Account receivable management method and device based on multi-layer perceptron and medium

Abstract

The embodiment of the invention discloses an account receivable management method and device based on a multi-layer perceptron and a medium, and relates to the technical field of prediction, and the method comprises the steps: obtaining multi-dimensional account data corresponding to account receivable, carrying out the preprocessing of the multi-dimensional account data through the pre-constructed dimension correlation thermodynamic data, and obtaining a pre-processed account data; determining a multi-dimensional account feature vector; according to the multi-dimensional account feature vector and a pre-trained account receivable prediction model, determining account receivable prediction data corresponding to the account receivable, so as to determine a key risk factor set corresponding to each account receivable based on the account receivable prediction data; performing causal chain reasoning on the account receivable prediction data through the key risk factor set, and determining risk causal chain data corresponding to the account receivable; based on the risk causal chain data and the account receivable prediction data, visual risk display is performed on the account receivable, risk management of the account receivable is realized, and the limitation that only a prediction result is output depending on single data and a model is broken through.

Description

Technical Field

[0001] This specification relates to the field of prediction technologies, and particularly to an accounts receivable management method, device, and medium based on a multi-layer perceptron. Background Art

[0002] Risk management of accounts receivable is a key link in an enterprise's cash flow and financial soundness. Enterprise accounts receivable management is crucial for financial health and cash flow prediction. Traditional accounts receivable prediction methods often rely on single historical accounts data. However, there are various uncertainties and risk factors in accounts receivable, such as external factors like market factors and customer dimensions. Traditional prediction methods often struggle to fully explore and integrate multi-dimensional data, resulting in low prediction accuracy.

[0003] In addition, existing models mostly have a "black box" structure and only output prediction results such as risk probabilities. Enterprise management users cannot locate core risk driving factors, such as "the contribution weight of the customer payment delay rate to overdue", leading to insufficient pertinence of risk response measures. Currently, risk scores are mostly presented through tables or two-dimensional charts, lacking visual dynamic tracking of the causes of risks, making it difficult to intuitively identify the core risk causes and their transmission paths, resulting in low risk disposal efficiency.

[0004] Therefore, traditional accounts receivable prediction methods often rely on single historical accounts data, ignoring the influence of external factors, resulting in low prediction accuracy, and only outputting prediction results, lacking visual dynamic tracking of the causes of risks, which affects subsequent risk disposal efficiency. Summary of the Invention

[0005] One or more embodiments of this specification provide an accounts receivable management method, device, and medium based on a multi-layer perceptron to solve the following technical problems: Traditional accounts receivable prediction methods often rely on single historical accounts data, ignoring the influence of external factors, resulting in low prediction accuracy, and only outputting prediction results, lacking visual dynamic tracking of the causes of risks, which affects subsequent risk disposal efficiency.

[0006] One or more embodiments of this specification adopt the following technical solutions:

[0007] One or more embodiments of this specification provide an accounts receivable management method based on a multi-layer perceptron. The method includes: obtaining multi-dimensional account data corresponding to accounts receivable, preprocessing the multi-dimensional account data by using pre-constructed dimension correlation heat data, and determining a multi-dimensional account feature vector. Among them, the multi-dimensional account data includes account customer dimension data, account product dimension data, account market dimension data, and enterprise internal management data; determining the accounts receivable prediction data corresponding to the accounts receivable according to the multi-dimensional account feature vector and a pre-trained accounts receivable prediction model, and determining a set of key risk factors corresponding to each accounts receivable based on the accounts receivable prediction data; performing causal chain reasoning on the accounts receivable prediction data through the set of key risk factors to determine the risk causal chain data corresponding to the accounts receivable; performing visual risk display on the accounts receivable based on the risk causal chain data and the accounts receivable prediction data to achieve risk management of the accounts receivable.

[0008] One or more embodiments of this specification provide an accounts receivable management device based on a multi-layer perceptron, including:

[0009] At least one processor; and,

[0010] A memory communicatively connected to the at least one processor; wherein,

[0011] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above method.

[0012] A non-volatile computer storage medium provided by one or more embodiments of this specification stores computer-executable instructions, and the computer-executable instructions are set to: execute the above method.

[0013] The above at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects: Through the embodiments of this specification, multi-dimensional accounts receivable data such as accounts receivable customer dimension, product dimension, market dimension, and enterprise internal management data are comprehensively collected, changing the limitation of traditional methods that only rely on single historical accounts receivable data. By using the pre-constructed dimension correlation heat data to preprocess the multi-dimensional data, potential connections between the data are mined, and more representative multi-dimensional accounts receivable feature vectors are determined, providing rich and high-quality data for the subsequent prediction model, enabling the model to better capture various factors affecting accounts receivable, thus significantly improving the accuracy of accounts receivable prediction; Based on the multi-dimensional accounts receivable feature vectors, the accounts receivable prediction data is determined with the help of a pre-trained accounts receivable prediction model, and further the key risk factor set is determined, breaking through the limitation of existing "black box" models that only output prediction results, and being able to clarify the core risk driving factors affecting accounts receivable. By analyzing the key risk factor set, the contribution weights of factors such as the customer payment delay rate to overdue can be accurately known; Using the key risk factor set to conduct causal chain reasoning on the accounts receivable prediction data to obtain risk causal chain data; Clearly showing the causes and transmission paths of risks, no longer looking at risk factors and prediction results in isolation, and it is possible to intuitively understand how different risk factors interact with each other and ultimately affect accounts receivable. The clear presentation of the risk transmission path helps to comprehensively master the risk formation mechanism and prevent and block the spread of risks in advance; Combining the risk causal chain data and the accounts receivable prediction data to conduct visual risk display of accounts receivable. Different from the traditional method of only displaying risk scores through tables or two-dimensional charts, this method can dynamically track the causes of risks, can more intuitively identify the core risk causes and their transmission paths, quickly locate high-risk areas and key risk links, and improve the efficiency of risk disposal. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments recorded in this specification. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings. In the drawings:

[0015] Figure 1 It is a schematic flowchart of an accounts receivable management method based on a multi-layer perceptron provided by the embodiments of this specification;

[0016] Figure 2 It is an example diagram of a dimension correlation heat data provided by the embodiments of this specification;

[0017] Figure 3A schematic structural diagram of an accounts receivable management device based on a multi-layer perceptron provided by an embodiment of this specification. Detailed implementation manners

[0018] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all of the embodiments. Based on the embodiments of this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this specification.

