0033] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0034] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0035] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0036] First, the technical terms used in this article are explained and clarified as follows.

[0037] Prompt engineering is a technique that guides a pre-trained language model to generate the desired output by designing specific input templates or prompts. It is often used to improve the controllability and task adaptability of large models.

[0038] Counterfactual prediction: an analytical method that simulates the possible outcomes of a system in the absence of a certain event, thereby assessing the causal impact of that event.

[0039] Instrumental variables (IVs) are a class of external variables used in statistical modeling to address the endogeneity problem in causal inference. They must satisfy the conditions of correlation and independence.

[0040] Reinforcement Learning (RL) is a learning method based on the interaction between an agent and its environment. The agent learns optimal behavioral strategies by attempting to obtain reward signals. In reinforcement learning or control systems, the state space is the set of all possible states describing the current state of the system, usually represented by a vector. The action space is the set of all possible actions the agent can take in a given state, and is an important component of reinforcement learning. The reward function, used in reinforcement learning, measures the merit of an action in a given state, providing numerical feedback to guide the learning direction.

[0041] Elastic Weight Consolidation (EWC) is a technique to prevent neural networks from forgetting old knowledge during incremental learning by weighting important parameters to limit their update range.

[0042] With the widespread application of big data, artificial intelligence, and other technologies in the financial sector, banks are gradually shifting from traditional experience-driven to data-driven and model-driven approaches in their asset and liability management practices. Especially against the backdrop of interest rate liberalization, diversified financial products, and increasingly stringent regulatory requirements, using intelligent methods to model and analyze changes in customer behavior and market fluctuations, and then dynamically adjusting the maturity structure of assets and liabilities, has become crucial for banks to improve their risk management capabilities and operational stability. However, existing technologies still face several technical bottlenecks in the intelligent implementation of asset and liability management, making it difficult to meet the demands for response speed, forecast accuracy, and flexibility in structural adjustments under the current complex environment.

[0043] Firstly, in terms of modeling customer financial behavior, current systems mostly employ statistical modeling methods based on static feature extraction or rule-driven scoring logic, which struggle to capture high-frequency trading behavior, short-term preference fluctuations, and nonlinear behavioral changes under the impact of macroeconomic variables. Such models lack the ability to model time-series dependencies and market state transition mechanisms, leading to lags or errors in predicting customer fund inflows and outflows, thus affecting the timeliness and accuracy of asset-liability matching strategies.

[0044] Secondly, in the process of adjusting the maturity structure of assets and liabilities, traditional practices often rely on macroeconomic interest rate trends, regulatory red lines, or rules of thumb for coarse-grained allocation, failing to effectively link with dynamic changes in micro-level customer behavior. For example, an increase in customer liquidity demand may indicate a decline in the stability of the liability side, but if changes in their preferences cannot be identified in real time and linked to asset reallocation, it may lead to the accumulation of short-term liquidity risks or duration mismatch problems.

[0045] Furthermore, current systems generally employ symmetric assumptions in perceiving and modeling influencing factors on both the asset and liability sides, making it difficult to address asymmetric shocks caused by changes in the market environment. In reality, factors such as interest rate changes, liquidity tightening, and reversals in market expectations often have different paths and intensities of impact on the asset side (e.g., loans, bond investments) and the liability side (e.g., customer deposits, interbank liabilities). If a modeling framework with the ability to differentiate between these factors cannot be introduced, it can easily lead to an imbalance in asset and liability adjustment strategies, increasing interest rate spread risk and maturity mismatch risk.

[0046] Based on this, embodiments of this application provide a resource scheduling method, including: acquiring multi-source heterogeneous data, wherein the multi-source heterogeneous data includes at least customer-side, environment-side, and system-side data; constructing a multimodal feature representation based on the multi-source heterogeneous data; inputting the multimodal feature representation into a customer behavior prediction model to perform multi-task prediction and obtain prediction state information; obtaining resource allocation strategy information based on the prediction state information and resource scheduling objectives; and performing resource scheduling based on the resource allocation strategy information. By introducing multi-source heterogeneous data fusion and a multi-task customer behavior prediction model, the asymmetric propagation path of market shocks on the asset and liability sides can be identified, and optimized scheduling strategies can be generated in a timely manner. This improves the system's computing performance while enhancing the real-time performance and intelligence of resource scheduling. Simultaneously, the prediction model can capture customer behavior change trends in advance, enabling resource scheduling to shift from passive response to proactive adjustment, thereby significantly shortening resource scheduling response time and improving the system's real-time adaptability. Furthermore, dynamic prediction and precise scheduling based on customer behavior can provide service resources that better meet individual needs, thereby enhancing customer trust and satisfaction with the system's services.

[0047] It should be noted that the resource scheduling method, apparatus, equipment, medium, and program products defined in this application relate to the application of large-scale models in fintech scenarios. They can be used in the fields of big data technology and artificial intelligence technology, as well as in the fintech field and various other fields besides big data technology, artificial intelligence technology, and fintech. The application fields of the resource scheduling method, apparatus, equipment, medium, and program products provided in the embodiments of this application are not limited.

[0048] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0049] In scenarios involving automated decision-making using personal information, the methods, devices, and systems provided in this application all offer users corresponding entry points for choosing to agree to or reject the automated decision-making results. If the user chooses to reject, the process proceeds to the expert decision-making stage. Here, "automated decision-making" refers to the activity of automatically analyzing and evaluating an individual's behavioral habits, interests, or economic, health, and credit status through computer programs, and then making a decision. Here, "expert decision-making" refers to the activity of making decisions by personnel who specialize in a particular field, possess specialized experience, knowledge, and skills, and have reached a certain level of professional expertise.

