WO2019061187A1 - Credit evaluation method and apparatus and gradient boosting decision tree parameter adjustment method and apparatus - Google Patents

Abstract

A credit evaluation method and apparatus and a gradient boosting decision tree parameter adjustment method and apparatus, the credit evaluation method comprising: respectively inputting first sample data into at least two gradient boosting decision tree GBDT models to obtain a first overdue credit probability set, the first sample data being credit data of a first user set; respectively inputting second sample data into at least two GBDT models to obtain a second overdue credit probability set, the second sample data being credit data of a second user set; the GBDT parameters of the at least two GBDT models are different; on the basis of the first overdue credit probability set and the second overdue credit probability set, implementing KS value calculation and, on the basis of the calculation result, determining a target GBDT model from amongst the at least two GBDT models; and, on the basis of the target GBDT model, performing a credit evaluation of a user.

Description

Credit evaluation method and device, and gradient progressive decision tree parameter adjustment method and device Technical field

The present disclosure relates to the field of information processing technologies, for example, to a credit evaluation method and apparatus, and a gradient progressive decision tree parameter adjustment method and apparatus.

Background technique

The Gradient Boost Decision Tree (GBDT) is a commonly used algorithm for solving classification problems and regression problems. The advantage is that it has strong fitting ability and classification ability, but the strong fitting ability may be tested. There has been a fitting phenomenon on the set.

In the related art, when the GBDT model is used for credit evaluation of users, it is usually necessary to manually adjust multiple parameters in the GBDT model one by one, so that the credit overdue probability outputted by the GBDT model is close to the user's true credit overdue probability, but in GBDT In the parameter adjustment process, the parameters are often adjusted based on the artificially determined parameter values, the accuracy of the parameters is not high, the model obtained by the parameter tuning method is unstable, the parameter adjustment efficiency is low, and the accuracy of the credit evaluation for the user is performed. Lower.

Summary of the invention

The present disclosure provides a credit evaluation method and apparatus and a gradient progressive decision tree parameter adjustment method and apparatus, which can improve the parameter adjustment efficiency of the GBDT model, improve the stability of the GBDT model, and ensure the accuracy of credit evaluation for users.

An embodiment provides a credit evaluation method, which may include:

The first sample data is respectively input into at least two gradient progressive decision tree GBDT models to obtain a first credit overdue probability set, and the first sample data is credit data of the first user set;

Entering the second sample data into the at least two GBDT models respectively to obtain a second credit overdue summary a rate set, the second sample data is credit data of a second user set; and the GBDT parameters of the at least two GBDT models are different;

Performing a KS value calculation according to the first credit overdue probability set and the second credit overdue probability set, determining a target GBDT model from the at least two GBDT models according to the calculation result; and pairing the user according to the target GBDT model Conduct a credit evaluation.

An embodiment provides a credit evaluation apparatus, which may include:

The first credit overdue probability obtaining module is configured to input the first sample data into the at least two gradient progressive decision tree GBDT models respectively, to obtain a first credit overdue probability set, where the first sample data is the first user set Credit data

a second credit overdue probability obtaining module, configured to input second sample data into the at least two GBDT models respectively to obtain a second credit overdue probability set, where the second sample data is credit data of the second user set; The GBDT parameters of at least two GBDT models are different;

a model determining module, configured to perform a KS value calculation according to the first credit overdue probability set and the second credit overdue probability set, and determine a target GBDT model from the at least two GBDT models according to the calculation result; and an evaluation module And set to perform credit evaluation on the user according to the target GBDT model.

An embodiment provides a gradient progressive decision tree parameter adjustment method, which may include:

Determining the domain dimension and the domain range of the particle swarm optimization algorithm according to the number of adjustment parameters in the gradient progressive decision tree and the range of values corresponding to each parameter;

Setting an initial parameter of the particle swarm optimization algorithm, obtaining a trajectory of each particle in the particle group according to the particle swarm optimization algorithm, the domain dimension, and the domain definition; determining the best advantage according to the trajectory The parameter value of the gradient progressive decision tree.

An embodiment provides a gradient progressive decision tree parameter adjustment apparatus, including:

Mapping module, set to adjust the number of parameters in the progressive decision tree according to the gradient and each parameter pair The range of values to be determined determines the domain dimensions of the particle swarm optimization algorithm and the scope of the domain;

a trajectory most advantageous determining module, configured to set an initial parameter of the particle swarm optimization algorithm, and obtain a trajectory of each particle in the particle group according to the particle swarm optimization algorithm, the domain dimension, and the domain definition range ;

The parameter determination module is configured to determine a parameter value of the gradient progressive decision tree according to the most advantageous of the trajectory.

An embodiment provides a computer readable storage medium storing computer executable instructions for performing any of the methods described above.

An embodiment further provides a data processing device comprising one or more processors, a memory, and one or more programs, the one or more programs being stored in the memory when processed by one or more When the device is executed, perform any of the above methods.

An embodiment further provides a computer program product comprising a computer program stored on a non-transitory computer readable storage medium, the computer program comprising program instructions, when the program instructions are executed by a computer Having the computer perform any of the methods described above.

The present disclosure can improve the parameter adjustment efficiency of the gradient progressive decision tree, avoid the local optimal search that falls into a single region during the adjustment process, and have a wider search range for the parameter space.

DRAWINGS

FIG. 1a is a schematic flowchart of a credit evaluation method according to an embodiment; FIG.

FIG. 1b is a schematic diagram of a sub-flow of a credit evaluation method according to an embodiment; FIG.