[0019] An embodiment of this specification provides an accounts receivable management method based on a multi-layer perceptron. It should be noted that the execution subject in the embodiments of this specification can be a server or any device with data processing capabilities. Figure 1 A flowchart of an accounts receivable management method based on a multi-layer perceptron provided by an embodiment of this specification, as Figure 1 shown, mainly includes the following steps:

[0020] Step S101: Obtain multi-dimensional account data corresponding to accounts receivable, and use the pre-constructed dimension correlation heat data to preprocess the multi-dimensional account data to determine the multi-dimensional account feature vector.

[0021] In an embodiment of this specification, the accounts receivable information in the system is monitored in real time to obtain the multi-dimensional account data corresponding to the accounts receivable. Among them, the multi-dimensional account data includes account customer dimension data, account product dimension data, account market dimension data, and enterprise internal management data. It should be noted that the multi-dimensional account data here is the relevant data affecting the accounts receivable. The account customer dimension data includes customer credit ratings, payment records, industry dynamics, etc., the account product dimension data includes sales amounts, product types, sales cycles, etc., the account market dimension data includes macroeconomic indicators, industry prosperity indices, competition situations, etc.; the enterprise internal management data includes changes in sales policies, adjustments to credit policies, collection efforts, etc.

[0022] The multi-dimensional accounts receivable data is preprocessed by using the pre-constructed dimension correlation thermal data to determine the multi-dimensional accounts receivable feature vector, specifically including: determining the historical multi-dimensional accounts receivable time series data, and extracting features from the historical multi-dimensional accounts receivable time series data according to each accounts receivable dimension to determine the historical dimension feature data corresponding to each accounts receivable dimension, wherein the accounts receivable dimension includes customer dimension, product dimension, market dimension and internal management dimension; performing feature correlation analysis on the historical dimension feature data corresponding to each accounts receivable dimension to determine the feature correlation coefficient between the accounts receivable dimensions, and constructing dimension correlation thermal data based on the feature correlation coefficient; determining the missing dimensions in the multi-dimensional accounts receivable data, and filling the missing dimensions with data through the dimension correlation thermal data to determine the current multi-dimensional accounts receivable data, and extracting features from the current multi-dimensional accounts receivable data to determine the multi-dimensional accounts receivable feature vector.

[0023] In one embodiment of the present specification, after collecting multi-dimensional accounts receivable data, it is necessary to perform pre-processing operations such as cleaning, missing value processing, outlier detection and correction, and time series alignment on the original data. The historical multi-dimensional accounts receivable time series data constructed during model training is used to analyze the historical multi-dimensional accounts receivable time series data to obtain the feature correlations between different accounts receivable dimensions. It should be noted that the accounts receivable dimensions here include customer dimensions, product dimensions, market dimensions, and internal management dimensions. The customer dimension includes secondary dimensions such as customer credit ratings, payment records, and industry dynamics. The product dimension includes secondary dimensions such as sales, product types, and sales cycles. The market dimension includes secondary dimensions such as macroeconomic indicators, industry prosperity indexes, and competitive situations. The internal management data of the enterprise includes secondary dimensions such as sales policy changes, credit policy adjustments, and collection efforts. The above dimensions are only used as examples. Use the Pearson correlation coefficient to analyze the correlation between the features of each dimension, determine the feature correlation coefficient between the accounts receivable dimensions, and construct dimension correlation thermal data based on the feature correlation coefficient. Figure 2 An example diagram of dimensional correlation thermal data provided in the embodiments of this specification, such as Figure 2 As shown, the correlation coefficient between the unit's credit rating and historical payment records is -0.25.

[0024] Determine the missing dimensions in the multi-dimensional accounts receivable data, fill in the missing dimensions with data through the thermal data of the dimension correlation, and determine the current multi-dimensional accounts receivable data. If a dimension is missing (such as industry dynamics), estimate the missing value according to the linear regression model based on the data of the same period with the highest correlation dimension in the thermal map (such as the unit credit rating), and fill it in. Then extract features from the current multi-dimensional accounts receivable data to determine the multi-dimensional accounts receivable feature vector.

[0025] Through the above technical solutions, multi-dimensional account receivable data is comprehensively collected, covering aspects such as customers, products, markets, and enterprise internal management, ensuring the richness and comprehensiveness of the data, and providing sufficient information for accurate analysis of accounts receivable; using the dimension correlation heat data to fill in the missing dimensions can better retain the internal relationship between data and reduce information loss caused by data missing compared with simple data filling methods (such as mean and median filling), improving the integrity and accuracy of the data; constructing the dimension correlation heat data through feature extraction and correlation analysis of historical multi-dimensional account receivable time series data helps to explore the potential relationship between different dimension data, provides more valuable information for the subsequent prediction model, enables the model to more accurately capture the influencing factors of accounts receivable, thereby improving the accuracy of prediction and reducing prediction errors.

[0026] Step S102: According to the multi-dimensional account receivable feature vector and the pre-trained accounts receivable prediction model, determine the accounts receivable prediction data corresponding to the accounts receivable, and based on the accounts receivable prediction data, determine the set of key risk factors corresponding to each accounts receivable.

[0027] Before determining the accounts receivable prediction data corresponding to the accounts receivable according to the multi-dimensional account receivable feature vector and the pre-trained accounts receivable prediction model, the method further includes: collecting historical accounts receivable data and the corresponding historical multi-dimensional account receivable time series data, where the historical accounts receivable data includes accounts receivable amount data, aging distribution data, and collection cycle data; performing abnormal data processing and missing value filling on the historical accounts receivable data and the historical multi-dimensional account receivable time series data to determine the accounts receivable data set; using the accounts receivable data set, through a preset optimization strategy, training the pre-constructed initial multi-layer perceptron model to determine an accounts receivable prediction model that meets the preset requirements, where the number of hidden layers of the initial multi-layer perceptron model is determined based on one-shot neural architecture search combined with weight sharing.