[0050] Figure 1 The illustration shows an application scenario diagram of the resource scheduling method, apparatus, device, medium, and program product according to embodiments of this application.

[0051] like Figure 1 As shown, application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0052] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0053] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smart mobile terminals, tablet computers, laptop computers, and desktop computers.

[0054] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0055] It should be noted that the resource scheduling method provided in this application embodiment can generally be executed by server 105. Correspondingly, the resource scheduling device provided in this application embodiment can generally be located in server 105. The resource scheduling method provided in this application embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the resource scheduling device provided in this application embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0056] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0057] The following will be based on Figure 1 The described scene, through Figures 2-4 The resource scheduling method of the disclosed embodiments will be described in detail.

[0058] Figure 2 A flowchart illustrating a resource scheduling method according to an embodiment of this application is shown.

[0059] like Figure 2 As shown, the resource scheduling method 200 of this embodiment includes operations S210 to S240.

[0060] In operation S210, multi-source heterogeneous data is acquired, which includes at least data from the client side, environment side, and system side. A multimodal feature representation is constructed based on the multi-source heterogeneous data.

[0061] In embodiments of this application, customer-side data may include customer behavior data. For example, customer behavior data may include deposit behavior data, such as recording the proportion of funds transferred out to reveal their daily cash flow habits.

[0062] For example, customer behavior data may include loan behavior data, including analyzing the correlation between mortgage prepayment rates and changes in the loan prime rate (LPR) and future housing price expectations.

[0063] For example, customer behavior data may also include channel behavior data, which comprehensively understands customers’ interaction preferences and demand signals by recording and analyzing customers’ operation logs on online channels such as mobile banking and online banking (e.g., the frequency of their use of the “deposit product comparison” function) and offline branch service demand data (e.g., the surge in inquiries about large-denomination certificates of deposit).

[0064] It should be noted that, to ensure data compliance and security, the embodiments of this application have strictly followed current laws and regulations and the regulatory requirements of the financial industry in the process of acquiring customer-related data. All data involving personal information has been obtained with the explicit authorization and consent of the customer in accordance with the law. Throughout the entire lifecycle of data collection, storage, processing, and transmission, the embodiments of this application employ multiple security protection measures, including data encryption, access control, anonymization or pseudonymization, and desensitization mechanisms, thereby reducing the risk of data leakage, misuse, and unauthorized access, and comprehensively protecting the integrity, confidentiality, and controllability of customer privacy and data assets.

[0065] In embodiments of this application, environmental data may include market environment data, such as interest rate indicators and market sentiment signals. System-side data may include internal data of financial institutions, such as core financial statements like a bank's balance sheet, profit and loss statement, and cash flow statement.

[0066] In the embodiments of this application, after acquiring multi-source heterogeneous data, a multimodal feature representation can be constructed, that is, the multi-source heterogeneous data is processed into a unified feature vector, thereby improving the model prediction accuracy and policy generation quality.

[0067] For example, time series decomposition can be performed on various behavioral data of customers, such as deposit withdrawals, loan repayments, and wealth management transactions. This allows for the identification of long-term trends (e.g., year-on-year growth in total deposits), short-term cycles (e.g., concentrated withdrawals at the end of the month or before holidays), and residuals (i.e., random fluctuations that cannot be explained by trends and cycles). Time series decomposition effectively extracts typical behavioral patterns, such as identifying patterns of concentrated withdrawals at the end of the month or capturing the characteristic of a consumption peak within three days after salary is deposited. Furthermore, to better capture long-term dependencies in transaction sequences, a Transformer architecture can be used to positionally encode the transaction sequences. This encoding method not only preserves the temporal order of transactions but also effectively captures the correlation between long-distance transaction events, such as identifying fund flow patterns spanning months or even years, thereby enhancing the model's accuracy in predicting future customer behavior.

[0068] For example, a comprehensive analysis of customer behavior can be conducted to assign customers descriptive tags with business insights, thereby enabling a high-level summary of specific customer attributes or preferences. For instance, if the system detects that a customer has recently been frequently logging into mobile banking and repeatedly viewing information related to "large-denomination certificate of deposit interest rates," that customer will be labeled as an "interest rate sensitive customer."

[0069] For example, contrastive learning methods can be used to map heterogeneous textual data (such as textual information extracted from customer service call logs and customer complaints) and numerical data (such as quantitative indicators like customer deposit attrition rate and delinquency rate) to the same high-dimensional vector space. The goal of this mapping is to bring semantically similar text and numerical data with similar behavioral patterns closer together in the vector space, thereby constructing a joint representation of behavior and the market. For instance, through cross-modal alignment, the model can understand that textual content such as "poor service attitude" mentioned in customer complaints may be strongly correlated with subsequent "deposit attrition" behavior, thus enhancing the model's robustness and prediction accuracy.

[0070] In operation S220, the multimodal feature representation is input into the customer behavior prediction model to perform multi-task prediction and obtain prediction state information.

