CN116911994B

CN116911994B - External trade risk early warning system

Info

Publication number: CN116911994B
Application number: CN202310898174.3A
Authority: CN
Inventors: 陈超; 李青; 周金; 曹文娇
Original assignee: Shandong Institute Of Standardization (wto/tbt Shandong Consulting Workstation)
Current assignee: Shandong Institute Of Standardization (wto/tbt Shandong Consulting Workstation)
Priority date: 2023-07-21
Filing date: 2023-07-21
Publication date: 2024-03-26
Anticipated expiration: 2043-07-21
Also published as: CN116911994A

Abstract

The invention relates to the technical field of external trade, in particular to an external trade risk early warning system. The system comprises a data collection part, a risk assessment part and a trade scene simulation model, wherein the risk assessment decision tree model is established by using a C4.5 decision tree algorithm through collecting external trade data, and the trade scene simulation model is used for simulation and prediction, so that the accuracy and efficiency of external trade risk prediction are effectively improved. The system can timely discover and early warn high-risk trade, provide decision support for enterprises, and reduce economic loss.

Description

External trade risk early warning system

Technical Field

The invention belongs to the technical field of external trade, and particularly relates to an external trade risk early warning system.

Background

Foreign trade refers to commodity and service transactions conducted across countries. With the continuous acceleration of globalization process, foreign trade activities are increasingly frequent, and trade sizes are also increasing year by year. Foreign trade relates to a plurality of links such as purchasing, quality inspection, payment settlement and the like of goods, and certain transaction risks exist. How to quickly and accurately evaluate transaction risks and discover and avoid potential risks in time becomes one of the important problems facing foreign trade enterprises.

Currently, foreign trade risk assessment mainly adopts manual assessment and rule-based methods. The manual evaluation requires expertise and abundant experience, consumes time and has high labor cost, and the result is easily influenced by subjective factors, so that the accuracy and consistency of the evaluation result cannot be ensured. Compared with a manual evaluation method, the evaluation result is more objective, but a great deal of expertise and experience are needed to formulate rules, and the rules are difficult to cover all transaction conditions.

In recent years, with the development of artificial intelligence and big data technology, data-based risk assessment methods are becoming research hotspots. The data-driven risk assessment method can automatically learn an assessment model based on a large amount of historical data, can adapt to different transaction conditions, and has higher accuracy and expandability. The data-driven risk assessment method mainly comprises methods such as decision trees, neural networks, support vector machines, random forests and the like.

For example, patent document US20160363562A1 discloses a foreign trade transaction risk assessment method based on machine learning. According to the method, a risk assessment model is built by using a support vector machine algorithm, and training and optimization are carried out through a large amount of historical transaction data so as to realize rapid and accurate assessment of foreign trade risks. Patent document CN110149535a discloses a foreign trade transaction risk assessment method based on a deep neural network. According to the method, massive historical transaction data are learned and optimized through a deep neural network algorithm, so that prediction and early warning of foreign trade transaction risks are realized.

However, some problems remain in the prior art. Firstly, the prior art mainly relies on learning and optimizing historical data, and effective prediction and early warning of emerging transaction risks cannot be carried out. Thus, when new risks arise, the prior art may not be able to identify and deal effectively, resulting in the possibility that the enterprise may suffer. Second, the prior art typically requires extensive data preprocessing and feature engineering, which requires significant time and effort by professionals, and typically involves some complex algorithmic and mathematical knowledge that is difficult for non-professionals to master and apply. Furthermore, the prior art generally requires the use of a large number of computing resources for model training and prediction, and thus may lack sufficient computing resources for small businesses to apply these techniques. Finally, the prior art generally fails to provide interpretable and visual results, meaning that the user may not understand how the model makes predictions, nor is the predicted results interpreted and validated.

Accordingly, in response to these problems with the prior art, there is a need to develop a transaction risk early warning system that better processes large-scale data, predicts transaction risk more accurately, and is easier to use and interpret.

Disclosure of Invention

The invention mainly aims to provide an external trade risk early warning system which can perform risk assessment and early warning in real time in the foreign trade transaction process, effectively reduce the foreign trade risk and improve the competitiveness and profitability of enterprises.

In order to solve the problems, the technical scheme of the invention is realized as follows:

there is provided an external trade risk early warning system, the system comprising:

the data collection part is used for collecting external trade data and carrying out data preprocessing to obtain preprocessed data;

the risk assessment part is used for establishing a risk assessment decision tree model based on a C4.5 decision tree algorithm, dividing the preprocessing data into a training set and a testing set, training the risk assessment decision tree model by using the training set data, selecting the characteristics by using a weighted information gain ratio for the nodes of each decision tree in the established risk assessment decision tree model, and selecting the characteristics with the largest weighted information gain ratio as the judging conditions of the nodes;

the trade scenario simulation part is used for simulating by using a trade scenario simulation model based on the test set to generate a plurality of possible trade scenarios, and inputting each trade scenario into the risk assessment decision tree model to obtain a risk value of each scenario;

And the risk judging part is used for evaluating the risk value by using the risk evaluation model, calculating the risk weight of each trade scene, sequencing according to the risk weights and determining the high-risk trade scene.

Further, the time span of the external trade data collected by the data collecting part exceeds a set time thresholdThe method comprises the steps of carrying out a first treatment on the surface of the The time threshold->The value range of the product is 365 days to 1100 days.