[0028] In one embodiment of this specification, historical accounts receivable data in recent years is extracted from the financial database, including amount, aging, and collection cycle, and historical accounts receivable data is collected by associating with multi-dimensional data in the same period (such as customer industry dynamics, enterprise sales policy change records), including but not limited to core indicators such as accounts receivable amount, aging distribution, and collection cycle in each period; historical multi-dimensional accounts time series data affecting accounts receivable is obtained, including but not limited to customer dimensions such as customer credit rating, historical payment records, and industry dynamics; product dimensions such as sales volume, product type, and sales cycle; market dimensions such as macroeconomic indicators, industry prosperity index, and competition situation; internal management dimensions such as sales policy changes, credit policy adjustments, and collection efforts. Preprocessing operations such as data cleaning, missing value processing, outlier detection and correction, and time series alignment are performed on the collected raw data. The density-based clustering method (Density-Based Spatial Clustering of Applications with Noise, DBSCAN) can identify regions with low density in the dataset, and the points in these regions may be outliers. For missing value and outlier processing, if the total data volume is large, outliers can be directly deleted, and in the case of a small total data volume, the exponential smoothing method of time series can be used to estimate the replacement value of outliers.

[0029] A multi-layer fully connected neural network model is constructed, including an input layer, a hidden layer, and an output layer. The input layer receives the processed multi-dimensional features, the hidden layer performs multi-level information processing through non-linear activation functions, and the output layer generates the predicted value of future accounts receivable. MLP (Multilayer Perceptron), that is, a multi-layer perceptron, is a feedforward artificial neural network (Feedforward Artificial Neural Network). It contains multiple hidden layers and an output layer, each layer consists of multiple neurons, and the neurons between adjacent layers are fully connected. The basic working principle of MLP is to receive data at the input layer, and then propagate and process information layer by layer through non-linear transformation, and finally generate a prediction result at the output layer.

[0030] The input layer is used to receive the original feature data and use it as the input of the neurons in the first layer. The neurons in each hidden layer apply a non-linear activation function (such as Sigmoid, ReLU, Tanh, etc.) to the output of the previous layer, then multiply it by the weight matrix of this layer and add the bias term to produce the output of this layer. Setting a reasonable hidden layer is the key to the success of the model. In the embodiments of this specification, One-Shot Neural Architecture Search (One-Shot NAS) combined with the weight sharing technique is used to determine the number of hidden layers of the multi-layer perceptron (MLP). In traditional NAS, each candidate network architecture needs to be independently trained to evaluate its performance, which consumes a large amount of computing resources and time. While One-Shot NAS constructs a super network that contains all possible sub-network structures and all sub-networks share weights, so only this single super network needs to be trained. This approach greatly reduces the repeated training work in the search process. The training of the super network allows the simultaneous update of the architecture parameters and weights. Although it may bring the coupling problem between the weights and the architecture, through reasonable design and optimization, it can accelerate the convergence speed of the entire search process. Weight sharing means that different network structures can learn from each other's information during the search process, making the training more efficient. Even if the finally selected architecture is "pruned" from the super network, it can have a relatively high starting point because it inherits the pre-trained weights and can reach better performance faster during subsequent fine-tuning. In addition, the One-Shot NAS method does not require a separate training process for each potential architecture, making it easier to handle large-scale and complex search spaces. It can flexibly explore different levels of network components (such as convolution kernel size, number of channels, layer type, etc.), and in many cases, the optimal architecture obtained from the search can be directly extracted from the super network and further independently trained to obtain the optimal performance. The core advantage of One-Shot NAS combined with weight sharing is that it can effectively utilize limited computing resources and find a high-performance neural network architecture in a short time. It is very suitable for application in the prediction scenario of accounts receivable.

[0031] The output of the last hidden layer enters the output layer, where the activation function used is the Softmax function. The reason for using the Softmax function is that it can alleviate the vanishing gradient problem and is suitable for the output based on multi-class and multi-neuron. The Softmax function receives an arbitrary real-valued vector and converts it into a probability distribution. For a given input vector Z = (z1, z2,..., z_K), the Softmax function calculates the exponential function e^(zi) for each element, and then divides it by the sum of the exponential functions of all elements to ensure that the sum of all elements of the output vector is 1, satisfying the properties of a probability distribution. The Softmax function is applicable to multi-class problems, and the prediction scores for each class are converted into the probabilities of the corresponding classes. The highest probability corresponds to the most likely class. Due to the probability distribution characteristics of the output, the Softmax function ensures that the output classes are mutually exclusive, that is, for a sample, it can only belong to one class, namely the class with the highest probability.

[0032] The output result can be directly interpreted as the probabilities of each class, which helps with understanding and decision-making; in combination with the cross-entropy loss function, the Softmax function makes the model optimization a clear task of minimizing the probability of misclassification; the Softmax function is smooth and continuous, which is beneficial to the convergence of optimization algorithms such as gradient descent; the Softmax function can give a relative comparison between classes, and the output values not only represent the probabilities of the classes to which they belong, but also reflect the relative strengths of different classes relative to other classes; the derivative of the Softmax function is easy to calculate, which is beneficial to the backpropagation process of the neural network. Compared with the simple linear output layer and the Sigmoid function in multi-class problems, it solves the problems of vanishing gradient and exploding gradient.

[0033] During the model training process, the AdamW (Adam with weight decay correction) optimizer with weight decay is used to adjust the weights and biases to minimize the loss function between the predicted output and the actual labels. AdamW corrects the weight decay on the basis of the original Adam algorithm to better conform to the effect of L2 regularization. By modifying the implementation of weight decay, AdamW improves the handling of Adam in regularization and maintains the original efficient optimization characteristics of Adam. The Adabelief optimizer and the learning rate decay strategy are used to update the weights and biases, gradually improving the performance of the model on the training set. Adabelief can overcome the overconfidence problem of existing adaptive optimizers such as Adam in gradient belief in some cases. The Adabelief optimizer can converge quickly and has good stability, achieving comparable or even better performance than Adam and other advanced optimizers in various environments. Since Adabelief reduces the over-optimistic trust in gradient estimation, it helps to improve the generalization ability of the model, especially in adversarial training and noise-sensitive environments. The Adabelief optimizer better adapts to the gradient signal, improving the training efficiency and model quality.