[0071] Customer behavior prediction models can employ deep learning architectures that integrate structured and unstructured information, such as combining Transformer networks, graph neural networks, and multi-task learning frameworks, to achieve parallel modeling and prediction of different customer behavioral events. During model training, key behavioral variables such as potential liquidity risk, changing trends in financial preferences, or loan repayment intentions can be effectively identified by introducing behavioral sequence modeling, trend cycle decomposition, and financial domain embedding.

[0072] In the embodiments of this application, the customer behavior prediction model may include multiple sub-tasks, such as: customer deposit outflow probability prediction, which can predict the probability of potential fund transfers by customers within a certain period of time (e.g., 30 days) based on factors such as customers' recent transaction activity, product interest rate attractiveness, and social media sentiment index, and perform risk stratification; loan prepayment time prediction, which constructs a time risk curve based on survival analysis by analyzing customers' occupational stability, loan history, market interest rate expectations, etc., to provide early warning for loan asset allocation; and customer investment preference migration prediction, which models information such as customers' past holdings, risk assessment results, and macroeconomic fluctuations to output the probability distribution of customers' possible migration from money market products to bond, stock, and precious metal products, thereby supporting product recommendation and asset-side return management.

[0073] For example, customer behavior prediction can also be used to monitor differences in response to external shocks, such as predicting the degree of reaction of specific customer groups in different business scenarios after interest rate changes, regulatory policy adjustments, or major financial news releases, providing targeted support for subsequent resource allocation and strategy optimization. The system can also dynamically expand the task set according to strategic operational needs, such as predicting the dormant risk of new existing customers and predicting cross-selling response rates.

[0074] The output of predicted status information can be in the form of classification labels, probability distributions, risk scores, or time intervals. It can also be further mapped to asset liquidity warning levels, customer operating status levels, or intermediate variables for subsequent resource allocation, depending on business objectives. The system can be configured with prediction outputs of different granularities (such as customer level, cluster level, and region level) to adapt to diverse strategy formulation and resource allocation tasks.

[0075] In operation S230, resource configuration strategy information is obtained based on the predicted state information and resource scheduling target.

[0076] In the embodiments of this application, the resource allocation objective can represent the specific business purpose that a financial institution hopes to achieve through resource allocation. The resource allocation objective can be multi-dimensional, including, for example, maximizing overall returns, minimizing risk exposure, improving customer satisfaction, optimizing capital utilization efficiency, or meeting compliance requirements. The resource allocation objective can be statically preset or dynamically adjusted according to market conditions or the bank's operational situation.

[0077] In order to effectively transform predicted state information and resource scheduling objectives into executable strategies, embodiments of this application may employ a variety of intelligent decision-making methods, which can be selected or combined according to specific business needs, data availability, requirements for model interpretability, and system dynamism.

[0078] For example, reinforcement learning models can be utilized. The RL model is designed as an agent to learn and make decisions in a simulated financial environment.

[0079] For example, a rule engine and expert system can be used to generate strategies through a pre-defined set of business rules and an expert knowledge base. Business experts will transform all known resource scheduling logic, risk control principles, customer response thresholds, etc., into a series of "if-then" rules. The rule engine is responsible for parsing the predicted state information and matching the corresponding rules according to priority and logical relationships, ultimately outputting resource configuration strategy information.

[0080] For example, policy learning can be based on supervised learning. The mapping relationship between predicted state information, resource scheduling objectives, and effective resource allocation strategies actually adopted in historical data can be regarded as training samples for supervised learning. That is, the model learns what resource allocation operation should be taken given a certain predicted state and scheduling objective. Supervised learning models such as decision trees, random forests, support vector machines, or neural networks can be used for training to learn decision patterns from historical decision data.

[0081] In operation S240, resource scheduling is performed based on the resource configuration strategy information.

[0082] In embodiments of this application, resource allocation strategy information may include optimized resource allocation schemes tailored to specific customer groups, market environments, and internal organizational conditions. Resource allocation strategy information may be personalized product recommendations for individual customers, or it may be macro-level, such as adjusting the overall asset-liability structure or the capital allocation for specific business lines. Upon receiving resource allocation strategy information, the system can parse it and transform it into a series of atomic scheduling instructions to drive the execution of the actual business system.

[0083] In some embodiments, to ensure high timeliness and responsiveness of resource scheduling, a streaming processing engine can be used to handle policy execution. For example, a streaming processing framework can be used to receive and parse policy instructions and the latest business data streams in real time. To ensure timely policy response, the streaming processing engine can be configured with a rolling window (e.g., a 10-second or 1-minute rolling window, the specific duration of which can be flexibly configured according to business scenarios and latency requirements) to quickly aggregate and analyze real-time transaction and behavioral data, ensuring that the policy can be applied to the latest business status with extremely low latency.

[0084] In some embodiments, an execution priority queue can be introduced to dynamically sort scheduling instructions based on the impact weight of each strategy or the urgency of the business scenario it targets. For example, instructions aimed at replenishing liquidity can be given the highest priority; instructions aimed at optimizing revenue but not time-sensitive may be ranked with lower priority. The execution priority queue can ensure that critical strategy instructions are processed and executed first, especially when resources are limited or system load is high, avoiding delays to important business operations.