Further, the risk assessment unitDividing the preprocessing data into training set and test set by setting a time valueDividing the pretreatment data into a training set and a test set, and setting the initial time in the pretreatment data to be +.>Will->Preprocessing data in a time range as a training set, and +.>The preprocessed data after the time range is used as a test set; the time value->The following constraint relationship must be satisfied:。

Further, assume that the test set hasIndividual foreign trade data samples, each sample consisting of +.>A characteristic composition in whichIndicate->Personal characteristic value->A label representing the sample to indicate whether the sample is a high risk trade; establishing a risk assessment decision tree model by using a C4.5 decision tree algorithm, wherein the risk assessment decision tree model is assumed to be shared +. >Individual nodes, each node->All have a judgment condition->Indicating if the characteristics of the sample meet +.>Then the sample is assigned to node +.>In (a) and (b); every node->Are all provided with two child nodes->And->Respectively representing two possibilities of yes and no; when simulation is performed using a trade scenario simulation model, it is assumed that +.>Trade scenes, each trade scene +.>All have a feature vector +.>The method comprises the steps of carrying out a first treatment on the surface of the Risk assessment for each trade scenario using cross entropy function, assuming +.>The risk weight of the individual trade scenario is +.>。

Further, when calculating the weighted information gain ratio in the risk assessment decision tree model, the method comprises the following steps: the weighted information gain ratio is calculated using the following formula:

；

wherein,representing the sample set contained by the node,>representing a candidate feature set, ++>Representation feature->The number of possible values, +.>Representation feature->The value is +.>Sample subset of->Representation feature->Sample set->Is used for the information gain of (a),representation feature->Intrinsic value of (2);Representing weighting informationGain ratio.

Further, the method for calculating the information gain includes: the information gain is calculated using the following formula:

；

wherein the method comprises the steps ofRepresenting sample set +.>Empirical entropy of >Is characterized by->Under the condition of (1) sample set->Is a rule of thumb condition entropy.

Further, the calculation formulas of the empirical entropy and the empirical conditional entropy are as follows:

；

wherein the method comprises the steps ofRepresenting the number of categories>The representation belongs to->Sample number of individual categories, +.>Is characterized by->The value is +.>Under the condition of (1) sample set->Is a rule of thumb.

Further, the risk assessment model is expressed using the following formula:

；

wherein the method comprises the steps ofIndicate->True risk weight for individual trade scenario, +.>Representing the risk weight predicted by the model; to ensure that the sum of risk weights is 1, normalization was performed using the softmax function, i.e.:

；

wherein the method comprises the steps ofIndicate->Risk values of individual trade scenarios are calculated by using a weighted average mode; the model training is performed by using a weighted cross entropy function, and the formula is as follows:

；

wherein the method comprises the steps ofIndicate->The weights of the trade scenarios are set using the weights of the samples to increase the importance of high risk trade.

Further, the executing process of the trade scenario simulation model comprises the following steps: for each featureCalculate the average value thereofAnd standard deviation->The method comprises the steps of carrying out a first treatment on the surface of the For each trade scenario->Generate->Random number->Satisfy->The method comprises the steps of carrying out a first treatment on the surface of the The generated random number is combined into a feature vector +. >Inputting into a risk assessment decision tree model for prediction to obtain a risk value +.>The method comprises the steps of carrying out a first treatment on the surface of the Repeating the above process->Next, a plurality of possible trade scenarios are generated.

Further, the risk weights of all trade scenes are weighted and averaged to obtain the risk weight of each node, and the formula is as follows:

；

wherein the method comprises the steps ofIndicating the function, returning to 1 when the condition is true, otherwise returning to 0;Representing the risk weight of each node.

The external trade risk early warning system has the following beneficial effects:

firstly, the risk assessment model is established by adopting the decision tree algorithm, and the high-risk trade can be accurately identified by training and predicting the external trade data, so that the risk control capability of the external trade is improved. Compared with the traditional risk assessment method based on artificial experience, the risk assessment model provided by the invention can assess trade risk more objectively and accurately, avoids subjectivity and misjudgment caused by artificial assessment, and improves the accuracy and timeliness of risk early warning.

And secondly, the invention adopts a trade scenario simulation model to generate a plurality of possible trade scenarios, and calculates the risk weight of each node in a weighted average mode, so that the actual trade situation and the risk degree can be better reflected. Through simulation and weighted average of a plurality of trade scenes, deviation and misjudgment of a single trade scene can be avoided, and accuracy and reliability of risk early warning are improved.

Thirdly, the invention adopts the cross entropy function to carry out model training, and simultaneously uses the weight of the sample to carry out weighted average, thereby being capable of better adapting to the data distribution under different risk conditions and improving the robustness and generalization capability of the model. Compared with the traditional loss function, the cross entropy function can better process the multi-classification problem, reduces the transmission and accumulation of errors, and improves the learning efficiency and accuracy of the model.

Fourth, the invention adopts the trade scene simulation model and the weighted cross entropy function to carry out model training, can better consider the weight and the risk degree of the sample, and improves the recognition capability and the early warning effect of high risk trade. Compared with the traditional average loss function, the weighted cross entropy function can better process the importance and the weight of different samples, avoid the missed judgment and the misjudgment of the samples, and improve the accuracy and the reliability of risk early warning.