[0034] The obtained feature vectors are used as the model input and trained using the labels corresponding to the historical accounts receivable data; the Leave-One-Out Cross-Validation (LOOCV) method is used to optimize the model parameters to ensure good generalization performance of the model. LOOCV is more accurate than the conventional k-fold cross-validation because only one sample is left for testing each time, and all other samples are used for training, making the most of the limited data. The data is fully utilized, so that almost all of the data volume (N - 1 observations) can be used for training each time. When the sample size is relatively small and computing resources permit, LOOCV can give a more accurate estimate, providing nearly N times of cross-validation. Since the dataset of the historical data of accounts receivable is not large, this method is very suitable. After training, the prediction performance of the model on the validation set is evaluated, such as metrics like mean squared error (MSE), mean absolute error (MAE), coefficient of determination (R 2 ) etc. In the above manner, an accounts receivable prediction model is obtained.

[0035] In an embodiment of this specification, the trained accounts receivable prediction model is used to input the multi-dimensional accounts receivable feature vectors into the model to predict the accounts receivable for a future time period containing the latest multi-dimensional feature information, and determine the accounts receivable prediction data corresponding to the accounts receivable.

[0036] Based on the accounts receivable prediction data, determine the set of key risk factors corresponding to each such accounts receivable, specifically including: calculating the contribution weight of each account feature in the multi-dimensional account feature vector to the accounts receivable prediction data through a preset gradient approximation algorithm; screening among multiple such account features according to the contribution weight of each such account feature to the accounts receivable prediction data and a preset threshold to determine at least one key risk feature, so as to construct the set of key risk factors.

[0037] Calculating the contribution weight of each account feature in the multi-dimensional account feature vector to the accounts receivable prediction data through a preset gradient approximation algorithm specifically includes: performing backpropagation on the output of the accounts receivable prediction model to calculate the input feature gradient matrix corresponding to the multi-dimensional account feature vector; based on the input feature gradient matrix, using the integral path method to calculate the contribution weight of each account feature in the multi-dimensional account feature vector to the accounts receivable prediction data.

[0038] In one embodiment of the present specification, after the model training is completed, a backpropagation operation is performed on the output of the model. During the backpropagation process, according to the calculation logic of the model, the input feature gradient matrix corresponding to the multi-dimensional account feature vector is calculated. The input feature gradient matrix reflects the gradient change of the model output corresponding to each input feature (i.e., account feature), and reflects the local influence degree of each account feature on the prediction result. Based on the calculated input feature gradient matrix, the integral path method is used to calculate the contribution weight of each account feature to the accounts receivable prediction data. The integral path method performs integral operations on the gradient information by selecting a suitable integral path in the feature space, so as to obtain a weight value that more comprehensively and accurately reflects the overall contribution of the account feature to the prediction result. In actual operation, a path from the point of taking the average value of all features to the target sample point can be selected for integration to comprehensively consider the influence of the change of the feature from the initial state to the current state on the prediction result.

[0039] Set a reasonable preset threshold. The determination of the threshold needs to comprehensively consider factors such as the enterprise's tolerance for accounts receivable risks, the industry average level, and the analysis results of historical data. Compare the contribution weight of each account feature with the preset threshold, and screen out the account features with contribution weights greater than the preset threshold among multiple account features. These screened account features are the key risk features. Integrate all the determined key risk features to construct the set of key risk factors. Each key risk feature in the set of key risk factors has an important impact on the accounts receivable prediction data and represents the risk factors that the enterprise needs to focus on during the accounts receivable management process.

[0040] Through the above technical solutions, starting from multi-dimensional account data, it comprehensively covers aspects such as customers, products, markets, and enterprise internal management; uses the gradient approximation algorithm and backpropagation to calculate the contribution weights, and comprehensively considers the impact of the characteristics of each dimension of accounts receivable on the predicted data of accounts receivable; this multi-dimensional analysis and precise calculation method avoid the limitations of single-dimensional analysis, can accurately locate the key risk characteristics that have a significant impact on the predicted data, enable enterprises to clearly master the core factors affecting accounts receivable, and provide accurate targets for subsequent risk prevention and control; based on the contribution weights and preset thresholds, screen the key risk characteristics, construct a set of key risk factors, and according to this set, focus limited resources on key risk points and formulate targeted prevention and control strategies; strengthen credit monitoring for customers with a greater impact on credit ratings; adjust sales strategies for products with prominent fluctuations in the sales cycle. Compared with the traditional comprehensive but lack-of-focus prevention and control methods, it greatly improves the efficiency of risk prevention and control, reduces management costs, and effectively reduces potential bad debt losses.

[0041] Step S103, through the set of key risk factors, conduct causal chain reasoning on the predicted data of accounts receivable to determine the risk causal chain data corresponding to the accounts receivable.

[0042] Collect the expert rule description text, and based on natural language processing technology, convert this expert rule description text to generate expert description rules, where the expert description rules include description rule conduction data and preset rule strength coefficients; collect the historical accounts receivable data tables corresponding to historical risk cases, where the historical accounts receivable data tables include customer dimension fields, accounts receivable product dimension fields, accounts receivable market dimension fields, enterprise internal management fields, and risk result fields; according to this historical accounts receivable data table, conduct association rule mining to obtain the mining rule conduction data of the specified feature combination and the risk result, and determine the conduction strength coefficient for the regression relationship between the feature changes and the risk result in this historical accounts receivable data table to determine the mining association rules; through this mining association rule and this expert description rule, construct a risk conduction rule library.

[0043] In one embodiment of this specification, by means of organizing expert seminars, interviewing industry veterans, and collecting relevant professional literature, etc., expert rule discourse texts are widely collected, covering professional knowledge and experience in the field of accounts receivable risks, including but not limited to the understanding of the interaction of different risk factors, common paths of risk transmission, and judgment criteria for the severity of risks, etc. For example, if the industry prosperity index of a customer has been declining for 3 consecutive months and its accounts payable turnover rate is lower than 20% of the industry average, then the probability of overdue in the next 6 months will increase by 40%. Use the pre-trained BERT model for dependency syntactic analysis to identify the conditional terms (such as "the industry prosperity index has been declining for 3 consecutive months") and result terms (such as "the probability of overdue increases by 40%") in the rule text. Map the variables in the conditional terms (such as "industry prosperity index") to the preset dimensional fields (such as "industry prosperity index field" under the accounts market dimension), and extract the constraint parameters (such as "has been declining for 3 consecutive months" is converted into the time window parameter t_w = 3 and the trend judgment function F(trend ≤ -0.1)). Extract the numerical value of the probability change value (such as "increases by 40%") in the result term to generate the rule strength coefficient (denoted as R_s = 0.4), and convert it into the preset rule strength base value through regularization processing (such as 0.4 × time decay factor). Store the structured conditions and results in the triple format:

[0044] {

[0045] "Antecedent Conditions":

[0046] {"Dimension": "Market Dimension", "Field": "Industry Prosperity Index", "Constraint": "trend ≤ -0.1 over t_w = 3"},

[0047] {"Dimension": "Customer Dimension", "Field": "Accounts Payable Turnover Rate", "Constraint": "value < industry average × 0.8"}

[0048] ,

[0049] "Conclusion": "The probability of overdue increases",

[0050] "Strength Coefficient": "R_s = 0.4 × exp(-0.1t)",

[0051] "Transmission Path": "Market Risk → Customer Liquidity Risk → Overdue"

[0052] }

[0053] In addition to the above methods, natural language processing techniques can also be used to process the collected text. First, perform text cleaning to remove noise information in the text, such as irrelevant punctuation marks, special characters, and redundant whitespace characters, etc., to improve the readability and processing efficiency of the text. Then, use lexical analysis techniques to decompose the text into words or phrases, label the part of speech, and identify named entities, such as customer names, product types, market indicators, etc. Next, through syntactic analysis and semantic understanding, parse the sentence structure and understand the semantic logic of the text. Use a pre-trained language model or a custom rule template to extract key information from the processed text and generate discourse rule conduction data. The discourse rule conduction data describes the conduction relationship between risk factors. For example, "If the customer credit rating decreases and the market competition situation intensifies, then the risk of overdue accounts receivable increases." At the same time, according to the description of the importance and influence degree of different rules in the expert text, combined with industry experience and expert judgment, determine the preset rule strength coefficient. This coefficient is used to quantify the relative importance of each rule in risk assessment, and the value range can be set according to the actual situation, such as between 0 and 1. The larger the value, the stronger the influence of the rule.

[0054] Collect the historical accounts receivable data tables corresponding to historical risk cases to ensure the integrity and accuracy of the data. The data table includes customer dimension fields (such as customer credit rating, historical payment records, industry affiliation, etc.), accounts receivable product dimension fields (sales amount, product type, sales cycle, etc.), accounts receivable market dimension fields (macroeconomic indicators, industry prosperity index, market share, etc.), enterprise internal management fields (sales policy changes, credit policy adjustments, collection efforts, etc.), and risk result fields (whether overdue, overdue amount, bad debt situation, etc.). Clean and preprocess the collected data, handle missing values and outliers, and unify the data format to ensure that the data quality meets the mining requirements. Use association rule mining algorithms, such as the Apriori algorithm or the FP-Growth algorithm, to analyze the historical accounts receivable data table. Set appropriate support (for example, the minimum support threshold is set to 0.05) and confidence threshold. Support represents the frequency of the rule appearing in the data set, and confidence represents the probability that the conclusion holds under the premise condition. Through the algorithm, mine the potential relationship between the specified feature combination and the risk result to obtain the mined rule conduction data. For example, the example mining result: {Industry prosperity index decreases, accounts payable turnover rate < 0.5} → High risk (support = 0.08).

[0055] For the mined association rules, further analyze the regression relationship between the characteristic changes and the risk results in the historical accounts receivable data table to determine the conduction intensity coefficient. Linear regression, logistic regression, or other appropriate regression analysis methods can be used to quantify the impact of characteristic changes on the risk results. If the probability of accounts receivable overdue increases by 10% for each level decrease in the customer credit rating, this ratio can be used as the conduction intensity coefficient of this association rule. Based on the support, confidence, and conduction intensity coefficient, screen out the association rules with high reliability and influence to determine the mined association rules.

[0056] Integrate the generated expert-discussed rules and the mined association rules. During the integration process, coordinate and optimize the duplicate or conflicting rules. If both the expert rules and the mined rules involve the impact of the customer credit rating on the accounts receivable risk but are slightly different in expression, comprehensively consider the advantages and rationality of both to form a unified rule expression. If there is a conflict between "the number of customer overdue times > 3 times → high risk" in the expert rules and "the number of overdue times > 2 times" in the data rules, retain the rule with higher intensity according to the intensity coefficient priority (expert rule Rs = 0.5 vs. data rule Sm = 0.7). Store the fused rules in the graph database (Neo4j), where the nodes represent the rule conditions (such as "the industry prosperity index decreases"), the edges represent the conduction paths (such as "leading to the overdue risk"), and the edge weights are the intensity coefficients. Assign a unique identifier to each rule for convenient query and management. The rule library can be stored and managed using a database management system to ensure data security and scalability. At the same time, establish a rule update mechanism. With the accumulation of new expert knowledge, historical data, and the change of the business environment, regularly update and maintain the rule library to ensure the timeliness and accuracy of the rules. When the new risk case data exceeds the threshold, trigger the FP-Growth algorithm to re-mine the association rules and update the corresponding rule intensity.

[0057] Through the above technical solutions, by integrating expert knowledge rules and data-driven association rules, a multi-dimensional risk assessment system is constructed. The expert rules refine the experience logic of senior practitioners (such as "the continuous decline of the industry index triggers the overdue probability"), while the mined association rules reveal potential statistical associations (such as "the accounts payable turnover rate is significantly correlated with the risk"); use the graph database to store the rule nodes (such as "the industry index decreases") and the conduction edges (such as "leading to the deterioration of customer liquidity"), and the edge weights are the association intensity coefficients. When a risk event is triggered, an automatic visual conduction chain is generated (example: raw material price increase → customer gross profit margin decreases → accounts overdue →The provision for bad debts is increased), and the influence weights of key nodes are marked (such as the contribution ratio of the customer dimension being 62%). Compared with the traditional two-dimensional risk score, the efficiency of tracing the causes of risks is effectively improved, and targeted collection strategies can be customized according to the conduction path.