[0085] In the embodiments of this application, resource scheduling based on resource allocation strategy information, i.e., triggering corresponding business execution operations, may include: sending personalized marketing information or product recommendations to the customer management system; automatically adjusting the fund allocation ratio in the fund transfer system; updating risk limits, early warning parameters, or compliance verification rules in the risk management system to adapt to changes in risk exposure caused by strategy adjustments and ensure that risks are within an acceptable range; triggering the asset and liability management system to perform asset portfolio rebalancing, liability structure optimization, or interest rate risk hedging operations to adapt to market changes and regulatory requirements; or pushing operation guidelines, standard scripts, or decision support information to the counter operating system, customer service call center system, or automated service robot to guide frontline personnel in providing accurate services and handling business.

[0086] It should be noted that business execution operations can be integrated with existing financial business systems through standardized interfaces or message queues, thereby automating the implementation of strategies. The entire scheduling process can incorporate built-in security auditing and rollback mechanisms to ensure that every resource scheduling operation is traceable and verifiable, and can be corrected or revoked when necessary, thereby minimizing the impact of operational risks, system failures, or strategy deviations.

[0087] The resource scheduling method of this application will be specifically described below by way of preferred embodiments.

[0088] Figure 3 A flowchart illustrating a method for performing multi-task prediction according to some exemplary embodiments of this application is shown schematically.

[0089] like Figure 3 As shown, the method for performing multi-task prediction includes operations S310 to S320.

[0090] In operation S310, based on the multimodal features, multi-task prediction is performed using multiple task sub-models in the customer behavior prediction model. Specifically, based on the first task sub-model, the customer's deposit loss probability is predicted, and the loss risk level is output. Based on the second task sub-model, the customer's loan repayment time is predicted, and the risk function information is output. Based on the third task sub-model, the customer's financial migration is predicted, and the migration probability distribution is output.

[0091] In the embodiments of this application, the customer behavior prediction model can be a holistic system or framework capable of multi-task prediction. The holistic system can consist of multiple task sub-models sharing underlying components, or it can be a single complex model integrating multi-task outputs. By performing multi-task predictions through multiple task sub-models within the customer behavior prediction model, it is possible to gain parallel insights into customers' potential behaviors and risks across multiple dimensions in the financial field, thereby providing refined predictive basis.

[0092] Specifically, the first task sub-model can be used to predict the probability of customer deposit outflow. By analyzing factors such as the frequency of customer fund inflows and outflows, account balance fluctuations, product interest rate sensitivity, and changes in external market interest rates, the first task sub-model can predict the probability of a customer experiencing large-scale fund outflows or account inactivity within a specific future time period (e.g., the next month or quarter). Its output is the customer's churn risk level, which can be a continuous probability value or a discrete risk classification (e.g., high, medium, low risk).

[0093] The second task sub-model can be used to predict a customer's loan repayment time. It comprehensively assesses a customer's income stability, historical repayment record, and asset-liability status, and can incorporate macroeconomic indicators (such as regional unemployment rates and industry prosperity indices) to predict the customer's repayment probability and potential delinquency risk at a specific future point in time or within a future time window. Its output is a risk function, which quantifies the customer's combined repayment ability and willingness, for example, providing the probability of delinquency within the next N days or the distribution of remaining repayment life.

[0094] The third task sub-model can be used to predict clients' financial migration. By analyzing clients' current financial product holdings, risk preferences, historical subscription and redemption behavior, and combining this with factors such as market investment trends, the attractiveness of peer products, and changes in clients' asset allocation needs, the third task sub-model can predict the probability distribution of clients transferring funds from their current financial products to other products. For example, it can predict which type of product to migrate to (e.g., from money market funds to equity funds), the potential strength of the migration, and the time window in which the migration may occur.

[0095] It should be noted that the above task sub-models can employ different machine learning or deep learning architectures, such as deep neural networks, recurrent neural networks, long short-term memory networks, or Transformer architectures, to collaboratively mine deep patterns of customer behavior from multimodal features.

[0096] In operation S320, the churn risk level, the risk function information, and the migration probability distribution are fused to obtain the predicted state information.

[0097] For example, the fusion process can use a weighted average based on the importance of each prediction task, or utilize a fusion network to combine heterogeneous prediction results. Alternatively, an attention mechanism can be applied to dynamically allocate weights to different prediction results, allowing them to focus more on information from a specific key prediction dimension in a particular context. The goal of fusion is to eliminate potential redundancy between predictions from different tasks and to uncover their complementary information, thereby generating a comprehensive, unified, and multi-dimensional prediction state.

[0098] Predictive status information can be an abstract representation of a customer's comprehensive behavioral profile and risk situation over a future period of time, including key information such as the tendency of customer churn, the stability of repayment, and potential changes in asset allocation.

[0099] Furthermore, embodiments of this application can improve the training process when constructing and applying a customer behavior prediction model to ensure that the model can accurately capture complex customer behavior patterns and adapt to the ever-changing financial market environment. The training process may include a combination of at least one or more of the following methods to maximize the model's predictive accuracy and robustness.

[0100] For example, target training samples can be constructed based on financial knowledge graphs. As a structured form of knowledge representation, financial knowledge graphs can organize financial entities (such as customers, products, transactions, market events, regulatory policies, etc.) and their complex relationships (such as causality, attributes, affiliation, temporal order, etc.) in a graph form. Embodiments of this application utilize financial knowledge graphs to automatically or semi-automatically construct high-quality target training samples. Specifically, the target training samples not only include basic customer behavioral data but can also include at least causal relationship information between customer behavior and market events, thereby providing the model with deep-level causal correlation features and improving the model's interpretability and predictive accuracy in complex financial scenarios.