Drawings

Fig. 1 is a schematic system structure diagram of an external trade risk early warning system according to an embodiment of the present invention.

Detailed Description

In order that those skilled in the art will better understand the present invention, a technical solution in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in which it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the present invention without making any inventive effort, shall fall within the scope of the present invention.

The following will describe in detail.

The terms "first," "second," "third," "fourth" and the like in the description and in the claims and in the above drawings, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged where appropriate such that the embodiments described herein may be implemented in other sequences than those illustrated or otherwise described herein.

Example 1

The embodiment provides an external trade risk early warning system, the system includes:

international trade related data, including but not limited to commodity prices, transaction amounts, exchange rates, credit ratings of trading partners, and other macroscopic economic data, is collected by linking to an external database or API interface. These data may be preprocessed, such as to remove missing or anomalous data, to convert category data to virtual variables, and to normalize the numerical variables to accommodate model training requirements.

In this section, a risk assessment model is built using a C4.5 decision tree algorithm. The model may be characterized by factors such as commodity price fluctuations, transaction volume changes, and credit rating changes for the transaction partner, with the goal being a binary variable such as whether or not an offer has occurred. The data may be divided into a training set for training the model and a test set for evaluating the performance of the model.

When training a risk assessment decision tree model using training set data, an initial decision tree is first generated using a C4.5 decision tree algorithm. The initial decision tree comprises a root node representing a set of all samples and a number of child nodes representing a subset of the samples. In the process of generating the decision tree, the algorithm gradually selects the optimal characteristics for division to generate a decision tree with better generalization performance.

The training set data are input into the decision tree model one by one for prediction, and the error between the predicted result and the actual result is calculated (for example, cross entropy function calculation can be used). And then pruning the decision tree according to the error value, namely deleting some unnecessary nodes, so that the decision tree is simpler and the risk of overfitting is reduced.

And finally, predicting the test set by using the trained decision tree model, and calculating evaluation indexes such as accuracy, recall rate and the like of the model to evaluate the performance of the model.

various trade scenarios, such as price fluctuation, trade object credit decline, international political situation change, etc., are generated by using the trade scenario simulation model. These scenes are then input into a risk assessment model, and the risk value of each scene is calculated.

In the partial execution process, for each trade scenario, the characteristic vector is calculatedAnd inputting the prediction result into a decision tree model for prediction. Specifically, starting from the root node, according to the judgment condition +.>Assigning the trade scenario to a corresponding child nodeIs a kind of medium. The decision tree is then recursively traversed down until the leaf nodes are reached. Finally, the risk value corresponding to the leaf node +.>I.e., the risk value for that trade scenario.

Next, risk assessment is performed for each trade scenario using a cross entropy function. The cross entropy function is a function that measures the difference between two probability distributions, which can be used to calculate the difference between the model predicted outcome and the true outcome. Specifically, assume the first The true risk weight for individual trade scenarios is +.>The risk weight predicted by the model is +.>The cross entropy loss for this trade scenario is:

。

in order to ensure that the sum of the risk weights is 1, the risk weights predicted by the model need to be normalized, and a common method is to use a softmax function. Specifically, assume the firstThe risk value of the individual trade scenario is +.>The model predictive risk weight for the trade scenario is:

。

and finally, comprehensively evaluating the risk value by using a weighted cross entropy function, and calculating the risk weight of each trade scene. To take into account the importance of high risk trade, a weighted cross entropy function can be used for model training, formulated as follows:

。

wherein:indicate->The weights for the individual trade scenarios may be set using the weights for the samples to increase the importance of high risk trade.

When a large-scale national company carries out external trade, the system is used for acquiring required trade data through a data collecting part and then training a trade risk prediction model through a risk assessment part. In practice, the company may predict different trade scenarios through the trade scenario simulation part, and then calculate risk values for these scenarios using a risk assessment model. Finally, the risk judging part sorts various trade scenes according to the magnitude of the risk value, so as to assist the company to make an optimal trade decision.

Example 2

On the basis of the above embodiment, the time span of the external trade data collected by the data collecting part exceeds a set time thresholdThe method comprises the steps of carrying out a first treatment on the surface of the The time threshold->The value range of the product is 365 days to 1100 days.

Specifically, the time threshold F has a value ranging from 365 days to 1100 days. This means that the collected out-of-pair trade data needs to cover a time span of at least one year, up to three years. The setting of this time span is critical to the accuracy of the risk assessment.

In a specific practical application, the system can adjust the specific value of the time threshold value F according to different service requirements and data conditions. For example, if a company's major trade business is seasonal (such as agricultural trade), it may be necessary to set a relatively long time threshold (approximately 1100 days) in order to capture a more complete business cycle.

Conversely, if a business of a company changes faster or the business situation at hand changes significantly (e.g., due to policy adjustments or technical innovations), a relatively short time threshold (approaching 365 days) may need to be set to more accurately reflect the recent business situation.

Example 3

On the basis of the above embodiment, the risk assessment section divides the preprocessing data into a training set and a test set by setting a time valueDividing the pretreatment data into a training set and a test set, and setting the initial time in the pretreatment data to be +.>Will->Preprocessing data in a time range as a training set, and +.>The preprocessed data after the time range is used as a test set; the time value->The following constraint relationship must be satisfied:

。

specifically, the risk assessment section divides the preprocessed data into a training set and a test set by setting a time valueThe preprocessed data is divided into training and test sets. Specifically, the start time in the pre-processed data is +.>Will->Preprocessing data in a time range as a training set, and +.>The preprocessed data after the time frame is used as a test set. In this way, future data can be prevented from being used for training the model, so that the generalization capability of the model to unknown data is ensured.