FIG. 1c is a schematic diagram of another sub-flow of a credit evaluation method according to an embodiment;

2a is a schematic flow chart of a credit evaluation provided by an embodiment;

2b is a schematic diagram of a sub-flow of credit evaluation provided by an embodiment;

FIG. 3 is a schematic structural diagram of a credit evaluation apparatus according to an embodiment; FIG.

4 is a schematic flow chart of a gradient progressive decision tree parameter adjustment method according to an embodiment;

FIG. 5 is a schematic flowchart diagram of another gradient progressive decision tree parameter adjustment method according to an embodiment; FIG.

FIG. 6 is a schematic flowchart diagram of another gradient progressive decision tree parameter adjustment method according to an embodiment; FIG.

FIG. 7 is a schematic structural diagram of a gradient progressive decision tree parameter adjusting apparatus according to an embodiment; FIG.

FIG. 8 is a schematic diagram of a hardware structure of a data processing device according to an embodiment.

Detailed ways

FIG. 1a is a schematic flowchart of a credit evaluation method according to an embodiment. The method may be applied to a data processing device, such as a computing device. As shown in FIG. 1a, the method may include steps 110-140.

In step 110, the first sample data is respectively input into at least two gradient progressive decision tree GBDT models to obtain a first credit overdue probability set, and the first sample data is credit data of the first user set.

In step 120, the second sample data is separately input into the at least two GBDT models to obtain a second credit overdue probability set, the second sample data is credit data of the second user set; the at least two GBDTs The model's GBDT parameters are different.

For example, the user's credit data may include information such as the user's performance capability, long-term data, credit duration, total amount of arrears, and behavioral preference. After inputting the sample data into the GBDT model, the user's credit overdue probability may be obtained.

The performance capability may include user history overdue records, such as historical maximum overdue days and 90 days or 180 days of overdue times; the multi-head data may include the user's past 30 days, 60 days, 90 days, 120 days, and 180 days. Within, information on the number of borrowings on the financial platform and non-financial platforms; credit duration may include information such as the length of time the user opens an account, the start time of the first transaction, and the length of time the mobile phone is on the network; the total amount of the arrears may include the current pocket of the individual user. The total amount or the total amount of loans in the organization and the total amount of loans outside the organization; behavioral preferences may include information such as whether the user is browsing on multiple types of web pages or purchasing consumer goods, the user withdraws cash, virtual transactions, or the proportion of e-commerce transactions in the online registration.

In this embodiment, the first sample data and the second sample data respectively include credit data of a plurality of users, and all credit data of each user in the first sample data is input into each GBDT model to obtain the first The credit expected probability of the plurality of users in the sample data constitutes the first credit overdue probability set, and similarly, the second credit overdue probability set is obtained.

In step 130, the KS value calculation is performed according to the first credit overdue probability set and the second credit overdue probability set, and the target GBDT model is determined from the at least two GBDT models according to the calculation result.

Optionally, as shown in FIG. 1b, the foregoing step 130 may include steps 1310 to 1330.

In step 1310, the KS value calculation is performed according to the first credit overdue probability set and the first actual credit overdue probability set corresponding to the first user set, to obtain a first KS set.

In step 1320, the KS value calculation is performed according to the second credit overdue probability set and the second actual credit overdue probability set corresponding to the second user set, to obtain a second KS set.

For example, according to the first credit overdue probability set and the second credit overdue probability set, a corresponding probability threshold may be selected, and according to the calculation principle of the KS curve, the KS value corresponding to each GBDT model of the first sample data input is obtained. The first KS set is formed, and the second KS set is obtained in the same manner.

In step 1330, a comparison calculation is performed on the first KS set and the second KS set, and the target GBDT model is determined from the at least two GBDT models according to a calculation result.

For example, the first sample may be determined as a training sample, and the second sample may be determined as a test sample, and the KS value in the first KS set may be represented as KS_train, and the KS value in the second KS set may be represented as KS_test.

Optionally, as shown in FIG. 1c, the above step 1330 may include step 1332 - step 1336.

In step 1322, the KS value of the first KS set obtained according to the same GBDT model and the KS value of the second KS set are subjected to a minimum value calculation to obtain a third KS set.

For example, the KS value of the first KS set calculated according to the same GBDT model and the KS value of the second KS set may be calculated by using the min function (KS_train, KS_test) to obtain a minimum value, which is a third value. KS set.

In step 1334, the maximum value of the KS value included in the third KS set is calculated to obtain a target KS value.

For example, the plurality of KS values in the third KS set are calculated according to the function max(min(KS_train, KS_test)), and the KS value of the target is obtained, that is, the KS maximum value in the third KS set is calculated, that is, the target KS value.

In step 1336, the GBDT model corresponding to the target KS value in the at least two GBDT models is determined as the target GBDT model.

In step 140, the user is credited according to the target GBDT model.

For example, the newly input user credit data is input into the target GBDT model to obtain the credit overdue probability of the user, and the user's credit condition can be evaluated according to the credit overdue probability of the user.

Optionally, as shown in FIG. 2a, before step 110, step 100 is further included.

In step 100, the GBDT parameters of the at least two GBDT models are determined according to a particle swarm optimization PSO algorithm.

Optionally, as shown in FIG. 2b, step 100 may include: step 1010 - step 1050.

In step 1010, the number of parameters in the GBDT model is mapped to the domain dimension of the PSO algorithm.