[0101] For example, supervised training samples can be constructed based on prompt word engineering. With the widespread application of large-scale pre-trained models in the field of natural language processing, embodiments of this application utilize prompt word engineering to transform raw unstructured or semi-structured data into supervised training samples that are easy for the model to understand and process. Specifically, pre-designed prompt word templates (such as "Please determine whether the following customer behavior description [customer behavior] belongs to [financial scenario label]?") can transform raw customer behavior records, customer service dialogues, market news, etc., into samples with clear question-and-answer or classification structures. The construction of supervised training samples allows the model to learn the mapping relationship between specific financial scenario labels and customer behavior, thereby efficiently utilizing a small number of manually or semi-automatically labeled labels to improve the model's classification and recognition capabilities in specific financial business scenarios.

[0102] For example, customer behavior prediction models can also be trained by fine-tuning the parameters of a large model. The parameter fine-tuning process involves inputting the aforementioned constructed financial knowledge graph samples or engineered prompt word samples into the large model and updating only a subset of the model's parameters. Specifically, the fine-tuning process can focus on updating the matrix parameters of the large model's attention mechanism that have pre-defined rank constraints. For instance, in the Transformer architecture, the key, value, and query matrices in the attention mechanism are high-dimensional. By performing low-rank decomposition on these matrices and updating only the low-rank portions, the number of parameters requiring training can be significantly reduced. This effectively avoids overfitting while maintaining the model's powerful performance, improves training efficiency, and allows the model to adapt more quickly to the complex patterns and terminology specific to the financial field.

[0103] In the embodiments of this application, the customer behavior prediction model can also be trained using some customer behavior prediction and causal reasoning methods.

[0104] For example, counterfactual intervention scenarios can be constructed. Traditional predictive models often only identify correlations between variables but cannot determine causal relationships, which may lead to unexpected results during actual intervention. Therefore, embodiments of this application construct a series of counterfactual intervention scenarios to simulate potential changes in customer behavior under specific conditions if a certain action is taken. For example, the system can simulate scenarios such as "how the customer's fund outflow rate will change if a bank significantly increases deposit interest rates at a specific point in time." For each constructed counterfactual scenario, the average treatment effect (ATE) of the counterfactual intervention on customer behavior can be calculated. ATE quantifies the net impact of a specific intervention on the target customer's behavior. Subsequently, the model can be trained based on ATE to improve the model's ability to simulate the effects of different strategies and evaluate the effects of real interventions.

[0105] For example, instrumental variables can also be introduced. In complex financial market environments, customer behavior is often influenced by multiple factors, some of which may simultaneously affect both customer behavior and bank intervention measures, creating confounding factors that are difficult to separate. For instance, market sentiment may influence both customer investment decisions and bank marketing strategies, obscuring the relationship between directly observed customer behavior and strategies. To address this issue, embodiments of this application can introduce instrumental variables during the training process of the customer behavior prediction model. An instrumental variable is a special type of variable that is related to bank intervention measures but does not directly affect customer behavior through confounding factors, nor is it directly influenced by customer behavior. For example, one possible instrumental variable is a confounding factor that includes at least market sentiment, but the instrumental variable itself is not market sentiment; rather, it is an external signal that can indirectly influence bank strategies and is independent of market sentiment. By incorporating instrumental variables into model training, embodiments of this application can control or counteract the influence of confounding factors, thereby more accurately identifying the true causal relationship between bank intervention strategies and customer behavior. This enables the model to more reliably predict the actual effects of resource allocation strategies when generating them, reducing decision-making bias caused by confounding factors.

[0106] Figure 4 A flowchart illustrating a method for obtaining resource configuration strategy information according to some exemplary embodiments of this application is shown.

[0107] like Figure 4 As shown, the method for obtaining spatial consistency verification includes operations S410 to S440.

[0108] In operation S410, the predicted state information is mapped into a state space vector.

[0109] In the embodiments of this application, in order for the RL model to understand and process the predicted state information, it needs to be converted into a numerical form that the model can recognize, namely a high-dimensional state space vector.

[0110] In operating S420, define the action space of the reinforcement learning model.

[0111] In the embodiments of this application, the action space defines all possible resource allocation operations that the RL model can perform. These operations can be specific, executable business instructions, representing the intervention measures that a bank can take when facing different customer states and market environments. For example, actions may include: adjusting deposit product interest rates for specific customer groups, increasing or decreasing the allocation of a certain type of asset, adjusting marketing budget allocation, or optimizing the scheduling of branch service personnel, etc. The granularity of the action space can be refined according to business needs; for example, different levels of interest rate adjustment, different types of asset allocation ratios, or differentiated strategies for customers with different risk levels can be defined.

[0112] To ensure operational compliance and security, the scope of action can be limited by hard constraints, such as ensuring that the final capital adequacy ratio does not fall below the regulatory threshold, or that the loan concentration of a single customer does not exceed a preset limit. Any operation that violates these hard constraints will be directly excluded by the model, thereby guaranteeing the feasibility and security of the strategy.

[0113] In operation S430, a multi-objective reward function is constructed based on the resource scheduling objective.

[0114] In the embodiments of this application, the reward function can quantify the achievement of business objectives by the RL model after performing specific resource allocation operations. Since bank business objectives are often multifaceted and may conflict with each other, the reward function can be designed as a multi-objective form to comprehensively consider different business indicators.