Time valueThe following constraint relationship must be satisfied:

。

wherein,is a set time threshold representing the time span of the external trade data collected by the data collection portion. This constraint can ensure that the time span of the training set and the test set is not too small or too large in order to evaluate the performance of the model. If the time value +. >Too small, too short a time span of the training set and the test set may cause problems with over-fitting and under-fitting of the model. If the time value +.>Too large a time span of the training set and the test set may result in insufficient generalization ability of the model to future data.

Assume that a trading company wishes to use the risk early warning system to assess its trade risk for the next year. First, the system collects past external trade data of the company, including the goods, price, quantity, date of the transaction, and credit ratings of the transaction partners, etc. Assume that the time threshold F is set to 365 days, i.e., data is collected for the past year.

The system then sets a time value d to divide the data into training and test sets. D should be specified to be between 1/5F and 1/2F, i.e., between 73 days (about 2.5 months) and 182.5 days (about half a year). The d value set here should be based on the actual business and market conditions.

For example, assuming that the past market conditions are relatively stable and that the market conditions for the next year are not expected to change too much, it is possible to set d to a value close to 1/2F, i.e., about half a year. Thus, the time span of the training set is larger, and longer-term market rules can be reflected.

The system then trains a risk assessment model using the data of the training set, and uses the data of the test set to assess the performance of the model. For example, the model may find that trade risk may increase if the price of the goods being traded is too high or the credit rating of the trading partner is decreasing.

Finally, the system uses a risk assessment model to predict trade risk in the next year based on future possible trade scenarios (e.g., changes in commodity prices, changes in credit ratings of trading partners, etc.), and ranks these scenarios by risk value.

Through such a process, the trading company may pre-warn of possible risk scenarios and prepare in advance, such as adjusting the trading strategy, or finding new trading partners, etc. Thus, the risk of trade can be effectively reduced, and the stability and profit of trade are improved.

Example 4

Based on the above embodiment, it is assumed that the test set hasIndividual foreign trade data samples, each sample consisting of +.>A characteristic composition of->Indicate->Personal characteristic value->A label representing the sample to indicate whether the sample is a high risk trade; establishing a risk assessment decision tree model by using a C4.5 decision tree algorithm, wherein the risk assessment decision tree model is assumed to be shared +. >Individual nodes, each node->All have a judgment condition->Indicating if the characteristics of the sample meet +.>Then the sample is assigned to the nodeIn (a) and (b); every node->Are all provided with two child nodes->And->Respectively representing two possibilities of yes and no; when simulation is performed using a trade scenario simulation model, it is assumed that +.>Trade scenes, each trade scene +.>All have a feature vector +.>The method comprises the steps of carrying out a first treatment on the surface of the Risk assessment for each trade scenario using cross entropy function, assuming +.>The risk weight of the individual trade scenario is +.>。

In the risk assessment stage, the system builds a risk assessment decision tree model by using a C4.5 decision tree algorithm. In this model, it is assumed that the test set has N foreign trade data samples, each sample having m featuresEach feature is represented as. Label of each sample->Is used to indicate whether the sample is a high risk trade. The decision tree model consists of T nodes, each node has a judgment condition +.>If the characteristics of the sample satisfy +.>Then the sample will be assigned to node t. Each node has two child nodes, < ->And->They represent the two possibilities "yes" and "no", respectively.

In the trade scenario simulation stage, the system uses the trade scenario simulation model to generate K trade scenarios, each scenario having a feature vector . The system then uses the cross entropy function to perform a risk assessment for each trade scenario, assuming the risk weight for the kth trade scenario is +.>。

In practice, for example, a specific scenario may be that the credit rating of a trade partner decreases, the price of goods increases, or an unstable factor occurs in the global market, etc. These scenarios can all be translated into feature vectors. The system can then predict these conditions through the already trained decision tree modelThe prospect may be trade risk.

At the same time, risk weighting for each trade scenarioAnd also calculated. This weight value can be used to evaluate the degree of risk under different scenarios, and also to rank the various scenarios to find the scenario that is likely to result in the greatest risk. This allows for the trading company to identify potentially high risk scenarios at an initial stage, thereby enabling early adjustment and planning to reduce trade risk.

Example 5

On the basis of the above embodiment, when calculating the weighted information gain ratio in the risk assessment decision tree model, the method includes: the weighted information gain ratio is calculated using the following formula:

；

wherein, Representing the sample set contained by the node,>representing a candidate feature set, ++>Representation feature->The number of possible values, +.>Representation feature->The value is +.>Sample subset of->Representation feature->Sample set->Information gain of->Representation feature->Intrinsic value of (2);Representing the weighted information gain ratio.

Specifically, for each node, feature selection is performed using a weighted information gain ratio, and a feature having the largest weighted information gain ratio is selected as a judgment condition for the node. When the weighted information gain ratio is calculated, the information gain of each feature is calculated first, then the inherent value is calculated, and finally the weighted information gain ratio is obtained by dividing the information gain of each feature by the inherent value.