[0058] Through this set of key risk factors, causal chain reasoning is performed on the predicted data of this accounts receivable to determine the risk causal chain data corresponding to this accounts receivable, specifically including: presetting a risk conduction rule library, where the risk conduction rule library includes multiple preset rules, and each of these rules includes a rule antecedent condition; traversing the preset rules in the risk conduction rule library according to the multiple key risk factors in this set of key risk factors, and determining at least one currently triggered rule and the triggering risk factor corresponding to each of these currently triggered rules through the matching status of the key risk factor and the rule antecedent condition; generating the action intensity corresponding to each of these currently triggered rules through the intensity coefficient of each of these currently triggered rules preset and the contribution weight corresponding to the triggering risk factor; sorting these currently triggered rules in the order of the magnitude of the action intensity to generate causal chain list data described in natural language.

[0059] In an embodiment of this specification, through multiple key risk factors in the set of key risk factors, a distributed computing framework (Apache Spark) is used to perform parallel scanning on the rule library. The input data features are compared item by item with the rule antecedent conditions. For example, the key risk factors include that the accounts payable turnover rate in the customer dimension = 0.65 (the critical value is the industry average × 0.8), the industry prosperity index in the market dimension has decreased by 6.2% continuously for 4 months, and the current sales gross profit margin in the product dimension has decreased by 12% compared with the previous quarter. Taking rule R_2023_001 as an example, it is judged whether all the following conditions are met: whether the industry prosperity index meets "decrease for ≥ 3 months and the monthly decline ≥ is 5%" (the current input meets a 6.2% decline for 4 months, and the trigger is successful) whether the accounts payable turnover rate < the industry average × 0.7 (the input value is 0.65, if the industry average is 0.9, the threshold is 0.63, and this condition is met here). If all the antecedent conditions of a certain rule match, it is marked as a "currently triggered rule". For the above case, rule R_2023_001 is triggered.

[0060] For each triggering rule, based on the contribution weights corresponding to the key risk factors and the strength coefficient of the current triggering rule, generate the action strength corresponding to each current triggering rule. For example, if the strength coefficient is 0.8 and the contribution weights of two triggering risk factors are 0.7 and 0.3 respectively, the action strength of this rule is 0.8×(0.7 + 0.3) = 0.8. Sort all the triggering rules in descending order of action strength to ensure that the sorted list can clearly show the differences in the impact degrees of different rules on accounts receivable risks. The rule with a greater action strength is ranked higher. Generate a natural language description, automatically generate readable text through the NLP template engine, and identify the conduction path with arrows or other symbols. For example, "Since the credit rating of customer A has dropped from A to C within one month and a major negative policy restricting the industry development has been issued in its industry in the past three months --- the overdue risk of the accounts receivable of this customer has increased significantly, and the action strength is 0.8". In the description, elaborate on the triggering risk factors, risk results, and action strength to make the results easy to understand. Integrate the converted natural language descriptions into a list of risk causal chain data in sequence.

[0061] Through the above technical solutions, through the matching mechanism of the pre-set risk conduction rule library and the set of key risk factors, the risk factors affecting accounts receivable and their conduction paths can be comprehensively and accurately identified; compared with traditional risk assessment methods, it is no longer limited to single factors or simple associations, and can deeply explore complex risk causal relationships; by traversing the rule library, accurately find the current triggering rules and triggering risk factors, clearly present the risk sources and propagation links, accurately grasp the core of the risks, and provide precise targets for subsequent risk responses; the generated list data of causal chains in natural language descriptions is intuitive and easy to understand, providing a clear risk panorama for enterprise management and financial personnel; according to the results sorted by action strength, the severity and priority of different risks can be quickly judged, and scientific decisions can be made in formulating credit policies, sales strategies, collection plans, etc. For risks with high action strength, increase the collection efforts or adjust the credit limit; for risks with low action strength, reasonably allocate resources for monitoring to improve resource utilization efficiency and balance business development and risk control.

[0062] Step S104, based on the risk causal chain data and accounts receivable prediction data, conduct a visual risk display of accounts receivable to achieve the risk management of accounts receivable.

[0063] Based on the risk causal chain data and the accounts receivable prediction data, a visual risk display of the accounts receivable is performed, specifically including: obtaining the causal chain list data in the risk causal chain data, where the causal chain list data includes multiple risk causal chains arranged in sequence; mapping the causal nodes in each risk causal chain to entity nodes, mapping the conduction direction in the risk causal chain to a directed edge connection, and setting the directed edge display parameters of the directed edge connection according to the list order corresponding to the risk causal chain; setting the entity display parameters of each entity node using the color scale coding rule according to the risk prediction result in the accounts receivable prediction data; constructing a risk causal map of the accounts receivable through the directed edge display parameters and the entity display parameters to perform a visual risk display of the accounts receivable.

[0064] In one embodiment of this specification, for each risk causal chain, the causal nodes are mapped to entity nodes. The causal nodes may include various risk factors and risk events such as a decline in customer credit rating, intensified market competition, and adjustment of the enterprise's internal sales policy. Each entity node represents a specific risk-related factor or event, presented in a graphical manner in the visualization map, such as a circle, a square, or other custom shapes. A unique identifier is assigned to each entity node for accurate identification and management in subsequent operations. The conduction direction in the risk causal chain is mapped to a directed edge connecting each entity node, clearly showing the propagation path of the risk. The directed edge display parameters of the directed edge connection are set according to the list order corresponding to the risk causal chain, including the color, thickness, transparency, etc. of the directed edge. The directed edges corresponding to the risk causal chains ranked in the front can be set to have brighter colors and thicker lines to highlight their importance; while the directed edges corresponding to the subsequent causal chains are adjusted to have lighter colors and thinner lines accordingly. In this way, users can intuitively understand the relative importance and sequence of different risk causal chains in the overall risk. According to the risk prediction result in the accounts receivable prediction data, the entity display parameters of each entity node are set using the color scale coding rule. For example, if the risk prediction result shows that the accounts receivable overdue risk of a certain customer is relatively high, the entity node representing the risk factors related to this customer is set to a color in the red series, and the depth of the color is adjusted according to the risk level, the higher the risk, the deeper the color; if the risk is relatively low, it is set to a color in the green series. At the same time, the size of the entity node can also be adjusted according to the risk prediction result, the greater the risk, the larger the node, further enhancing the visualization effect and enabling users to quickly identify high-risk areas.