[0115] In embodiments of this application, penalties or rewards based on soft constraints can be introduced to flexibly guide the model to weigh and optimize among multiple objectives. Even if some objectives are not perfectly achieved, the model can learn to minimize negative impacts or pursue suboptimal solutions within permissible limits to find an equilibrium point under complex constraints. For example, positive rewards may be related to improved net interest margins, decreased non-performing loan ratios, increased customer retention rates, optimized asset allocation efficiency, and improved customer satisfaction; while negative penalties may be related to customer churn, increased funding costs, irregular operations, or failure to meet specific performance indicators.

[0116] In operation S440, the resource allocation strategy information is generated using the reinforcement learning model based on the state space vector, the action space, and the multi-objective reward function.

[0117] In the embodiments of this application, the RL model continuously interacts with a simulated financial environment, explores different actions in a given state space, and learns the optimal strategy based on feedback from the reward function to generate resource allocation strategy information. This resource allocation strategy information can guide the efficient and intelligent allocation and scheduling of various resources such as funds, products, and personnel when faced with specific customer behavior predictions and market environments.

[0118] In the embodiments of this application, after obtaining the resource configuration strategy information, further processing can be performed to ensure the interpretability and risk controllability of the strategy, and to achieve continuous iterative optimization.

[0119] Specifically, interpretability analysis can be performed on resource allocation strategy information to obtain explanatory information. For example, interpretability analysis methods can be used to quantify the influence weights and directional contributions of different input features on the final strategy decision. Explanatory information not only helps business experts understand and trust the model, but also identifies potential business logic loopholes or model biases, providing a basis for subsequent strategy adjustments.

[0120] Furthermore, the risk level of resource allocation strategy information can be determined based on market volatility. Embodiments of this application can introduce market volatility indicators as key risk factors; for example, real-time monitoring of the fear index, stock index futures volatility, and implied volatility in the bond market. When the market volatility indicator reaches a preset threshold, the system can adjust the risk level of the resource allocation strategy accordingly, such as limiting the allocation ratio of high-risk assets or reducing the overall aggressiveness of the strategy. This automatically switches to a more prudent and robust strategy execution mode when market uncertainty increases, effectively preventing systemic risks and market shocks.

[0121] Based on the aforementioned explanatory information and risk levels, resource allocation strategies can be iteratively optimized to ensure continuous learning and improvement. Explanatory information obtained through interpretability analysis can help identify potential flaws, biases, or unexpected behavioral patterns in the strategy; the risk level determined by market volatility can influence the direction and magnitude of strategy adjustments.

[0122] In embodiments of this application, online learning of the model can also be introduced. For example, incremental customer behavior data can be acquired based on preset time intervals (e.g., daily or weekly), and the incremental customer behavior data can be used to incrementally train the customer behavior prediction model. To effectively prevent catastrophic forgetting that may occur during incremental training (i.e., the model forgets old knowledge while learning new knowledge), an elastic weight consolidation algorithm can be adopted. By calculating the importance of model parameters, constraints are imposed on important parameters when updating parameters, thereby effectively consolidating the model's existing knowledge while learning new data patterns, ensuring the continued high accuracy and stability of the prediction model.

[0123] In the embodiments of this application, a strategy comparison experiment (i.e., A / B testing) can be introduced to mitigate the risks that may arise from a direct full rollout. Specifically, the original resource allocation strategy and an updated resource allocation strategy based on the resource allocation strategy information can be deployed in parallel. This means that for a period of time, some customer groups or business scenarios will continue to use the existing strategy, while others will apply the new strategy. Within a preset evaluation period (e.g., one week or one month), the system will continuously compare the performance of the two strategies on key operational indicators. These key operational indicators may include the increase in net interest margin, changes in the non-performing loan ratio, a decrease in customer churn rate, an improvement in liquidity coverage ratio, and the sales conversion rate of specific products. Only when the updated resource allocation strategy improves the operational indicators relative to the original resource allocation strategy within the preset evaluation period to a preset threshold (e.g., an increase in net interest margin ≥ 5% or a decrease in the non-performing loan ratio ≥ X%) will the system perform resource scheduling based on the updated resource allocation strategy. This rollout mechanism ensures that only new strategies that have been fully validated and proven effective will be fully rolled out, thereby guaranteeing the bank's business stability and the scientific nature of its decision-making.

[0124] Furthermore, embodiments of this application can also utilize real-time dashboards for continuous and visual monitoring of key risk indicators. By integrating professional visualization tools, real-time dashboards enable business decision-makers and risk managers to intuitively view the dynamic changes of multiple core indicators in real time. These include, for example: position changes (e.g., real-time changes in the size and composition of various assets and liabilities); derivative exposure, i.e., the bank's real-time risk exposure to various derivative contracts; and capital consumption rate, i.e., the rate at which the bank's capital is consumed in operations and risk management.

[0125] To address unforeseen risks and prevent potential negative impacts from strategy execution, an automatic circuit breaker mechanism can be introduced. This mechanism is an intelligent risk mitigation measure that automatically suspends the execution of subsequent strategies and immediately triggers a manual review process when specific risk indicators reach preset danger thresholds. For example, the system can continuously monitor the bank's liquidity coverage ratio, a key liquidity indicator. If, within a specific time window after the resource allocation strategy is executed (e.g., within one hour), the liquidity coverage ratio experiences a significant decline exceeding a preset percentage (e.g., a decline exceeding 5%), the automatic circuit breaker mechanism will be activated immediately. Once the circuit breaker is triggered, all subsequent unexecuted intelligent strategy instructions will be immediately suspended, and the sending of new scheduling instructions to the business system will cease.