In step 1020, the range of values of each of the parameters in the GBDT model is mapped to the domain of the PSO algorithm.

In step 1030, at least two sets of dimension value data are extracted from the domain of the domain corresponding to the domain dimension as at least two particles.

In step 1040, the trajectory of the at least two particles is calculated by the PSO algorithm.

The most advantageous of the trajectory refers to a point in the trajectory of the particle that makes the objective function reach a maximum value, and the objective function is that the KS value in the first KS set and the KS value in the second KS set are taken. The function of the minimum and

In step 1050, the dimension value data corresponding to the trajectory of the at least two particles is mapped back into the GBDT model to obtain at least two sets of GBDT parameters.

Among them, the PSO algorithm belongs to the particle swarm theory, which defines the particles xi=(x1,x2,...,xN) in the N-dimensional space, and the flying speed of the particles in space is vi=(v1,v2,...,vN ), each particle has a fitness value determined by the objective function, and each particle follows the optimal particle in the entire particle group to search in space, and finds the best position in the whole space after multiple iterations. .

Optionally, performing credit evaluation on the user according to the target GBDT model, including:

Entering the credit data of the user into the target GBDT model to obtain a credit overdue probability of the user; and comparing the credit overdue probability of the user with a preset credit overdue probability threshold to obtain a credit evaluation result of the user .

For example, the corresponding credit overdue probability threshold may be set, for example, when the user's credit overdue probability is greater than or equal to 80%, the user credit is determined to be poor; when the user's credit overdue probability is less than 80%, greater than or equal to 50%, the user credit is generally determined. When the user's credit overdue probability is less than 50%, greater than or equal to 10%, it is determined that the user's credit is good, and when the user's credit overdue probability is less than 10%, it is determined that the user credit is excellent.

FIG. 3 is a schematic structural diagram of a credit evaluation apparatus according to an embodiment. The apparatus may perform the credit evaluation method provided by the foregoing embodiment. The function of the module in this embodiment may refer to the method steps provided in the foregoing embodiment. As shown in Figure 3, the device can include:

The first credit overdue probability obtaining module 310 is configured to input the first sample data into the at least two gradient progressive decision tree GBDT models respectively, to obtain a first credit overdue probability set, where the first sample data is Credit data of the first user set;

The second credit overdue probability obtaining module 320 is configured to input the second sample data into the at least two GBDT models respectively to obtain a second credit overdue probability set, where the second sample data is the credit data of the second user set; The GBDT parameters of the at least two GBDT models are different;

The model determining module 330 is configured to perform KS value calculation according to the first credit overdue probability set and the second credit overdue probability set, and determine a target GBDT model from the at least two GBDT models according to the calculation result;

The evaluation module 340 is configured to perform credit evaluation on the user according to the target GBDT model.

Optionally, the apparatus may further include a parameter determining module 300 configured to determine the at least two according to the particle swarm optimization PSO algorithm before inputting the first sample data into the at least two gradient progressive decision tree GBDT models, respectively. GBDT parameters of the GBDT model.

FIG. 4 is a flowchart of a gradient progressive decision tree parameter adjustment method according to an embodiment. The method is applicable to parameters in a gradient progressive decision tree when performing calculations such as modeling or machine learning using a gradient progressive decision tree. In the case of adjustment, the method may be performed by a computing device such as a computer, or may be performed by a gradient progressive decision tree parameter adjustment device, and the gradient progressive decision tree parameter adjustment device may be implemented by at least one of software and hardware, such as As shown in FIG. 4, the method can include steps 410-430.

In step 410, the domain dimension and the domain scope of the particle swarm optimization algorithm are determined according to the number of adjustment parameters in the gradient progressive decision tree and the range of values corresponding to each parameter.

For example, there are 8 adjustment parameters in the gradient progressive decision tree, namely n_estimators, learning_rate, subsample, max_features, max_depth, min_samples_split, min_samples_leaf, and random_state. Correspondingly, the domain dimension of the particle swarm optimization algorithm is 8 dimensions.

Among them, n_estimators refers to the maximum number of iterations of the weak learner. If the value of n_estimators is too small, it is easy to underfit. The value of n_estimators is too large and easy to overfit. Adjust the size of the value of n_estimators to choose a moderate value. The value of n_estimators The range can be defined as [1,1000]. Learning_rate refers to the weight reduction coefficient of each weak learner, also called the step size. For the same training set fitting effect, the smaller step size indicates that more weak learner iterations are needed, and the learning_rate value range can be Defined as (0,1). Subsample refers to subsampling, with a range of (0,1). Max_features refers to the maximum number of features, the value range can be set to (0,1). Max_depth refers to the maximum depth of the decision tree, which can be any integer in (0, 10). Min_samples_split refers to the minimum number of samples required for internal node division. This value limits the condition for subtree to continue to divide. If the number of samples of a node is less than min_samples_split, then it will not continue to try to select the optimal feature for partitioning. The min_samples_split is taken. The value range can be set to [1,1000]. Min_samples_leaf refers to the minimum number of samples of leaf nodes. If the number of leaf nodes is less than the minimum number of samples mentioned above, the leaf nodes will be pruned together with the sibling nodes. When the sample size is not large, the value plays a small role when the sample size If the order of magnitude is very large, adjust the value adaptively. The random_state parameter is used to randomly divide the training samples (ie, modeled samples) and test samples, and the range of values can be defined as [1,1000].