The information gain represents the degree of importance of a feature to the classification of samples, with a larger value indicating that the feature is more able to distinguish between different classes of samples. The eigenvalue represents the complexity of the feature itself, and its smaller value represents the simpler the feature, the less impact on the classification of the sample. Therefore, the weighted information gain ratio comprehensively considers the information gain and the complexity of the features, and important features can be better selected for classification.

In calculating the weighted information gain ratio, it is necessary to traverse all possible features and values and calculate the information gain and eigenvalues for each feature. This is a time consuming process, but the computation can be accelerated by optimization algorithms and parallel computation.

Specifically, the following optimization algorithm is provided to accelerate the calculation:

and updating the information gain and the eigenvalue of the new sample by using the existing calculation result. In the decision tree generation process, the calculation result of each node can be multiplexed, and the information gain and the eigenvalue of the new sample are only relevant to the current node, so the following algorithm can be used to update the values rapidly.

Taking the information gain as an example, assume that the sample set included in the current node isWherein the number of positive cases is +.>The number of negative cases is->. If a certain feature is->Dividing into values, for each value +.>Sample set +.>Dividing into subsets->Wherein the characteristic->The value of (2) is +.>. In this algorithm, if the feature +.>Information gain of->And add a sample +>The information gain can be updated quickly by the following formula:

；

wherein,indicating the addition of sample->Information gain after (I/O)>Indicating the addition of sample->Gain of previous information, ++>And->Respectively represent the sample at the characteristic->The upper value is a subset of the left and right subtrees, |d_l| and |d_r| represent the sizes of the two subsets, respectively. By this formula, the information gain of the whole sample set can be avoided from being recalculated, and the calculation result only needs to be updated incrementally.

Specifically, in the decision tree algorithm, feature selection refers to how to select an optimal feature as a judging condition of a node on each node of the decision tree, so that the classifying effect of a sample is best under the condition of the feature. The core idea of the feature selection algorithm is to calculate the contribution of each feature to the sample classification, and then select the feature with the largest contribution as the judging condition of the node.

In particular, consider the example of a simplified question that is being studied that includes two features, namely "trade amount" and "national credit rating of the trading partner". Here, the sample set D contained by the node may include a plurality of transaction records, while the candidate feature set a contains both features of "trade amount" and "national credit rating of trade partner".

Assume that the information gain of the feature "trade amount" is being calculated. First, all transaction records need to be classified into three categories, "low", "medium", "high", i.e., the sample subset $D_v$ is calculated. On this basis, the proportion of risk transactions in each subset may be calculated. For example, in a transaction record of a "low" amount, there may be 10% of transactions at high risk; in a transaction record of a "medium" amount, there may be 20% of transactions at high risk; in a transaction record of a "high" amount, 30% of the transactions may be at high risk.

These proportions may then be used to calculate the information gain for the feature "trade amount". The information gain is calculated by comparing the risk transaction proportion in each subset with the overall risk transaction proportion. If the "trade amount" feature has a high degree of accuracy in predicting risk transactions, the information gain of this feature is large.

After the information gain for all features is calculated, the weighted information gain ratio for each feature can be calculated using the above formula. The weighted information gain ratio is calculated by dividing the information gain of each feature by the eigenvalue of that feature. An eigenvalue may be understood as the degree of dispersion of a feature, which would be larger if it were able to disperse data into multiple subsets.

Finally, the feature with the largest weighted information gain ratio can be selected as the splitting feature of the decision tree. In this example, assuming that the weighted information gain ratio of "trade amount" is maximum, then "trade amount" is selected as a split feature of the decision tree to predict risk of the transaction.

Computing featuresIntrinsic value of +.>The formula is as follows:

；

selecting a feature from a candidate feature set So that its weighted information gain ratio is maximized, namely:

；

and then characterized byAs a judgment condition of the node, a sample set +.>The sub-tree is recursively built for each sub-set, respectively, divided into several sub-sets, until a certain stop condition is met.

Feature selection is an important step in the C4.5 decision tree algorithm. By feature selection, optimal features can be selected as judgment conditions for nodes, so that the sample set is divided into purer subsets. The weighted information gain ratio is a common feature selection method, which can solve the problems of unbalanced samples and different feature value numbers, so that the weighted information gain ratio is widely applied to decision tree algorithms.

Specifically, the weighted information gain ratio is based on the information gain, and a penalty term of the characteristic value number is added, so that the importance of different characteristics is compared more fairly under the condition that the characteristic value number is different. For each node, a weighted information gain ratio of each feature is first calculated, and then a feature with the largest weighted information gain ratio is selected as a judging condition of the node.

In the decision tree algorithm, the judgment condition of each nodeIs the value of a feature. For example, in the binary classification problem, the node +. >The judgment condition of (2) is expressed as->Wherein->And->The feature value is represented as "yes" or "no", respectively. When the characteristic value of the sample meets the node +.>When the judgment condition of (2) is satisfied, the sample is allocated to the node +.>Or else is allocated to another child node. By continuing recursively until the leaf nodes, a decision tree model can be obtained.

In summary, feature selection is a key step in the decision tree algorithm, and the weighted information gain ratio can help to select the optimal feature as the judging condition of the node, so that the accuracy and reliability of the decision tree algorithm are improved.