[0126] Corresponding to the above-described resource scheduling method, embodiments of this application also provide a resource scheduling device.

[0127] Figure 5 A schematic block diagram of a resource scheduling apparatus according to an embodiment of this application is shown.

[0128] like Figure 5 As shown, the resource scheduling device 500 in this embodiment includes a data acquisition module 510, a multi-task prediction module 520, a resource configuration strategy acquisition module 530, and a scheduling module 540.

[0129] The data acquisition module 510 can be used to acquire multi-source heterogeneous data, which includes at least data from the client side, environment side, and system side, and construct a multimodal feature representation based on the multi-source heterogeneous data. In one embodiment, the data acquisition module 510 can be used to perform the operation S210 described above, which will not be repeated here.

[0130] The multi-task prediction module 520 can be used to input the multimodal feature representation into the customer behavior prediction model to perform multi-task prediction and obtain prediction state information. In one embodiment, the multi-task prediction module 520 can be used to perform the operation S220 described above, which will not be repeated here.

[0131] The resource allocation strategy acquisition module 530 can be used to obtain resource allocation strategy information based on the predicted state information and resource scheduling target. In one embodiment, the resource allocation strategy acquisition module 530 can be used to execute the operation S230 described above, which will not be repeated here.

[0132] The scheduling module 540 can be used to perform resource scheduling based on the resource configuration strategy information. In one embodiment, the scheduling module 540 can be used to execute the operation S240 described above, which will not be repeated here.

[0133] According to an embodiment of this application, the multi-task prediction module 520 can also be used to perform multi-task prediction based on the multimodal features and using multiple task sub-models in the customer behavior prediction model. Specifically, it predicts the probability of customer deposit churn based on the first task sub-model and outputs a churn risk level; it predicts the loan repayment time of customer based on the second task sub-model and outputs risk function information; it predicts the financial migration of customer based on the third task sub-model and outputs a migration probability distribution; and it fuses the churn risk level, the risk function information, and the migration probability distribution to obtain the prediction state information.

[0134] According to an embodiment of this application, the resource allocation strategy acquisition module 530 can also be used to map the predicted state information into a state space vector; define the action space of the reinforcement learning model; construct a multi-objective reward function based on the resource scheduling objective; and generate the resource allocation strategy information using the reinforcement learning model based on the state space vector, the action space, and the multi-objective reward function.

[0135] According to an embodiment of this application, the resource allocation strategy acquisition module 530 can also be used to perform interpretability analysis on the resource allocation strategy information to obtain explanatory information of the resource allocation strategy information; determine the risk level of the resource allocation strategy information based on market volatility; and iteratively optimize the resource allocation strategy information based on the explanatory information and the risk level.

[0136] According to an embodiment of this application, the scheduling module 540 can also be used to deploy the original resource configuration strategy and the updated resource configuration strategy based on the resource configuration strategy information in parallel. In response to the updated resource configuration strategy's improvement in operational indicators relative to the original resource configuration strategy within a preset evaluation period reaching a preset threshold, resource scheduling is performed based on the updated resource configuration strategy.

[0137] According to an embodiment of this application, the resource scheduling device 500 may further include a training module. The training module can be used to construct target training samples based on a financial knowledge graph, and train the customer behavior prediction model based on the target training samples, wherein the target training samples at least include causal relationship information between customer behavior and market events; construct supervised training samples based on cue word engineering, and train the customer behavior prediction model based on the supervised training samples, wherein the supervised training samples at least include financial scenario labels; and train the customer behavior prediction model by fine-tuning the parameters of a large model, wherein the parameter fine-tuning includes updating the matrix parameters with preset rank constraints in the attention mechanism of the large model.

[0138] According to embodiments of this application, the training module can also be used to construct counterfactual intervention scenarios, calculate the average treatment effect of the counterfactual intervention scenarios on customer behavior, train the customer behavior prediction model based on the average treatment effect, and train the customer behavior prediction model based on instrumental variables, wherein the instrumental variables are used to control for confounding factors including at least market sentiment.

[0139] According to an embodiment of this application, the training module can also be used to acquire incremental customer behavior data based on a preset time interval, and use the incremental customer behavior data and the elastic weight consolidation algorithm to incrementally train the customer behavior prediction model.

[0140] According to embodiments of this application, any multiple modules among the data acquisition module 510, multi-task prediction module 520, resource allocation strategy acquisition module 530, and scheduling module 540 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the data acquisition module 510, multi-task prediction module 520, resource allocation strategy acquisition module 530, and scheduling module 540 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the data acquisition module 510, the multi-task prediction module 520, the resource configuration strategy acquisition module 530, and the scheduling module 540 can be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0141] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a resource scheduling method according to an embodiment of this application.

[0142] like Figure 6As shown, an electronic device 600 according to an embodiment of this application includes a processor 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage portion 608 into a random access memory (RAM) 603. The processor 601 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 601 may also include onboard memory for caching purposes. The processor 601 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0143] RAM 603 stores various programs and data required for the operation of electronic device 600. Processor 601, ROM 602, and RAM 603 are interconnected via bus 604. Processor 601 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 602 and / or RAM 603. It should be noted that the programs may also be stored in one or more memories other than ROM 602 and RAM 603. Processor 601 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0144] According to embodiments of this application, the electronic device 600 may further include an input / output (I / O) interface 605, which is also connected to a bus 604. The electronic device 600 may also include one or more of the following components connected to the input / output (I / O) interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the input / output (I / O) interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 610 as needed so that computer programs read from it can be installed into the storage section 608 as needed.