The above parameters and corresponding range of values are mapped into the domain of the particle swarm optimization algorithm to obtain the domain dimension and the domain scope of the particle swarm optimization algorithm. Among them, Particle Swarm Optimization (PSO) is a population-based stochastic optimization algorithm that can mimic the clustering behavior of insects, herds, flocks and fish groups. These groups are in a cooperative way. Looking for food, each member of the group constantly changes its search model by learning its own experience and the experience of other members. In this embodiment, the particle swarm optimization algorithm is selected to adjust the decision tree parameters as an example, and other stochastic optimization algorithms may also be used to adjust the decision tree parameters.

In step 420, setting initial parameters of the particle swarm optimization algorithm according to the particle group optimization The algorithm, the domain dimension, and the scope of the domain are derived to obtain the most advantageous trajectory of each particle in the particle group.

Among them, the initial parameters of the particle swarm optimization algorithm can be set to (ω, φ ₁ , φ ₂ ), where ω is the impulse term, the value is between (0, 1) (can be defined as 0.5), and the size of φ ₁ can be Custom, if defined as 0.5, φ ₂ size is the setting parameter of PSO, which can be defined as 0.5, the number of specified particle populations (popsize) is 100, and the speed and position of 100 particles are randomly assigned through the particles. The current position and the current speed are updated by the particle position, and the speed of the particle is updated according to the value of the objective function. For example, the PSO algorithm updates the next and next positions of the particle based on the best merit of the trajectory that each particle has traveled and the global orbital best of the 100 particles combined with the speed of the current particle. The formula is as follows:

为该粒子的轨迹最优点，代表粒子曾经走过的使目标函数达到最大值的点，

为全局最优点，即所有粒子走过的点中使目标函数达到最大值的点，x_i表示粒子的当前位置，x_i+1表示粒子的下一位置，。Where v _i+1 represents the next velocity of the particle, v _i represents the current velocity of the particle, ω is the impulse term, and U(0, φ ₁ ) is a random number uniformly distributed between (0, φ ₁ ), U (0, φ ₂ ) is a random number uniformly distributed between (0, φ ₂ ),

The best point for the trajectory of the particle, which represents the point at which the particle has reached the maximum value of the objective function.

For the global best, the point at which all particles travel through the point where the objective function reaches its maximum value, x _i represents the current position of the particle, and x _i+1 represents the next position of the particle.

Optionally, the trajectory of each of the 100 particles calculated by the PSO algorithm is recorded as the most advantageous.

In step 430, the parameter values of the gradient progressive decision tree are determined based on the trajectory best.

The parameter values of the final gradient progressive decision tree are determined based on the trajectory of the recorded particles. Wherein, the objective function is the minimum function of the KS value of the training sample and the KS value of the test sample, ie min(KS_train, KS_test), and the most advantageous trajectory of the particle is obtained according to the maximum objective function of the particle swarm optimization algorithm, that is, according to max(min) (KS_test, KS_train)) function is obtained.

The parameter value of the gradient progressive decision tree is determined according to the best advantage of the trajectory and the value of the corresponding objective function, and the objective function is a minimum function of the KS value of the training sample and the KS value of the test sample. Among them, the KS value is an evaluation index used to distinguish the degree of separation between positive and negative samples in the model. The value range of the KS value is [0, 1], indicating the separation ability of the model.

The GBDT model in this embodiment can be used as a credit scoring model, and the sample data can be the user's credit information, such as the user's performance capability, long data, credit duration, total amount of arrears, and behavioral preferences, and the sample data is input into the GBDT model. After that, you can get the user's credit overdue probability. The process of adjusting the gradient progressive decision tree parameters in this embodiment may include steps 11 - 18:

In step 11, according to the number of parameters in the GBDT and the range of values of each parameter, the domain is mapped to the domain of the PSO algorithm, and the domain dimension and the domain scope of the PSO algorithm are obtained.

In step 12, 100 sets of data, that is, the above 100 particles, can be randomly extracted in the domain dimension and the domain range of the PSO algorithm.

In step 13, according to the trajectory of the extracted 100 particles, and the best advantage of the global trajectory, the calculation is performed according to the above formula (1), and the next position of each particle is updated until the adaptation according to each particle The fitness value comparison determines the most advantageous trajectory of each particle.

For example, the above particles can be:

[n_estimators, learning_rate, subsample, max_features, max_depth, min_samples_split, min_samples_leaf, random_state], updating the position of the particle can be understood as the position of the particle in the previous step: [50, 0.1, 0.8, 0.7, 5, 900, 500, 70 ], according to the formula of PSO, the position of the particle can be updated to another position: [52, 0.096, 0.73, 0.65, 4, 903, 495, 69].

In step 14, according to the dimension value corresponding to the trajectory of the above 100 particles, the mapping is returned to the GBDT, and the corresponding 100 sets of GBDT parameters are obtained.

In S15, the 100 sets of GBDT parameters obtained above are grouped into the GBDT model for credit card scoring, and the training sample data and the test sample data are respectively substituted to obtain the credit overdue probability value of the corresponding user.

In step 16, according to the real credit overdue probability of the user and the credit overdue probability obtained according to the GBDT model, the KS value is calculated for each group of users' credit overdue probability values, and 100 KS values of the training sample data are obtained (ie, KS-test) And test the sample data for 100 KS values (ie KS-train).

In step 17, the target KS-test value is obtained according to max(min(KS-train, KS-test)).