[0065] Based on the above-set entity nodes and directed edge connections, use visualization drawing tools (such as Graphviz, D3.js, etc.) to construct a risk causal map of accounts receivable. During the construction process, reasonably layout the positions of entity nodes to avoid overlap and chaos between nodes and ensure the readability of the map. Algorithms such as hierarchical layout and force-directed layout can be used to make the map structure clearer and the risk conduction path obvious at a glance. Display the constructed risk causal map on the visualization interface to provide users with an intuitive risk visualization view. Design interactive functions to enhance the user experience. Generate an interactive map on the front-end interface, support the expansion of secondary causal chains by clicking on nodes, and associate and display the multi-dimensional account feature vectors and gradient contribution weights corresponding to the nodes; call the WebGL dynamic rendering engine to synchronously update the real-time prediction data output by the accounts receivable prediction model with the node attributes of the visualization map; in response to the filtering conditions input by the user, dynamically filter the nodes in the visualization map, and only retain the causal chain paths with an overdue probability exceeding the preset threshold or an influence strength greater than 0.2; extract the key rule list of the causal chain path to generate a risk report document, and complete the log data docking with the financial audit system interface. In addition, set up a filtering function to allow users to filter and display some risk causal chains according to specific conditions (such as risk type, risk degree, etc.) to focus on key risks. Since the accounts receivable risk is dynamically changing, the risk causal map needs to be updated regularly. According to the newly obtained risk causal chain data and accounts receivable prediction data, recalculate the display parameters of entity nodes and directed edges, and update the map content.

[0066] Through the above technical solutions, the risk causal chain is transformed into a visual risk causal map, which graphically displays risk factors (such as a decline in customer credit rating, increased market competition, etc.), risk events, and their conduction paths. The intuitive presentation method enables enterprise personnel, especially financial and management personnel without a technical background, to quickly understand complex risk relationships without spending a lot of time analyzing and interpreting abstract data and written reports, improving the risk awareness efficiency; by setting the display parameters of directed edges, adjust the color, thickness, and transparency of directed edges according to the order of the risk causal chain list, allowing users to intuitively understand the relative importance and sequence of different risk causal chains; it helps to prioritize the risk factors that have a greater impact on the overall risk and are transmitted earlier, reasonably allocate risk management resources, avoid ignoring key issues due to focusing on minor risks, and enhance the pertinence and effectiveness of risk management.

[0067] Through the embodiments of this specification, multi-dimensional accounts receivable data such as accounts receivable customer dimension, product dimension, market dimension, and enterprise internal management data are comprehensively collected, changing the limitation of traditional methods that only rely on single historical accounts receivable data. By using the pre-constructed dimension correlation heat data to preprocess the multi-dimensional data, potential connections between data are mined, and more representative multi-dimensional accounts receivable feature vectors are determined, providing rich and high-quality data for the subsequent prediction model, enabling the model to better capture various factors affecting accounts receivable, thereby significantly improving the accuracy of accounts receivable prediction; based on the multi-dimensional accounts receivable feature vectors, the accounts receivable prediction data is determined with the help of a pre-trained accounts receivable prediction model, and further a set of key risk factors is determined, breaking through the limitation of existing "black box" models that only output prediction results, and being able to clarify the core risk driving factors affecting accounts receivable. By analyzing the set of key risk factors, the contribution weights of factors such as the customer payment delay rate to overdue can be accurately known; the accounts receivable prediction data is subjected to causal chain reasoning using the set of key risk factors to obtain risk causal chain data; clearly showing the causes and transmission paths of risks, no longer looking at risk factors and prediction results in isolation, and being able to intuitively understand how different risk factors interact with each other and ultimately affect accounts receivable. The clear presentation of the risk transmission path helps to comprehensively master the risk formation mechanism and prevent and block the spread of risks in advance; combining the risk causal chain data and the accounts receivable prediction data, a visual risk display of accounts receivable is carried out. Different from the traditional way of only showing risk scores through tables or two-dimensional charts, it can dynamically track the causes of risks, can more intuitively identify the core risk causes and their transmission paths, quickly locate high-risk areas and key risk links, and improve the efficiency of risk disposal.

[0068] The embodiments of this specification also provide an accounts receivable management device based on a multi-layer perceptron, as Figure 3 shown. The device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the above method.

[0069] The embodiments of this specification also provide a non-volatile computer storage medium storing computer-executable instructions, and the computer-executable instructions are set to: execute the above method.

[0070] The various embodiments in this specification are all described in a progressive manner. The same or similar parts among the various embodiments can be referred to each other, and the key points of each embodiment are the differences from other embodiments. In particular, for the embodiments of the device, equipment, and non-volatile computer storage medium, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiments.

[0071] The above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the acts or steps recited in the claims may be performed in a different order than in the embodiments and still achieve the desired results. Additionally, the processes depicted in the figures do not necessarily require the particular order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0072] The devices and media provided by the embodiments of this specification correspond one-to-one with the methods. Therefore, the devices and media also have beneficial technical effects similar to those of their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be elaborated here.

[0073] Those skilled in the art should understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification can take the form of an all-hardware embodiment, an all-software embodiment, or an embodiment combining software and hardware aspects. Moreover, this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0074] This specification is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing device produce means for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0075] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means that implement the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0076] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one Figure 1 process or multiple processes and / or blocks Figure 1 or steps for implementing the functions specified in one block or multiple blocks.

[0077] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0078] The memory may include non-permanent memory in the computer-readable medium, in the form of random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM). Memory is an example of a computer-readable medium.

[0079] Computer-readable media includes permanent and non-permanent, removable and non-removable media and can be implemented by any method or technology for information storage. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media, such as modulated data signals and carrier waves.

[0080] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, commodity or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "including one …… " does not exclude the presence of additional identical elements in the process, method, commodity or device including said element.

[0081] The above description is only for one or more embodiments of this specification and is not intended to limit this specification. For those skilled in the art, various changes and modifications can be made to one or more embodiments of this specification. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of one or more embodiments of this specification shall be included within the scope of the claims of this specification.

Claims

1. A method for managing accounts receivable based on a multi-layer perceptron, characterized in that: The method comprises: Acquire multi-dimensional account data corresponding to accounts receivable, and pre-process the multi-dimensional account data using pre-built dimensional correlation thermal data to determine a multi-dimensional account feature vector, wherein the multi-dimensional account data includes account customer dimension data, account product dimension data, account market dimension data, and enterprise internal management data; Determine accounts receivable forecast data corresponding to the accounts receivable according to the multi-dimensional accounts receivable feature vector and the pre-trained accounts receivable forecast model, and determine a set of key risk factors corresponding to each of the accounts receivable based on the accounts receivable forecast data; Performing causal chain reasoning on the accounts receivable forecast data through the key risk factor set to determine risk causal chain data corresponding to the accounts receivable; Based on the risk causal chain data and the accounts receivable forecast data, the accounts receivable are visualized for risk display to achieve risk management of the accounts receivable.