[0145] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0146] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include ROM 602 and / or RAM 603 and / or one or more memories other than ROM 602 and RAM 603 described above.

[0147] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to enable the computer system to implement the resource scheduling method provided in the embodiments of this application.

[0148] When the computer program is executed by the processor 601, it performs the functions defined in the system / apparatus of this application embodiment. According to the embodiments of this application, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0149] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 609, and / or installed from the removable medium 611. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0150] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 609, and / or installed from the removable medium 611. When the computer program is executed by the processor 601, it performs the functions defined in the system of this application embodiment. According to the embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0151] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented 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 the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via 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).

[0152] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0153] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

[0154] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Without departing from the scope of this application, those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. A resource scheduling method, characterized in that, The method includes: Acquire multi-source heterogeneous data, which includes at least data from the customer side, environment side, and system side, and construct multimodal feature representations based on the multi-source heterogeneous data; The multimodal feature representation is input into the customer behavior prediction model to perform multi-task prediction and obtain prediction state information; Based on the predicted state information and resource scheduling objectives, resource allocation strategy information is obtained; and Resource scheduling is performed based on the resource configuration strategy information.

2. The method according to claim 1, characterized in that, The step of inputting the multimodal feature representation into the customer behavior prediction model to perform multi-task prediction and obtain prediction state information includes: Based on the aforementioned multimodal features, multi-task prediction is performed using multiple task sub-models in the customer behavior prediction model. Specifically, the first task sub-model predicts the probability of customer deposit churn and outputs the churn risk level; the second task sub-model predicts the customer's loan repayment time and outputs risk function information; and the third task sub-model predicts customer financial migration and outputs a migration probability distribution. The predicted state information is obtained by fusing the churn risk level, the risk function information, and the migration probability distribution.

3. The method according to claim 2, characterized in that, The training process of the customer behavior prediction model includes at least one of the following: The target training samples are constructed based on a financial knowledge graph, and the customer behavior prediction model is trained based on the target training samples. The target training samples include at least causal relationship information between customer behavior and market events. A supervised training sample is constructed based on the prompt word engineering, and the customer behavior prediction model is trained based on the supervised training sample, wherein the supervised training sample includes at least financial scenario labels; The customer behavior prediction model is trained by fine-tuning the parameters of a large model, wherein the fine-tuning includes updating the matrix parameters with preset rank constraints in the attention mechanism of the large model.

4. The method according to claim 2 or 3, characterized in that, The training process of the customer behavior prediction model includes at least one of the following: Construct counterfactual intervention scenarios, calculate the average treatment effect of the counterfactual intervention scenarios on customer behavior, and train the customer behavior prediction model based on the average treatment effect; The customer behavior prediction model is trained based on instrumental variables, wherein the instrumental variables are used to control for confounding factors, including at least market sentiment.

5. The method according to any one of claims 1 to 3, characterized in that, The step of obtaining resource allocation strategy information based on the predicted state information and resource scheduling objectives includes: The predicted state information is mapped into a state space vector; Define the action space of the reinforcement learning model; Based on the resource scheduling objective, a multi-objective reward function is constructed; and Based on the state space vector, the action space, and the multi-objective reward function, the resource allocation strategy information is generated using the reinforcement learning model.

6. The method according to claim 5, characterized in that, The constraints of the action space include hard constraints, which are used to limit the target resource allocation operations that can meet preset regulatory indicator requirements; and / or The multi-objective reward function includes soft constraints, which are used to penalize resource allocation operations that do not meet the objective optimization requirements, or to reward resource allocation operations that exceed the objective optimization requirements.

7. The method according to claim 5, characterized in that, The method further includes: Perform interpretability analysis on the resource allocation strategy information to obtain interpretive information about the resource allocation strategy information; Determining the risk level of the resource allocation strategy information based on market volatility; and The resource allocation strategy information is iteratively optimized based on the explanatory information and the risk level.

8. The method according to claim 1, characterized in that, The method further includes at least one of the following: Incremental customer behavior data is acquired based on a preset time interval, and the customer behavior prediction model is incrementally trained using the incremental customer behavior data and the elastic weight consolidation algorithm. The original resource configuration strategy and the updated resource configuration strategy based on the resource configuration strategy information are deployed in parallel. In response to the improvement of the operation indicators of the updated resource configuration strategy relative to the original resource configuration strategy within a preset evaluation period reaching a preset threshold, resource scheduling is performed based on the updated resource configuration strategy.

9. A resource scheduling device, characterized in that, The device includes: The data acquisition module is used to: acquire multi-source heterogeneous data, which includes at least data from the client side, the environment side, and the system side, and construct a multimodal feature representation based on the multi-source heterogeneous data; The multi-task prediction module is used to: input the multimodal feature representation into the customer behavior prediction model to perform multi-task prediction and obtain prediction state information; The resource allocation strategy acquisition module is used to: obtain resource allocation strategy information based on the predicted state information and the resource scheduling target; and The scheduling module is used to: perform resource scheduling based on the resource configuration strategy information.

10. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 8.

12. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 8.

Images