KS-train is the KS value calculated based on the training sample data, KS-test is the KS value calculated according to the test sample data, and for a set of GBDT parameters, corresponding to a KS-train value and a KS-test value, In this embodiment, 100 sets of particles are set in the PSO algorithm, so there are 100 sets of GBDT parameters, corresponding to 100 KS-trains and 100 KS-test values, and KS-train and KS-test corresponding to each set of GBDT parameters, The comparison calculation is performed according to max(min(KS-train, KS-test)), thereby obtaining the target KS-test value.

For example, comparing KS-train and KS-test corresponding to 100 sets of GBDT parameters, according to min (KS-train, KS-test), 100 smaller KS values are obtained, and 100 smaller KS values are selected. The maximum KS value, resulting in the target KS-test value.

In step 18, the user is credit evaluated using the target GBDT model corresponding to the target KS-test value.

For example, the GBDT parameter value corresponding to the target KS-test is used as the parameter value of the GBDT model to obtain the target GBDT model, and the credit information of the new user is input into the target GBDT model, and the credit overdue probability of the new user is obtained, and the overdue period can be set. The probability threshold, when the user's credit overdue probability reaches the probability threshold, then the user's credit is lower. You can also set multiple credit overdue probability ranges and corresponding letters. Use rank.

In this embodiment, the parameter value corresponding to the trajectory best advantage when the objective function value is the largest value is selected as the parameter value of the decision tree, thereby maximizing the KS value of the training sample and the KS value of the test sample. The selected objective function is to maximize min(KS_train, KS_test), which can effectively prevent the test sample KS from being higher than the training sample KS, and can well make the KS values of the training and test samples close, thereby obtaining a model with strong generalization ability. .

Optionally, the original data set is classified into training samples and test samples, wherein the original data set may be modeled sample data for predicting credit overdue probability.

Define popsize=100, generation=100, ω=0.5, φ ₁ =0.5, φ ₂ =0.5 in the PSO algorithm, and calculate the parameter value corresponding to the most advantageous trajectory with the largest objective value (the fitness value) in the most advantageous set of trajectories. As follows (where the fitness value is 0.44368566870386):

N_estimators=89.9755412363669, learning_rate=0.255267311338214, Subsample=0.861905071771738, max_features=0.786393083477439, max_depth=5.51493470652752, min_samples_split=788.538534238246, min_samples_leaf=318.682482373024, random_state=678.303928724576.

When the above parameters corresponding to the most advantageous trajectory are mapped back to the decision tree parameter, the parameters that need to be rounded are automatically rounded. For example, the value of the parameter n_estimators must be an integer. Then, the parameter value is rounded down to obtain the result. Is 89.

The related technology cannot perform global search when adjusting the parameters of the decision tree, and the accuracy of the adjusted parameters is not high, and manual parameter adjustment needs to manually set the parameter values of the gradient decision tree, and then perform multiple adjustments one by one according to the results. The embodiment provides a gradient progressive decision tree parameter adjustment method, which can avoid the local optimal search in a single region, without manually determining the value of the parameter and performing the parameter test one by one, and the parameter value of the GBDT provided by the embodiment. Adjustment method is more than manual adjustment in the test sample The upper KS value is higher and the resulting model is more stable.

FIG. 5 is a flowchart of another gradient progressive decision tree parameter adjustment method according to an embodiment. As shown in FIG. 5, the method provided in this embodiment may include steps 510-530.

In step 510, the domain dimension and the domain scope of the particle swarm optimization algorithm are determined according to the number of adjustment parameters in the gradient progressive decision tree and the range of values corresponding to each parameter.

In step 520, an initial parameter of the particle swarm optimization algorithm is set, and a trajectory of each particle in the particle group is obtained according to the particle swarm optimization algorithm, the domain dimension, and the domain definition.

In step 530, the corresponding peripheral point is determined according to the trajectory best advantage, and the parameter value of the gradient progressive decision tree is determined according to the magnitude of the value of the objective function corresponding to the peripheral point.

Wherein, the peripheral point of the most advantageous trajectory is obtained according to the Hill Climbing algorithm with the best advantage of the trajectory as a starting point, and the objective function is a minimum function of the KS value of the training sample and the test sample, for example, the most advantageous trajectory The surrounding points are obtained by the hill climbing algorithm to maximize the objective function (ie, max(KS_test, KS_train)), so that the parameters of the determined gradient progressive decision tree are better. Mountain climbing algorithm is a local optimization method. Heuristic method is an improvement on depth-first search. The algorithm uses feedback information to generate solution decisions. Since there may be better trajectory points in the peripheral points of the most advantageous trajectory in this embodiment, the hill climbing algorithm is used to perform operations to find peripheral points that are superior to the trajectory best.

For example, define the step size of the 8 parameters in the hill climbing algorithm, which can be the step size as follows:

The n_estimators step size is 1, the learning_rate step size is 0.01, the Subsample step size is 0.01, the max_features step size is 0.01, the max_depth step size is 1, the min_samples_split step size is 20, the min_samples_leaf step size is 20, and the random_state step size is 1.

According to the step size defined above, test the peripheral points of the most advantageous trajectory one by one, during the test, find The maximum point of the increase of the objective function value is the starting point of the next step. If there is no point where the value of the objective function increases, the operation is stopped, and the corresponding peripheral point when the operation is stopped is the best advantage of the track.

The embodiment provides a gradient progressive decision tree parameter adjustment method, and determines a corresponding peripheral point according to the trajectory best advantage, and determines a parameter value of the gradient progressive decision tree according to the value of the objective function corresponding to the surrounding point, thereby improving parameter adjustment. result.