2. The method for managing accounts receivable based on a multi-layer perceptron according to claim 1, characterized in that: Before determining the accounts receivable prediction data corresponding to the accounts receivable according to the multi-dimensional accounts receivable feature vector and the pre-trained accounts receivable prediction model, the method further includes: Collect historical accounts receivable data and historical multi-dimensional accounts receivable time series data corresponding to the historical accounts receivable data, wherein the historical accounts receivable data includes accounts receivable amount data, account age distribution data and collection cycle data; Performing abnormal data processing and missing value filling on the historical accounts receivable data and the historical multi-dimensional accounts receivable time series data to determine an accounts receivable data set; Using the accounts receivable dataset, a pre-built initial multi-layer perception model is trained through a preset optimization strategy to determine an accounts receivable prediction model that meets preset requirements, wherein the number of hidden layers of the initial multi-layer perception model is determined based on a one-time neural architecture search combined with weight sharing.

3. The method for managing accounts receivable based on a multi-layer perceptron according to claim 2, characterized in that: The multi-dimensional accounts receivable data is preprocessed using pre-built dimensional correlation thermal data to determine a multi-dimensional accounts receivable feature vector, specifically including: Determine the historical multi-dimensional accounts receivable time series data, and perform feature extraction on the historical multi-dimensional accounts receivable time series data according to each accounts receivable dimension to determine the historical dimension feature data corresponding to each accounts receivable dimension, wherein the accounts receivable dimension includes a customer dimension, a product dimension, a market dimension, and an internal management dimension; Performing feature correlation analysis on the historical dimension feature data corresponding to each account receivable dimension, determining the feature correlation coefficient between the account receivable dimensions, and constructing dimension correlation thermal data based on the feature correlation coefficient; Determine the missing dimensions in the multi-dimensional accounts receivable data, fill the missing dimensions with data through the dimension correlation thermal data, determine the current multi-dimensional accounts receivable data, extract features from the current multi-dimensional accounts receivable data, and determine a multi-dimensional accounts receivable feature vector.

4. The method for managing accounts receivable based on a multi-layer perceptron according to claim 1, characterized in that: Based on the accounts receivable forecast data, a set of key risk factors corresponding to each of the accounts receivable is determined, specifically including: Calculating the contribution weight of each account feature in the multi-dimensional account feature vector to the accounts receivable forecast data by using a preset gradient approximation algorithm; According to the contribution weight of each account feature to the accounts receivable forecast data and a preset threshold, a plurality of account features are screened to determine at least one key risk feature to construct the key risk factor set.

5. The method for managing accounts receivable based on a multi-layer perceptron according to claim 4, characterized in that: The contribution weight of each account feature in the multi-dimensional account feature vector to the accounts receivable forecast data is calculated by a preset gradient approximation algorithm, specifically including: Back-propagating the output of the accounts receivable prediction model to calculate the input feature gradient matrix corresponding to the multi-dimensional accounts receivable feature vector; Based on the input feature gradient matrix, the integral path method is used to calculate the contribution weight of each account feature in the multi-dimensional account feature vector to the accounts receivable forecast data.

6. The method for managing accounts receivable based on a multi-layer perceptron according to claim 1, characterized in that: By using the key risk factor set, causal chain reasoning is performed on the accounts receivable forecast data to determine the risk causal chain data corresponding to the accounts receivable, specifically including: Pre-setting a risk transmission rule base, wherein the risk transmission rule base includes a plurality of preset rules, each of which includes a rule antecedent condition; According to the multiple key risk factors in the key risk factor set, the preset rules in the risk transmission rule base are traversed, and at least one current triggering rule and a triggering risk factor corresponding to each current triggering rule are determined through the matching status of the key risk factors and the rule antecedent conditions; Generate the action intensity corresponding to each current trigger rule by presetting the intensity coefficient of each current trigger rule and the contribution weight corresponding to the trigger risk factor; The current trigger rules are sorted according to the order of the magnitude of the action strength to generate causal chain list data described in natural language.

7. The method for managing accounts receivable based on a multi-layer perceptron according to claim 1, characterized in that: The method further comprises: Collecting expert rule discussion texts, converting the expert rule discussion texts based on natural language processing technology, and generating expert discussion rules, wherein the expert discussion rules include discussion rule conduction data and preset rule strength coefficients; Collecting historical accounts receivable data tables corresponding to historical risk cases, wherein the historical accounts receivable data tables include customer dimension fields, accounts receivable product dimension fields, accounts receivable market dimension fields, enterprise internal management fields, and risk result fields; According to the historical accounts receivable data table, association rule mining is performed to obtain mining rule transmission data of the specified feature combination and risk results, and the transmission intensity coefficient is determined for the regression relationship between the feature changes and the risk results in the historical accounts receivable data table to determine the mining association rules; A risk transmission rule base is constructed by using the mining association rules and the expert discussion rules.

8. The method for managing accounts receivable based on a multi-layer perceptron according to claim 1, characterized in that: Based on the risk causal chain data and the accounts receivable forecast data, the accounts receivable are visualized for risk display, specifically including: Acquire causal chain list data in the risk causal chain data, wherein the causal chain list data includes a plurality of risk causal chains arranged in sequence; Mapping the causal nodes in each of the risk causal chains to entity nodes, mapping the conduction directions in the risk causal chains to directed edge connections, and setting directed edge display parameters of the directed edge connections according to the list order corresponding to the risk causal chains; According to the risk prediction results in the accounts receivable prediction data, the entity display parameters of each entity node are set using color coding rules; A risk causal graph of the accounts receivable is constructed through the directed edge display parameters and the entity display parameters to visualize the risk of the accounts receivable.

9. An accounts receivable management device based on a multi-layer perceptron, characterized in that: The device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the method according to any one of claims 1 to 8.

10. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions are configured to execute the method according to any one of claims 1 to 8.

Images