Example 6

On the basis of the above embodiment, the method for calculating the information gain includes: the information gain is calculated using the following formula:

；

wherein the method comprises the steps ofRepresenting sample set +.>Empirical entropy of>Is characterized by->Under the condition of (1) sample set->Is a rule of thumb condition entropy.

Specifically, in an actual trade risk early warning system, a large amount of trade data, such as transaction amount, transaction time, transaction product category, transaction location, transaction counterpart reputation, etc., is collected. All of these data constitute a data setThese features constitute a feature set +. >。

The main role of this formula is to help find out which features are most critical in trade risk early warning. Looking at a practical example, assume that there is a feature "transaction amount". The information gain of the feature "transaction amount" can be calculated according to the formula。

In this formula of the present invention,representing the current uncertainty of trade risk, it is a baseline that does not take into account any characteristic information. Then, calculation is required>This represents how much uncertainty is on the risk of trade after knowing the information "transaction amount". If the feature "transaction amount" is helpful for the prediction of trade risk, then knowing this information, the uncertainty about trade risk should be reduced, that is +.>Should be less than->Is small.

And then re-usingMinus->The information gain of the feature "transaction amount" is obtained. If this value is large, then it is stated that the "transaction amount" feature is important for the prediction of trade risk, and this feature should be prioritized when building the early warning model.

By calculating the information gain of all the features, the features which are most important for trade risk early warning can be found out, so that a more effective trade risk early warning model is established.

Example 7

On the basis of the above embodiment, the calculation formulas of the empirical entropy and the empirical conditional entropy are respectively as follows:

；

wherein the method comprises the steps ofRepresenting the number of categories>The representation belongs to->Sample number of individual categories, +.>Is characterized in thatThe value is +.>Under the condition of (1) sample set->Is a rule of thumb.

In particular, empirical entropy and empirical conditional entropy are important concepts used in decision tree algorithms to calculate information gain, and are used to measure uncertainty of a sample set and influence of features on sample classification.

Entropy of experienceIs a sample set +.>Is representative of the average amount of information when classifying the sample set.

When all samples belong to the same class, the empirical entropy reaches a minimumThe uncertainty representing the sample set is +.>The method comprises the steps of carrying out a first treatment on the surface of the When the samples are distributed uniformly in the individual classes, the empirical entropy reaches a maximum +.>Indicating that the uncertainty of the sample set is greatest. />

Entropy of experience conditionIs at the given level->Sample set->Is indicative of the characteristic ++>The degree of influence on the classification of the sample.

Wherein,representation feature->The number of possible values,/->Representation feature->The value is +.>Is a function of the size of the sample subset,is characterized by- >The value is +.>Under the condition of (1) sample subset->Is a rule of thumb.

The calculation of the empirical entropy and the empirical conditional entropy can be used to calculate the gain of the information, i.e. the characteristicsSample set->Is a classification capability of (c). The information gain is calculated by the following formula:

；

In the process of calculating the empirical entropy and the empirical conditional entropy, it is necessary to divide a sample set and calculate the probability of each division, thereby calculating the entropy value. The larger the values of the empirical entropy and the empirical conditional entropy, the higher the uncertainty representing the sample. In the decision tree algorithm, the purpose of feature selection is to select features that will minimize the uncertainty after node partitioning. Therefore, the calculation of the empirical entropy and the empirical conditional entropy can help to select optimal characteristics for node division, so that the accuracy and generalization capability of the decision tree model are improved.

In particular, empirical entropy and empirical conditional entropy may be used to calculate the information gain, thereby helping the decision tree algorithm to select the optimal partitioning characteristics. The information gain represents the variation of entropy before and after node division, namely the reduction degree of sample uncertainty after node division. In the decision tree algorithm, the feature with the maximum information gain is selected as the node dividing feature each time until all the samples of the leaf nodes belong to the same category or reach a preset stopping condition.

The calculation of the empirical entropy and the empirical conditional entropy can also be used to evaluate the importance of the feature. For each feature, the information gain can be calculated according to the division condition of the feature on the sample set, and the larger the information gain is, the larger the influence of the feature on the sample set is, and the higher importance is achieved. In practical application, feature selection can be performed according to the importance of the features, and features which do not contribute to model prediction are removed, so that the accuracy and efficiency of the model are improved.

At actual risk of tradeIn the early warning system, trade data is divided into different categories, such as "high risk", "medium risk" and "low risk", which are in the formula. Then, the number of samples belonging to each category is calculated, which is +.>。

Looking back at the formulasThis is a concept of empirical entropy that represents the degree of confusion or uncertainty of the current data set. If all trade data belong to the same category, e.g. "low risk", then +.>The value of (2) will be 0 indicating no confusion or uncertainty. Conversely, if trade data is uniformly distributed in each category, +.>The value of (2) is the largest and represents the greatest degree of confusion or uncertainty.

Next, seeThis is a concept of conditional entropy. This formula is in the known feature +.>In the case of the value of (2), the data set is calculated +.>Is a degree of confusion of (a). The smaller this value, the description feature +.>The more can help reduce uncertainty in trade risk.

Taking actual trade risk early warning as an example, assume a specialThe sign "A" represents the reputation of the transaction partner.There are two possible values, good and bad, respectively. Then it can be calculated at +.>In case of a value of "good" and "bad", the data set +.>Is then according to the formula +.>They are weighted and summed to give the data set +.>Is a degree of confusion of (a). If this value is compared +.>There is a significant reduction that suggests that this feature of the reputation of the transaction partner is helpful in reducing uncertainty in the risk of trade, and should be emphasized in the model.