For example, for the same GBDT model, the KS_train value obtained by manual tuning is 58.19%, the KS_test value is 41.57%, the KS_train value determined by the PSO algorithm is 45.19%, and the KS_test value is 44.12%. The PSO algorithm is used to add the hill climbing algorithm. The KS_train value is 50.37% and the KS_test value is 45.22%. It can be seen that the KS value determined by the PSO algorithm plus the hill climbing algorithm is higher than the KS value obtained by the PSO algorithm, and the PSO algorithm and the PSO and the hill climbing algorithm are adopted. The difference between the obtained KS value of the training sample and the KS value of the test sample is smaller than the difference between the KS value of the training sample obtained by the manual adjustment and the KS value of the test sample.

For example, the hill-climbing algorithm can be further optimized according to the global best obtained by the PSO algorithm, and the corresponding KS_train value is 45.54%, and the KS_test value is 44.46%. The effect is between using only the PSO algorithm and using the PSO algorithm combined with the hill-climbing algorithm. .

FIG. 6 is a flowchart of another gradient progressive decision tree parameter adjustment method according to an embodiment. As shown in FIG. 6, the method provided in this embodiment may include steps 610-630.

In step 610, the domain dimension and the domain scope of the particle swarm optimization algorithm are determined according to the number of adjustment parameters in the gradient progressive decision tree and the range of values corresponding to each parameter.

In step 620, an initial parameter of the particle swarm optimization algorithm is set, and a trajectory of each particle in the particle group is obtained according to the particle swarm optimization algorithm, the domain dimension, and the domain definition.

In step 630, a corresponding peripheral point is determined according to the most advantageous trajectory, and the peripheral point pair is determined. The value of the value of the objective function should be sorted, and the parameter value corresponding to the peripheral point corresponding to the value of the largest objective function is selected as the parameter value of the gradient progressive decision tree.

Optionally, the value of the value of the objective function corresponding to the peripheral point of the most advantageous trajectory is sorted, and the parameter value corresponding to the peripheral point corresponding to the value of the largest objective function is selected as the parameter value of the gradient progressive decision tree. The parameter values corresponding to the peripheral points having the largest target function value in the sort result are selected by automatic sorting. For example, the parameter values corresponding to the peripheral points having the largest target function value are as follows (where the fitness value is 0.456814121199906):

N_estimators=89.944668235715, learning_rate=0.253604654375516, subsample=0.84938040034035, max_features=0.791557099759923, max_depth=5.52083587628895, min_samples_split=785.648574406732, min_samples_leaf=323.345684890637, random_state=683.655366674717.

It is also possible to select only the point where the objective function value is the largest among the trajectories, and the climbing algorithm obtains the surrounding points, and the values of the respective dimensions corresponding to the surrounding points are determined as the parameter values of the gradient progressive decision tree.

The gradient progressive decision tree parameter adjustment method provided in this embodiment can improve the parameter adjustment efficiency of the gradient progressive decision tree, avoids the local optimal search that falls into a single region during the adjustment process, and has a wider search range for the parameter space.

Most bank credit card scoring model development frameworks are based on mathematical statistics theory. Variables (ie, parameters) must play a role in the model. Variables and output variables are statistically significant, and data volume and variable information strength are very high.

Compared with the statistical methods in the related art, the gradient progressive decision tree has stronger fitting ability and classification ability when solving classification problems and regression problems, and can more effectively utilize the weak variable information in the sample data, but A strong fit may result in over-fitting on the test set. In order to overcome the over-fitting phenomenon, algorithm parameter selection is very important. In practice, a large amount of practice relies on manual selection of parameters. This example provides a scheme for automatically selecting parameters.

FIG. 7 is a schematic structural diagram of a gradient progressive decision tree parameter adjustment apparatus according to an embodiment. The apparatus may perform the gradient progressive decision tree parameter adjustment method provided by the foregoing embodiment, and has a corresponding functional module and a beneficial effect of the execution method. As shown in FIG. 7, the apparatus may include: a mapping module 701, a trajectory best advantage determining module 702, and a parameter determining module 703.

The mapping module 701 is configured to determine a domain dimension and a domain range of the particle group optimization algorithm according to the number of adjustment parameters in the gradient progressive decision tree and the value range corresponding to each parameter;

The trajectory most advantageous determination module 702 is configured to set initial parameters of the particle swarm optimization algorithm, and obtain the trajectory of each particle in the particle group according to the particle swarm optimization algorithm, the domain dimension, and the domain definition range. ;

The parameter determination module 703 is arranged to determine a parameter value of the gradient progressive decision tree based on the trajectory best.

In this embodiment, determining the domain dimension and the domain range of the particle swarm optimization algorithm according to the number of adjustment parameters in the gradient progressive decision tree and the range of values corresponding to each parameter, and setting initial parameters of the particle swarm optimization algorithm Determining the trajectory of each particle in the particle group according to the particle swarm optimization algorithm, the domain dimension and the domain definition, determining the parameter value of the gradient progressive tree according to the trajectory best, and improving the gradient The parameter adjustment efficiency of the progressive decision tree avoids the local optimal search that falls into a single region during the adjustment process, and has a wider search range for the parameter space.

Optionally, the parameter determining module 703 is configured to:

A parameter value of the gradient progressive decision tree is determined according to the trajectory best advantage and the magnitude of the value of the corresponding objective function, the objective function being a minimum function of the KS value of the training sample and the test sample.

Optionally, the parameter determining module 703 is configured to:

Determining corresponding peripheral points according to the most advantageous trajectory, and the peripheral points of the most advantageous trajectory are as described The most advantageous advantage of the trajectory is that the starting point is obtained according to the hill climbing algorithm;

The parameter value of the gradient progressive decision tree is determined according to the magnitude of the value of the objective function corresponding to the peripheral point, and the objective function is a minimum function of the KS value of the training sample and the test sample.

Optionally, the parameter determining module 703 is configured to:

The size of the value of the objective function corresponding to the peripheral point is sorted, and the parameter value corresponding to the peripheral point corresponding to the value of the largest objective function is selected as the parameter value of the gradient progressive decision tree.

Optionally, the number of adjustment parameters of the gradient progressive decision tree is 8, and the range of the definition domain is a range from a minimum value to a maximum value of each adjustment parameter.

An embodiment further provides a computer readable storage medium storing computer executable instructions for performing any of the credit evaluation methods described above.

An embodiment further provides a storage medium containing computer executable instructions that, when executed by a computer processor, can perform any of the gradient progressive decision tree parameter adjustment methods provided by the above embodiments.

The above storage medium may be a different type of memory device or storage device. These may include: a mounting medium such as a CD-ROM, a floppy disk or a tape device; a computer system memory or a random access memory such as DRAM, DDR RAM, SRAM, EDO RAM, Rambus RAM, etc.; non-volatile memory , such as flash memory, magnetic media (such as hard disk or optical storage); registers or other similar types of memory components, and the like. The storage medium may also include other types of memory or combinations. Additionally, the storage medium may be located in a first computer system in which the program is executed, or may be located in a different second computer system, the second computer system being coupled to the first computer system via a network, such as the Internet. The second computer system can provide program instructions to the first computer for execution. Storage media may also include two or more storage media that reside in different locations (eg, in different computer systems connected through a network). A storage medium may store program instructions executable by one or more processors (eg, computer programs) sequence).

An embodiment provides a data processing device, which may be a filler, as shown in FIG. 8, is a hardware structure diagram of a data processing device provided by an embodiment, and the data processing device may include: a processor (processor) 810 and memory 820; may also include a communication interface 830 and a bus 840.

The processor 810, the memory 820, and the communication interface 830 can complete communication with each other through the bus 840. Communication interface 830 can be used for information transfer. Processor 810 can invoke logic instructions in memory 820 to perform any of the methods of the above-described embodiments.

The memory 820 may include a storage program area and a storage data area, and the storage program area may store an operating system and an application required for at least one function. The storage data area can store data and the like created according to the use of the data processing device. Further, the memory may include, for example, a volatile memory of a random access memory, and may also include a non-volatile memory. For example, at least one disk storage device, flash memory device, or other non-transitory solid state storage device.

Moreover, when the logic instructions in the memory 820 described above can be implemented in the form of software functional units and sold or used as separate products, the logic instructions can be stored in a computer readable storage medium. The technical solution of the present disclosure may be embodied in the form of a computer software product, which may be stored in a storage medium, and includes a plurality of instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) All or part of the steps of the method described in this embodiment are performed.

All or part of the processes in the foregoing embodiment may be completed by a computer program indicating related hardware, and the program may be stored in a non-transitory computer readable storage medium, and when the program is executed, may include the above The flow of an embodiment of the method.

Industrial applicability

The credit evaluation method and device provided by the present disclosure and the gradient progressive decision tree parameter adjustment method and device can improve the parameter adjustment efficiency of the GBDT model and improve the stability of the GBDT model.

Claims (18)