Example 8

On the basis of the above embodiment, the risk assessment model is expressed using the following formula:

；

wherein the method comprises the steps ofIndicate->True risk weight for individual trade scenario, +.>Representing the risk weight predicted by the model; to ensure the sum of risk weightsFor 1, normalization was performed using the softmax function, i.e.:

；

In particular, the model has important application in actual trade risk early warning. For example, when a series of trade data is collected, the data may be separated into different trade scenarios. Each trade scenario has its unique characteristics such as trade amount, reputation of trade object, category of merchandise, etc.

By means of the model, a risk weight can be calculated for each trade scenarioThis risk weight expresses the model's prediction of the risk that the trade scenario is at.

In the model training phase, a weighted cross entropy function is used as a loss function, wherein the weight of each trade scenarioThe settings may be based on the weight of the sample to highlight the importance of high risk commerce. For example, if a trade scenario involves a large amount of money, it may be given a higher weight because if the trade scenario is truly risky, the company will be more affected.

Finally, this model predicts the risk weight for each trade scenarioThereby providing risk early warning for the company. If the risk weight of a trade scenario +.>High, the company needs to pay attention to this and take corresponding risk control measures.

Therefore, through the model, the company can early warn the potential trade risk in advance, so that measures can be timely taken, and the loss caused by the trade risk is avoided or reduced.

Example 9

On the basis of the above embodiment, the executing process of the trade scenario simulation model includes: for each featureCalculate its mean->And standard deviation->The method comprises the steps of carrying out a first treatment on the surface of the For each trade scenario->Generate->Random number->Satisfies the following conditionsThe method comprises the steps of carrying out a first treatment on the surface of the The generated random number is combined into a feature vector +.>Inputting into a risk assessment decision tree model for prediction to obtain a risk value +.>The method comprises the steps of carrying out a first treatment on the surface of the Repeating the above process->Next, a plurality of possible trade scenarios are generated.

Specifically, for each featureCalculate its mean->And standard deviation->. The mean represents the mean of the feature in the sample set and the standard deviation represents the degree of dispersion of the feature in the sample set.

For each trade scenarioGenerate->Random number- >Satisfy->. Here->Expressed as +.>For mean value->Is a normal distribution of standard deviation. Generated random number->Feature vectors for modeling each trade scenario.

Combining the generated random number into a feature vectorInputting into decision tree model for prediction to obtain risk value +.>. Feature vector->By->Random number composition corresponding to +.>And features. Feature vector +.>Inputting into decision tree model for prediction to obtain trade scenario +.>Risk value of->。

Repeating the above processNext, a plurality of possible trade scenarios are generated.Representing the number of generated trade scenarios.

Weighting risk weights for all trade scenariosThe risk weight of each node can be obtained by averaging, and the formula is as follows. Wherein->Representing trade scenario->Risk weight of->Representing trade scenario->Assigned to node->Judging condition of->Representing trade scenario->Is a real tag of (a).Indicating an indication function, when the condition isReturns 1 when true, otherwise returns 0.Representing node->And represents the risk level of the sample under the node.

Specifically, in actual trade risk early warning, this step is a key link for risk prediction based on statistical data. To understand its use in real trade, let a simplified scenario be assumed:

Assume an international trading company that needs to assess risk of trading with different countries. They collect various characteristic data including trade amount, date of trade, type of goods traded, country of trade partner, etc. These features may help companies predict risk levels for trade.

In this model, each featureAll have an average value +.>And standard deviation->. For example, the average of the trade amounts may be $500 ten thousand and the standard deviation may be $100 ten thousand.

The company then uses the trade scenario simulation model to predict future trade conditions. In each simulated trade scenario, a company generates a series of random numbersThese random numbers conform to a normal distribution, the mean and standard deviation of which are the mean +.>And standard deviation->。

For example, for a simulated trade, the company may generate a random trade amount that is generated around a normal distribution of $ 500 ten thousand on average, 100 ten thousand on standard deviation. This allows the possible trade amount to be simulated.

The same procedure may also be applied to other features such as date of transaction, type of goods being transacted, country of the trading partner, etc. The combination of all these eigenvalues constitutes a simulated trade scenario.

The simulated trade scenario is then input into a risk assessment model that generates a predicted risk value for the scenario. This risk value reflects the risk level that this simulated scenario may present.

Finally, by repeating this process, a large number of simulated trade scenarios and corresponding risk values may be generated, which may enable risk assessment for various trade scenarios possible in the future, thereby better circumventing the risk in actual trade.