A credit evaluation method, including: The first sample data is respectively input into at least two gradient progressive decision tree GBDT models to obtain a first credit overdue probability set, and the first sample data is credit data of the first user set; Entering the second sample data into the at least two GBDT models respectively to obtain a second credit overdue probability set, the second sample data being credit data of the second user set; the GBDT parameters of the at least two GBDT models are different ; Performing a KS value calculation according to the first credit overdue probability set and the second credit overdue probability set, and determining a target GBDT model from the at least two GBDT models according to the calculation result; The user is credit evaluated according to the target GBDT model. The method according to claim 1, wherein the KS value calculation is performed according to the first credit overdue probability set and the second credit overdue probability set, and the target GBDT model is determined from the at least two GBDT models according to the calculation result, include: And performing KS value calculation according to the first credit overdue probability set and the first actual credit overdue probability set corresponding to the first user set, to obtain a first KS set; And performing a KS value calculation according to the second credit overdue probability set and the second actual credit overdue probability set corresponding to the second user set, to obtain a second KS set; Performing a comparison calculation on the first KS set and the second KS set, and determining the target GBDT model from the at least two GBDT models according to a calculation result. The method according to claim 2, wherein the first KS set and the second KS set are compared and calculated, and the target GBDT model is determined from the at least two GBDT models according to a calculation result, including : Calculating a minimum value of the KS value of the first KS set obtained from the same GBDT model and the KS value of the second KS set to obtain a third KS set; Calculating a maximum value of the KS value included in the third KS set to obtain a target KS value; A GBDT model corresponding to the target KS value among the at least two GBDT models is determined as the target GBDT model. The method according to claim 2, wherein before the first sample data is separately input into the at least two gradient progressive decision tree GBDT models, the method further comprises: The GBDT parameters of the at least two GBDT models are determined according to a particle swarm optimization PSO algorithm. The method according to claim 4, wherein the GBDT parameters of the at least two GBDT models are determined according to a particle swarm optimization algorithm PSO algorithm, including: Mapping the number of parameters in the GBDT model to the domain dimension of the PSO algorithm; Mapping the range of values of each of the parameters in the GBDT model to the domain of the PSO algorithm; Extracting at least two sets of dimension value data from the domain of the domain corresponding to the domain dimension as at least two particles; Calculating the most advantageous trajectory of the at least two particles by a PSO algorithm; wherein the trajectory most advantageous is a point in the trajectory through which the particle passes to maximize the objective function, the objective function being the first KS a function of the concentrated KS value and the minimum value of the KS value in the second KS set and Mapping the dimension value data corresponding to the trajectory of the at least two particles to the GBDT model to obtain at least two sets of GBDT parameters. The method of claim 1, wherein the credit evaluation of the user based on the target GBDT model comprises: Entering the credit data of the user into the target GBDT model to obtain a credit overdue probability of the user; Comparing the credit overdue probability of the user with a preset credit overdue probability threshold to obtain a credit evaluation result of the user. A credit evaluation device comprising: The first credit overdue probability obtaining module is configured to input the first sample data into the at least two gradient progressive decision tree GBDT models respectively, to obtain a first credit overdue probability set, where the first sample data is the first user set Credit data a second credit overdue probability obtaining module, configured to input second sample data into the at least two GBDT models respectively to obtain a second credit overdue probability set, where the second sample data is credit data of the second user set; The GBDT parameters of at least two GBDT models are different; a model determining module, configured to perform a KS value calculation according to the first credit overdue probability set and the second credit overdue probability set, and determine a target GBDT model from the at least two GBDT models according to the calculation result; The evaluation module is configured to perform credit evaluation on the user according to the target GBDT model. A gradient progressive decision tree parameter adjustment method includes: Determining the domain dimension and the domain range of the particle swarm optimization algorithm according to the number of adjustment parameters in the gradient progressive decision tree and the range of values corresponding to each parameter; Setting an initial parameter of the particle swarm optimization algorithm, and obtaining a trajectory of each particle in the particle group according to the particle swarm optimization algorithm, the domain dimension, and the domain definition; and The parameter values of the gradient progressive decision tree are determined according to the best merit of the trajectory. The method of claim 8 wherein determining a parameter value of the gradient progressive decision tree based on the best merit of the trajectory comprises: Determining a parameter value of the gradient progressive decision tree according to the maximum merit of the trajectory and the value of the corresponding objective function, the objective function being the minimum value of the Kolmogorov-Smirnov KS value of the training sample and the test sample Value function. The method of claim 8 wherein the gradient is determined based on the most advantageous trajectory The parameter values of the decision tree, including: Determining a corresponding peripheral point according to the most advantageous trajectory, wherein the peripheral point of the most advantageous trajectory is obtained according to the hill climbing algorithm with the best advantage of the trajectory as a starting point; The parameter value of the gradient progressive decision tree is determined according to the magnitude of the value of the objective function corresponding to the peripheral point, and the objective function is a minimum function of the KS value of the training sample and the test sample. The method according to claim 10, wherein determining a parameter value of the gradient progressive decision tree according to a magnitude of a value of the objective function corresponding to the peripheral point comprises: The size of the value of the objective function corresponding to the peripheral point is sorted, and the parameter value corresponding to the corresponding peripheral point when the value of the objective function takes the maximum value is selected as the parameter value of the gradient progressive decision tree. The method according to any of claims 8-11, wherein the number of adjustment parameters of the gradient progressive decision tree is 8, the range of the definition domain being the interval from the minimum value to the maximum value of each adjustment parameter. A gradient progressive decision tree parameter adjustment device, comprising: The mapping module is configured to determine a domain dimension and a domain range of the particle swarm optimization algorithm according to the number of adjustment parameters in the gradient progressive decision tree and the value range corresponding to each parameter; a trajectory most advantageous determining module, configured to set an initial parameter of the particle swarm optimization algorithm, and obtain a trajectory of each particle in the particle group according to the particle swarm optimization algorithm, the domain dimension, and the domain definition range ;as well as The parameter determination module is configured to determine a parameter value of the gradient progressive decision tree according to the most advantageous of the trajectory. The apparatus of claim 13 wherein said parameter determination module is configured to: A parameter value of the gradient progressive decision tree is determined according to the trajectory best advantage and the magnitude of the value of the corresponding objective function, the objective function being a minimum function of the KS value of the training sample and the test sample. The apparatus of claim 13 wherein said parameter determination module is configured to: Determining a corresponding peripheral point according to the most advantageous trajectory, wherein the peripheral point of the most advantageous trajectory is obtained according to the hill climbing algorithm with the best advantage of the trajectory as a starting point; Determining a parameter of the gradient progressive decision tree according to a magnitude of a value of the objective function corresponding to the peripheral point, the objective function being a minimum function of the KS value of the training sample and the test sample. The apparatus of claim 15 wherein said parameter determination module is configured to: The size of the value of the objective function corresponding to the peripheral point is sorted, and the parameter value corresponding to the peripheral point corresponding to the value of the largest objective function is selected as the parameter value of the gradient progressive decision tree. Apparatus according to any one of claims 13-16, wherein the number of adjustment parameters of the gradient progressive decision tree is 8, the range of the definition domain being the interval from the minimum value to the maximum value of each adjustment parameter. A computer readable storage medium storing computer executable instructions for performing the method of any of claims 1-6 and 8-12.

Images