Example 10

On the basis of the above embodiment, the risk weights of all trade situations are weighted and averaged, so that the risk weight of each node can be obtained, and the formula is as follows:

；

The risk assessment method based on the decision tree is used, has interpretability and comprehensiveness, and can help users to understand the process and the result of trade risk assessment in depth, and the trust degree of the users on the predicted result is improved. And secondly, the invention uses a trade scene simulation model to simulate risk assessment under different trade scenes, thereby enhancing the robustness and reliability of the model. The trade scenario simulation model can help the user to deeply understand risk prediction results under different conditions, so that risk management and decision making are better performed. Thirdly, the invention uses the weighted cross entropy function to carry out model training, which can effectively improve the importance of high risk trade and reduce the misjudgment rate and the missed judgment rate. Meanwhile, the weighted cross entropy function can also avoid the problem of sample unbalance, and the generalization capability and accuracy of the model are improved. Fourth, the invention uses the method of feature standardization, can effectively reduce the influence between the features, raise robustness and reliability of the model. Feature normalization can help users to know the influence of different features on risk prediction in depth, and the effect and efficiency of risk management are improved. Fifth, the invention has higher practicability and applicability. The method is not only suitable for external trade risk management and decision making of enterprises, but also can be used for trade supervision and risk management. Meanwhile, the invention can be applied to risk assessment and decision making in other fields, and has wide application prospect and market potential. In conclusion, the invention adopts a plurality of advanced technical means and methods, and has high innovation and practicability. The application of the invention can effectively improve the effect and efficiency of external trade risk management and decision making, reduce the loss and influence caused by trade risk, and has important economic and social significance.

It will be clear to those skilled in the art that, for convenience and brevity of description, specific working procedures of the above-described systems, apparatuses and units may refer to corresponding procedures in the foregoing method embodiments, which are not repeated herein.

In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods may be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative, e.g., the division of the units is merely a logical function division, and there may be additional divisions when actually implemented, e.g., multiple units or components may be combined or integrated into another system, or some features may be omitted or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be an indirect coupling or communication connection via some interfaces, devices or units, which may be in electrical, mechanical or other form.

The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in the embodiments of the present invention may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit. The integrated units may be implemented in hardware or in software functional units.

The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied essentially or in part or all of the technical solution or in part in the form of a software product stored in a storage medium, including instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-only memory (ROM), a random access memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

The above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. External trade risk early warning system, its characterized in that, the system includes:

the risk judging part is used for evaluating the risk value by using the risk evaluation model, calculating the risk weight of each trade scene, sequencing according to the risk weight and determining the high-risk trade scene;

when the weighted information gain ratio is calculated in the risk assessment decision tree model, the method comprises the following steps: the weighted information gain ratio is calculated using the following formula:

；

wherein,representing the sample set contained by the node,>representing a candidate feature set, ++>Representing candidate feature set +.>The number of possible values, +.>Representing candidate feature set +.>The value is +.>Vectors corresponding to sample subsets +.>Representing candidate feature set +.>Sample set->Information gain of->Representing candidate feature set +.>Intrinsic value of (2);Representing a weighted information gain ratio;

assuming that the test set hasIndividual foreign trade data samples, each sample consisting of +.>A characteristic composition of->Indicate->Personal characteristic value->A label representing the sample to indicate whether the sample is a high risk trade; establishing a risk assessment decision tree model by using a C4.5 decision tree algorithm, wherein the risk assessment decision tree model is assumed to be shared +. >Individual nodes, each node->All have a judgment condition->Indicating if the characteristics of the sample meet +.>Then the sample is assigned to node +.>In (a) and (b); every node->Are all provided with two child nodes->And->Respectively representing two possibilities of yes and no; when simulation is performed using a trade scenario simulation model, it is assumed that +.>Trade scenes, each trade scene +.>All have a feature vector +.>The method comprises the steps of carrying out a first treatment on the surface of the Performing risk assessment on each trade scenario using a cross entropy function;

the risk assessment model is expressed using the following formula:

；

2. The system of claim 1, wherein the time span of the external trade data collected by the data collection portion exceeds a set time threshold The method comprises the steps of carrying out a first treatment on the surface of the The time threshold->The value range of the product is 365 days to 1100 days.

3. The system of claim 2, wherein the risk assessment section divides the preprocessed data into the training set and the test set by setting a time valueDividing the pretreatment data into a training set and a test set, and setting the initial time in the pretreatment data to be +.>Will->Preprocessing data in a time range as a training set, and +.>The preprocessed data after the time range is used as a test set; the time value->The following constraint relationship must be satisfied:。

4. The system of claim 3, wherein the information gain calculation method comprises: the information gain is calculated using the following formula:

；

wherein the method comprises the steps ofRepresenting sample set +.>Empirical entropy of>Expressed in candidate feature set +.>Under the condition of (1) sample setIs a rule of thumb condition entropy; the calculation formulas of the empirical entropy and the empirical conditional entropy are respectively as follows:

；

wherein the method comprises the steps ofRepresenting the number of categories>The representation belongs to->Sample number of individual categories, +.>Expressed in candidate feature set +.>The value is +.>Under the condition of (1) sample set->Is a rule of thumb.

5. The system of claim 4, wherein the execution of the trade scenario simulation model comprises: for each feature Calculate its mean->And standard deviation->The method comprises the steps of carrying out a first treatment on the surface of the For each trade scenario->Generate->Random number->Satisfies the following conditionsThe method comprises the steps of carrying out a first treatment on the surface of the The generated random number is combined into a feature vector +.>Inputting into a risk assessment decision tree model for prediction to obtain a risk value +.>The method comprises the steps of carrying out a first treatment on the surface of the Repeating the above process->Next, a plurality of possible trade scenarios are generated.

6. The system of claim 5, wherein the risk weights for each node are obtained by weighted averaging the risk weights for all trade scenarios, as follows:

；